Merge "Update to bbb828223e9c8f83f0e84db1e98b116029e62765."
diff --git a/Android.bp b/Android.bp
index 368fec2..408ee3c 100644
--- a/Android.bp
+++ b/Android.bp
@@ -78,40 +78,46 @@
instruction_set: "arm",
// ARM v7 NEON
srcs: [
- "simd/arm/arm/jsimd.c",
- "simd/arm/arm/jsimd_neon.S",
- "simd/arm/common/jccolor-neon.c",
- "simd/arm/common/jcgray-neon.c",
- "simd/arm/common/jcsample-neon.c",
- "simd/arm/common/jdcolor-neon.c",
- "simd/arm/common/jdmerge-neon.c",
- "simd/arm/common/jdsample-neon.c",
- "simd/arm/common/jfdctfst-neon.c",
- "simd/arm/common/jfdctint-neon.c",
- "simd/arm/common/jidctfst-neon.c",
- "simd/arm/common/jidctint-neon.c",
- "simd/arm/common/jidctred-neon.c",
- "simd/arm/common/jquanti-neon.c",
+ "simd/arm/aarch32/jchuff-neon.c",
+ "simd/arm/aarch32/jsimd.c",
+ "simd/arm/jccolor-neon.c",
+ "simd/arm/jcgray-neon.c",
+ "simd/arm/jcphuff-neon.c",
+ "simd/arm/jcsample-neon.c",
+ "simd/arm/jdcolor-neon.c",
+ "simd/arm/jdmerge-neon.c",
+ "simd/arm/jdsample-neon.c",
+ "simd/arm/jfdctfst-neon.c",
+ "simd/arm/jfdctint-neon.c",
+ "simd/arm/jidctfst-neon.c",
+ "simd/arm/jidctint-neon.c",
+ "simd/arm/jidctred-neon.c",
+ "simd/arm/jquanti-neon.c",
],
+ cflags: ["-DNEON_INTRINSICS"],
+ local_include_dirs: ["simd/arm"],
},
arm64: {
// ARM v8 64-bit NEON
srcs: [
- "simd/arm/arm64/jsimd.c",
- "simd/arm/arm64/jsimd_neon.S",
- "simd/arm/common/jccolor-neon.c",
- "simd/arm/common/jcgray-neon.c",
- "simd/arm/common/jcsample-neon.c",
- "simd/arm/common/jdcolor-neon.c",
- "simd/arm/common/jdmerge-neon.c",
- "simd/arm/common/jdsample-neon.c",
- "simd/arm/common/jfdctfst-neon.c",
- "simd/arm/common/jfdctint-neon.c",
- "simd/arm/common/jidctfst-neon.c",
- "simd/arm/common/jidctint-neon.c",
- "simd/arm/common/jidctred-neon.c",
- "simd/arm/common/jquanti-neon.c",
+ "simd/arm/aarch64/jchuff-neon.c",
+ "simd/arm/aarch64/jsimd.c",
+ "simd/arm/jccolor-neon.c",
+ "simd/arm/jcgray-neon.c",
+ "simd/arm/jcphuff-neon.c",
+ "simd/arm/jcsample-neon.c",
+ "simd/arm/jdcolor-neon.c",
+ "simd/arm/jdmerge-neon.c",
+ "simd/arm/jdsample-neon.c",
+ "simd/arm/jfdctfst-neon.c",
+ "simd/arm/jfdctint-neon.c",
+ "simd/arm/jidctfst-neon.c",
+ "simd/arm/jidctint-neon.c",
+ "simd/arm/jidctred-neon.c",
+ "simd/arm/jquanti-neon.c",
],
+ cflags: ["-DNEON_INTRINSICS"],
+ local_include_dirs: ["simd/arm"],
},
x86: {
// x86 MMX and SSE2
@@ -166,10 +172,7 @@
"-DPIC",
"-D__x86__",
],
- local_include_dirs: [
- "simd",
- "simd/nasm",
- ],
+ local_include_dirs: ["simd/nasm"],
},
x86_64: {
// x86-64 SSE2
@@ -207,10 +210,7 @@
"-DPIC",
"-D__x86_64__",
],
- local_include_dirs: [
- "simd",
- "simd/nasm",
- ],
+ local_include_dirs: ["simd/nasm"],
},
},
diff --git a/BUILD.gn b/BUILD.gn
index f9c483e..80bdf9b 100644
--- a/BUILD.gn
+++ b/BUILD.gn
@@ -152,40 +152,50 @@
]
} else if (current_cpu == "arm" && arm_version >= 7 &&
(arm_use_neon || arm_optionally_use_neon)) {
+ include_dirs += [ "simd/arm/" ]
sources = [
- "simd/arm/arm/jsimd.c",
- "simd/arm/arm/jsimd_neon.S",
- "simd/arm/common/jccolor-neon.c",
- "simd/arm/common/jcgray-neon.c",
- "simd/arm/common/jcsample-neon.c",
- "simd/arm/common/jdcolor-neon.c",
- "simd/arm/common/jdmerge-neon.c",
- "simd/arm/common/jdsample-neon.c",
- "simd/arm/common/jfdctfst-neon.c",
- "simd/arm/common/jfdctint-neon.c",
- "simd/arm/common/jidctfst-neon.c",
- "simd/arm/common/jidctint-neon.c",
- "simd/arm/common/jidctred-neon.c",
- "simd/arm/common/jquanti-neon.c",
+ "simd/arm/aarch32/jchuff-neon.c",
+ "simd/arm/aarch32/jsimd.c",
+ "simd/arm/jccolor-neon.c",
+ "simd/arm/jcgray-neon.c",
+ "simd/arm/jcphuff-neon.c",
+ "simd/arm/jcsample-neon.c",
+ "simd/arm/jdcolor-neon.c",
+ "simd/arm/jdmerge-neon.c",
+ "simd/arm/jdsample-neon.c",
+ "simd/arm/jfdctfst-neon.c",
+ "simd/arm/jfdctint-neon.c",
+ "simd/arm/jidctfst-neon.c",
+ "simd/arm/jidctint-neon.c",
+ "simd/arm/jidctred-neon.c",
+ "simd/arm/jquanti-neon.c",
+ ]
+ defines = [
+ "NEON_INTRINSICS"
]
configs -= [ "//build/config/compiler:default_optimization" ]
configs += [ "//build/config/compiler:optimize_speed" ]
} else if (current_cpu == "arm64") {
+ include_dirs += [ "simd/arm/" ]
sources = [
- "simd/arm/arm64/jsimd.c",
- "simd/arm/arm64/jsimd_neon.S",
- "simd/arm/common/jccolor-neon.c",
- "simd/arm/common/jcgray-neon.c",
- "simd/arm/common/jcsample-neon.c",
- "simd/arm/common/jdcolor-neon.c",
- "simd/arm/common/jdmerge-neon.c",
- "simd/arm/common/jdsample-neon.c",
- "simd/arm/common/jfdctfst-neon.c",
- "simd/arm/common/jfdctint-neon.c",
- "simd/arm/common/jidctfst-neon.c",
- "simd/arm/common/jidctint-neon.c",
- "simd/arm/common/jidctred-neon.c",
- "simd/arm/common/jquanti-neon.c",
+ "simd/arm/aarch64/jchuff-neon.c",
+ "simd/arm/aarch64/jsimd.c",
+ "simd/arm/jccolor-neon.c",
+ "simd/arm/jcgray-neon.c",
+ "simd/arm/jcphuff-neon.c",
+ "simd/arm/jcsample-neon.c",
+ "simd/arm/jdcolor-neon.c",
+ "simd/arm/jdmerge-neon.c",
+ "simd/arm/jdsample-neon.c",
+ "simd/arm/jfdctfst-neon.c",
+ "simd/arm/jfdctint-neon.c",
+ "simd/arm/jidctfst-neon.c",
+ "simd/arm/jidctint-neon.c",
+ "simd/arm/jidctred-neon.c",
+ "simd/arm/jquanti-neon.c",
+ ]
+ defines = [
+ "NEON_INTRINSICS"
]
configs -= [ "//build/config/compiler:default_optimization" ]
configs += [ "//build/config/compiler:optimize_speed" ]
@@ -212,7 +222,6 @@
"jccolor.c",
"jcdctmgr.c",
"jchuff.c",
- "jchuff.h",
"jcicc.c",
"jcinit.c",
"jcmainct.c",
@@ -232,7 +241,6 @@
"jdcolor.c",
"jddctmgr.c",
"jdhuff.c",
- "jdhuff.h",
"jdicc.c",
"jdinput.c",
"jdmainct.c",
@@ -244,7 +252,6 @@
"jdsample.c",
"jdtrans.c",
"jerror.c",
- "jerror.h",
"jfdctflt.c",
"jfdctfst.c",
"jfdctint.c",
@@ -254,13 +261,10 @@
"jidctred.c",
"jmemmgr.c",
"jmemnobs.c",
- "jmemsys.h",
"jpeg_nbits_table.c",
- "jpegint.h",
"jquant1.c",
"jquant2.c",
"jutils.c",
- "jversion.h",
]
defines = [
@@ -281,6 +285,12 @@
sources += [ "jsimd_none.c" ]
} else {
public_deps += [ ":simd" ]
+
+ if ((current_cpu == "arm" && arm_version >= 7 &&
+ (arm_use_neon || arm_optionally_use_neon)) ||
+ current_cpu == "arm64") {
+ defines += [ "NEON_INTRINSICS", ]
+ }
}
}
diff --git a/BUILDING.md b/BUILDING.md
index a4ae1e0..b5106ee 100644
--- a/BUILDING.md
+++ b/BUILDING.md
@@ -12,10 +12,7 @@
- [NASM](http://www.nasm.us) or [YASM](http://yasm.tortall.net)
(if building x86 or x86-64 SIMD extensions)
- * If using NASM, 2.10 or later is required.
- * If using NASM, 2.10 or later (except 2.11.08) is required for an x86-64 Mac
- build (2.11.08 does not work properly with libjpeg-turbo's x86-64 SIMD code
- when building macho64 objects.)
+ * If using NASM, 2.13 or later is required.
* If using YASM, 1.2.0 or later is required.
* If building on macOS, NASM or YASM can be obtained from
[MacPorts](http://www.macports.org/) or [Homebrew](http://brew.sh/).
@@ -49,10 +46,8 @@
- If building the TurboJPEG Java wrapper, JDK or OpenJDK 1.5 or later is
required. Most modern Linux distributions, as well as Solaris 10 and later,
- include JDK or OpenJDK. On OS X 10.5 and 10.6, it will be necessary to
- install the Java Developer Package, which can be downloaded from
- <http://developer.apple.com/downloads> (Apple ID required.) For other
- systems, you can obtain the Oracle Java Development Kit from
+ include JDK or OpenJDK. For other systems, you can obtain the Oracle Java
+ Development Kit from
<http://www.oracle.com/technetwork/java/javase/downloads>.
* If using JDK 11 or later, CMake 3.10.x or later must also be used.
@@ -62,22 +57,22 @@
- Microsoft Visual C++ 2005 or later
If you don't already have Visual C++, then the easiest way to get it is by
- installing the
- [Windows SDK](http://msdn.microsoft.com/en-us/windows/bb980924.aspx).
- The Windows SDK includes both 32-bit and 64-bit Visual C++ compilers and
- everything necessary to build libjpeg-turbo.
+ installing
+ [Visual Studio Community Edition](https://visualstudio.microsoft.com),
+ which includes everything necessary to build libjpeg-turbo.
- * You can also use Microsoft Visual Studio Express/Community Edition, which
- is a free download. (NOTE: versions prior to 2012 can only be used to
- build 32-bit code.)
+ * You can also download and install the standalone Windows SDK (for Windows 7
+ or later), which includes command-line versions of the 32-bit and 64-bit
+ Visual C++ compilers.
* If you intend to build libjpeg-turbo from the command line, then add the
appropriate compiler and SDK directories to the `INCLUDE`, `LIB`, and
`PATH` environment variables. This is generally accomplished by
- executing `vcvars32.bat` or `vcvars64.bat` and `SetEnv.cmd`.
- `vcvars32.bat` and `vcvars64.bat` are part of Visual C++ and are located in
- the same directory as the compiler. `SetEnv.cmd` is part of the Windows
- SDK. You can pass optional arguments to `SetEnv.cmd` to specify a 32-bit
- or 64-bit build environment.
+ executing `vcvars32.bat` or `vcvars64.bat`, which are located in the same
+ directory as the compiler.
+ * If built with Visual C++ 2015 or later, the libjpeg-turbo static libraries
+ cannot be used with earlier versions of Visual C++, and vice versa.
+ * The libjpeg API DLL (**jpeg{version}.dll**) will depend on the C run-time
+ DLLs corresponding to the version of Visual C++ that was used to build it.
... OR ...
@@ -333,7 +328,7 @@
-------------
-### 32-bit Build on 64-bit Linux/Unix/Mac
+### 32-bit Build on 64-bit Linux/Unix
Use export/setenv to set the following environment variables before running
CMake:
@@ -398,117 +393,23 @@
Building libjpeg-turbo for iOS
------------------------------
-iOS platforms, such as the iPhone and iPad, use ARM processors, and all
-currently supported models include NEON instructions. Thus, they can take
+iOS platforms, such as the iPhone and iPad, use Arm processors, and all
+currently supported models include Neon instructions. Thus, they can take
advantage of libjpeg-turbo's SIMD extensions to significantly accelerate JPEG
compression/decompression. This section describes how to build libjpeg-turbo
for these platforms.
-### Additional build requirements
+### Armv8 (64-bit)
-- For configurations that require [gas-preprocessor.pl]
- (https://raw.githubusercontent.com/libjpeg-turbo/gas-preprocessor/master/gas-preprocessor.pl),
- it should be installed in your `PATH`.
-
-
-### ARMv7 (32-bit)
-
-**gas-preprocessor.pl required**
-
-The following scripts demonstrate how to build libjpeg-turbo to run on the
-iPhone 3GS-4S/iPad 1st-3rd Generation and newer:
-
-#### Xcode 4.2 and earlier (LLVM-GCC)
-
- IOS_PLATFORMDIR=/Developer/Platforms/iPhoneOS.platform
- IOS_SYSROOT=($IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk)
- export CFLAGS="-mfloat-abi=softfp -march=armv7 -mcpu=cortex-a8 -mtune=cortex-a8 -mfpu=neon -miphoneos-version-min=3.0"
-
- cd {build_directory}
-
- cat <<EOF >toolchain.cmake
- set(CMAKE_SYSTEM_NAME Darwin)
- set(CMAKE_SYSTEM_PROCESSOR arm)
- set(CMAKE_C_COMPILER ${IOS_PLATFORMDIR}/Developer/usr/bin/arm-apple-darwin10-llvm-gcc-4.2)
- EOF
-
- cmake -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.cmake \
- -DCMAKE_OSX_SYSROOT=${IOS_SYSROOT[0]} \
- [additional CMake flags] {source_directory}
- make
-
-#### Xcode 4.3-4.6 (LLVM-GCC)
-
-Same as above, but replace the first line with:
-
- IOS_PLATFORMDIR=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform
-
-#### Xcode 5 and later (Clang)
-
- IOS_PLATFORMDIR=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform
- IOS_SYSROOT=($IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk)
- export CFLAGS="-mfloat-abi=softfp -arch armv7 -miphoneos-version-min=3.0"
- export ASMFLAGS="-no-integrated-as"
-
- cd {build_directory}
-
- cat <<EOF >toolchain.cmake
- set(CMAKE_SYSTEM_NAME Darwin)
- set(CMAKE_SYSTEM_PROCESSOR arm)
- set(CMAKE_C_COMPILER /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang)
- EOF
-
- cmake -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.cmake \
- -DCMAKE_OSX_SYSROOT=${IOS_SYSROOT[0]} \
- [additional CMake flags] {source_directory}
- make
-
-
-### ARMv7s (32-bit)
-
-**gas-preprocessor.pl required**
-
-The following scripts demonstrate how to build libjpeg-turbo to run on the
-iPhone 5/iPad 4th Generation and newer:
-
-#### Xcode 4.5-4.6 (LLVM-GCC)
-
- IOS_PLATFORMDIR=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform
- IOS_SYSROOT=($IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk)
- export CFLAGS="-Wall -mfloat-abi=softfp -march=armv7s -mcpu=swift -mtune=swift -mfpu=neon -miphoneos-version-min=6.0"
-
- cd {build_directory}
-
- cat <<EOF >toolchain.cmake
- set(CMAKE_SYSTEM_NAME Darwin)
- set(CMAKE_SYSTEM_PROCESSOR arm)
- set(CMAKE_C_COMPILER ${IOS_PLATFORMDIR}/Developer/usr/bin/arm-apple-darwin10-llvm-gcc-4.2)
- EOF
-
- cmake -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=toolchain.cmake \
- -DCMAKE_OSX_SYSROOT=${IOS_SYSROOT[0]} \
- [additional CMake flags] {source_directory}
- make
-
-#### Xcode 5 and later (Clang)
-
-Same as the ARMv7 build procedure for Xcode 5 and later, except replace the
-compiler flags as follows:
-
- export CFLAGS="-Wall -mfloat-abi=softfp -arch armv7s -miphoneos-version-min=6.0"
-
-
-### ARMv8 (64-bit)
-
-**gas-preprocessor.pl required if using Xcode < 6**
+**Xcode 5 or later required, Xcode 6.3.x or later recommended**
The following script demonstrates how to build libjpeg-turbo to run on the
iPhone 5S/iPad Mini 2/iPad Air and newer.
IOS_PLATFORMDIR=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform
IOS_SYSROOT=($IOS_PLATFORMDIR/Developer/SDKs/iPhoneOS*.sdk)
- export CFLAGS="-Wall -arch arm64 -miphoneos-version-min=7.0 -funwind-tables"
+ export CFLAGS="-Wall -arch arm64 -miphoneos-version-min=8.0 -funwind-tables"
cd {build_directory}
@@ -523,9 +424,6 @@
[additional CMake flags] {source_directory}
make
-Once built, lipo can be used to combine the ARMv7, v7s, and/or v8 variants into
-a universal library.
-
Building libjpeg-turbo for Android
----------------------------------
@@ -534,7 +432,9 @@
[Android NDK](https://developer.android.com/tools/sdk/ndk).
-### ARMv7 (32-bit)
+### Armv7 (32-bit)
+
+**NDK r19 or later with Clang recommended**
The following is a general recipe script that can be modified for your specific
needs.
@@ -559,7 +459,9 @@
make
-### ARMv8 (64-bit)
+### Armv8 (64-bit)
+
+**Clang recommended**
The following is a general recipe script that can be modified for your specific
needs.
@@ -735,44 +637,23 @@
make dmg
Create Mac package/disk image. This requires pkgbuild and productbuild, which
-are installed by default on OS X 10.7 and later and which can be obtained by
-installing Xcode 3.2.6 (with the "Unix Development" option) on OS X 10.6.
-Packages built in this manner can be installed on OS X 10.5 and later, but they
-must be built on OS X 10.6 or later.
+are installed by default on OS X/macOS 10.7 and later.
- make udmg
+In order to create a Mac package/disk image that contains universal
+x86-64/Arm binaries, set the following CMake variable:
-This creates a Mac package/disk image that contains universal x86-64/i386/ARM
-binaries. The following CMake variables control which architectures are
-included in the universal binaries. Setting any of these variables to an empty
-string excludes that architecture from the package.
+* `ARMV8_BUILD`: Directory containing an Armv8 (64-bit) iOS or macOS build of
+ libjpeg-turbo to include in the universal binaries
-* `OSX_32BIT_BUILD`: Directory containing an i386 (32-bit) Mac build of
- libjpeg-turbo (default: *{source_directory}*/osxx86)
-* `IOS_ARMV7_BUILD`: Directory containing an ARMv7 (32-bit) iOS build of
- libjpeg-turbo (default: *{source_directory}*/iosarmv7)
-* `IOS_ARMV7S_BUILD`: Directory containing an ARMv7s (32-bit) iOS build of
- libjpeg-turbo (default: *{source_directory}*/iosarmv7s)
-* `IOS_ARMV8_BUILD`: Directory containing an ARMv8 (64-bit) iOS build of
- libjpeg-turbo (default: *{source_directory}*/iosarmv8)
-
-You should first use CMake to configure i386, ARMv7, ARMv7s, and/or ARMv8
-sub-builds of libjpeg-turbo (see "Build Recipes" and "Building libjpeg-turbo
-for iOS" above) in build directories that match those specified in the
-aforementioned CMake variables. Next, configure the primary build of
-libjpeg-turbo as an out-of-tree build, and build it. Once the primary build
-has been built, run `make udmg` from the build directory. The packaging system
-will build the sub-builds, use lipo to combine them into a single set of
-universal binaries, then package the universal binaries in the same manner as
-`make dmg`.
-
-
-Cygwin
-------
-
- make cygwinpkg
-
-Build a Cygwin binary package.
+You should first use CMake to configure an Armv8 sub-build of libjpeg-turbo
+(see "Building libjpeg-turbo for iOS" above, if applicable) in a build
+directory that matches the one specified in the aforementioned CMake variable.
+Next, configure the primary (x86-64) build of libjpeg-turbo as an out-of-tree
+build, specifying the aforementioned CMake variable, and build it. Once the
+primary build has been built, run `make dmg` from the build directory. The
+packaging system will build the sub-build, use lipo to combine it with the
+primary build into a single set of universal binaries, then package the
+universal binaries.
Windows
diff --git a/ChangeLog.md b/ChangeLog.md
index 59fb2de..6eb06f0 100644
--- a/ChangeLog.md
+++ b/ChangeLog.md
@@ -1,3 +1,180 @@
+2.0.90 (2.1 beta1)
+==================
+
+### Significant changes relative to 2.0.6:
+
+1. The build system, x86-64 SIMD extensions, and accelerated Huffman codec now
+support the x32 ABI on Linux, which allows for using x86-64 instructions with
+32-bit pointers. The x32 ABI is generally enabled by adding `-mx32` to the
+compiler flags.
+
+ Caveats:
+ - CMake 3.9.0 or later is required in order for the build system to
+automatically detect an x32 build.
+ - Java does not support the x32 ABI, and thus the TurboJPEG Java API will
+automatically be disabled with x32 builds.
+
+2. Added Loongson MMI SIMD implementations of the RGB-to-grayscale, 4:2:2 fancy
+chroma upsampling, 4:2:2 and 4:2:0 merged chroma upsampling/color conversion,
+and fast integer DCT/IDCT algorithms. Relative to libjpeg-turbo 2.0.x, this
+speeds up:
+
+ - the compression of RGB source images into grayscale JPEG images by
+approximately 20%
+ - the decompression of 4:2:2 JPEG images by approximately 40-60% when
+using fancy upsampling
+ - the decompression of 4:2:2 and 4:2:0 JPEG images by approximately
+15-20% when using merged upsampling
+ - the compression of RGB source images by approximately 30-45% when using
+the fast integer DCT
+ - the decompression of JPEG images into RGB destination images by
+approximately 2x when using the fast integer IDCT
+
+ The overall decompression speedup for RGB images is now approximately
+2.3-3.7x (compared to 2-3.5x with libjpeg-turbo 2.0.x.)
+
+3. 32-bit (Armv7 or Armv7s) iOS builds of libjpeg-turbo are no longer
+supported, and the libjpeg-turbo build system can no longer be used to package
+such builds. 32-bit iOS apps cannot run in iOS 11 and later, and the App Store
+no longer allows them.
+
+4. 32-bit (i386) OS X/macOS builds of libjpeg-turbo are no longer supported,
+and the libjpeg-turbo build system can no longer be used to package such
+builds. 32-bit Mac applications cannot run in macOS 10.15 "Catalina" and
+later, and the App Store no longer allows them.
+
+5. The SSE2 (x86 SIMD) and C Huffman encoding algorithms have been
+significantly optimized, resulting in a measured average overall compression
+speedup of 12-28% for 64-bit code and 22-52% for 32-bit code on various Intel
+and AMD CPUs, as well as a measured average overall compression speedup of
+0-23% on platforms that do not have a SIMD-accelerated Huffman encoding
+implementation.
+
+6. The block smoothing algorithm that is applied by default when decompressing
+progressive Huffman-encoded JPEG images has been improved in the following
+ways:
+
+ - The algorithm is now more fault-tolerant. Previously, if a particular
+scan was incomplete, then the smoothing parameters for the incomplete scan
+would be applied to the entire output image, including the parts of the image
+that were generated by the prior (complete) scan. Visually, this had the
+effect of removing block smoothing from lower-frequency scans if they were
+followed by an incomplete higher-frequency scan. libjpeg-turbo now applies
+block smoothing parameters to each iMCU row based on which scan generated the
+pixels in that row, rather than always using the block smoothing parameters for
+the most recent scan.
+ - When applying block smoothing to DC scans, a Gaussian-like kernel with a
+5x5 window is used to reduce the "blocky" appearance.
+
+7. Added SIMD acceleration for progressive Huffman encoding on Arm platforms.
+This speeds up the compression of full-color progressive JPEGs by about 30-40%
+on average (relative to libjpeg-turbo 2.0.x) when using modern Arm CPUs.
+
+8. Added configure-time and run-time auto-detection of Loongson MMI SIMD
+instructions, so that the Loongson MMI SIMD extensions can be included in any
+MIPS64 libjpeg-turbo build.
+
+9. Added fault tolerance features to djpeg and jpegtran, mainly to demonstrate
+methods by which applications can guard against the exploits of the JPEG format
+described in the report
+["Two Issues with the JPEG Standard"](https://libjpeg-turbo.org/pmwiki/uploads/About/TwoIssueswiththeJPEGStandard.pdf).
+
+ - Both programs now accept a `-maxscans` argument, which can be used to
+limit the number of allowable scans in the input file.
+ - Both programs now accept a `-strict` argument, which can be used to
+treat all warnings as fatal.
+
+10. CMake package config files are now included for both the libjpeg and
+TurboJPEG API libraries. This facilitates using libjpeg-turbo with CMake's
+`find_package()` function. For example:
+
+ find_package(libjpeg-turbo CONFIG REQUIRED)
+
+ add_executable(libjpeg_program libjpeg_program.c)
+ target_link_libraries(libjpeg_program PUBLIC libjpeg-turbo::jpeg)
+
+ add_executable(libjpeg_program_static libjpeg_program.c)
+ target_link_libraries(libjpeg_program_static PUBLIC
+ libjpeg-turbo::jpeg-static)
+
+ add_executable(turbojpeg_program turbojpeg_program.c)
+ target_link_libraries(turbojpeg_program PUBLIC
+ libjpeg-turbo::turbojpeg)
+
+ add_executable(turbojpeg_program_static turbojpeg_program.c)
+ target_link_libraries(turbojpeg_program_static PUBLIC
+ libjpeg-turbo::turbojpeg-static)
+
+11. Since the Unisys LZW patent has long expired, cjpeg and djpeg can now
+read/write both LZW-compressed and uncompressed GIF files (feature ported from
+jpeg-6a and jpeg-9d.)
+
+12. jpegtran now includes the `-wipe` and `-drop` options from jpeg-9a and
+jpeg-9d, as well as the ability to expand the image size using the `-crop`
+option. Refer to jpegtran.1 or usage.txt for more details.
+
+13. Added a complete intrinsics implementation of the Arm Neon SIMD extensions,
+thus providing SIMD acceleration on Arm platforms for all of the algorithms
+that are SIMD-accelerated on x86 platforms. This new implementation is
+significantly faster in some cases than the old GAS implementation--
+depending on the algorithms used, the type of CPU core, and the compiler. GCC,
+as of this writing, does not provide a full or optimal set of Neon intrinsics,
+so for performance reasons, the default when building libjpeg-turbo with GCC is
+to continue using the GAS implementation of the following algorithms:
+
+ - 32-bit RGB-to-YCbCr color conversion
+ - 32-bit fast and accurate inverse DCT
+ - 64-bit RGB-to-YCbCr and YCbCr-to-RGB color conversion
+ - 64-bit accurate forward and inverse DCT
+ - 64-bit Huffman encoding
+
+ A new CMake variable (`NEON_INTRINSICS`) can be used to override this
+default.
+
+ Since the new intrinsics implementation includes SIMD acceleration
+for merged upsampling/color conversion, 1.5.1[5] is no longer necessary and has
+been reverted.
+
+14. The Arm Neon SIMD extensions can now be built using Visual Studio.
+
+15. The build system can now be used to generate a universal x86-64 + Armv8
+libjpeg-turbo SDK package for both iOS and macOS.
+
+
+2.0.6
+=====
+
+### Significant changes relative to 2.0.5:
+
+1. Fixed "using JNI after critical get" errors that occurred on Android
+platforms when using any of the YUV encoding/compression/decompression/decoding
+methods in the TurboJPEG Java API.
+
+2. Fixed or worked around multiple issues with `jpeg_skip_scanlines()`:
+
+ - Fixed segfaults or "Corrupt JPEG data: premature end of data segment"
+errors in `jpeg_skip_scanlines()` that occurred when decompressing 4:2:2 or
+4:2:0 JPEG images using merged (non-fancy) upsampling/color conversion (that
+is, when setting `cinfo.do_fancy_upsampling` to `FALSE`.) 2.0.0[6] was a
+similar fix, but it did not cover all cases.
+ - `jpeg_skip_scanlines()` now throws an error if two-pass color
+quantization is enabled. Two-pass color quantization never worked properly
+with `jpeg_skip_scanlines()`, and the issues could not readily be fixed.
+ - Fixed an issue whereby `jpeg_skip_scanlines()` always returned 0 when
+skipping past the end of an image.
+
+3. The Arm 64-bit (Armv8) Neon SIMD extensions can now be built using MinGW
+toolchains targetting Arm64 (AArch64) Windows binaries.
+
+4. Fixed unexpected visual artifacts that occurred when using
+`jpeg_crop_scanline()` and interblock smoothing while decompressing only the DC
+scan of a progressive JPEG image.
+
+5. Fixed an issue whereby libjpeg-turbo would not build if 12-bit-per-component
+JPEG support (`WITH_12BIT`) was enabled along with libjpeg v7 or libjpeg v8
+API/ABI emulation (`WITH_JPEG7` or `WITH_JPEG8`.)
+
+
2.0.5
=====
@@ -64,7 +241,7 @@
(unlike the decompressor) is not generally exposed to arbitrary data exploits,
this issue did not likely pose a security risk.
-6. The ARM 64-bit (ARMv8) NEON SIMD assembly code now stores constants in a
+6. The Arm 64-bit (Armv8) Neon SIMD assembly code now stores constants in a
separate read-only data section rather than in the text section, to support
execute-only memory layouts.
@@ -246,7 +423,7 @@
1. Added AVX2 SIMD implementations of the colorspace conversion, chroma
downsampling and upsampling, integer quantization and sample conversion, and
-slow integer DCT/IDCT algorithms. When using the slow integer DCT/IDCT
+accurate integer DCT/IDCT algorithms. When using the accurate integer DCT/IDCT
algorithms on AVX2-equipped CPUs, the compression of RGB images is
approximately 13-36% (avg. 22%) faster (relative to libjpeg-turbo 1.5.x) with
64-bit code and 11-21% (avg. 17%) faster with 32-bit code, and the
@@ -350,16 +527,16 @@
now produces bitwise-identical results to the unmerged algorithms.
12. The SIMD function symbols for x86[-64]/ELF, MIPS/ELF, macOS/x86[-64] (if
-libjpeg-turbo is built with YASM), and iOS/ARM[64] builds are now private.
+libjpeg-turbo is built with YASM), and iOS/Arm[64] builds are now private.
This prevents those symbols from being exposed in applications or shared
libraries that link statically with libjpeg-turbo.
13. Added Loongson MMI SIMD implementations of the RGB-to-YCbCr and
YCbCr-to-RGB colorspace conversion, 4:2:0 chroma downsampling, 4:2:0 fancy
-chroma upsampling, integer quantization, and slow integer DCT/IDCT algorithms.
-When using the slow integer DCT/IDCT, this speeds up the compression of RGB
-images by approximately 70-100% and the decompression of RGB images by
-approximately 2-3.5x.
+chroma upsampling, integer quantization, and accurate integer DCT/IDCT
+algorithms. When using the accurate integer DCT/IDCT, this speeds up the
+compression of RGB images by approximately 70-100% and the decompression of RGB
+images by approximately 2-3.5x.
14. Fixed a build error when building with older MinGW releases (regression
caused by 1.5.1[7].)
@@ -409,9 +586,9 @@
`jpeg_consume_input()` would return `JPEG_SUSPENDED` rather than
`JPEG_REACHED_EOI`.
-9. `jpeg_crop_scanlines()` now works correctly when decompressing grayscale
-JPEG images that were compressed with a sampling factor other than 1 (for
-instance, with `cjpeg -grayscale -sample 2x2`).
+9. `jpeg_crop_scanline()` now works correctly when decompressing grayscale JPEG
+images that were compressed with a sampling factor other than 1 (for instance,
+with `cjpeg -grayscale -sample 2x2`).
1.5.2
@@ -435,7 +612,7 @@
5. Fixed build and runtime errors on Windows that occurred when building
libjpeg-turbo with libjpeg v7 API/ABI emulation and the in-memory
source/destination managers. Due to an oversight, the `jpeg_skip_scanlines()`
-and `jpeg_crop_scanlines()` functions were not being included in jpeg7.dll when
+and `jpeg_crop_scanline()` functions were not being included in jpeg7.dll when
libjpeg-turbo was built with `-DWITH_JPEG7=1` and `-DWITH_MEMSRCDST=1`.
6. Fixed "Bogus virtual array access" error that occurred when using the
@@ -691,8 +868,8 @@
disabled by setting the `JSIMD_NOHUFFENC` environment variable to `1`.
13. Added ARM 64-bit (ARMv8) NEON SIMD implementations of the commonly-used
-compression algorithms (including the slow integer forward DCT and h2v2 & h2v1
-downsampling algorithms, which are not accelerated in the 32-bit NEON
+compression algorithms (including the accurate integer forward DCT and h2v2 &
+h2v1 downsampling algorithms, which are not accelerated in the 32-bit NEON
implementation.) This speeds up the compression of full-color JPEGs by about
75% on average on a Cavium ThunderX processor and by about 2-2.5x on average on
Cortex-A53 and Cortex-A57 cores.
@@ -823,8 +1000,8 @@
7. Fixed an extremely rare bug in the Huffman encoder that caused 64-bit
builds of libjpeg-turbo to incorrectly encode a few specific test images when
-quality=98, an optimized Huffman table, and the slow integer forward DCT were
-used.
+quality=98, an optimized Huffman table, and the accurate integer forward DCT
+were used.
8. The Windows (CMake) build system now supports building only static or only
shared libraries. This is accomplished by adding either `-DENABLE_STATIC=0` or
@@ -983,8 +1160,8 @@
The accuracy of this implementation now matches the accuracy of the SSE/SSE2
implementation. Note, however, that the floating point DCT/IDCT algorithms are
mainly a legacy feature. They generally do not produce significantly better
-accuracy than the slow integer DCT/IDCT algorithms, and they are quite a bit
-slower.
+accuracy than the accurate integer DCT/IDCT algorithms, and they are quite a
+bit slower.
8. Added a new output colorspace (`JCS_RGB565`) to the libjpeg API that allows
for decompressing JPEG images into RGB565 (16-bit) pixels. If dithering is not
@@ -1394,8 +1571,8 @@
2. Despite the above, the fast integer forward DCT still degrades somewhat for
JPEG qualities greater than 95, so the TurboJPEG wrapper will now automatically
-use the slow integer forward DCT when generating JPEG images of quality 96 or
-greater. This reduces compression performance by as much as 15% for these
+use the accurate integer forward DCT when generating JPEG images of quality 96
+or greater. This reduces compression performance by as much as 15% for these
high-quality images but is necessary to ensure that the images are perceptually
lossless. It also ensures that the library can avoid the performance pitfall
created by [1].
diff --git a/README.chromium b/README.chromium
index fca4ea6..1aa67cf 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
Name: libjpeg-turbo
URL: https://github.com/libjpeg-turbo/libjpeg-turbo/
-Version: 2.0.5
+Version: 2.0.90 (2.1 Beta 1)
License: Custom license
License File: LICENSE.md
Security Critical: yes
@@ -8,19 +8,24 @@
Description:
This consists of the components:
-* libjpeg-turbo 2.0.5
+* libjpeg-turbo 2.0.90 (2.1 Beta 1)
* This file (README.chromium)
* A build file (BUILD.gn)
* An OWNERS file
* A codereview.settings file
* Patched header files used by Chromium
-* Cherry picked a fix from upstream master to enable AArch64 Windows builds:
- https://github.com/libjpeg-turbo/libjpeg-turbo/commit/6ee5d5f568fda1a7c6a49dd8995f2d89866ee42d
+* Cherry-picked two additional patches from upstream master to fix bugs found
+ by fuzzers:
+ https://github.com/libjpeg-turbo/libjpeg-turbo/commit/ccaba5d7894ecfb5a8f11e48d3f86e1f14d5a469
+ https://github.com/libjpeg-turbo/libjpeg-turbo/commit/c7ca521bc85b57d41d3ad4963c13fc0100481084
* Deleted unused directories: ci, cmakescripts, doc, java, release, sharedlib,
simd/loongson, simd/mips, simd/powerpc, and win
* Deleted unused files: appveyor.yml, CMakeLists.txt, doxygen.config,
- doxygen-extra.css, .gitattributes, tjexample.c, tjexampletest.java.in,
- .travis.yml
+ doxygen-extra.css, .gitattributes, tjexample.c, tjexampletest.in,
+ tjexampletest.java.in, .travis.yml
+* Deleted legacy Arm Neon assembly files (supporting old compiler versions that
+ do not generate performant code from intrinsics):
+ simd/arm/aarch32/jsimd_neon.S, simd/arm/aarch64/jsimd_neon.S.
This libjpeg-turbo can replace our libjpeg-6b without any modifications in the
Chromium code.
@@ -35,10 +40,11 @@
decoder fallback to slow decoding if the Huffman decoding bit sentinel > 16,
this to match the exact behavior of jpeg_huff_decode().
http://crbug.com/398235
- The patch in the above bug removed "& 0xFF". It has been restored from upstream
+ The patch in the above bug removed "& 0xFF". It has been restored from
+ upstream:
https://github.com/libjpeg-turbo/libjpeg-turbo/commit/fa1d18385d904d530b4aec83ab7757a33397de6e
-* Configuration files jconfig.h and jconfigint.h were generated and then altered
- manually to be compatible on all of Chromium's platforms.
+* Configuration files jconfig.h, jconfigint.h and neon-compat.h were generated
+ and then altered manually to be compatible on all of Chromium's platforms.
http://crbug.com/608347
* Fix static const data duplication of jpeg_nbits_table. A unique copy
was in the jchuff.obj and jcphuff.obj resulting in an added 65k in
@@ -51,53 +57,21 @@
lld) arising from attempts to reference the table from assembler on
32-bit x86. This only affects shared libraries, but that's important
for downstream Android builds.
-* Arm NEON patches to improve performance and maintainability. These changes
- are tracked by the following Chromium issue: https://crbug.com/922430
- - Add memory alignment size check in jmemmgr.c
- - Add 32-byte memory alignment check for Arm NEON
- - Add Arm NEON implementation of h2v2_fancy_upsample
- - Add SIMD function stubs for h1v2_fancy_upsample
- - Add Arm NEON implementation of h1v2_fancy_upsample
- - Add Arm NEON implementation of h2v1_fancy_upsample
- - Add extra guard for loop unroll pragma on AArch64
- - Add Arm NEON implementation of h2v1_upsample
- - Add Arm NEON implementation of h2v2_upsample
- - Implement YCbCr->RGB using Arm NEON intrinsics
- - Implement YCbCr->RGB565 using Arm NEON intrinsics
- - Add Arm NEON implementation of h2v1_merged_upsample
- - Add Arm NEON implementation of h2v2_merged_upsample
- - Implement 2x2 IDCT using Arm NEON intrinsics
- - Implement 4x4 IDCT using Arm NEON intrinsics
- - Implement slow IDCT using Arm NEON intrinsics
- - Precompute DCT block output pointers in IDCT functions
- - Implement fast IDCT using Arm NEON intrinsics
- - Add Arm NEON implementation of h2v1_downsample
- - Add Arm NEON implementation of h2v2_downsample
- - Implement RGB->YCbCr using Arm NEON intrinsics
- - Add Arm NEON implementation of RGB->Grayscale
- - Add compiler-independent alignment macro
- - Implement sample conversion using Arm NEON intrinsics
- - Implement quantization using Arm NEON intrinsics
- - Implement fast DCT using Arm NEON intrinsics
- - Implement accurate DCT using Arm NEON intrinsics
-* Patches to enable running the upstream unit tests through gtest.
+* Patches to enable running the upstream unit tests through GTest.
The upstream unit tests are defined here under the section 'TESTS':
https://github.com/libjpeg-turbo/libjpeg-turbo/blob/master/CMakeLists.txt
These changes are tracked by Chromium issue: https://crbug.com/993876
- Refactor tjunittest.c to provide test interface
- - Add gtest wrapper for tjunittests
- Move tjunittest logs from stdout to stderr
- Refactor tjbench.c to provide test interface
- - Add gtest wrapper for tjbench tests
- Move tbench logs from stdout to stderr
- Write tjunittest output files to sdcard on Android
- Refactor cjpeg.c to provide test interface
- - Add gtest wrapper for cjpeg tests
- Refactor jpegtran.c to provide test interface
- Add input JPEG images for djpeg and jpegtran tests
- - Add gtest wrapper for jpegtran tests
- Refactor djpeg.c to provide test interface
- - Add gtest wrapper for djpeg tests
+ A new gtest directory contains GTest wrappers (and associated utilities) for
+ each of tjunittest, tjbench, cjpeg, djpeg and jpegtran.
Refer to working-with-nested-repos [1] for details of how to setup your git
svn client to update the code (for making local changes, cherry picking from
diff --git a/README.ijg b/README.ijg
index 2e39f96..9453c19 100644
--- a/README.ijg
+++ b/README.ijg
@@ -128,7 +128,7 @@
fitness for a particular purpose. This software is provided "AS IS", and you,
its user, assume the entire risk as to its quality and accuracy.
-This software is copyright (C) 1991-2016, Thomas G. Lane, Guido Vollbeding.
+This software is copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
All Rights Reserved except as specified below.
Permission is hereby granted to use, copy, modify, and distribute this
@@ -159,19 +159,6 @@
assumed by the product vendor.
-The IJG distribution formerly included code to read and write GIF files.
-To avoid entanglement with the Unisys LZW patent (now expired), GIF reading
-support has been removed altogether, and the GIF writer has been simplified
-to produce "uncompressed GIFs". This technique does not use the LZW
-algorithm; the resulting GIF files are larger than usual, but are readable
-by all standard GIF decoders.
-
-We are required to state that
- "The Graphics Interchange Format(c) is the Copyright property of
- CompuServe Incorporated. GIF(sm) is a Service Mark property of
- CompuServe Incorporated."
-
-
REFERENCES
==========
@@ -223,12 +210,12 @@
A PDF file of the older JFIF 1.02 specification is available at
http://www.w3.org/Graphics/JPEG/jfif3.pdf.
-The TIFF 6.0 file format specification can be obtained by FTP from
-ftp://ftp.sgi.com/graphics/tiff/TIFF6.ps.gz. The JPEG incorporation scheme
-found in the TIFF 6.0 spec of 3-June-92 has a number of serious problems.
-IJG does not recommend use of the TIFF 6.0 design (TIFF Compression tag 6).
-Instead, we recommend the JPEG design proposed by TIFF Technical Note #2
-(Compression tag 7). Copies of this Note can be obtained from
+The TIFF 6.0 file format specification can be obtained from
+http://mirrors.ctan.org/graphics/tiff/TIFF6.ps.gz. The JPEG incorporation
+scheme found in the TIFF 6.0 spec of 3-June-92 has a number of serious
+problems. IJG does not recommend use of the TIFF 6.0 design (TIFF Compression
+tag 6). Instead, we recommend the JPEG design proposed by TIFF Technical Note
+#2 (Compression tag 7). Copies of this Note can be obtained from
http://www.ijg.org/files/. It is expected that the next revision
of the TIFF spec will replace the 6.0 JPEG design with the Note's design.
Although IJG's own code does not support TIFF/JPEG, the free libtiff library
@@ -243,14 +230,8 @@
directory "files".
The JPEG FAQ (Frequently Asked Questions) article is a source of some
-general information about JPEG.
-It is available on the World Wide Web at http://www.faqs.org/faqs/jpeg-faq/
-and other news.answers archive sites, including the official news.answers
-archive at rtfm.mit.edu: ftp://rtfm.mit.edu/pub/usenet/news.answers/jpeg-faq/.
-If you don't have Web or FTP access, send e-mail to mail-server@rtfm.mit.edu
-with body
- send usenet/news.answers/jpeg-faq/part1
- send usenet/news.answers/jpeg-faq/part2
+general information about JPEG. It is available at
+http://www.faqs.org/faqs/jpeg-faq.
FILE FORMAT COMPATIBILITY
diff --git a/README.md b/README.md
index e7ff743..01e391e 100644
--- a/README.md
+++ b/README.md
@@ -2,8 +2,8 @@
==========
libjpeg-turbo is a JPEG image codec that uses SIMD instructions to accelerate
-baseline JPEG compression and decompression on x86, x86-64, ARM, PowerPC, and
-MIPS systems, as well as progressive JPEG compression on x86 and x86-64
+baseline JPEG compression and decompression on x86, x86-64, Arm, PowerPC, and
+MIPS systems, as well as progressive JPEG compression on x86, x86-64, and Arm
systems. On such systems, libjpeg-turbo is generally 2-6x as fast as libjpeg,
all else being equal. On other types of systems, libjpeg-turbo can still
outperform libjpeg by a significant amount, by virtue of its highly-optimized
@@ -179,8 +179,8 @@
NOTE: As of this writing, extensive research has been conducted into the
usefulness of DCT scaling as a means of data reduction and SmartScale as a
-means of quality improvement. The reader is invited to peruse the research at
-<http://www.libjpeg-turbo.org/About/SmartScale> and draw his/her own conclusions,
+means of quality improvement. Readers are invited to peruse the research at
+<http://www.libjpeg-turbo.org/About/SmartScale> and draw their own conclusions,
but it is the general belief of our project that these features have not
demonstrated sufficient usefulness to justify inclusion in libjpeg-turbo.
@@ -287,12 +287,13 @@
(and slightly faster) floating point IDCT algorithm introduced in libjpeg
v8a as opposed to the algorithm used in libjpeg v6b. It should be noted,
however, that this algorithm basically brings the accuracy of the floating
- point IDCT in line with the accuracy of the slow integer IDCT. The floating
- point DCT/IDCT algorithms are mainly a legacy feature, and they do not
- produce significantly more accuracy than the slow integer algorithms (to put
- numbers on this, the typical difference in PNSR between the two algorithms
- is less than 0.10 dB, whereas changing the quality level by 1 in the upper
- range of the quality scale is typically more like a 1.0 dB difference.)
+ point IDCT in line with the accuracy of the accurate integer IDCT. The
+ floating point DCT/IDCT algorithms are mainly a legacy feature, and they do
+ not produce significantly more accuracy than the accurate integer algorithms
+ (to put numbers on this, the typical difference in PNSR between the two
+ algorithms is less than 0.10 dB, whereas changing the quality level by 1 in
+ the upper range of the quality scale is typically more like a 1.0 dB
+ difference.)
- If the floating point algorithms in libjpeg-turbo are not implemented using
SIMD instructions on a particular platform, then the accuracy of the
@@ -340,7 +341,7 @@
correct results whenever the fast integer forward DCT is used along with a JPEG
quality of 98-100. Thus, libjpeg-turbo must use the non-SIMD quantization
function in those cases. This causes performance to drop by as much as 40%.
-It is therefore strongly advised that you use the slow integer forward DCT
+It is therefore strongly advised that you use the accurate integer forward DCT
whenever encoding images with a JPEG quality of 98 or higher.
diff --git a/cderror.h b/cderror.h
index 4f2c7a3..a386b69 100644
--- a/cderror.h
+++ b/cderror.h
@@ -42,7 +42,7 @@
#ifdef BMP_SUPPORTED
JMESSAGE(JERR_BMP_BADCMAP, "Unsupported BMP colormap format")
-JMESSAGE(JERR_BMP_BADDEPTH, "Only 8- and 24-bit BMP files are supported")
+JMESSAGE(JERR_BMP_BADDEPTH, "Only 8-, 24-, and 32-bit BMP files are supported")
JMESSAGE(JERR_BMP_BADHEADER, "Invalid BMP file: bad header length")
JMESSAGE(JERR_BMP_BADPLANES, "Invalid BMP file: biPlanes not equal to 1")
JMESSAGE(JERR_BMP_COLORSPACE, "BMP output must be grayscale or RGB")
@@ -50,9 +50,9 @@
JMESSAGE(JERR_BMP_EMPTY, "Empty BMP image")
JMESSAGE(JERR_BMP_NOT, "Not a BMP file - does not start with BM")
JMESSAGE(JERR_BMP_OUTOFRANGE, "Numeric value out of range in BMP file")
-JMESSAGE(JTRC_BMP, "%ux%u 24-bit BMP image")
+JMESSAGE(JTRC_BMP, "%ux%u %d-bit BMP image")
JMESSAGE(JTRC_BMP_MAPPED, "%ux%u 8-bit colormapped BMP image")
-JMESSAGE(JTRC_BMP_OS2, "%ux%u 24-bit OS2 BMP image")
+JMESSAGE(JTRC_BMP_OS2, "%ux%u %d-bit OS2 BMP image")
JMESSAGE(JTRC_BMP_OS2_MAPPED, "%ux%u 8-bit colormapped OS2 BMP image")
#endif /* BMP_SUPPORTED */
@@ -84,23 +84,6 @@
JMESSAGE(JTRC_PPM_TEXT, "%ux%u text PPM image")
#endif /* PPM_SUPPORTED */
-#ifdef RLE_SUPPORTED
-JMESSAGE(JERR_RLE_BADERROR, "Bogus error code from RLE library")
-JMESSAGE(JERR_RLE_COLORSPACE, "RLE output must be grayscale or RGB")
-JMESSAGE(JERR_RLE_DIMENSIONS, "Image dimensions (%ux%u) too large for RLE")
-JMESSAGE(JERR_RLE_EMPTY, "Empty RLE file")
-JMESSAGE(JERR_RLE_EOF, "Premature EOF in RLE header")
-JMESSAGE(JERR_RLE_MEM, "Insufficient memory for RLE header")
-JMESSAGE(JERR_RLE_NOT, "Not an RLE file")
-JMESSAGE(JERR_RLE_TOOMANYCHANNELS, "Cannot handle %d output channels for RLE")
-JMESSAGE(JERR_RLE_UNSUPPORTED, "Cannot handle this RLE setup")
-JMESSAGE(JTRC_RLE, "%ux%u full-color RLE file")
-JMESSAGE(JTRC_RLE_FULLMAP, "%ux%u full-color RLE file with map of length %d")
-JMESSAGE(JTRC_RLE_GRAY, "%ux%u grayscale RLE file")
-JMESSAGE(JTRC_RLE_MAPGRAY, "%ux%u grayscale RLE file with map of length %d")
-JMESSAGE(JTRC_RLE_MAPPED, "%ux%u colormapped RLE file with map of length %d")
-#endif /* RLE_SUPPORTED */
-
#ifdef TARGA_SUPPORTED
JMESSAGE(JERR_TGA_BADCMAP, "Unsupported Targa colormap format")
JMESSAGE(JERR_TGA_BADPARMS, "Invalid or unsupported Targa file")
diff --git a/cdjpeg.c b/cdjpeg.c
index e0e382d..5278c1d 100644
--- a/cdjpeg.c
+++ b/cdjpeg.c
@@ -3,8 +3,8 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1997, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2019, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -25,26 +25,37 @@
* Optional progress monitor: display a percent-done figure on stderr.
*/
-#ifdef PROGRESS_REPORT
-
METHODDEF(void)
progress_monitor(j_common_ptr cinfo)
{
cd_progress_ptr prog = (cd_progress_ptr)cinfo->progress;
- int total_passes = prog->pub.total_passes + prog->total_extra_passes;
- int percent_done =
- (int)(prog->pub.pass_counter * 100L / prog->pub.pass_limit);
- if (percent_done != prog->percent_done) {
- prog->percent_done = percent_done;
- if (total_passes > 1) {
- fprintf(stderr, "\rPass %d/%d: %3d%% ",
- prog->pub.completed_passes + prog->completed_extra_passes + 1,
- total_passes, percent_done);
- } else {
- fprintf(stderr, "\r %3d%% ", percent_done);
+ if (prog->max_scans != 0 && cinfo->is_decompressor) {
+ int scan_no = ((j_decompress_ptr)cinfo)->input_scan_number;
+
+ if (scan_no > (int)prog->max_scans) {
+ fprintf(stderr, "Scan number %d exceeds maximum scans (%d)\n", scan_no,
+ prog->max_scans);
+ exit(EXIT_FAILURE);
}
- fflush(stderr);
+ }
+
+ if (prog->report) {
+ int total_passes = prog->pub.total_passes + prog->total_extra_passes;
+ int percent_done =
+ (int)(prog->pub.pass_counter * 100L / prog->pub.pass_limit);
+
+ if (percent_done != prog->percent_done) {
+ prog->percent_done = percent_done;
+ if (total_passes > 1) {
+ fprintf(stderr, "\rPass %d/%d: %3d%% ",
+ prog->pub.completed_passes + prog->completed_extra_passes + 1,
+ total_passes, percent_done);
+ } else {
+ fprintf(stderr, "\r %3d%% ", percent_done);
+ }
+ fflush(stderr);
+ }
}
}
@@ -57,6 +68,8 @@
progress->pub.progress_monitor = progress_monitor;
progress->completed_extra_passes = 0;
progress->total_extra_passes = 0;
+ progress->max_scans = 0;
+ progress->report = FALSE;
progress->percent_done = -1;
cinfo->progress = &progress->pub;
}
@@ -73,8 +86,6 @@
}
}
-#endif
-
/*
* Case-insensitive matching of possibly-abbreviated keyword switches.
diff --git a/cdjpeg.h b/cdjpeg.h
index 9868a0b..71b0c4f 100644
--- a/cdjpeg.h
+++ b/cdjpeg.h
@@ -3,8 +3,9 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 2019 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2017, D. R. Commander.
+ * Copyright (C) 2017, 2019, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -56,9 +57,9 @@
void (*finish_output) (j_decompress_ptr cinfo, djpeg_dest_ptr dinfo);
/* Re-calculate buffer dimensions based on output dimensions (for use with
partial image decompression.) If this is NULL, then the output format
- does not support partial image decompression (BMP and RLE, in particular,
- cannot support partial decompression because they use an inversion buffer
- to write the image in bottom-up order.) */
+ does not support partial image decompression (BMP, in particular, cannot
+ support partial decompression because it uses an inversion buffer to write
+ the image in bottom-up order.) */
void (*calc_buffer_dimensions) (j_decompress_ptr cinfo,
djpeg_dest_ptr dinfo);
@@ -87,6 +88,9 @@
struct jpeg_progress_mgr pub; /* fields known to JPEG library */
int completed_extra_passes; /* extra passes completed */
int total_extra_passes; /* total extra */
+ JDIMENSION max_scans; /* abort if the number of scans exceeds this
+ value and the value is non-zero */
+ boolean report; /* whether or not to report progress */
/* last printed percentage stored here to avoid multiple printouts */
int percent_done;
};
@@ -101,11 +105,9 @@
EXTERN(djpeg_dest_ptr) jinit_write_bmp(j_decompress_ptr cinfo, boolean is_os2,
boolean use_inversion_array);
EXTERN(cjpeg_source_ptr) jinit_read_gif(j_compress_ptr cinfo);
-EXTERN(djpeg_dest_ptr) jinit_write_gif(j_decompress_ptr cinfo);
+EXTERN(djpeg_dest_ptr) jinit_write_gif(j_decompress_ptr cinfo, boolean is_lzw);
EXTERN(cjpeg_source_ptr) jinit_read_ppm(j_compress_ptr cinfo);
EXTERN(djpeg_dest_ptr) jinit_write_ppm(j_decompress_ptr cinfo);
-EXTERN(cjpeg_source_ptr) jinit_read_rle(j_compress_ptr cinfo);
-EXTERN(djpeg_dest_ptr) jinit_write_rle(j_decompress_ptr cinfo);
EXTERN(cjpeg_source_ptr) jinit_read_targa(j_compress_ptr cinfo);
EXTERN(djpeg_dest_ptr) jinit_write_targa(j_decompress_ptr cinfo);
@@ -125,7 +127,6 @@
/* common support routines (in cdjpeg.c) */
-EXTERN(void) enable_signal_catcher(j_common_ptr cinfo);
EXTERN(void) start_progress_monitor(j_common_ptr cinfo,
cd_progress_ptr progress);
EXTERN(void) end_progress_monitor(j_common_ptr cinfo);
diff --git a/change.log b/change.log
index f090d77..e4d0ddc 100644
--- a/change.log
+++ b/change.log
@@ -6,6 +6,25 @@
CHANGE LOG for Independent JPEG Group's JPEG software
+Version 9d 12-Jan-2020
+-----------------------
+
+Restore GIF read and write support from libjpeg version 6a.
+Thank to Wolfgang Werner (W.W.) Heinz for suggestion.
+
+Add jpegtran -drop option; add options to the crop extension and wipe
+to fill the extra area with content from the source image region,
+instead of gray out.
+
+
+Version 9c 14-Jan-2018
+-----------------------
+
+jpegtran: add an option to the -wipe switch to fill the region
+with the average of adjacent blocks, instead of gray out.
+Thank to Caitlyn Feddock and Maddie Ziegler for inspiration.
+
+
Version 9b 17-Jan-2016
-----------------------
@@ -13,6 +32,13 @@
Thank to Michele Martone for suggestion.
+Version 9a 19-Jan-2014
+-----------------------
+
+Add jpegtran -wipe option and extension for -crop.
+Thank to Andrew Senior, David Clunie, and Josef Schmid for suggestion.
+
+
Version 9 13-Jan-2013
----------------------
@@ -138,11 +164,6 @@
Huffman tables are checked for validity much more carefully than before.
-To avoid the Unisys LZW patent, djpeg's GIF output capability has been
-changed to produce "uncompressed GIFs", and cjpeg's GIF input capability
-has been removed altogether. We're not happy about it either, but there
-seems to be no good alternative.
-
The configure script now supports building libjpeg as a shared library
on many flavors of Unix (all the ones that GNU libtool knows how to
build shared libraries for). Use "./configure --enable-shared" to
diff --git a/cjpeg.1 b/cjpeg.1
index a3e47ba..569dc3f 100644
--- a/cjpeg.1
+++ b/cjpeg.1
@@ -1,4 +1,4 @@
-.TH CJPEG 1 "18 March 2017"
+.TH CJPEG 1 "4 November 2020"
.SH NAME
cjpeg \- compress an image file to a JPEG file
.SH SYNOPSIS
@@ -16,8 +16,7 @@
compresses the named image file, or the standard input if no file is
named, and produces a JPEG/JFIF file on the standard output.
The currently supported input file formats are: PPM (PBMPLUS color
-format), PGM (PBMPLUS grayscale format), BMP, Targa, and RLE (Utah Raster
-Toolkit format). (RLE is supported only if the URT library is available.)
+format), PGM (PBMPLUS grayscale format), BMP, GIF, and Targa.
.SH OPTIONS
All switch names may be abbreviated; for example,
.B \-grayscale
@@ -42,10 +41,10 @@
.TP
.B \-grayscale
Create monochrome JPEG file from color input. Be sure to use this switch when
-compressing a grayscale BMP file, because
+compressing a grayscale BMP or GIF file, because
.B cjpeg
-isn't bright enough to notice whether a BMP file uses only shades of gray.
-By saying
+isn't bright enough to notice whether a BMP or GIF file uses only shades of
+gray. By saying
.BR \-grayscale,
you'll get a smaller JPEG file that takes less time to process.
.TP
@@ -161,31 +160,40 @@
unable to view an arithmetic coded JPEG file at all.
.TP
.B \-dct int
-Use integer DCT method (default).
+Use accurate integer DCT method (default).
.TP
.B \-dct fast
-Use fast integer DCT (less accurate).
-In libjpeg-turbo, the fast method is generally about 5-15% faster than the int
-method when using the x86/x86-64 SIMD extensions (results may vary with other
-SIMD implementations, or when using libjpeg-turbo without SIMD extensions.)
+Use less accurate integer DCT method [legacy feature].
+When the Independent JPEG Group's software was first released in 1991, the
+compression time for a 1-megapixel JPEG image on a mainstream PC was measured
+in minutes. Thus, the \fBfast\fR integer DCT algorithm provided noticeable
+performance benefits. On modern CPUs running libjpeg-turbo, however, the
+compression time for a 1-megapixel JPEG image is measured in milliseconds, and
+thus the performance benefits of the \fBfast\fR algorithm are much less
+noticeable. On modern x86/x86-64 CPUs that support AVX2 instructions, the
+\fBfast\fR and \fBint\fR methods have similar performance. On other types of
+CPUs, the \fBfast\fR method is generally about 5-15% faster than the \fBint\fR
+method.
+
For quality levels of 90 and below, there should be little or no perceptible
-difference between the two algorithms. For quality levels above 90, however,
-the difference between the fast and the int methods becomes more pronounced.
-With quality=97, for instance, the fast method incurs generally about a 1-3 dB
-loss (in PSNR) relative to the int method, but this can be larger for some
-images. Do not use the fast method with quality levels above 97. The
-algorithm often degenerates at quality=98 and above and can actually produce a
-more lossy image than if lower quality levels had been used. Also, in
-libjpeg-turbo, the fast method is not fully accelerated for quality levels
-above 97, so it will be slower than the int method.
+quality difference between the two algorithms. For quality levels above 90,
+however, the difference between the \fBfast\fR and \fBint\fR methods becomes
+more pronounced. With quality=97, for instance, the \fBfast\fR method incurs
+generally about a 1-3 dB loss in PSNR relative to the \fBint\fR method, but
+this can be larger for some images. Do not use the \fBfast\fR method with
+quality levels above 97. The algorithm often degenerates at quality=98 and
+above and can actually produce a more lossy image than if lower quality levels
+had been used. Also, in libjpeg-turbo, the \fBfast\fR method is not fully
+accelerated for quality levels above 97, so it will be slower than the
+\fBint\fR method.
.TP
.B \-dct float
-Use floating-point DCT method.
-The float method is mainly a legacy feature. It does not produce significantly
-more accurate results than the int method, and it is much slower. The float
-method may also give different results on different machines due to varying
-roundoff behavior, whereas the integer methods should give the same results on
-all machines.
+Use floating-point DCT method [legacy feature].
+The \fBfloat\fR method does not produce significantly more accurate results
+than the \fBint\fR method, and it is much slower. The \fBfloat\fR method may
+also give different results on different machines due to varying roundoff
+behavior, whereas the integer methods should give the same results on all
+machines.
.TP
.BI \-icc " file"
Embed ICC color management profile contained in the specified file.
@@ -215,6 +223,9 @@
way of testing the in-memory destination manager (jpeg_mem_dest()), but it is
also useful for benchmarking, since it reduces the I/O overhead.
.TP
+.BI \-report
+Report compression progress.
+.TP
.B \-verbose
Enable debug printout. More
.BR \-v 's
@@ -341,11 +352,6 @@
relevant to libjpeg-turbo, to wordsmith certain sections, and to describe
features not present in libjpeg.
.SH ISSUES
-Support for GIF input files was removed in cjpeg v6b due to concerns over
-the Unisys LZW patent. Although this patent expired in 2006, cjpeg still
-lacks GIF support, for these historical reasons. (Conversion of GIF files to
-JPEG is usually a bad idea anyway, since GIF is a 256-color format.)
-.PP
Not all variants of BMP and Targa file formats are supported.
.PP
The
diff --git a/cjpeg.c b/cjpeg.c
index e80d16b..1c2d632 100644
--- a/cjpeg.c
+++ b/cjpeg.c
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1998, Thomas G. Lane.
* Modified 2003-2011 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2010, 2013-2014, 2017, D. R. Commander.
+ * Copyright (C) 2010, 2013-2014, 2017, 2019-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -69,9 +69,9 @@
* 2) assume we can push back more than one character (works in
* some C implementations, but unportable);
* 3) provide our own buffering (breaks input readers that want to use
- * stdio directly, such as the RLE library);
+ * stdio directly);
* or 4) don't put back the data, and modify the input_init methods to assume
- * they start reading after the start of file (also breaks RLE library).
+ * they start reading after the start of file.
* #1 is attractive for MS-DOS but is untenable on Unix.
*
* The most portable solution for file types that can't be identified by their
@@ -117,10 +117,6 @@
case 'P':
return jinit_read_ppm(cinfo);
#endif
-#ifdef RLE_SUPPORTED
- case 'R':
- return jinit_read_rle(cinfo);
-#endif
#ifdef TARGA_SUPPORTED
case 0x00:
return jinit_read_targa(cinfo);
@@ -146,7 +142,8 @@
static const char *progname; /* program name for error messages */
static char *icc_filename; /* for -icc switch */
static char *outfilename; /* for -outfile switch */
-boolean memdst; /* for -memdst switch */
+static boolean memdst; /* for -memdst switch */
+static boolean report; /* for -report switch */
LOCAL(void)
@@ -179,15 +176,15 @@
fprintf(stderr, " -arithmetic Use arithmetic coding\n");
#endif
#ifdef DCT_ISLOW_SUPPORTED
- fprintf(stderr, " -dct int Use integer DCT method%s\n",
+ fprintf(stderr, " -dct int Use accurate integer DCT method%s\n",
(JDCT_DEFAULT == JDCT_ISLOW ? " (default)" : ""));
#endif
#ifdef DCT_IFAST_SUPPORTED
- fprintf(stderr, " -dct fast Use fast integer DCT (less accurate)%s\n",
+ fprintf(stderr, " -dct fast Use less accurate integer DCT method [legacy feature]%s\n",
(JDCT_DEFAULT == JDCT_IFAST ? " (default)" : ""));
#endif
#ifdef DCT_FLOAT_SUPPORTED
- fprintf(stderr, " -dct float Use floating-point DCT method%s\n",
+ fprintf(stderr, " -dct float Use floating-point DCT method [legacy feature]%s\n",
(JDCT_DEFAULT == JDCT_FLOAT ? " (default)" : ""));
#endif
fprintf(stderr, " -icc FILE Embed ICC profile contained in FILE\n");
@@ -200,6 +197,7 @@
#if JPEG_LIB_VERSION >= 80 || defined(MEM_SRCDST_SUPPORTED)
fprintf(stderr, " -memdst Compress to memory instead of file (useful for benchmarking)\n");
#endif
+ fprintf(stderr, " -report Report compression progress\n");
fprintf(stderr, " -verbose or -debug Emit debug output\n");
fprintf(stderr, " -version Print version information and exit\n");
fprintf(stderr, "Switches for wizards:\n");
@@ -244,6 +242,7 @@
icc_filename = NULL;
outfilename = NULL;
memdst = FALSE;
+ report = FALSE;
cinfo->err->trace_level = 0;
/* Scan command line options, adjust parameters */
@@ -395,6 +394,9 @@
qtablefile = argv[argn];
/* We postpone actually reading the file in case -quality comes later. */
+ } else if (keymatch(arg, "report", 3)) {
+ report = TRUE;
+
} else if (keymatch(arg, "restart", 1)) {
/* Restart interval in MCU rows (or in MCUs with 'b'). */
long lval;
@@ -509,9 +511,7 @@
{
struct jpeg_compress_struct cinfo;
struct jpeg_error_mgr jerr;
-#ifdef PROGRESS_REPORT
struct cdjpeg_progress_mgr progress;
-#endif
int file_index;
cjpeg_source_ptr src_mgr;
FILE *input_file;
@@ -632,9 +632,10 @@
fclose(icc_file);
}
-#ifdef PROGRESS_REPORT
- start_progress_monitor((j_common_ptr)&cinfo, &progress);
-#endif
+ if (report) {
+ start_progress_monitor((j_common_ptr)&cinfo, &progress);
+ progress.report = report;
+ }
/* Figure out the input file format, and set up to read it. */
src_mgr = select_file_type(&cinfo, input_file);
@@ -680,9 +681,8 @@
if (output_file != stdout && output_file != NULL)
fclose(output_file);
-#ifdef PROGRESS_REPORT
- end_progress_monitor((j_common_ptr)&cinfo);
-#endif
+ if (report)
+ end_progress_monitor((j_common_ptr)&cinfo);
if (memdst) {
fprintf(stderr, "Compressed size: %lu bytes\n", outsize);
diff --git a/croptest.in b/croptest.in
new file mode 100644
index 0000000..7e3c293
--- /dev/null
+++ b/croptest.in
@@ -0,0 +1,95 @@
+#!/bin/bash
+
+set -u
+set -e
+trap onexit INT
+trap onexit TERM
+trap onexit EXIT
+
+onexit()
+{
+ if [ -d $OUTDIR ]; then
+ rm -rf $OUTDIR
+ fi
+}
+
+runme()
+{
+ echo \*\*\* $*
+ $*
+}
+
+IMAGE=vgl_6548_0026a.bmp
+WIDTH=128
+HEIGHT=95
+IMGDIR=@CMAKE_CURRENT_SOURCE_DIR@/testimages
+OUTDIR=`mktemp -d /tmp/__croptest_output.XXXXXX`
+EXEDIR=@CMAKE_CURRENT_BINARY_DIR@
+
+if [ -d $OUTDIR ]; then
+ rm -rf $OUTDIR
+fi
+mkdir -p $OUTDIR
+
+exec >$EXEDIR/croptest.log
+
+echo "============================================================"
+echo "$IMAGE ($WIDTH x $HEIGHT)"
+echo "============================================================"
+echo
+
+for PROGARG in "" -progressive; do
+
+ cp $IMGDIR/$IMAGE $OUTDIR
+ basename=`basename $IMAGE .bmp`
+ echo "------------------------------------------------------------"
+ echo "Generating test images"
+ echo "------------------------------------------------------------"
+ echo
+ runme $EXEDIR/cjpeg $PROGARG -grayscale -outfile $OUTDIR/${basename}_GRAY.jpg $IMGDIR/${basename}.bmp
+ runme $EXEDIR/cjpeg $PROGARG -sample 2x2 -outfile $OUTDIR/${basename}_420.jpg $IMGDIR/${basename}.bmp
+ runme $EXEDIR/cjpeg $PROGARG -sample 2x1 -outfile $OUTDIR/${basename}_422.jpg $IMGDIR/${basename}.bmp
+ runme $EXEDIR/cjpeg $PROGARG -sample 1x2 -outfile $OUTDIR/${basename}_440.jpg $IMGDIR/${basename}.bmp
+ runme $EXEDIR/cjpeg $PROGARG -sample 1x1 -outfile $OUTDIR/${basename}_444.jpg $IMGDIR/${basename}.bmp
+ echo
+
+ for NSARG in "" -nosmooth; do
+
+ for COLORSARG in "" "-colors 256 -dither none -onepass"; do
+
+ for Y in {0..16}; do
+
+ for H in {1..16}; do
+
+ X=$(( (Y*16)%128 ))
+ W=$(( WIDTH-X-7 ))
+ if [ $Y -le 15 ]; then
+ CROPSPEC="${W}x${H}+${X}+${Y}"
+ else
+ Y2=$(( HEIGHT-H ));
+ CROPSPEC="${W}x${H}+${X}+${Y2}"
+ fi
+
+ echo "------------------------------------------------------------"
+ echo $PROGARG $NSARG $COLORSARG -crop $CROPSPEC
+ echo "------------------------------------------------------------"
+ echo
+ for samp in GRAY 420 422 440 444; do
+ $EXEDIR/djpeg $NSARG $COLORSARG -rgb -outfile $OUTDIR/${basename}_${samp}_full.ppm $OUTDIR/${basename}_${samp}.jpg
+ convert -crop $CROPSPEC $OUTDIR/${basename}_${samp}_full.ppm $OUTDIR/${basename}_${samp}_ref.ppm
+ runme $EXEDIR/djpeg $NSARG $COLORSARG -crop $CROPSPEC -rgb -outfile $OUTDIR/${basename}_${samp}.ppm $OUTDIR/${basename}_${samp}.jpg
+ runme cmp $OUTDIR/${basename}_${samp}.ppm $OUTDIR/${basename}_${samp}_ref.ppm
+ done
+ echo
+
+ done
+
+ done
+
+ done
+
+ done
+
+done
+
+echo SUCCESS!
diff --git a/djpeg.1 b/djpeg.1
index e4204b2..31431b9 100644
--- a/djpeg.1
+++ b/djpeg.1
@@ -1,4 +1,4 @@
-.TH DJPEG 1 "13 November 2017"
+.TH DJPEG 1 "4 November 2020"
.SH NAME
djpeg \- decompress a JPEG file to an image file
.SH SYNOPSIS
@@ -15,8 +15,7 @@
.B djpeg
decompresses the named JPEG file, or the standard input if no file is named,
and produces an image file on the standard output. PBMPLUS (PPM/PGM), BMP,
-GIF, Targa, or RLE (Utah Raster Toolkit) output format can be selected.
-(RLE is supported only if the URT library is available.)
+GIF, or Targa output format can be selected.
.SH OPTIONS
All switch names may be abbreviated; for example,
.B \-grayscale
@@ -81,9 +80,20 @@
format is emitted.
.TP
.B \-gif
-Select GIF output format. Since GIF does not support more than 256 colors,
+Select GIF output format (LZW-compressed). Since GIF does not support more
+than 256 colors,
.B \-colors 256
-is assumed (unless you specify a smaller number of colors).
+is assumed (unless you specify a smaller number of colors). If you specify
+.BR \-fast,
+the default number of colors is 216.
+.TP
+.B \-gif0
+Select GIF output format (uncompressed). Since GIF does not support more than
+256 colors,
+.B \-colors 256
+is assumed (unless you specify a smaller number of colors). If you specify
+.BR \-fast,
+the default number of colors is 216.
.TP
.B \-os2
Select BMP output format (OS/2 1.x flavor). 8-bit colormapped format is
@@ -100,9 +110,6 @@
.B \-grayscale
is specified; otherwise PPM is emitted.
.TP
-.B \-rle
-Select RLE output format. (Requires URT library.)
-.TP
.B \-targa
Select Targa output format. Grayscale format is emitted if the JPEG file is
grayscale or if
@@ -114,32 +121,40 @@
Switches for advanced users:
.TP
.B \-dct int
-Use integer DCT method (default).
+Use accurate integer DCT method (default).
.TP
.B \-dct fast
-Use fast integer DCT (less accurate).
-In libjpeg-turbo, the fast method is generally about 5-15% faster than the int
-method when using the x86/x86-64 SIMD extensions (results may vary with other
-SIMD implementations, or when using libjpeg-turbo without SIMD extensions.) If
-the JPEG image was compressed using a quality level of 85 or below, then there
-should be little or no perceptible difference between the two algorithms. When
-decompressing images that were compressed using quality levels above 85,
-however, the difference between the fast and int methods becomes more
-pronounced. With images compressed using quality=97, for instance, the fast
-method incurs generally about a 4-6 dB loss (in PSNR) relative to the int
-method, but this can be larger for some images. If you can avoid it, do not
-use the fast method when decompressing images that were compressed using
-quality levels above 97. The algorithm often degenerates for such images and
-can actually produce a more lossy output image than if the JPEG image had been
-compressed using lower quality levels.
+Use less accurate integer DCT method [legacy feature].
+When the Independent JPEG Group's software was first released in 1991, the
+decompression time for a 1-megapixel JPEG image on a mainstream PC was measured
+in minutes. Thus, the \fBfast\fR integer DCT algorithm provided noticeable
+performance benefits. On modern CPUs running libjpeg-turbo, however, the
+decompression time for a 1-megapixel JPEG image is measured in milliseconds,
+and thus the performance benefits of the \fBfast\fR algorithm are much less
+noticeable. On modern x86/x86-64 CPUs that support AVX2 instructions, the
+\fBfast\fR and \fBint\fR methods have similar performance. On other types of
+CPUs, the \fBfast\fR method is generally about 5-15% faster than the \fBint\fR
+method.
+
+If the JPEG image was compressed using a quality level of 85 or below, then
+there should be little or no perceptible quality difference between the two
+algorithms. When decompressing images that were compressed using quality
+levels above 85, however, the difference between the \fBfast\fR and \fBint\fR
+methods becomes more pronounced. With images compressed using quality=97, for
+instance, the \fBfast\fR method incurs generally about a 4-6 dB loss in PSNR
+relative to the \fBint\fR method, but this can be larger for some images. If
+you can avoid it, do not use the \fBfast\fR method when decompressing images
+that were compressed using quality levels above 97. The algorithm often
+degenerates for such images and can actually produce a more lossy output image
+than if the JPEG image had been compressed using lower quality levels.
.TP
.B \-dct float
-Use floating-point DCT method.
-The float method is mainly a legacy feature. It does not produce significantly
-more accurate results than the int method, and it is much slower. The float
-method may also give different results on different machines due to varying
-roundoff behavior, whereas the integer methods should give the same results on
-all machines.
+Use floating-point DCT method [legacy feature].
+The \fBfloat\fR method does not produce significantly more accurate results
+than the \fBint\fR method, and it is much slower. The \fBfloat\fR method may
+also give different results on different machines due to varying roundoff
+behavior, whereas the integer methods should give the same results on all
+machines.
.TP
.B \-dither fs
Use Floyd-Steinberg dithering in color quantization.
@@ -190,6 +205,19 @@
.B \-max 4m
selects 4000000 bytes. If more space is needed, an error will occur.
.TP
+.BI \-maxscans " N"
+Abort if the JPEG image contains more than
+.I N
+scans. This feature demonstrates a method by which applications can guard
+against denial-of-service attacks instigated by specially-crafted malformed
+JPEG images containing numerous scans with missing image data or image data
+consisting only of "EOB runs" (a feature of progressive JPEG images that allows
+potentially hundreds of thousands of adjoining zero-value pixels to be
+represented using only a few bytes.) Attempting to decompress such malformed
+JPEG images can cause excessive CPU activity, since the decompressor must fully
+process each scan (even if the scan is corrupt) before it can proceed to the
+next scan.
+.TP
.BI \-outfile " name"
Send output image to the named file, not to standard output.
.TP
@@ -197,6 +225,9 @@
Load input file into memory before decompressing. This feature was implemented
mainly as a way of testing the in-memory source manager (jpeg_mem_src().)
.TP
+.BI \-report
+Report decompression progress.
+.TP
.BI \-skip " Y0,Y1"
Decompress all rows of the JPEG image except those between Y0 and Y1
(inclusive.) Note that if decompression scaling is being used, then Y0 and Y1
@@ -210,6 +241,12 @@
scaled image dimensions. Currently this option only works with the
PBMPLUS (PPM/PGM), GIF, and Targa output formats.
.TP
+.BI \-strict
+Treat all warnings as fatal. This feature also demonstrates a method by which
+applications can guard against attacks instigated by specially-crafted
+malformed JPEG images. Enabling this option will cause the decompressor to
+abort if the JPEG image contains incomplete or corrupt image data.
+.TP
.B \-verbose
Enable debug printout. More
.BR \-v 's
@@ -253,12 +290,6 @@
.B \-dither none
may give acceptable results in two-pass mode, but is seldom tolerable in
one-pass mode.
-.PP
-If you are fortunate enough to have very fast floating point hardware,
-\fB\-dct float\fR may be even faster than \fB\-dct fast\fR. But on most
-machines \fB\-dct float\fR is slower than \fB\-dct int\fR; in this case it is
-not worth using, because its theoretical accuracy advantage is too small to be
-significant in practice.
.SH ENVIRONMENT
.TP
.B JPEGMEM
@@ -287,10 +318,3 @@
This file was modified by The libjpeg-turbo Project to include only information
relevant to libjpeg-turbo, to wordsmith certain sections, and to describe
features not present in libjpeg.
-.SH ISSUES
-Support for compressed GIF output files was removed in djpeg v6b due to
-concerns over the Unisys LZW patent. Although this patent expired in 2006,
-djpeg still lacks compressed GIF support, for these historical reasons.
-(Conversion of JPEG files to GIF is usually a bad idea anyway, since GIF is a
-256-color format.) The uncompressed GIF files that djpeg generates are larger
-than they should be, but they are readable by standard GIF decoders.
diff --git a/djpeg.c b/djpeg.c
index 2489ccc..cc2eb9d 100644
--- a/djpeg.c
+++ b/djpeg.c
@@ -3,9 +3,9 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1997, Thomas G. Lane.
- * Modified 2013 by Guido Vollbeding.
+ * Modified 2013-2019 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2010-2011, 2013-2017, D. R. Commander.
+ * Copyright (C) 2010-2011, 2013-2017, 2019-2020, D. R. Commander.
* Copyright (C) 2015, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
@@ -68,10 +68,10 @@
typedef enum {
FMT_BMP, /* BMP format (Windows flavor) */
- FMT_GIF, /* GIF format */
+ FMT_GIF, /* GIF format (LZW-compressed) */
+ FMT_GIF0, /* GIF format (uncompressed) */
FMT_OS2, /* BMP format (OS/2 flavor) */
FMT_PPM, /* PPM/PGM (PBMPLUS formats) */
- FMT_RLE, /* RLE format */
FMT_TARGA, /* Targa format */
FMT_TIFF /* TIFF format */
} IMAGE_FORMATS;
@@ -94,11 +94,14 @@
static const char *progname; /* program name for error messages */
static char *icc_filename; /* for -icc switch */
+static JDIMENSION max_scans; /* for -maxscans switch */
static char *outfilename; /* for -outfile switch */
-boolean memsrc; /* for -memsrc switch */
+static boolean memsrc; /* for -memsrc switch */
+static boolean report; /* for -report switch */
boolean skip, crop;
JDIMENSION skip_start, skip_end;
JDIMENSION crop_x, crop_y, crop_width, crop_height;
+static boolean strict; /* for -strict switch */
#define INPUT_BUF_SIZE 4096
@@ -127,8 +130,10 @@
(DEFAULT_FMT == FMT_BMP ? " (default)" : ""));
#endif
#ifdef GIF_SUPPORTED
- fprintf(stderr, " -gif Select GIF output format%s\n",
+ fprintf(stderr, " -gif Select GIF output format (LZW-compressed)%s\n",
(DEFAULT_FMT == FMT_GIF ? " (default)" : ""));
+ fprintf(stderr, " -gif0 Select GIF output format (uncompressed)%s\n",
+ (DEFAULT_FMT == FMT_GIF0 ? " (default)" : ""));
#endif
#ifdef BMP_SUPPORTED
fprintf(stderr, " -os2 Select BMP output format (OS/2 style)%s\n",
@@ -138,25 +143,21 @@
fprintf(stderr, " -pnm Select PBMPLUS (PPM/PGM) output format%s\n",
(DEFAULT_FMT == FMT_PPM ? " (default)" : ""));
#endif
-#ifdef RLE_SUPPORTED
- fprintf(stderr, " -rle Select Utah RLE output format%s\n",
- (DEFAULT_FMT == FMT_RLE ? " (default)" : ""));
-#endif
#ifdef TARGA_SUPPORTED
fprintf(stderr, " -targa Select Targa output format%s\n",
(DEFAULT_FMT == FMT_TARGA ? " (default)" : ""));
#endif
fprintf(stderr, "Switches for advanced users:\n");
#ifdef DCT_ISLOW_SUPPORTED
- fprintf(stderr, " -dct int Use integer DCT method%s\n",
+ fprintf(stderr, " -dct int Use accurate integer DCT method%s\n",
(JDCT_DEFAULT == JDCT_ISLOW ? " (default)" : ""));
#endif
#ifdef DCT_IFAST_SUPPORTED
- fprintf(stderr, " -dct fast Use fast integer DCT (less accurate)%s\n",
+ fprintf(stderr, " -dct fast Use less accurate integer DCT method [legacy feature]%s\n",
(JDCT_DEFAULT == JDCT_IFAST ? " (default)" : ""));
#endif
#ifdef DCT_FLOAT_SUPPORTED
- fprintf(stderr, " -dct float Use floating-point DCT method%s\n",
+ fprintf(stderr, " -dct float Use floating-point DCT method [legacy feature]%s\n",
(JDCT_DEFAULT == JDCT_FLOAT ? " (default)" : ""));
#endif
fprintf(stderr, " -dither fs Use F-S dithering (default)\n");
@@ -171,14 +172,16 @@
fprintf(stderr, " -onepass Use 1-pass quantization (fast, low quality)\n");
#endif
fprintf(stderr, " -maxmemory N Maximum memory to use (in kbytes)\n");
+ fprintf(stderr, " -maxscans N Maximum number of scans to allow in input file\n");
fprintf(stderr, " -outfile name Specify name for output file\n");
#if JPEG_LIB_VERSION >= 80 || defined(MEM_SRCDST_SUPPORTED)
fprintf(stderr, " -memsrc Load input file into memory before decompressing\n");
#endif
-
+ fprintf(stderr, " -report Report decompression progress\n");
fprintf(stderr, " -skip Y0,Y1 Decompress all rows except those between Y0 and Y1 (inclusive)\n");
fprintf(stderr, " -crop WxH+X+Y Decompress only a rectangular subregion of the image\n");
fprintf(stderr, " [requires PBMPLUS (PPM/PGM), GIF, or Targa output format]\n");
+ fprintf(stderr, " -strict Treat all warnings as fatal\n");
fprintf(stderr, " -verbose or -debug Emit debug output\n");
fprintf(stderr, " -version Print version information and exit\n");
exit(EXIT_FAILURE);
@@ -203,10 +206,13 @@
/* Set up default JPEG parameters. */
requested_fmt = DEFAULT_FMT; /* set default output file format */
icc_filename = NULL;
+ max_scans = 0;
outfilename = NULL;
memsrc = FALSE;
+ report = FALSE;
skip = FALSE;
crop = FALSE;
+ strict = FALSE;
cinfo->err->trace_level = 0;
/* Scan command line options, adjust parameters */
@@ -224,7 +230,7 @@
arg++; /* advance past switch marker character */
if (keymatch(arg, "bmp", 1)) {
- /* BMP output format. */
+ /* BMP output format (Windows flavor). */
requested_fmt = FMT_BMP;
} else if (keymatch(arg, "colors", 1) || keymatch(arg, "colours", 1) ||
@@ -295,9 +301,13 @@
cinfo->do_fancy_upsampling = FALSE;
} else if (keymatch(arg, "gif", 1)) {
- /* GIF output format. */
+ /* GIF output format (LZW-compressed). */
requested_fmt = FMT_GIF;
+ } else if (keymatch(arg, "gif0", 4)) {
+ /* GIF output format (uncompressed). */
+ requested_fmt = FMT_GIF0;
+
} else if (keymatch(arg, "grayscale", 2) ||
keymatch(arg, "greyscale", 2)) {
/* Force monochrome output. */
@@ -351,6 +361,12 @@
lval *= 1000L;
cinfo->mem->max_memory_to_use = lval * 1000L;
+ } else if (keymatch(arg, "maxscans", 4)) {
+ if (++argn >= argc) /* advance to next argument */
+ usage();
+ if (sscanf(argv[argn], "%u", &max_scans) != 1)
+ usage();
+
} else if (keymatch(arg, "nosmooth", 3)) {
/* Suppress fancy upsampling */
cinfo->do_fancy_upsampling = FALSE;
@@ -383,9 +399,8 @@
/* PPM/PGM output format. */
requested_fmt = FMT_PPM;
- } else if (keymatch(arg, "rle", 1)) {
- /* RLE output format. */
- requested_fmt = FMT_RLE;
+ } else if (keymatch(arg, "report", 2)) {
+ report = TRUE;
} else if (keymatch(arg, "scale", 2)) {
/* Scale the output image by a fraction M/N. */
@@ -413,6 +428,9 @@
usage();
crop = TRUE;
+ } else if (keymatch(arg, "strict", 2)) {
+ strict = TRUE;
+
} else if (keymatch(arg, "targa", 1)) {
/* Targa output format. */
requested_fmt = FMT_TARGA;
@@ -444,7 +462,7 @@
ERREXIT(cinfo, JERR_CANT_SUSPEND);
}
datasrc->bytes_in_buffer--;
- return GETJOCTET(*datasrc->next_input_byte++);
+ return *datasrc->next_input_byte++;
}
@@ -499,6 +517,19 @@
}
+METHODDEF(void)
+my_emit_message(j_common_ptr cinfo, int msg_level)
+{
+ if (msg_level < 0) {
+ /* Treat warning as fatal */
+ cinfo->err->error_exit(cinfo);
+ } else {
+ if (cinfo->err->trace_level >= msg_level)
+ cinfo->err->output_message(cinfo);
+ }
+}
+
+
/*
* The main program.
*/
@@ -512,9 +543,7 @@
{
struct jpeg_decompress_struct cinfo;
struct jpeg_error_mgr jerr;
-#ifdef PROGRESS_REPORT
struct cdjpeg_progress_mgr progress;
-#endif
int file_index;
djpeg_dest_ptr dest_mgr = NULL;
FILE *input_file;
@@ -561,6 +590,9 @@
file_index = parse_switches(&cinfo, argc, argv, 0, FALSE);
+ if (strict)
+ jerr.emit_message = my_emit_message;
+
#ifdef TWO_FILE_COMMANDLINE
/* Must have either -outfile switch or explicit output file name */
if (outfilename == NULL) {
@@ -607,9 +639,11 @@
output_file = write_stdout();
}
-#ifdef PROGRESS_REPORT
- start_progress_monitor((j_common_ptr)&cinfo, &progress);
-#endif
+ if (report || max_scans != 0) {
+ start_progress_monitor((j_common_ptr)&cinfo, &progress);
+ progress.report = report;
+ progress.max_scans = max_scans;
+ }
/* Specify data source for decompression */
#if JPEG_LIB_VERSION >= 80 || defined(MEM_SRCDST_SUPPORTED)
@@ -657,7 +691,10 @@
#endif
#ifdef GIF_SUPPORTED
case FMT_GIF:
- dest_mgr = jinit_write_gif(&cinfo);
+ dest_mgr = jinit_write_gif(&cinfo, TRUE);
+ break;
+ case FMT_GIF0:
+ dest_mgr = jinit_write_gif(&cinfo, FALSE);
break;
#endif
#ifdef PPM_SUPPORTED
@@ -665,11 +702,6 @@
dest_mgr = jinit_write_ppm(&cinfo);
break;
#endif
-#ifdef RLE_SUPPORTED
- case FMT_RLE:
- dest_mgr = jinit_write_rle(&cinfo);
- break;
-#endif
#ifdef TARGA_SUPPORTED
case FMT_TARGA:
dest_mgr = jinit_write_targa(&cinfo);
@@ -712,7 +744,12 @@
dest_mgr->buffer_height);
(*dest_mgr->put_pixel_rows) (&cinfo, dest_mgr, num_scanlines);
}
- jpeg_skip_scanlines(&cinfo, skip_end - skip_start + 1);
+ if ((tmp = jpeg_skip_scanlines(&cinfo, skip_end - skip_start + 1)) !=
+ skip_end - skip_start + 1) {
+ fprintf(stderr, "%s: jpeg_skip_scanlines() returned %d rather than %d\n",
+ progname, tmp, skip_end - skip_start + 1);
+ return EXIT_FAILURE;
+ }
while (cinfo.output_scanline < cinfo.output_height) {
num_scanlines = jpeg_read_scanlines(&cinfo, dest_mgr->buffer,
dest_mgr->buffer_height);
@@ -748,13 +785,24 @@
cinfo.output_height = tmp;
/* Process data */
- jpeg_skip_scanlines(&cinfo, crop_y);
+ if ((tmp = jpeg_skip_scanlines(&cinfo, crop_y)) != crop_y) {
+ fprintf(stderr, "%s: jpeg_skip_scanlines() returned %d rather than %d\n",
+ progname, tmp, crop_y);
+ return EXIT_FAILURE;
+ }
while (cinfo.output_scanline < crop_y + crop_height) {
num_scanlines = jpeg_read_scanlines(&cinfo, dest_mgr->buffer,
dest_mgr->buffer_height);
(*dest_mgr->put_pixel_rows) (&cinfo, dest_mgr, num_scanlines);
}
- jpeg_skip_scanlines(&cinfo, cinfo.output_height - crop_y - crop_height);
+ if ((tmp =
+ jpeg_skip_scanlines(&cinfo,
+ cinfo.output_height - crop_y - crop_height)) !=
+ cinfo.output_height - crop_y - crop_height) {
+ fprintf(stderr, "%s: jpeg_skip_scanlines() returned %d rather than %d\n",
+ progname, tmp, cinfo.output_height - crop_y - crop_height);
+ return EXIT_FAILURE;
+ }
/* Normal full-image decompress */
} else {
@@ -769,12 +817,11 @@
}
}
-#ifdef PROGRESS_REPORT
/* Hack: count final pass as done in case finish_output does an extra pass.
* The library won't have updated completed_passes.
*/
- progress.pub.completed_passes = progress.pub.total_passes;
-#endif
+ if (report || max_scans != 0)
+ progress.pub.completed_passes = progress.pub.total_passes;
if (icc_filename != NULL) {
FILE *icc_file;
@@ -813,9 +860,8 @@
if (output_file != stdout)
fclose(output_file);
-#ifdef PROGRESS_REPORT
- end_progress_monitor((j_common_ptr)&cinfo);
-#endif
+ if (report || max_scans != 0)
+ end_progress_monitor((j_common_ptr)&cinfo);
if (memsrc)
free(inbuffer);
diff --git a/jccolext.c b/jccolext.c
index 19c955c..303b322 100644
--- a/jccolext.c
+++ b/jccolext.c
@@ -48,9 +48,9 @@
outptr2 = output_buf[2][output_row];
output_row++;
for (col = 0; col < num_cols; col++) {
- r = GETJSAMPLE(inptr[RGB_RED]);
- g = GETJSAMPLE(inptr[RGB_GREEN]);
- b = GETJSAMPLE(inptr[RGB_BLUE]);
+ r = inptr[RGB_RED];
+ g = inptr[RGB_GREEN];
+ b = inptr[RGB_BLUE];
inptr += RGB_PIXELSIZE;
/* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
* must be too; we do not need an explicit range-limiting operation.
@@ -100,9 +100,9 @@
outptr = output_buf[0][output_row];
output_row++;
for (col = 0; col < num_cols; col++) {
- r = GETJSAMPLE(inptr[RGB_RED]);
- g = GETJSAMPLE(inptr[RGB_GREEN]);
- b = GETJSAMPLE(inptr[RGB_BLUE]);
+ r = inptr[RGB_RED];
+ g = inptr[RGB_GREEN];
+ b = inptr[RGB_BLUE];
inptr += RGB_PIXELSIZE;
/* Y */
outptr[col] = (JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
@@ -135,9 +135,9 @@
outptr2 = output_buf[2][output_row];
output_row++;
for (col = 0; col < num_cols; col++) {
- outptr0[col] = GETJSAMPLE(inptr[RGB_RED]);
- outptr1[col] = GETJSAMPLE(inptr[RGB_GREEN]);
- outptr2[col] = GETJSAMPLE(inptr[RGB_BLUE]);
+ outptr0[col] = inptr[RGB_RED];
+ outptr1[col] = inptr[RGB_GREEN];
+ outptr2[col] = inptr[RGB_BLUE];
inptr += RGB_PIXELSIZE;
}
}
diff --git a/jccolor.c b/jccolor.c
index 036f601..bdc563c 100644
--- a/jccolor.c
+++ b/jccolor.c
@@ -392,11 +392,11 @@
outptr3 = output_buf[3][output_row];
output_row++;
for (col = 0; col < num_cols; col++) {
- r = MAXJSAMPLE - GETJSAMPLE(inptr[0]);
- g = MAXJSAMPLE - GETJSAMPLE(inptr[1]);
- b = MAXJSAMPLE - GETJSAMPLE(inptr[2]);
+ r = MAXJSAMPLE - inptr[0];
+ g = MAXJSAMPLE - inptr[1];
+ b = MAXJSAMPLE - inptr[2];
/* K passes through as-is */
- outptr3[col] = inptr[3]; /* don't need GETJSAMPLE here */
+ outptr3[col] = inptr[3];
inptr += 4;
/* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
* must be too; we do not need an explicit range-limiting operation.
@@ -438,7 +438,7 @@
outptr = output_buf[0][output_row];
output_row++;
for (col = 0; col < num_cols; col++) {
- outptr[col] = inptr[0]; /* don't need GETJSAMPLE() here */
+ outptr[col] = inptr[0];
inptr += instride;
}
}
@@ -497,7 +497,7 @@
inptr = *input_buf;
outptr = output_buf[ci][output_row];
for (col = 0; col < num_cols; col++) {
- outptr[col] = inptr[ci]; /* don't need GETJSAMPLE() here */
+ outptr[col] = inptr[ci];
inptr += nc;
}
}
diff --git a/jcdctmgr.c b/jcdctmgr.c
index c04058e..7dae17a 100644
--- a/jcdctmgr.c
+++ b/jcdctmgr.c
@@ -381,19 +381,19 @@
elemptr = sample_data[elemr] + start_col;
#if DCTSIZE == 8 /* unroll the inner loop */
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
#else
{
register int elemc;
for (elemc = DCTSIZE; elemc > 0; elemc--)
- *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+ *workspaceptr++ = (*elemptr++) - CENTERJSAMPLE;
}
#endif
}
@@ -533,20 +533,19 @@
for (elemr = 0; elemr < DCTSIZE; elemr++) {
elemptr = sample_data[elemr] + start_col;
#if DCTSIZE == 8 /* unroll the inner loop */
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
- *workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
#else
{
register int elemc;
for (elemc = DCTSIZE; elemc > 0; elemc--)
- *workspaceptr++ = (FAST_FLOAT)
- (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+ *workspaceptr++ = (FAST_FLOAT)((*elemptr++) - CENTERJSAMPLE);
}
#endif
}
diff --git a/jchuff.c b/jchuff.c
index cb05055..8ea48b8 100644
--- a/jchuff.c
+++ b/jchuff.c
@@ -4,8 +4,10 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2009-2011, 2014-2016, 2018-2019, D. R. Commander.
+ * Copyright (C) 2009-2011, 2014-2016, 2018-2020, D. R. Commander.
* Copyright (C) 2015, Matthieu Darbois.
+ * Copyright (C) 2018, Matthias Räncker.
+ * Copyright (C) 2020, Arm Limited.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -34,10 +36,10 @@
* memory footprint by 64k, which is important for some mobile applications
* that create many isolated instances of libjpeg-turbo (web browsers, for
* instance.) This may improve performance on some mobile platforms as well.
- * This feature is enabled by default only on ARM processors, because some x86
+ * This feature is enabled by default only on Arm processors, because some x86
* chips have a slow implementation of bsr, and the use of clz/bsr cannot be
* shown to have a significant performance impact even on the x86 chips that
- * have a fast implementation of it. When building for ARMv6, you can
+ * have a fast implementation of it. When building for Armv6, you can
* explicitly disable the use of clz/bsr by adding -mthumb to the compiler
* flags (this defines __thumb__).
*/
@@ -65,32 +67,43 @@
* but must not be updated permanently until we complete the MCU.
*/
+#if defined(__x86_64__) && defined(__ILP32__)
+typedef unsigned long long bit_buf_type;
+#else
+typedef size_t bit_buf_type;
+#endif
+
+/* NOTE: The more optimal Huffman encoding algorithm is only used by the
+ * intrinsics implementation of the Arm Neon SIMD extensions, which is why we
+ * retain the old Huffman encoder behavior when using the GAS implementation.
+ */
+#if defined(WITH_SIMD) && !(defined(__arm__) || defined(__aarch64__) || \
+ defined(_M_ARM) || defined(_M_ARM64))
+typedef unsigned long long simd_bit_buf_type;
+#else
+typedef bit_buf_type simd_bit_buf_type;
+#endif
+
+#if (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 8) || defined(_WIN64) || \
+ (defined(__x86_64__) && defined(__ILP32__))
+#define BIT_BUF_SIZE 64
+#elif (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 4) || defined(_WIN32)
+#define BIT_BUF_SIZE 32
+#else
+#error Cannot determine word size
+#endif
+#define SIMD_BIT_BUF_SIZE (sizeof(simd_bit_buf_type) * 8)
+
typedef struct {
- size_t put_buffer; /* current bit-accumulation buffer */
- int put_bits; /* # of bits now in it */
+ union {
+ bit_buf_type c;
+ simd_bit_buf_type simd;
+ } put_buffer; /* current bit accumulation buffer */
+ int free_bits; /* # of bits available in it */
+ /* (Neon GAS: # of bits now in it) */
int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
} savable_state;
-/* This macro is to work around compilers with missing or broken
- * structure assignment. You'll need to fix this code if you have
- * such a compiler and you change MAX_COMPS_IN_SCAN.
- */
-
-#ifndef NO_STRUCT_ASSIGN
-#define ASSIGN_STATE(dest, src) ((dest) = (src))
-#else
-#if MAX_COMPS_IN_SCAN == 4
-#define ASSIGN_STATE(dest, src) \
- ((dest).put_buffer = (src).put_buffer, \
- (dest).put_bits = (src).put_bits, \
- (dest).last_dc_val[0] = (src).last_dc_val[0], \
- (dest).last_dc_val[1] = (src).last_dc_val[1], \
- (dest).last_dc_val[2] = (src).last_dc_val[2], \
- (dest).last_dc_val[3] = (src).last_dc_val[3])
-#endif
-#endif
-
-
typedef struct {
struct jpeg_entropy_encoder pub; /* public fields */
@@ -123,6 +136,7 @@
size_t free_in_buffer; /* # of byte spaces remaining in buffer */
savable_state cur; /* Current bit buffer & DC state */
j_compress_ptr cinfo; /* dump_buffer needs access to this */
+ int simd;
} working_state;
@@ -201,8 +215,17 @@
}
/* Initialize bit buffer to empty */
- entropy->saved.put_buffer = 0;
- entropy->saved.put_bits = 0;
+ if (entropy->simd) {
+ entropy->saved.put_buffer.simd = 0;
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+ entropy->saved.free_bits = 0;
+#else
+ entropy->saved.free_bits = SIMD_BIT_BUF_SIZE;
+#endif
+ } else {
+ entropy->saved.put_buffer.c = 0;
+ entropy->saved.free_bits = BIT_BUF_SIZE;
+ }
/* Initialize restart stuff */
entropy->restarts_to_go = cinfo->restart_interval;
@@ -334,94 +357,94 @@
/* Outputting bits to the file */
-/* These macros perform the same task as the emit_bits() function in the
- * original libjpeg code. In addition to reducing overhead by explicitly
- * inlining the code, additional performance is achieved by taking into
- * account the size of the bit buffer and waiting until it is almost full
- * before emptying it. This mostly benefits 64-bit platforms, since 6
- * bytes can be stored in a 64-bit bit buffer before it has to be emptied.
+/* Output byte b and, speculatively, an additional 0 byte. 0xFF must be
+ * encoded as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the
+ * byte is 0xFF. Otherwise, the output buffer pointer is advanced by 1, and
+ * the speculative 0 byte will be overwritten by the next byte.
*/
-
-#define EMIT_BYTE() { \
- JOCTET c; \
- put_bits -= 8; \
- c = (JOCTET)GETJOCTET(put_buffer >> put_bits); \
- *buffer++ = c; \
- if (c == 0xFF) /* need to stuff a zero byte? */ \
- *buffer++ = 0; \
+#define EMIT_BYTE(b) { \
+ buffer[0] = (JOCTET)(b); \
+ buffer[1] = 0; \
+ buffer -= -2 + ((JOCTET)(b) < 0xFF); \
}
-#define PUT_BITS(code, size) { \
- put_bits += size; \
- put_buffer = (put_buffer << size) | code; \
-}
+/* Output the entire bit buffer. If there are no 0xFF bytes in it, then write
+ * directly to the output buffer. Otherwise, use the EMIT_BYTE() macro to
+ * encode 0xFF as 0xFF 0x00.
+ */
+#if BIT_BUF_SIZE == 64
-#if SIZEOF_SIZE_T != 8 && !defined(_WIN64)
-
-#define CHECKBUF15() { \
- if (put_bits > 15) { \
- EMIT_BYTE() \
- EMIT_BYTE() \
+#define FLUSH() { \
+ if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \
+ EMIT_BYTE(put_buffer >> 56) \
+ EMIT_BYTE(put_buffer >> 48) \
+ EMIT_BYTE(put_buffer >> 40) \
+ EMIT_BYTE(put_buffer >> 32) \
+ EMIT_BYTE(put_buffer >> 24) \
+ EMIT_BYTE(put_buffer >> 16) \
+ EMIT_BYTE(put_buffer >> 8) \
+ EMIT_BYTE(put_buffer ) \
+ } else { \
+ buffer[0] = (JOCTET)(put_buffer >> 56); \
+ buffer[1] = (JOCTET)(put_buffer >> 48); \
+ buffer[2] = (JOCTET)(put_buffer >> 40); \
+ buffer[3] = (JOCTET)(put_buffer >> 32); \
+ buffer[4] = (JOCTET)(put_buffer >> 24); \
+ buffer[5] = (JOCTET)(put_buffer >> 16); \
+ buffer[6] = (JOCTET)(put_buffer >> 8); \
+ buffer[7] = (JOCTET)(put_buffer); \
+ buffer += 8; \
} \
}
-#endif
-
-#define CHECKBUF31() { \
- if (put_bits > 31) { \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- } \
-}
-
-#define CHECKBUF47() { \
- if (put_bits > 47) { \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- EMIT_BYTE() \
- } \
-}
-
-#if !defined(_WIN32) && !defined(SIZEOF_SIZE_T)
-#error Cannot determine word size
-#endif
-
-#if SIZEOF_SIZE_T == 8 || defined(_WIN64)
-
-#define EMIT_BITS(code, size) { \
- CHECKBUF47() \
- PUT_BITS(code, size) \
-}
-
-#define EMIT_CODE(code, size) { \
- temp2 &= (((JLONG)1) << nbits) - 1; \
- CHECKBUF31() \
- PUT_BITS(code, size) \
- PUT_BITS(temp2, nbits) \
-}
-
#else
-#define EMIT_BITS(code, size) { \
- PUT_BITS(code, size) \
- CHECKBUF15() \
-}
-
-#define EMIT_CODE(code, size) { \
- temp2 &= (((JLONG)1) << nbits) - 1; \
- PUT_BITS(code, size) \
- CHECKBUF15() \
- PUT_BITS(temp2, nbits) \
- CHECKBUF15() \
+#define FLUSH() { \
+ if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \
+ EMIT_BYTE(put_buffer >> 24) \
+ EMIT_BYTE(put_buffer >> 16) \
+ EMIT_BYTE(put_buffer >> 8) \
+ EMIT_BYTE(put_buffer ) \
+ } else { \
+ buffer[0] = (JOCTET)(put_buffer >> 24); \
+ buffer[1] = (JOCTET)(put_buffer >> 16); \
+ buffer[2] = (JOCTET)(put_buffer >> 8); \
+ buffer[3] = (JOCTET)(put_buffer); \
+ buffer += 4; \
+ } \
}
#endif
+/* Fill the bit buffer to capacity with the leading bits from code, then output
+ * the bit buffer and put the remaining bits from code into the bit buffer.
+ */
+#define PUT_AND_FLUSH(code, size) { \
+ put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \
+ FLUSH() \
+ free_bits += BIT_BUF_SIZE; \
+ put_buffer = code; \
+}
+
+/* Insert code into the bit buffer and output the bit buffer if needed.
+ * NOTE: We can't flush with free_bits == 0, since the left shift in
+ * PUT_AND_FLUSH() would have undefined behavior.
+ */
+#define PUT_BITS(code, size) { \
+ free_bits -= size; \
+ if (free_bits < 0) \
+ PUT_AND_FLUSH(code, size) \
+ else \
+ put_buffer = (put_buffer << size) | code; \
+}
+
+#define PUT_CODE(code, size) { \
+ temp &= (((JLONG)1) << nbits) - 1; \
+ temp |= code << nbits; \
+ nbits += size; \
+ PUT_BITS(temp, nbits) \
+}
+
/* Although it is exceedingly rare, it is possible for a Huffman-encoded
* coefficient block to be larger than the 128-byte unencoded block. For each
@@ -444,6 +467,7 @@
#define STORE_BUFFER() { \
if (localbuf) { \
+ size_t bytes, bytestocopy; \
bytes = buffer - _buffer; \
buffer = _buffer; \
while (bytes > 0) { \
@@ -466,20 +490,46 @@
LOCAL(boolean)
flush_bits(working_state *state)
{
- JOCTET _buffer[BUFSIZE], *buffer;
- size_t put_buffer; int put_bits;
- size_t bytes, bytestocopy; int localbuf = 0;
+ JOCTET _buffer[BUFSIZE], *buffer, temp;
+ simd_bit_buf_type put_buffer; int put_bits;
+ int localbuf = 0;
- put_buffer = state->cur.put_buffer;
- put_bits = state->cur.put_bits;
+ if (state->simd) {
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+ put_bits = state->cur.free_bits;
+#else
+ put_bits = SIMD_BIT_BUF_SIZE - state->cur.free_bits;
+#endif
+ put_buffer = state->cur.put_buffer.simd;
+ } else {
+ put_bits = BIT_BUF_SIZE - state->cur.free_bits;
+ put_buffer = state->cur.put_buffer.c;
+ }
+
LOAD_BUFFER()
- /* fill any partial byte with ones */
- PUT_BITS(0x7F, 7)
- while (put_bits >= 8) EMIT_BYTE()
+ while (put_bits >= 8) {
+ put_bits -= 8;
+ temp = (JOCTET)(put_buffer >> put_bits);
+ EMIT_BYTE(temp)
+ }
+ if (put_bits) {
+ /* fill partial byte with ones */
+ temp = (JOCTET)((put_buffer << (8 - put_bits)) | (0xFF >> put_bits));
+ EMIT_BYTE(temp)
+ }
- state->cur.put_buffer = 0; /* and reset bit-buffer to empty */
- state->cur.put_bits = 0;
+ if (state->simd) { /* and reset bit buffer to empty */
+ state->cur.put_buffer.simd = 0;
+#if defined(__aarch64__) && !defined(NEON_INTRINSICS)
+ state->cur.free_bits = 0;
+#else
+ state->cur.free_bits = SIMD_BIT_BUF_SIZE;
+#endif
+ } else {
+ state->cur.put_buffer.c = 0;
+ state->cur.free_bits = BIT_BUF_SIZE;
+ }
STORE_BUFFER()
return TRUE;
@@ -493,7 +543,7 @@
c_derived_tbl *dctbl, c_derived_tbl *actbl)
{
JOCTET _buffer[BUFSIZE], *buffer;
- size_t bytes, bytestocopy; int localbuf = 0;
+ int localbuf = 0;
LOAD_BUFFER()
@@ -509,53 +559,41 @@
encode_one_block(working_state *state, JCOEFPTR block, int last_dc_val,
c_derived_tbl *dctbl, c_derived_tbl *actbl)
{
- int temp, temp2, temp3;
- int nbits;
- int r, code, size;
+ int temp, nbits, free_bits;
+ bit_buf_type put_buffer;
JOCTET _buffer[BUFSIZE], *buffer;
- size_t put_buffer; int put_bits;
- int code_0xf0 = actbl->ehufco[0xf0], size_0xf0 = actbl->ehufsi[0xf0];
- size_t bytes, bytestocopy; int localbuf = 0;
+ int localbuf = 0;
- put_buffer = state->cur.put_buffer;
- put_bits = state->cur.put_bits;
+ free_bits = state->cur.free_bits;
+ put_buffer = state->cur.put_buffer.c;
LOAD_BUFFER()
/* Encode the DC coefficient difference per section F.1.2.1 */
- temp = temp2 = block[0] - last_dc_val;
+ temp = block[0] - last_dc_val;
/* This is a well-known technique for obtaining the absolute value without a
* branch. It is derived from an assembly language technique presented in
* "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by
- * Agner Fog.
+ * Agner Fog. This code assumes we are on a two's complement machine.
*/
- temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
- temp ^= temp3;
- temp -= temp3;
-
- /* For a negative input, want temp2 = bitwise complement of abs(input) */
- /* This code assumes we are on a two's complement machine */
- temp2 += temp3;
+ nbits = temp >> (CHAR_BIT * sizeof(int) - 1);
+ temp += nbits;
+ nbits ^= temp;
/* Find the number of bits needed for the magnitude of the coefficient */
- nbits = JPEG_NBITS(temp);
+ nbits = JPEG_NBITS(nbits);
- /* Emit the Huffman-coded symbol for the number of bits */
- code = dctbl->ehufco[nbits];
- size = dctbl->ehufsi[nbits];
- EMIT_BITS(code, size)
-
- /* Mask off any extra bits in code */
- temp2 &= (((JLONG)1) << nbits) - 1;
-
- /* Emit that number of bits of the value, if positive, */
- /* or the complement of its magnitude, if negative. */
- EMIT_BITS(temp2, nbits)
+ /* Emit the Huffman-coded symbol for the number of bits.
+ * Emit that number of bits of the value, if positive,
+ * or the complement of its magnitude, if negative.
+ */
+ PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits])
/* Encode the AC coefficients per section F.1.2.2 */
- r = 0; /* r = run length of zeros */
+ {
+ int r = 0; /* r = run length of zeros */
/* Manually unroll the k loop to eliminate the counter variable. This
* improves performance greatly on systems with a limited number of
@@ -563,51 +601,46 @@
*/
#define kloop(jpeg_natural_order_of_k) { \
if ((temp = block[jpeg_natural_order_of_k]) == 0) { \
- r++; \
+ r += 16; \
} else { \
- temp2 = temp; \
/* Branch-less absolute value, bitwise complement, etc., same as above */ \
- temp3 = temp >> (CHAR_BIT * sizeof(int) - 1); \
- temp ^= temp3; \
- temp -= temp3; \
- temp2 += temp3; \
- nbits = JPEG_NBITS_NONZERO(temp); \
+ nbits = temp >> (CHAR_BIT * sizeof(int) - 1); \
+ temp += nbits; \
+ nbits ^= temp; \
+ nbits = JPEG_NBITS_NONZERO(nbits); \
/* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
- while (r > 15) { \
- EMIT_BITS(code_0xf0, size_0xf0) \
- r -= 16; \
+ while (r >= 16 * 16) { \
+ r -= 16 * 16; \
+ PUT_BITS(actbl->ehufco[0xf0], actbl->ehufsi[0xf0]) \
} \
/* Emit Huffman symbol for run length / number of bits */ \
- temp3 = (r << 4) + nbits; \
- code = actbl->ehufco[temp3]; \
- size = actbl->ehufsi[temp3]; \
- EMIT_CODE(code, size) \
+ r += nbits; \
+ PUT_CODE(actbl->ehufco[r], actbl->ehufsi[r]) \
r = 0; \
} \
}
- /* One iteration for each value in jpeg_natural_order[] */
- kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3);
- kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18);
- kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26);
- kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27);
- kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21);
- kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57);
- kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15);
- kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58);
- kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39);
- kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47);
- kloop(55); kloop(62); kloop(63);
+ /* One iteration for each value in jpeg_natural_order[] */
+ kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3);
+ kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18);
+ kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26);
+ kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27);
+ kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21);
+ kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57);
+ kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15);
+ kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58);
+ kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39);
+ kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47);
+ kloop(55); kloop(62); kloop(63);
- /* If the last coef(s) were zero, emit an end-of-block code */
- if (r > 0) {
- code = actbl->ehufco[0];
- size = actbl->ehufsi[0];
- EMIT_BITS(code, size)
+ /* If the last coef(s) were zero, emit an end-of-block code */
+ if (r > 0) {
+ PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+ }
}
- state->cur.put_buffer = put_buffer;
- state->cur.put_bits = put_bits;
+ state->cur.put_buffer.c = put_buffer;
+ state->cur.free_bits = free_bits;
STORE_BUFFER()
return TRUE;
@@ -654,8 +687,9 @@
/* Load up working state */
state.next_output_byte = cinfo->dest->next_output_byte;
state.free_in_buffer = cinfo->dest->free_in_buffer;
- ASSIGN_STATE(state.cur, entropy->saved);
+ state.cur = entropy->saved;
state.cinfo = cinfo;
+ state.simd = entropy->simd;
/* Emit restart marker if needed */
if (cinfo->restart_interval) {
@@ -694,7 +728,7 @@
/* Completed MCU, so update state */
cinfo->dest->next_output_byte = state.next_output_byte;
cinfo->dest->free_in_buffer = state.free_in_buffer;
- ASSIGN_STATE(entropy->saved, state.cur);
+ entropy->saved = state.cur;
/* Update restart-interval state too */
if (cinfo->restart_interval) {
@@ -723,8 +757,9 @@
/* Load up working state ... flush_bits needs it */
state.next_output_byte = cinfo->dest->next_output_byte;
state.free_in_buffer = cinfo->dest->free_in_buffer;
- ASSIGN_STATE(state.cur, entropy->saved);
+ state.cur = entropy->saved;
state.cinfo = cinfo;
+ state.simd = entropy->simd;
/* Flush out the last data */
if (!flush_bits(&state))
@@ -733,7 +768,7 @@
/* Update state */
cinfo->dest->next_output_byte = state.next_output_byte;
cinfo->dest->free_in_buffer = state.free_in_buffer;
- ASSIGN_STATE(entropy->saved, state.cur);
+ entropy->saved = state.cur;
}
diff --git a/jcinit.c b/jcinit.c
index 78aa465..157353a 100644
--- a/jcinit.c
+++ b/jcinit.c
@@ -1,8 +1,10 @@
/*
* jcinit.c
*
+ * This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1997, Thomas G. Lane.
- * This file is part of the Independent JPEG Group's software.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -19,6 +21,7 @@
#define JPEG_INTERNALS
#include "jinclude.h"
#include "jpeglib.h"
+#include "jpegcomp.h"
/*
diff --git a/jconfig.h b/jconfig.h
index 91e2fd3..7d89c0e 100644
--- a/jconfig.h
+++ b/jconfig.h
@@ -4,10 +4,10 @@
#define JPEG_LIB_VERSION 62
/* libjpeg-turbo version */
-#define LIBJPEG_TURBO_VERSION 2.0.1
+#define LIBJPEG_TURBO_VERSION 2.0.90
/* libjpeg-turbo version in integer form */
-#define LIBJPEG_TURBO_VERSION_NUMBER 2000001
+#define LIBJPEG_TURBO_VERSION_NUMBER 2000090
/* Support arithmetic encoding */
#define C_ARITH_CODING_SUPPORTED 1
@@ -30,7 +30,7 @@
* We do not support run-time selection of data precision, sorry.
*/
-#define BITS_IN_JSAMPLE 8 /* use 8 or 12 */
+#define BITS_IN_JSAMPLE 8 /* use 8 or 12 */
/* Define to 1 if you have the <locale.h> header file. */
#define HAVE_LOCALE_H 1
@@ -61,11 +61,6 @@
unsigned. */
/* #undef RIGHT_SHIFT_IS_UNSIGNED */
-/* Define to 1 if type `char' is unsigned and you are not using gcc. */
-#ifndef __CHAR_UNSIGNED__
-/* # undef __CHAR_UNSIGNED__ */
-#endif
-
/* Define to empty if `const' does not conform to ANSI C. */
/* #undef const */
diff --git a/jconfig.h.in b/jconfig.h.in
index 18a69a4..d4284d9 100644
--- a/jconfig.h.in
+++ b/jconfig.h.in
@@ -61,11 +61,6 @@
unsigned. */
#cmakedefine RIGHT_SHIFT_IS_UNSIGNED 1
-/* Define to 1 if type `char' is unsigned and you are not using gcc. */
-#ifndef __CHAR_UNSIGNED__
- #cmakedefine __CHAR_UNSIGNED__ 1
-#endif
-
/* Define to empty if `const' does not conform to ANSI C. */
/* #undef const */
diff --git a/jconfig.txt b/jconfig.txt
index 90cd724..21f35c1 100644
--- a/jconfig.txt
+++ b/jconfig.txt
@@ -42,12 +42,6 @@
*/
/* #define const */
-/* Define this if an ordinary "char" type is unsigned.
- * If you're not sure, leaving it undefined will work at some cost in speed.
- * If you defined HAVE_UNSIGNED_CHAR then the speed difference is minimal.
- */
-#undef __CHAR_UNSIGNED__
-
/* Define this if your system has an ANSI-conforming <stddef.h> file.
*/
#define HAVE_STDDEF_H
@@ -118,7 +112,6 @@
#define BMP_SUPPORTED /* BMP image file format */
#define GIF_SUPPORTED /* GIF image file format */
#define PPM_SUPPORTED /* PBMPLUS PPM/PGM image file format */
-#undef RLE_SUPPORTED /* Utah RLE image file format */
#define TARGA_SUPPORTED /* Targa image file format */
/* Define this if you want to name both input and output files on the command
diff --git a/jconfigint.h b/jconfigint.h
index 974890d..62b41fc 100644
--- a/jconfigint.h
+++ b/jconfigint.h
@@ -26,7 +26,7 @@
#define PACKAGE_NAME "libjpeg-turbo"
/* Version number of package */
-#define VERSION "2.0.5"
+#define VERSION "2.0.90"
/* The size of `size_t', as computed by sizeof. */
#if __WORDSIZE==64 || defined(_WIN64)
@@ -52,12 +52,3 @@
#define HAVEBITSCANFORWARD
#endif
#endif
-
-/* How to obtain memory alignment for structures and variables. */
-#if defined(_MSC_VER)
-#define ALIGN(ALIGNMENT) __declspec(align((ALIGNMENT)))
-#elif defined(__clang__) || defined(__GNUC__)
-#define ALIGN(ALIGNMENT) __attribute__((aligned(ALIGNMENT)))
-#else
-#error "Unknown compiler"
-#endif
diff --git a/jcphuff.c b/jcphuff.c
index 8c4efaf..a8b94be 100644
--- a/jcphuff.c
+++ b/jcphuff.c
@@ -43,10 +43,10 @@
* memory footprint by 64k, which is important for some mobile applications
* that create many isolated instances of libjpeg-turbo (web browsers, for
* instance.) This may improve performance on some mobile platforms as well.
- * This feature is enabled by default only on ARM processors, because some x86
+ * This feature is enabled by default only on Arm processors, because some x86
* chips have a slow implementation of bsr, and the use of clz/bsr cannot be
* shown to have a significant performance impact even on the x86 chips that
- * have a fast implementation of it. When building for ARMv6, you can
+ * have a fast implementation of it. When building for Armv6, you can
* explicitly disable the use of clz/bsr by adding -mthumb to the compiler
* flags (this defines __thumb__).
*/
diff --git a/jcsample.c b/jcsample.c
index bd27b84..e8515eb 100644
--- a/jcsample.c
+++ b/jcsample.c
@@ -6,7 +6,7 @@
* libjpeg-turbo Modifications:
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
* Copyright (C) 2014, MIPS Technologies, Inc., California.
- * Copyright (C) 2015, D. R. Commander.
+ * Copyright (C) 2015, 2019, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -103,7 +103,7 @@
if (numcols > 0) {
for (row = 0; row < num_rows; row++) {
ptr = image_data[row] + input_cols;
- pixval = ptr[-1]; /* don't need GETJSAMPLE() here */
+ pixval = ptr[-1];
for (count = numcols; count > 0; count--)
*ptr++ = pixval;
}
@@ -174,7 +174,7 @@
for (v = 0; v < v_expand; v++) {
inptr = input_data[inrow + v] + outcol_h;
for (h = 0; h < h_expand; h++) {
- outvalue += (JLONG)GETJSAMPLE(*inptr++);
+ outvalue += (JLONG)(*inptr++);
}
}
*outptr++ = (JSAMPLE)((outvalue + numpix2) / numpix);
@@ -237,8 +237,7 @@
inptr = input_data[outrow];
bias = 0; /* bias = 0,1,0,1,... for successive samples */
for (outcol = 0; outcol < output_cols; outcol++) {
- *outptr++ =
- (JSAMPLE)((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1]) + bias) >> 1);
+ *outptr++ = (JSAMPLE)((inptr[0] + inptr[1] + bias) >> 1);
bias ^= 1; /* 0=>1, 1=>0 */
inptr += 2;
}
@@ -277,8 +276,7 @@
bias = 1; /* bias = 1,2,1,2,... for successive samples */
for (outcol = 0; outcol < output_cols; outcol++) {
*outptr++ =
- (JSAMPLE)((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
- GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]) + bias) >> 2);
+ (JSAMPLE)((inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1] + bias) >> 2);
bias ^= 3; /* 1=>2, 2=>1 */
inptr0 += 2; inptr1 += 2;
}
@@ -337,33 +335,25 @@
below_ptr = input_data[inrow + 2];
/* Special case for first column: pretend column -1 is same as column 0 */
- membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
- GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
- neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
- GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
- GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
- GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
+ membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
+ neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+ inptr0[0] + inptr0[2] + inptr1[0] + inptr1[2];
neighsum += neighsum;
- neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
- GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
+ neighsum += above_ptr[0] + above_ptr[2] + below_ptr[0] + below_ptr[2];
membersum = membersum * memberscale + neighsum * neighscale;
*outptr++ = (JSAMPLE)((membersum + 32768) >> 16);
inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
for (colctr = output_cols - 2; colctr > 0; colctr--) {
/* sum of pixels directly mapped to this output element */
- membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
- GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+ membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
/* sum of edge-neighbor pixels */
- neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
- GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
- GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
- GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
+ neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+ inptr0[-1] + inptr0[2] + inptr1[-1] + inptr1[2];
/* The edge-neighbors count twice as much as corner-neighbors */
neighsum += neighsum;
/* Add in the corner-neighbors */
- neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
- GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
+ neighsum += above_ptr[-1] + above_ptr[2] + below_ptr[-1] + below_ptr[2];
/* form final output scaled up by 2^16 */
membersum = membersum * memberscale + neighsum * neighscale;
/* round, descale and output it */
@@ -372,15 +362,11 @@
}
/* Special case for last column */
- membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
- GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
- neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
- GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
- GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
- GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
+ membersum = inptr0[0] + inptr0[1] + inptr1[0] + inptr1[1];
+ neighsum = above_ptr[0] + above_ptr[1] + below_ptr[0] + below_ptr[1] +
+ inptr0[-1] + inptr0[1] + inptr1[-1] + inptr1[1];
neighsum += neighsum;
- neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
- GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
+ neighsum += above_ptr[-1] + above_ptr[1] + below_ptr[-1] + below_ptr[1];
membersum = membersum * memberscale + neighsum * neighscale;
*outptr = (JSAMPLE)((membersum + 32768) >> 16);
@@ -429,21 +415,18 @@
below_ptr = input_data[outrow + 1];
/* Special case for first column */
- colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
- GETJSAMPLE(*inptr);
- membersum = GETJSAMPLE(*inptr++);
- nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
- GETJSAMPLE(*inptr);
+ colsum = (*above_ptr++) + (*below_ptr++) + inptr[0];
+ membersum = *inptr++;
+ nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
neighsum = colsum + (colsum - membersum) + nextcolsum;
membersum = membersum * memberscale + neighsum * neighscale;
*outptr++ = (JSAMPLE)((membersum + 32768) >> 16);
lastcolsum = colsum; colsum = nextcolsum;
for (colctr = output_cols - 2; colctr > 0; colctr--) {
- membersum = GETJSAMPLE(*inptr++);
+ membersum = *inptr++;
above_ptr++; below_ptr++;
- nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
- GETJSAMPLE(*inptr);
+ nextcolsum = above_ptr[0] + below_ptr[0] + inptr[0];
neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
membersum = membersum * memberscale + neighsum * neighscale;
*outptr++ = (JSAMPLE)((membersum + 32768) >> 16);
@@ -451,7 +434,7 @@
}
/* Special case for last column */
- membersum = GETJSAMPLE(*inptr);
+ membersum = *inptr;
neighsum = lastcolsum + (colsum - membersum) + colsum;
membersum = membersum * memberscale + neighsum * neighscale;
*outptr = (JSAMPLE)((membersum + 32768) >> 16);
diff --git a/jctrans.c b/jctrans.c
index ce70a30..ab6a218 100644
--- a/jctrans.c
+++ b/jctrans.c
@@ -4,8 +4,8 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1995-1998, Thomas G. Lane.
* Modified 2000-2009 by Guido Vollbeding.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -17,6 +17,7 @@
#define JPEG_INTERNALS
#include "jinclude.h"
#include "jpeglib.h"
+#include "jpegcomp.h"
/* Forward declarations */
diff --git a/jdapistd.c b/jdapistd.c
index 2c808fa..695a620 100644
--- a/jdapistd.c
+++ b/jdapistd.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1994-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2010, 2015-2018, D. R. Commander.
+ * Copyright (C) 2010, 2015-2020, D. R. Commander.
* Copyright (C) 2015, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
@@ -21,6 +21,8 @@
#include "jinclude.h"
#include "jdmainct.h"
#include "jdcoefct.h"
+#include "jdmaster.h"
+#include "jdmerge.h"
#include "jdsample.h"
#include "jmemsys.h"
@@ -316,6 +318,10 @@
read_and_discard_scanlines(j_decompress_ptr cinfo, JDIMENSION num_lines)
{
JDIMENSION n;
+ my_master_ptr master = (my_master_ptr)cinfo->master;
+ JSAMPLE dummy_sample[1] = { 0 };
+ JSAMPROW dummy_row = dummy_sample;
+ JSAMPARRAY scanlines = NULL;
void (*color_convert) (j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
JDIMENSION input_row, JSAMPARRAY output_buf,
int num_rows) = NULL;
@@ -325,6 +331,10 @@
if (cinfo->cconvert && cinfo->cconvert->color_convert) {
color_convert = cinfo->cconvert->color_convert;
cinfo->cconvert->color_convert = noop_convert;
+ /* This just prevents UBSan from complaining about adding 0 to a NULL
+ * pointer. The pointer isn't actually used.
+ */
+ scanlines = &dummy_row;
}
if (cinfo->cquantize && cinfo->cquantize->color_quantize) {
@@ -332,8 +342,13 @@
cinfo->cquantize->color_quantize = noop_quantize;
}
+ if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
+ scanlines = &upsample->spare_row;
+ }
+
for (n = 0; n < num_lines; n++)
- jpeg_read_scanlines(cinfo, NULL, 1);
+ jpeg_read_scanlines(cinfo, scanlines, 1);
if (color_convert)
cinfo->cconvert->color_convert = color_convert;
@@ -353,6 +368,12 @@
{
JDIMENSION rows_left;
my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
+ my_master_ptr master = (my_master_ptr)cinfo->master;
+
+ if (master->using_merged_upsample && cinfo->max_v_samp_factor == 2) {
+ read_and_discard_scanlines(cinfo, rows);
+ return;
+ }
/* Increment the counter to the next row group after the skipped rows. */
main_ptr->rowgroup_ctr += rows / cinfo->max_v_samp_factor;
@@ -382,21 +403,27 @@
{
my_main_ptr main_ptr = (my_main_ptr)cinfo->main;
my_coef_ptr coef = (my_coef_ptr)cinfo->coef;
+ my_master_ptr master = (my_master_ptr)cinfo->master;
my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
JDIMENSION i, x;
int y;
JDIMENSION lines_per_iMCU_row, lines_left_in_iMCU_row, lines_after_iMCU_row;
JDIMENSION lines_to_skip, lines_to_read;
+ /* Two-pass color quantization is not supported. */
+ if (cinfo->quantize_colors && cinfo->two_pass_quantize)
+ ERREXIT(cinfo, JERR_NOTIMPL);
+
if (cinfo->global_state != DSTATE_SCANNING)
ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
/* Do not skip past the bottom of the image. */
if (cinfo->output_scanline + num_lines >= cinfo->output_height) {
+ num_lines = cinfo->output_height - cinfo->output_scanline;
cinfo->output_scanline = cinfo->output_height;
(*cinfo->inputctl->finish_input_pass) (cinfo);
cinfo->inputctl->eoi_reached = TRUE;
- return cinfo->output_height - cinfo->output_scanline;
+ return num_lines;
}
if (num_lines == 0)
@@ -445,8 +472,10 @@
main_ptr->buffer_full = FALSE;
main_ptr->rowgroup_ctr = 0;
main_ptr->context_state = CTX_PREPARE_FOR_IMCU;
- upsample->next_row_out = cinfo->max_v_samp_factor;
- upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ if (!master->using_merged_upsample) {
+ upsample->next_row_out = cinfo->max_v_samp_factor;
+ upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ }
}
/* Skipping is much simpler when context rows are not required. */
@@ -458,8 +487,10 @@
cinfo->output_scanline += lines_left_in_iMCU_row;
main_ptr->buffer_full = FALSE;
main_ptr->rowgroup_ctr = 0;
- upsample->next_row_out = cinfo->max_v_samp_factor;
- upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ if (!master->using_merged_upsample) {
+ upsample->next_row_out = cinfo->max_v_samp_factor;
+ upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ }
}
}
@@ -494,7 +525,8 @@
cinfo->output_iMCU_row += lines_to_skip / lines_per_iMCU_row;
increment_simple_rowgroup_ctr(cinfo, lines_to_read);
}
- upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ if (!master->using_merged_upsample)
+ upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
return num_lines;
}
@@ -506,6 +538,8 @@
* decoded coefficients. This is ~5% faster for large subsets, but
* it's tough to tell a difference for smaller images.
*/
+ if (!cinfo->entropy->insufficient_data)
+ cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
(*cinfo->entropy->decode_mcu) (cinfo, NULL);
}
}
@@ -535,7 +569,8 @@
* bit odd, since "rows_to_go" seems to be redundantly keeping track of
* output_scanline.
*/
- upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
+ if (!master->using_merged_upsample)
+ upsample->rows_to_go = cinfo->output_height - cinfo->output_scanline;
/* Always skip the requested number of lines. */
return num_lines;
diff --git a/jdarith.c b/jdarith.c
index 6002481..3c7ac57 100644
--- a/jdarith.c
+++ b/jdarith.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Developed 1997-2015 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2015-2018, D. R. Commander.
+ * Copyright (C) 2015-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -80,7 +80,7 @@
if (!(*src->fill_input_buffer) (cinfo))
ERREXIT(cinfo, JERR_CANT_SUSPEND);
src->bytes_in_buffer--;
- return GETJOCTET(*src->next_input_byte++);
+ return *src->next_input_byte++;
}
@@ -665,8 +665,14 @@
for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
int coefi, cindex = cinfo->cur_comp_info[ci]->component_index;
int *coef_bit_ptr = &cinfo->coef_bits[cindex][0];
+ int *prev_coef_bit_ptr =
+ &cinfo->coef_bits[cindex + cinfo->num_components][0];
if (cinfo->Ss && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+ for (coefi = MIN(cinfo->Ss, 1); coefi <= MAX(cinfo->Se, 9); coefi++) {
+ if (cinfo->input_scan_number > 1)
+ prev_coef_bit_ptr[coefi] = coef_bit_ptr[coefi];
+ }
for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
if (cinfo->Ah != expected)
@@ -727,6 +733,7 @@
entropy->c = 0;
entropy->a = 0;
entropy->ct = -16; /* force reading 2 initial bytes to fill C */
+ entropy->pub.insufficient_data = FALSE;
/* Initialize restart counter */
entropy->restarts_to_go = cinfo->restart_interval;
@@ -763,7 +770,7 @@
int *coef_bit_ptr, ci;
cinfo->coef_bits = (int (*)[DCTSIZE2])
(*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- cinfo->num_components * DCTSIZE2 *
+ cinfo->num_components * 2 * DCTSIZE2 *
sizeof(int));
coef_bit_ptr = &cinfo->coef_bits[0][0];
for (ci = 0; ci < cinfo->num_components; ci++)
diff --git a/jdcoefct.c b/jdcoefct.c
index 723a9ac..15e6cde 100644
--- a/jdcoefct.c
+++ b/jdcoefct.c
@@ -5,8 +5,8 @@
* Copyright (C) 1994-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
- * Copyright (C) 2010, 2015-2016, D. R. Commander.
- * Copyright (C) 2015, Google, Inc.
+ * Copyright (C) 2010, 2015-2016, 2019-2020, D. R. Commander.
+ * Copyright (C) 2015, 2020, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -102,6 +102,8 @@
/* Try to fetch an MCU. Entropy decoder expects buffer to be zeroed. */
jzero_far((void *)coef->MCU_buffer[0],
(size_t)(cinfo->blocks_in_MCU * sizeof(JBLOCK)));
+ if (!cinfo->entropy->insufficient_data)
+ cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
if (!(*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
/* Suspension forced; update state counters and exit */
coef->MCU_vert_offset = yoffset;
@@ -227,6 +229,8 @@
}
}
}
+ if (!cinfo->entropy->insufficient_data)
+ cinfo->master->last_good_iMCU_row = cinfo->input_iMCU_row;
/* Try to fetch the MCU. */
if (!(*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
/* Suspension forced; update state counters and exit */
@@ -326,19 +330,22 @@
#ifdef BLOCK_SMOOTHING_SUPPORTED
/*
- * This code applies interblock smoothing as described by section K.8
- * of the JPEG standard: the first 5 AC coefficients are estimated from
- * the DC values of a DCT block and its 8 neighboring blocks.
+ * This code applies interblock smoothing; the first 9 AC coefficients are
+ * estimated from the DC values of a DCT block and its 24 neighboring blocks.
* We apply smoothing only for progressive JPEG decoding, and only if
* the coefficients it can estimate are not yet known to full precision.
*/
-/* Natural-order array positions of the first 5 zigzag-order coefficients */
+/* Natural-order array positions of the first 9 zigzag-order coefficients */
#define Q01_POS 1
#define Q10_POS 8
#define Q20_POS 16
#define Q11_POS 9
#define Q02_POS 2
+#define Q03_POS 3
+#define Q12_POS 10
+#define Q21_POS 17
+#define Q30_POS 24
/*
* Determine whether block smoothing is applicable and safe.
@@ -356,8 +363,8 @@
int ci, coefi;
jpeg_component_info *compptr;
JQUANT_TBL *qtable;
- int *coef_bits;
- int *coef_bits_latch;
+ int *coef_bits, *prev_coef_bits;
+ int *coef_bits_latch, *prev_coef_bits_latch;
if (!cinfo->progressive_mode || cinfo->coef_bits == NULL)
return FALSE;
@@ -366,34 +373,47 @@
if (coef->coef_bits_latch == NULL)
coef->coef_bits_latch = (int *)
(*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- cinfo->num_components *
+ cinfo->num_components * 2 *
(SAVED_COEFS * sizeof(int)));
coef_bits_latch = coef->coef_bits_latch;
+ prev_coef_bits_latch =
+ &coef->coef_bits_latch[cinfo->num_components * SAVED_COEFS];
for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
ci++, compptr++) {
/* All components' quantization values must already be latched. */
if ((qtable = compptr->quant_table) == NULL)
return FALSE;
- /* Verify DC & first 5 AC quantizers are nonzero to avoid zero-divide. */
+ /* Verify DC & first 9 AC quantizers are nonzero to avoid zero-divide. */
if (qtable->quantval[0] == 0 ||
qtable->quantval[Q01_POS] == 0 ||
qtable->quantval[Q10_POS] == 0 ||
qtable->quantval[Q20_POS] == 0 ||
qtable->quantval[Q11_POS] == 0 ||
- qtable->quantval[Q02_POS] == 0)
+ qtable->quantval[Q02_POS] == 0 ||
+ qtable->quantval[Q03_POS] == 0 ||
+ qtable->quantval[Q12_POS] == 0 ||
+ qtable->quantval[Q21_POS] == 0 ||
+ qtable->quantval[Q30_POS] == 0)
return FALSE;
/* DC values must be at least partly known for all components. */
coef_bits = cinfo->coef_bits[ci];
+ prev_coef_bits = cinfo->coef_bits[ci + cinfo->num_components];
if (coef_bits[0] < 0)
return FALSE;
+ coef_bits_latch[0] = coef_bits[0];
/* Block smoothing is helpful if some AC coefficients remain inaccurate. */
- for (coefi = 1; coefi <= 5; coefi++) {
+ for (coefi = 1; coefi < SAVED_COEFS; coefi++) {
+ if (cinfo->input_scan_number > 1)
+ prev_coef_bits_latch[coefi] = prev_coef_bits[coefi];
+ else
+ prev_coef_bits_latch[coefi] = -1;
coef_bits_latch[coefi] = coef_bits[coefi];
if (coef_bits[coefi] != 0)
smoothing_useful = TRUE;
}
coef_bits_latch += SAVED_COEFS;
+ prev_coef_bits_latch += SAVED_COEFS;
}
return smoothing_useful;
@@ -412,17 +432,20 @@
JDIMENSION block_num, last_block_column;
int ci, block_row, block_rows, access_rows;
JBLOCKARRAY buffer;
- JBLOCKROW buffer_ptr, prev_block_row, next_block_row;
+ JBLOCKROW buffer_ptr, prev_prev_block_row, prev_block_row;
+ JBLOCKROW next_block_row, next_next_block_row;
JSAMPARRAY output_ptr;
JDIMENSION output_col;
jpeg_component_info *compptr;
inverse_DCT_method_ptr inverse_DCT;
- boolean first_row, last_row;
+ boolean change_dc;
JCOEF *workspace;
int *coef_bits;
JQUANT_TBL *quanttbl;
- JLONG Q00, Q01, Q02, Q10, Q11, Q20, num;
- int DC1, DC2, DC3, DC4, DC5, DC6, DC7, DC8, DC9;
+ JLONG Q00, Q01, Q02, Q03 = 0, Q10, Q11, Q12 = 0, Q20, Q21 = 0, Q30 = 0, num;
+ int DC01, DC02, DC03, DC04, DC05, DC06, DC07, DC08, DC09, DC10, DC11, DC12,
+ DC13, DC14, DC15, DC16, DC17, DC18, DC19, DC20, DC21, DC22, DC23, DC24,
+ DC25;
int Al, pred;
/* Keep a local variable to avoid looking it up more than once */
@@ -434,10 +457,10 @@
if (cinfo->input_scan_number == cinfo->output_scan_number) {
/* If input is working on current scan, we ordinarily want it to
* have completed the current row. But if input scan is DC,
- * we want it to keep one row ahead so that next block row's DC
+ * we want it to keep two rows ahead so that next two block rows' DC
* values are up to date.
*/
- JDIMENSION delta = (cinfo->Ss == 0) ? 1 : 0;
+ JDIMENSION delta = (cinfo->Ss == 0) ? 2 : 0;
if (cinfo->input_iMCU_row > cinfo->output_iMCU_row + delta)
break;
}
@@ -452,34 +475,53 @@
if (!compptr->component_needed)
continue;
/* Count non-dummy DCT block rows in this iMCU row. */
- if (cinfo->output_iMCU_row < last_iMCU_row) {
+ if (cinfo->output_iMCU_row < last_iMCU_row - 1) {
+ block_rows = compptr->v_samp_factor;
+ access_rows = block_rows * 3; /* this and next two iMCU rows */
+ } else if (cinfo->output_iMCU_row < last_iMCU_row) {
block_rows = compptr->v_samp_factor;
access_rows = block_rows * 2; /* this and next iMCU row */
- last_row = FALSE;
} else {
/* NB: can't use last_row_height here; it is input-side-dependent! */
block_rows = (int)(compptr->height_in_blocks % compptr->v_samp_factor);
if (block_rows == 0) block_rows = compptr->v_samp_factor;
access_rows = block_rows; /* this iMCU row only */
- last_row = TRUE;
}
/* Align the virtual buffer for this component. */
- if (cinfo->output_iMCU_row > 0) {
- access_rows += compptr->v_samp_factor; /* prior iMCU row too */
+ if (cinfo->output_iMCU_row > 1) {
+ access_rows += 2 * compptr->v_samp_factor; /* prior two iMCU rows too */
+ buffer = (*cinfo->mem->access_virt_barray)
+ ((j_common_ptr)cinfo, coef->whole_image[ci],
+ (cinfo->output_iMCU_row - 2) * compptr->v_samp_factor,
+ (JDIMENSION)access_rows, FALSE);
+ buffer += 2 * compptr->v_samp_factor; /* point to current iMCU row */
+ } else if (cinfo->output_iMCU_row > 0) {
buffer = (*cinfo->mem->access_virt_barray)
((j_common_ptr)cinfo, coef->whole_image[ci],
(cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
(JDIMENSION)access_rows, FALSE);
buffer += compptr->v_samp_factor; /* point to current iMCU row */
- first_row = FALSE;
} else {
buffer = (*cinfo->mem->access_virt_barray)
((j_common_ptr)cinfo, coef->whole_image[ci],
(JDIMENSION)0, (JDIMENSION)access_rows, FALSE);
- first_row = TRUE;
}
- /* Fetch component-dependent info */
- coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+ /* Fetch component-dependent info.
+ * If the current scan is incomplete, then we use the component-dependent
+ * info from the previous scan.
+ */
+ if (cinfo->output_iMCU_row > cinfo->master->last_good_iMCU_row)
+ coef_bits =
+ coef->coef_bits_latch + ((ci + cinfo->num_components) * SAVED_COEFS);
+ else
+ coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+
+ /* We only do DC interpolation if no AC coefficient data is available. */
+ change_dc =
+ coef_bits[1] == -1 && coef_bits[2] == -1 && coef_bits[3] == -1 &&
+ coef_bits[4] == -1 && coef_bits[5] == -1 && coef_bits[6] == -1 &&
+ coef_bits[7] == -1 && coef_bits[8] == -1 && coef_bits[9] == -1;
+
quanttbl = compptr->quant_table;
Q00 = quanttbl->quantval[0];
Q01 = quanttbl->quantval[Q01_POS];
@@ -487,25 +529,51 @@
Q20 = quanttbl->quantval[Q20_POS];
Q11 = quanttbl->quantval[Q11_POS];
Q02 = quanttbl->quantval[Q02_POS];
+ if (change_dc) {
+ Q03 = quanttbl->quantval[Q03_POS];
+ Q12 = quanttbl->quantval[Q12_POS];
+ Q21 = quanttbl->quantval[Q21_POS];
+ Q30 = quanttbl->quantval[Q30_POS];
+ }
inverse_DCT = cinfo->idct->inverse_DCT[ci];
output_ptr = output_buf[ci];
/* Loop over all DCT blocks to be processed. */
for (block_row = 0; block_row < block_rows; block_row++) {
buffer_ptr = buffer[block_row] + cinfo->master->first_MCU_col[ci];
- if (first_row && block_row == 0)
+
+ if (block_row > 0 || cinfo->output_iMCU_row > 0)
+ prev_block_row =
+ buffer[block_row - 1] + cinfo->master->first_MCU_col[ci];
+ else
prev_block_row = buffer_ptr;
+
+ if (block_row > 1 || cinfo->output_iMCU_row > 1)
+ prev_prev_block_row =
+ buffer[block_row - 2] + cinfo->master->first_MCU_col[ci];
else
- prev_block_row = buffer[block_row - 1];
- if (last_row && block_row == block_rows - 1)
+ prev_prev_block_row = prev_block_row;
+
+ if (block_row < block_rows - 1 || cinfo->output_iMCU_row < last_iMCU_row)
+ next_block_row =
+ buffer[block_row + 1] + cinfo->master->first_MCU_col[ci];
+ else
next_block_row = buffer_ptr;
+
+ if (block_row < block_rows - 2 ||
+ cinfo->output_iMCU_row < last_iMCU_row - 1)
+ next_next_block_row =
+ buffer[block_row + 2] + cinfo->master->first_MCU_col[ci];
else
- next_block_row = buffer[block_row + 1];
+ next_next_block_row = next_block_row;
+
/* We fetch the surrounding DC values using a sliding-register approach.
- * Initialize all nine here so as to do the right thing on narrow pics.
+ * Initialize all 25 here so as to do the right thing on narrow pics.
*/
- DC1 = DC2 = DC3 = (int)prev_block_row[0][0];
- DC4 = DC5 = DC6 = (int)buffer_ptr[0][0];
- DC7 = DC8 = DC9 = (int)next_block_row[0][0];
+ DC01 = DC02 = DC03 = DC04 = DC05 = (int)prev_prev_block_row[0][0];
+ DC06 = DC07 = DC08 = DC09 = DC10 = (int)prev_block_row[0][0];
+ DC11 = DC12 = DC13 = DC14 = DC15 = (int)buffer_ptr[0][0];
+ DC16 = DC17 = DC18 = DC19 = DC20 = (int)next_block_row[0][0];
+ DC21 = DC22 = DC23 = DC24 = DC25 = (int)next_next_block_row[0][0];
output_col = 0;
last_block_column = compptr->width_in_blocks - 1;
for (block_num = cinfo->master->first_MCU_col[ci];
@@ -513,18 +581,39 @@
/* Fetch current DCT block into workspace so we can modify it. */
jcopy_block_row(buffer_ptr, (JBLOCKROW)workspace, (JDIMENSION)1);
/* Update DC values */
- if (block_num < last_block_column) {
- DC3 = (int)prev_block_row[1][0];
- DC6 = (int)buffer_ptr[1][0];
- DC9 = (int)next_block_row[1][0];
+ if (block_num == cinfo->master->first_MCU_col[ci] &&
+ block_num < last_block_column) {
+ DC04 = (int)prev_prev_block_row[1][0];
+ DC09 = (int)prev_block_row[1][0];
+ DC14 = (int)buffer_ptr[1][0];
+ DC19 = (int)next_block_row[1][0];
+ DC24 = (int)next_next_block_row[1][0];
}
- /* Compute coefficient estimates per K.8.
- * An estimate is applied only if coefficient is still zero,
- * and is not known to be fully accurate.
+ if (block_num + 1 < last_block_column) {
+ DC05 = (int)prev_prev_block_row[2][0];
+ DC10 = (int)prev_block_row[2][0];
+ DC15 = (int)buffer_ptr[2][0];
+ DC20 = (int)next_block_row[2][0];
+ DC25 = (int)next_next_block_row[2][0];
+ }
+ /* If DC interpolation is enabled, compute coefficient estimates using
+ * a Gaussian-like kernel, keeping the averages of the DC values.
+ *
+ * If DC interpolation is disabled, compute coefficient estimates using
+ * an algorithm similar to the one described in Section K.8 of the JPEG
+ * standard, except applied to a 5x5 window rather than a 3x3 window.
+ *
+ * An estimate is applied only if the coefficient is still zero and is
+ * not known to be fully accurate.
*/
/* AC01 */
if ((Al = coef_bits[1]) != 0 && workspace[1] == 0) {
- num = 36 * Q00 * (DC4 - DC6);
+ num = Q00 * (change_dc ?
+ (-DC01 - DC02 + DC04 + DC05 - 3 * DC06 + 13 * DC07 -
+ 13 * DC09 + 3 * DC10 - 3 * DC11 + 38 * DC12 - 38 * DC14 +
+ 3 * DC15 - 3 * DC16 + 13 * DC17 - 13 * DC19 + 3 * DC20 -
+ DC21 - DC22 + DC24 + DC25) :
+ (-7 * DC11 + 50 * DC12 - 50 * DC14 + 7 * DC15));
if (num >= 0) {
pred = (int)(((Q01 << 7) + num) / (Q01 << 8));
if (Al > 0 && pred >= (1 << Al))
@@ -539,7 +628,12 @@
}
/* AC10 */
if ((Al = coef_bits[2]) != 0 && workspace[8] == 0) {
- num = 36 * Q00 * (DC2 - DC8);
+ num = Q00 * (change_dc ?
+ (-DC01 - 3 * DC02 - 3 * DC03 - 3 * DC04 - DC05 - DC06 +
+ 13 * DC07 + 38 * DC08 + 13 * DC09 - DC10 + DC16 -
+ 13 * DC17 - 38 * DC18 - 13 * DC19 + DC20 + DC21 +
+ 3 * DC22 + 3 * DC23 + 3 * DC24 + DC25) :
+ (-7 * DC03 + 50 * DC08 - 50 * DC18 + 7 * DC23));
if (num >= 0) {
pred = (int)(((Q10 << 7) + num) / (Q10 << 8));
if (Al > 0 && pred >= (1 << Al))
@@ -554,7 +648,10 @@
}
/* AC20 */
if ((Al = coef_bits[3]) != 0 && workspace[16] == 0) {
- num = 9 * Q00 * (DC2 + DC8 - 2 * DC5);
+ num = Q00 * (change_dc ?
+ (DC03 + 2 * DC07 + 7 * DC08 + 2 * DC09 - 5 * DC12 - 14 * DC13 -
+ 5 * DC14 + 2 * DC17 + 7 * DC18 + 2 * DC19 + DC23) :
+ (-DC03 + 13 * DC08 - 24 * DC13 + 13 * DC18 - DC23));
if (num >= 0) {
pred = (int)(((Q20 << 7) + num) / (Q20 << 8));
if (Al > 0 && pred >= (1 << Al))
@@ -569,7 +666,11 @@
}
/* AC11 */
if ((Al = coef_bits[4]) != 0 && workspace[9] == 0) {
- num = 5 * Q00 * (DC1 - DC3 - DC7 + DC9);
+ num = Q00 * (change_dc ?
+ (-DC01 + DC05 + 9 * DC07 - 9 * DC09 - 9 * DC17 +
+ 9 * DC19 + DC21 - DC25) :
+ (DC10 + DC16 - 10 * DC17 + 10 * DC19 - DC02 - DC20 + DC22 -
+ DC24 + DC04 - DC06 + 10 * DC07 - 10 * DC09));
if (num >= 0) {
pred = (int)(((Q11 << 7) + num) / (Q11 << 8));
if (Al > 0 && pred >= (1 << Al))
@@ -584,7 +685,10 @@
}
/* AC02 */
if ((Al = coef_bits[5]) != 0 && workspace[2] == 0) {
- num = 9 * Q00 * (DC4 + DC6 - 2 * DC5);
+ num = Q00 * (change_dc ?
+ (2 * DC07 - 5 * DC08 + 2 * DC09 + DC11 + 7 * DC12 - 14 * DC13 +
+ 7 * DC14 + DC15 + 2 * DC17 - 5 * DC18 + 2 * DC19) :
+ (-DC11 + 13 * DC12 - 24 * DC13 + 13 * DC14 - DC15));
if (num >= 0) {
pred = (int)(((Q02 << 7) + num) / (Q02 << 8));
if (Al > 0 && pred >= (1 << Al))
@@ -597,14 +701,96 @@
}
workspace[2] = (JCOEF)pred;
}
+ if (change_dc) {
+ /* AC03 */
+ if ((Al = coef_bits[6]) != 0 && workspace[3] == 0) {
+ num = Q00 * (DC07 - DC09 + 2 * DC12 - 2 * DC14 + DC17 - DC19);
+ if (num >= 0) {
+ pred = (int)(((Q03 << 7) + num) / (Q03 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ } else {
+ pred = (int)(((Q03 << 7) - num) / (Q03 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ pred = -pred;
+ }
+ workspace[3] = (JCOEF)pred;
+ }
+ /* AC12 */
+ if ((Al = coef_bits[7]) != 0 && workspace[10] == 0) {
+ num = Q00 * (DC07 - 3 * DC08 + DC09 - DC17 + 3 * DC18 - DC19);
+ if (num >= 0) {
+ pred = (int)(((Q12 << 7) + num) / (Q12 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ } else {
+ pred = (int)(((Q12 << 7) - num) / (Q12 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ pred = -pred;
+ }
+ workspace[10] = (JCOEF)pred;
+ }
+ /* AC21 */
+ if ((Al = coef_bits[8]) != 0 && workspace[17] == 0) {
+ num = Q00 * (DC07 - DC09 - 3 * DC12 + 3 * DC14 + DC17 - DC19);
+ if (num >= 0) {
+ pred = (int)(((Q21 << 7) + num) / (Q21 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ } else {
+ pred = (int)(((Q21 << 7) - num) / (Q21 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ pred = -pred;
+ }
+ workspace[17] = (JCOEF)pred;
+ }
+ /* AC30 */
+ if ((Al = coef_bits[9]) != 0 && workspace[24] == 0) {
+ num = Q00 * (DC07 + 2 * DC08 + DC09 - DC17 - 2 * DC18 - DC19);
+ if (num >= 0) {
+ pred = (int)(((Q30 << 7) + num) / (Q30 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ } else {
+ pred = (int)(((Q30 << 7) - num) / (Q30 << 8));
+ if (Al > 0 && pred >= (1 << Al))
+ pred = (1 << Al) - 1;
+ pred = -pred;
+ }
+ workspace[24] = (JCOEF)pred;
+ }
+ /* coef_bits[0] is non-negative. Otherwise this function would not
+ * be called.
+ */
+ num = Q00 *
+ (-2 * DC01 - 6 * DC02 - 8 * DC03 - 6 * DC04 - 2 * DC05 -
+ 6 * DC06 + 6 * DC07 + 42 * DC08 + 6 * DC09 - 6 * DC10 -
+ 8 * DC11 + 42 * DC12 + 152 * DC13 + 42 * DC14 - 8 * DC15 -
+ 6 * DC16 + 6 * DC17 + 42 * DC18 + 6 * DC19 - 6 * DC20 -
+ 2 * DC21 - 6 * DC22 - 8 * DC23 - 6 * DC24 - 2 * DC25);
+ if (num >= 0) {
+ pred = (int)(((Q00 << 7) + num) / (Q00 << 8));
+ } else {
+ pred = (int)(((Q00 << 7) - num) / (Q00 << 8));
+ pred = -pred;
+ }
+ workspace[0] = (JCOEF)pred;
+ } /* change_dc */
+
/* OK, do the IDCT */
(*inverse_DCT) (cinfo, compptr, (JCOEFPTR)workspace, output_ptr,
output_col);
/* Advance for next column */
- DC1 = DC2; DC2 = DC3;
- DC4 = DC5; DC5 = DC6;
- DC7 = DC8; DC8 = DC9;
- buffer_ptr++, prev_block_row++, next_block_row++;
+ DC01 = DC02; DC02 = DC03; DC03 = DC04; DC04 = DC05;
+ DC06 = DC07; DC07 = DC08; DC08 = DC09; DC09 = DC10;
+ DC11 = DC12; DC12 = DC13; DC13 = DC14; DC14 = DC15;
+ DC16 = DC17; DC17 = DC18; DC18 = DC19; DC19 = DC20;
+ DC21 = DC22; DC22 = DC23; DC23 = DC24; DC24 = DC25;
+ buffer_ptr++, prev_block_row++, next_block_row++,
+ prev_prev_block_row++, next_next_block_row++;
output_col += compptr->_DCT_scaled_size;
}
output_ptr += compptr->_DCT_scaled_size;
@@ -653,7 +839,7 @@
#ifdef BLOCK_SMOOTHING_SUPPORTED
/* If block smoothing could be used, need a bigger window */
if (cinfo->progressive_mode)
- access_rows *= 3;
+ access_rows *= 5;
#endif
coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
((j_common_ptr)cinfo, JPOOL_IMAGE, TRUE,
diff --git a/jdcoefct.h b/jdcoefct.h
index c4d1943..9a0e780 100644
--- a/jdcoefct.h
+++ b/jdcoefct.h
@@ -5,6 +5,7 @@
* Copyright (C) 1994-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2020, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*/
@@ -51,7 +52,7 @@
#ifdef BLOCK_SMOOTHING_SUPPORTED
/* When doing block smoothing, we latch coefficient Al values here */
int *coef_bits_latch;
-#define SAVED_COEFS 6 /* we save coef_bits[0..5] */
+#define SAVED_COEFS 10 /* we save coef_bits[0..9] */
#endif
} my_coef_controller;
diff --git a/jdcol565.c b/jdcol565.c
index 40068ef..53c7bd9 100644
--- a/jdcol565.c
+++ b/jdcol565.c
@@ -45,9 +45,9 @@
outptr = *output_buf++;
if (PACK_NEED_ALIGNMENT(outptr)) {
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[y + Crrtab[cr]];
g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
SCALEBITS))];
@@ -58,18 +58,18 @@
num_cols--;
}
for (col = 0; col < (num_cols >> 1); col++) {
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[y + Crrtab[cr]];
g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
SCALEBITS))];
b = range_limit[y + Cbbtab[cb]];
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[y + Crrtab[cr]];
g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
SCALEBITS))];
@@ -80,9 +80,9 @@
outptr += 4;
}
if (num_cols & 1) {
- y = GETJSAMPLE(*inptr0);
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ y = *inptr0;
+ cb = *inptr1;
+ cr = *inptr2;
r = range_limit[y + Crrtab[cr]];
g = range_limit[y + ((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
SCALEBITS))];
@@ -125,9 +125,9 @@
input_row++;
outptr = *output_buf++;
if (PACK_NEED_ALIGNMENT(outptr)) {
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
g = range_limit[DITHER_565_G(y +
((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
@@ -139,9 +139,9 @@
num_cols--;
}
for (col = 0; col < (num_cols >> 1); col++) {
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
g = range_limit[DITHER_565_G(y +
((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
@@ -150,9 +150,9 @@
d0 = DITHER_ROTATE(d0);
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr0++);
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ y = *inptr0++;
+ cb = *inptr1++;
+ cr = *inptr2++;
r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
g = range_limit[DITHER_565_G(y +
((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
@@ -165,9 +165,9 @@
outptr += 4;
}
if (num_cols & 1) {
- y = GETJSAMPLE(*inptr0);
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ y = *inptr0;
+ cb = *inptr1;
+ cr = *inptr2;
r = range_limit[DITHER_565_R(y + Crrtab[cr], d0)];
g = range_limit[DITHER_565_G(y +
((int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
@@ -202,32 +202,32 @@
input_row++;
outptr = *output_buf++;
if (PACK_NEED_ALIGNMENT(outptr)) {
- r = GETJSAMPLE(*inptr0++);
- g = GETJSAMPLE(*inptr1++);
- b = GETJSAMPLE(*inptr2++);
+ r = *inptr0++;
+ g = *inptr1++;
+ b = *inptr2++;
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr = (INT16)rgb;
outptr += 2;
num_cols--;
}
for (col = 0; col < (num_cols >> 1); col++) {
- r = GETJSAMPLE(*inptr0++);
- g = GETJSAMPLE(*inptr1++);
- b = GETJSAMPLE(*inptr2++);
+ r = *inptr0++;
+ g = *inptr1++;
+ b = *inptr2++;
rgb = PACK_SHORT_565(r, g, b);
- r = GETJSAMPLE(*inptr0++);
- g = GETJSAMPLE(*inptr1++);
- b = GETJSAMPLE(*inptr2++);
+ r = *inptr0++;
+ g = *inptr1++;
+ b = *inptr2++;
rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
WRITE_TWO_ALIGNED_PIXELS(outptr, rgb);
outptr += 4;
}
if (num_cols & 1) {
- r = GETJSAMPLE(*inptr0);
- g = GETJSAMPLE(*inptr1);
- b = GETJSAMPLE(*inptr2);
+ r = *inptr0;
+ g = *inptr1;
+ b = *inptr2;
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr = (INT16)rgb;
}
@@ -259,24 +259,24 @@
input_row++;
outptr = *output_buf++;
if (PACK_NEED_ALIGNMENT(outptr)) {
- r = range_limit[DITHER_565_R(GETJSAMPLE(*inptr0++), d0)];
- g = range_limit[DITHER_565_G(GETJSAMPLE(*inptr1++), d0)];
- b = range_limit[DITHER_565_B(GETJSAMPLE(*inptr2++), d0)];
+ r = range_limit[DITHER_565_R(*inptr0++, d0)];
+ g = range_limit[DITHER_565_G(*inptr1++, d0)];
+ b = range_limit[DITHER_565_B(*inptr2++, d0)];
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr = (INT16)rgb;
outptr += 2;
num_cols--;
}
for (col = 0; col < (num_cols >> 1); col++) {
- r = range_limit[DITHER_565_R(GETJSAMPLE(*inptr0++), d0)];
- g = range_limit[DITHER_565_G(GETJSAMPLE(*inptr1++), d0)];
- b = range_limit[DITHER_565_B(GETJSAMPLE(*inptr2++), d0)];
+ r = range_limit[DITHER_565_R(*inptr0++, d0)];
+ g = range_limit[DITHER_565_G(*inptr1++, d0)];
+ b = range_limit[DITHER_565_B(*inptr2++, d0)];
d0 = DITHER_ROTATE(d0);
rgb = PACK_SHORT_565(r, g, b);
- r = range_limit[DITHER_565_R(GETJSAMPLE(*inptr0++), d0)];
- g = range_limit[DITHER_565_G(GETJSAMPLE(*inptr1++), d0)];
- b = range_limit[DITHER_565_B(GETJSAMPLE(*inptr2++), d0)];
+ r = range_limit[DITHER_565_R(*inptr0++, d0)];
+ g = range_limit[DITHER_565_G(*inptr1++, d0)];
+ b = range_limit[DITHER_565_B(*inptr2++, d0)];
d0 = DITHER_ROTATE(d0);
rgb = PACK_TWO_PIXELS(rgb, PACK_SHORT_565(r, g, b));
@@ -284,9 +284,9 @@
outptr += 4;
}
if (num_cols & 1) {
- r = range_limit[DITHER_565_R(GETJSAMPLE(*inptr0), d0)];
- g = range_limit[DITHER_565_G(GETJSAMPLE(*inptr1), d0)];
- b = range_limit[DITHER_565_B(GETJSAMPLE(*inptr2), d0)];
+ r = range_limit[DITHER_565_R(*inptr0, d0)];
+ g = range_limit[DITHER_565_G(*inptr1, d0)];
+ b = range_limit[DITHER_565_B(*inptr2, d0)];
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr = (INT16)rgb;
}
diff --git a/jdcolext.c b/jdcolext.c
index 72a5301..863c7a2 100644
--- a/jdcolext.c
+++ b/jdcolext.c
@@ -53,9 +53,9 @@
input_row++;
outptr = *output_buf++;
for (col = 0; col < num_cols; col++) {
- y = GETJSAMPLE(inptr0[col]);
- cb = GETJSAMPLE(inptr1[col]);
- cr = GETJSAMPLE(inptr2[col]);
+ y = inptr0[col];
+ cb = inptr1[col];
+ cr = inptr2[col];
/* Range-limiting is essential due to noise introduced by DCT losses. */
outptr[RGB_RED] = range_limit[y + Crrtab[cr]];
outptr[RGB_GREEN] = range_limit[y +
@@ -93,7 +93,6 @@
inptr = input_buf[0][input_row++];
outptr = *output_buf++;
for (col = 0; col < num_cols; col++) {
- /* We can dispense with GETJSAMPLE() here */
outptr[RGB_RED] = outptr[RGB_GREEN] = outptr[RGB_BLUE] = inptr[col];
/* Set unused byte to 0xFF so it can be interpreted as an opaque */
/* alpha channel value */
@@ -128,7 +127,6 @@
input_row++;
outptr = *output_buf++;
for (col = 0; col < num_cols; col++) {
- /* We can dispense with GETJSAMPLE() here */
outptr[RGB_RED] = inptr0[col];
outptr[RGB_GREEN] = inptr1[col];
outptr[RGB_BLUE] = inptr2[col];
diff --git a/jdcolor.c b/jdcolor.c
index dc0e3b6..8da2b4e 100644
--- a/jdcolor.c
+++ b/jdcolor.c
@@ -341,9 +341,9 @@
input_row++;
outptr = *output_buf++;
for (col = 0; col < num_cols; col++) {
- r = GETJSAMPLE(inptr0[col]);
- g = GETJSAMPLE(inptr1[col]);
- b = GETJSAMPLE(inptr2[col]);
+ r = inptr0[col];
+ g = inptr1[col];
+ b = inptr2[col];
/* Y */
outptr[col] = (JSAMPLE)((ctab[r + R_Y_OFF] + ctab[g + G_Y_OFF] +
ctab[b + B_Y_OFF]) >> SCALEBITS);
@@ -550,9 +550,9 @@
input_row++;
outptr = *output_buf++;
for (col = 0; col < num_cols; col++) {
- y = GETJSAMPLE(inptr0[col]);
- cb = GETJSAMPLE(inptr1[col]);
- cr = GETJSAMPLE(inptr2[col]);
+ y = inptr0[col];
+ cb = inptr1[col];
+ cr = inptr2[col];
/* Range-limiting is essential due to noise introduced by DCT losses. */
outptr[0] = range_limit[MAXJSAMPLE - (y + Crrtab[cr])]; /* red */
outptr[1] = range_limit[MAXJSAMPLE - (y + /* green */
@@ -560,7 +560,7 @@
SCALEBITS)))];
outptr[2] = range_limit[MAXJSAMPLE - (y + Cbbtab[cb])]; /* blue */
/* K passes through unchanged */
- outptr[3] = inptr3[col]; /* don't need GETJSAMPLE here */
+ outptr[3] = inptr3[col];
outptr += 4;
}
}
@@ -571,11 +571,10 @@
* RGB565 conversion
*/
-#define PACK_SHORT_565_LE(r, g, b) ((((r) << 8) & 0xF800) | \
- (((g) << 3) & 0x7E0) | ((b) >> 3))
-#define PACK_SHORT_565_BE(r, g, b) (((r) & 0xF8) | ((g) >> 5) | \
- (((g) << 11) & 0xE000) | \
- (((b) << 5) & 0x1F00))
+#define PACK_SHORT_565_LE(r, g, b) \
+ ((((r) << 8) & 0xF800) | (((g) << 3) & 0x7E0) | ((b) >> 3))
+#define PACK_SHORT_565_BE(r, g, b) \
+ (((r) & 0xF8) | ((g) >> 5) | (((g) << 11) & 0xE000) | (((b) << 5) & 0x1F00))
#define PACK_TWO_PIXELS_LE(l, r) ((r << 16) | l)
#define PACK_TWO_PIXELS_BE(l, r) ((l << 16) | r)
diff --git a/jdhuff.c b/jdhuff.c
index 3880890..83478f3 100644
--- a/jdhuff.c
+++ b/jdhuff.c
@@ -5,6 +5,7 @@
* Copyright (C) 1991-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright (C) 2009-2011, 2016, 2018-2019, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -39,24 +40,6 @@
int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
} savable_state;
-/* This macro is to work around compilers with missing or broken
- * structure assignment. You'll need to fix this code if you have
- * such a compiler and you change MAX_COMPS_IN_SCAN.
- */
-
-#ifndef NO_STRUCT_ASSIGN
-#define ASSIGN_STATE(dest, src) ((dest) = (src))
-#else
-#if MAX_COMPS_IN_SCAN == 4
-#define ASSIGN_STATE(dest, src) \
- ((dest).last_dc_val[0] = (src).last_dc_val[0], \
- (dest).last_dc_val[1] = (src).last_dc_val[1], \
- (dest).last_dc_val[2] = (src).last_dc_val[2], \
- (dest).last_dc_val[3] = (src).last_dc_val[3])
-#endif
-#endif
-
-
typedef struct {
struct jpeg_entropy_decoder pub; /* public fields */
@@ -325,7 +308,7 @@
bytes_in_buffer = cinfo->src->bytes_in_buffer;
}
bytes_in_buffer--;
- c = GETJOCTET(*next_input_byte++);
+ c = *next_input_byte++;
/* If it's 0xFF, check and discard stuffed zero byte */
if (c == 0xFF) {
@@ -342,7 +325,7 @@
bytes_in_buffer = cinfo->src->bytes_in_buffer;
}
bytes_in_buffer--;
- c = GETJOCTET(*next_input_byte++);
+ c = *next_input_byte++;
} while (c == 0xFF);
if (c == 0) {
@@ -405,8 +388,8 @@
#define GET_BYTE { \
register int c0, c1; \
- c0 = GETJOCTET(*buffer++); \
- c1 = GETJOCTET(*buffer); \
+ c0 = *buffer++; \
+ c1 = *buffer; \
/* Pre-execute most common case */ \
get_buffer = (get_buffer << 8) | c0; \
bits_left += 8; \
@@ -423,7 +406,7 @@
} \
}
-#if SIZEOF_SIZE_T == 8 || defined(_WIN64)
+#if SIZEOF_SIZE_T == 8 || defined(_WIN64) || (defined(__x86_64__) && defined(__ILP32__))
/* Pre-fetch 48 bytes, because the holding register is 64-bit */
#define FILL_BIT_BUFFER_FAST \
@@ -568,7 +551,7 @@
/* Load up working state */
BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
- ASSIGN_STATE(state, entropy->saved);
+ state = entropy->saved;
for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
JBLOCKROW block = MCU_data ? MCU_data[blkn] : NULL;
@@ -653,7 +636,7 @@
/* Completed MCU, so update state */
BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
- ASSIGN_STATE(entropy->saved, state);
+ entropy->saved = state;
return TRUE;
}
@@ -671,7 +654,7 @@
/* Load up working state */
BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
buffer = (JOCTET *)br_state.next_input_byte;
- ASSIGN_STATE(state, entropy->saved);
+ state = entropy->saved;
for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
JBLOCKROW block = MCU_data ? MCU_data[blkn] : NULL;
@@ -741,7 +724,7 @@
br_state.bytes_in_buffer -= (buffer - br_state.next_input_byte);
br_state.next_input_byte = buffer;
BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
- ASSIGN_STATE(entropy->saved, state);
+ entropy->saved = state;
return TRUE;
}
diff --git a/jdhuff.h b/jdhuff.h
index 319f4d5..6a3fbb5 100644
--- a/jdhuff.h
+++ b/jdhuff.h
@@ -5,6 +5,7 @@
* Copyright (C) 1991-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright (C) 2010-2011, 2015-2016, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -78,6 +79,11 @@
typedef size_t bit_buf_type; /* type of bit-extraction buffer */
#define BIT_BUF_SIZE 64 /* size of buffer in bits */
+#elif defined(__x86_64__) && defined(__ILP32__)
+
+typedef unsigned long long bit_buf_type; /* type of bit-extraction buffer */
+#define BIT_BUF_SIZE 64 /* size of buffer in bits */
+
#else
typedef unsigned long bit_buf_type; /* type of bit-extraction buffer */
diff --git a/jdicc.c b/jdicc.c
index 7224695..a1a5b86 100644
--- a/jdicc.c
+++ b/jdicc.c
@@ -38,18 +38,18 @@
marker->marker == ICC_MARKER &&
marker->data_length >= ICC_OVERHEAD_LEN &&
/* verify the identifying string */
- GETJOCTET(marker->data[0]) == 0x49 &&
- GETJOCTET(marker->data[1]) == 0x43 &&
- GETJOCTET(marker->data[2]) == 0x43 &&
- GETJOCTET(marker->data[3]) == 0x5F &&
- GETJOCTET(marker->data[4]) == 0x50 &&
- GETJOCTET(marker->data[5]) == 0x52 &&
- GETJOCTET(marker->data[6]) == 0x4F &&
- GETJOCTET(marker->data[7]) == 0x46 &&
- GETJOCTET(marker->data[8]) == 0x49 &&
- GETJOCTET(marker->data[9]) == 0x4C &&
- GETJOCTET(marker->data[10]) == 0x45 &&
- GETJOCTET(marker->data[11]) == 0x0;
+ marker->data[0] == 0x49 &&
+ marker->data[1] == 0x43 &&
+ marker->data[2] == 0x43 &&
+ marker->data[3] == 0x5F &&
+ marker->data[4] == 0x50 &&
+ marker->data[5] == 0x52 &&
+ marker->data[6] == 0x4F &&
+ marker->data[7] == 0x46 &&
+ marker->data[8] == 0x49 &&
+ marker->data[9] == 0x4C &&
+ marker->data[10] == 0x45 &&
+ marker->data[11] == 0x0;
}
@@ -102,12 +102,12 @@
for (marker = cinfo->marker_list; marker != NULL; marker = marker->next) {
if (marker_is_icc(marker)) {
if (num_markers == 0)
- num_markers = GETJOCTET(marker->data[13]);
- else if (num_markers != GETJOCTET(marker->data[13])) {
+ num_markers = marker->data[13];
+ else if (num_markers != marker->data[13]) {
WARNMS(cinfo, JWRN_BOGUS_ICC); /* inconsistent num_markers fields */
return FALSE;
}
- seq_no = GETJOCTET(marker->data[12]);
+ seq_no = marker->data[12];
if (seq_no <= 0 || seq_no > num_markers) {
WARNMS(cinfo, JWRN_BOGUS_ICC); /* bogus sequence number */
return FALSE;
@@ -154,7 +154,7 @@
JOCTET FAR *src_ptr;
JOCTET *dst_ptr;
unsigned int length;
- seq_no = GETJOCTET(marker->data[12]);
+ seq_no = marker->data[12];
dst_ptr = icc_data + data_offset[seq_no];
src_ptr = marker->data + ICC_OVERHEAD_LEN;
length = data_length[seq_no];
diff --git a/jdmarker.c b/jdmarker.c
index c9c7ef6..b964c3a 100644
--- a/jdmarker.c
+++ b/jdmarker.c
@@ -151,7 +151,7 @@
#define INPUT_BYTE(cinfo, V, action) \
MAKESTMT( MAKE_BYTE_AVAIL(cinfo, action); \
bytes_in_buffer--; \
- V = GETJOCTET(*next_input_byte++); )
+ V = *next_input_byte++; )
/* As above, but read two bytes interpreted as an unsigned 16-bit integer.
* V should be declared unsigned int or perhaps JLONG.
@@ -159,10 +159,10 @@
#define INPUT_2BYTES(cinfo, V, action) \
MAKESTMT( MAKE_BYTE_AVAIL(cinfo, action); \
bytes_in_buffer--; \
- V = ((unsigned int)GETJOCTET(*next_input_byte++)) << 8; \
+ V = ((unsigned int)(*next_input_byte++)) << 8; \
MAKE_BYTE_AVAIL(cinfo, action); \
bytes_in_buffer--; \
- V += GETJOCTET(*next_input_byte++); )
+ V += *next_input_byte++; )
/*
@@ -608,18 +608,18 @@
JLONG totallen = (JLONG)datalen + remaining;
if (datalen >= APP0_DATA_LEN &&
- GETJOCTET(data[0]) == 0x4A &&
- GETJOCTET(data[1]) == 0x46 &&
- GETJOCTET(data[2]) == 0x49 &&
- GETJOCTET(data[3]) == 0x46 &&
- GETJOCTET(data[4]) == 0) {
+ data[0] == 0x4A &&
+ data[1] == 0x46 &&
+ data[2] == 0x49 &&
+ data[3] == 0x46 &&
+ data[4] == 0) {
/* Found JFIF APP0 marker: save info */
cinfo->saw_JFIF_marker = TRUE;
- cinfo->JFIF_major_version = GETJOCTET(data[5]);
- cinfo->JFIF_minor_version = GETJOCTET(data[6]);
- cinfo->density_unit = GETJOCTET(data[7]);
- cinfo->X_density = (GETJOCTET(data[8]) << 8) + GETJOCTET(data[9]);
- cinfo->Y_density = (GETJOCTET(data[10]) << 8) + GETJOCTET(data[11]);
+ cinfo->JFIF_major_version = data[5];
+ cinfo->JFIF_minor_version = data[6];
+ cinfo->density_unit = data[7];
+ cinfo->X_density = (data[8] << 8) + data[9];
+ cinfo->Y_density = (data[10] << 8) + data[11];
/* Check version.
* Major version must be 1, anything else signals an incompatible change.
* (We used to treat this as an error, but now it's a nonfatal warning,
@@ -634,24 +634,22 @@
cinfo->JFIF_major_version, cinfo->JFIF_minor_version,
cinfo->X_density, cinfo->Y_density, cinfo->density_unit);
/* Validate thumbnail dimensions and issue appropriate messages */
- if (GETJOCTET(data[12]) | GETJOCTET(data[13]))
- TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL,
- GETJOCTET(data[12]), GETJOCTET(data[13]));
+ if (data[12] | data[13])
+ TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL, data[12], data[13]);
totallen -= APP0_DATA_LEN;
- if (totallen !=
- ((JLONG)GETJOCTET(data[12]) * (JLONG)GETJOCTET(data[13]) * (JLONG)3))
+ if (totallen != ((JLONG)data[12] * (JLONG)data[13] * (JLONG)3))
TRACEMS1(cinfo, 1, JTRC_JFIF_BADTHUMBNAILSIZE, (int)totallen);
} else if (datalen >= 6 &&
- GETJOCTET(data[0]) == 0x4A &&
- GETJOCTET(data[1]) == 0x46 &&
- GETJOCTET(data[2]) == 0x58 &&
- GETJOCTET(data[3]) == 0x58 &&
- GETJOCTET(data[4]) == 0) {
+ data[0] == 0x4A &&
+ data[1] == 0x46 &&
+ data[2] == 0x58 &&
+ data[3] == 0x58 &&
+ data[4] == 0) {
/* Found JFIF "JFXX" extension APP0 marker */
/* The library doesn't actually do anything with these,
* but we try to produce a helpful trace message.
*/
- switch (GETJOCTET(data[5])) {
+ switch (data[5]) {
case 0x10:
TRACEMS1(cinfo, 1, JTRC_THUMB_JPEG, (int)totallen);
break;
@@ -662,8 +660,7 @@
TRACEMS1(cinfo, 1, JTRC_THUMB_RGB, (int)totallen);
break;
default:
- TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION,
- GETJOCTET(data[5]), (int)totallen);
+ TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION, data[5], (int)totallen);
break;
}
} else {
@@ -684,16 +681,16 @@
unsigned int version, flags0, flags1, transform;
if (datalen >= APP14_DATA_LEN &&
- GETJOCTET(data[0]) == 0x41 &&
- GETJOCTET(data[1]) == 0x64 &&
- GETJOCTET(data[2]) == 0x6F &&
- GETJOCTET(data[3]) == 0x62 &&
- GETJOCTET(data[4]) == 0x65) {
+ data[0] == 0x41 &&
+ data[1] == 0x64 &&
+ data[2] == 0x6F &&
+ data[3] == 0x62 &&
+ data[4] == 0x65) {
/* Found Adobe APP14 marker */
- version = (GETJOCTET(data[5]) << 8) + GETJOCTET(data[6]);
- flags0 = (GETJOCTET(data[7]) << 8) + GETJOCTET(data[8]);
- flags1 = (GETJOCTET(data[9]) << 8) + GETJOCTET(data[10]);
- transform = GETJOCTET(data[11]);
+ version = (data[5] << 8) + data[6];
+ flags0 = (data[7] << 8) + data[8];
+ flags1 = (data[9] << 8) + data[10];
+ transform = data[11];
TRACEMS4(cinfo, 1, JTRC_ADOBE, version, flags0, flags1, transform);
cinfo->saw_Adobe_marker = TRUE;
cinfo->Adobe_transform = (UINT8)transform;
diff --git a/jdmaster.c b/jdmaster.c
index b209064..cbc8774 100644
--- a/jdmaster.c
+++ b/jdmaster.c
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1997, Thomas G. Lane.
* Modified 2002-2009 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2009-2011, 2016, D. R. Commander.
+ * Copyright (C) 2009-2011, 2016, 2019, D. R. Commander.
* Copyright (C) 2013, Linaro Limited.
* Copyright (C) 2015, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
@@ -22,7 +22,6 @@
#include "jpeglib.h"
#include "jpegcomp.h"
#include "jdmaster.h"
-#include "jsimd.h"
/*
@@ -70,17 +69,6 @@
cinfo->comp_info[1]._DCT_scaled_size != cinfo->_min_DCT_scaled_size ||
cinfo->comp_info[2]._DCT_scaled_size != cinfo->_min_DCT_scaled_size)
return FALSE;
-#ifdef WITH_SIMD
- /* If YCbCr-to-RGB color conversion is SIMD-accelerated but merged upsampling
- isn't, then disabling merged upsampling is likely to be faster when
- decompressing YCbCr JPEG images. */
- if (!jsimd_can_h2v2_merged_upsample() && !jsimd_can_h2v1_merged_upsample() &&
- jsimd_can_ycc_rgb() && cinfo->jpeg_color_space == JCS_YCbCr &&
- (cinfo->out_color_space == JCS_RGB ||
- (cinfo->out_color_space >= JCS_EXT_RGB &&
- cinfo->out_color_space <= JCS_EXT_ARGB)))
- return FALSE;
-#endif
/* ??? also need to test for upsample-time rescaling, when & if supported */
return TRUE; /* by golly, it'll work... */
#else
@@ -580,6 +568,7 @@
*/
cinfo->master->first_iMCU_col = 0;
cinfo->master->last_iMCU_col = cinfo->MCUs_per_row - 1;
+ cinfo->master->last_good_iMCU_row = 0;
#ifdef D_MULTISCAN_FILES_SUPPORTED
/* If jpeg_start_decompress will read the whole file, initialize
diff --git a/jdmerge.c b/jdmerge.c
index dff5a35..3a456d6 100644
--- a/jdmerge.c
+++ b/jdmerge.c
@@ -5,7 +5,7 @@
* Copyright (C) 1994-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
- * Copyright (C) 2009, 2011, 2014-2015, D. R. Commander.
+ * Copyright (C) 2009, 2011, 2014-2015, 2020, D. R. Commander.
* Copyright (C) 2013, Linaro Limited.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
@@ -40,41 +40,13 @@
#define JPEG_INTERNALS
#include "jinclude.h"
#include "jpeglib.h"
+#include "jdmerge.h"
#include "jsimd.h"
#include "jconfigint.h"
#ifdef UPSAMPLE_MERGING_SUPPORTED
-/* Private subobject */
-
-typedef struct {
- struct jpeg_upsampler pub; /* public fields */
-
- /* Pointer to routine to do actual upsampling/conversion of one row group */
- void (*upmethod) (j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
- JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf);
-
- /* Private state for YCC->RGB conversion */
- int *Cr_r_tab; /* => table for Cr to R conversion */
- int *Cb_b_tab; /* => table for Cb to B conversion */
- JLONG *Cr_g_tab; /* => table for Cr to G conversion */
- JLONG *Cb_g_tab; /* => table for Cb to G conversion */
-
- /* For 2:1 vertical sampling, we produce two output rows at a time.
- * We need a "spare" row buffer to hold the second output row if the
- * application provides just a one-row buffer; we also use the spare
- * to discard the dummy last row if the image height is odd.
- */
- JSAMPROW spare_row;
- boolean spare_full; /* T if spare buffer is occupied */
-
- JDIMENSION out_row_width; /* samples per output row */
- JDIMENSION rows_to_go; /* counts rows remaining in image */
-} my_upsampler;
-
-typedef my_upsampler *my_upsample_ptr;
-
#define SCALEBITS 16 /* speediest right-shift on some machines */
#define ONE_HALF ((JLONG)1 << (SCALEBITS - 1))
#define FIX(x) ((JLONG)((x) * (1L << SCALEBITS) + 0.5))
@@ -189,7 +161,7 @@
LOCAL(void)
build_ycc_rgb_table(j_decompress_ptr cinfo)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
int i;
JLONG x;
SHIFT_TEMPS
@@ -232,7 +204,7 @@
METHODDEF(void)
start_pass_merged_upsample(j_decompress_ptr cinfo)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
/* Mark the spare buffer empty */
upsample->spare_full = FALSE;
@@ -254,7 +226,7 @@
JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
/* 2:1 vertical sampling case: may need a spare row. */
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
JSAMPROW work_ptrs[2];
JDIMENSION num_rows; /* number of rows returned to caller */
@@ -305,7 +277,7 @@
JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail)
/* 1:1 vertical sampling case: much easier, never need a spare row. */
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
/* Just do the upsampling. */
(*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr,
@@ -420,11 +392,10 @@
* RGB565 conversion
*/
-#define PACK_SHORT_565_LE(r, g, b) ((((r) << 8) & 0xF800) | \
- (((g) << 3) & 0x7E0) | ((b) >> 3))
-#define PACK_SHORT_565_BE(r, g, b) (((r) & 0xF8) | ((g) >> 5) | \
- (((g) << 11) & 0xE000) | \
- (((b) << 5) & 0x1F00))
+#define PACK_SHORT_565_LE(r, g, b) \
+ ((((r) << 8) & 0xF800) | (((g) << 3) & 0x7E0) | ((b) >> 3))
+#define PACK_SHORT_565_BE(r, g, b) \
+ (((r) & 0xF8) | ((g) >> 5) | (((g) << 11) & 0xE000) | (((b) << 5) & 0x1F00))
#define PACK_TWO_PIXELS_LE(l, r) ((r << 16) | l)
#define PACK_TWO_PIXELS_BE(l, r) ((l << 16) | r)
@@ -566,11 +537,11 @@
GLOBAL(void)
jinit_merged_upsampler(j_decompress_ptr cinfo)
{
- my_upsample_ptr upsample;
+ my_merged_upsample_ptr upsample;
- upsample = (my_upsample_ptr)
+ upsample = (my_merged_upsample_ptr)
(*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- sizeof(my_upsampler));
+ sizeof(my_merged_upsampler));
cinfo->upsample = (struct jpeg_upsampler *)upsample;
upsample->pub.start_pass = start_pass_merged_upsample;
upsample->pub.need_context_rows = FALSE;
diff --git a/jdmerge.h b/jdmerge.h
new file mode 100644
index 0000000..b583396
--- /dev/null
+++ b/jdmerge.h
@@ -0,0 +1,47 @@
+/*
+ * jdmerge.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, D. R. Commander.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+#define JPEG_INTERNALS
+#include "jpeglib.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+/* Private subobject */
+
+typedef struct {
+ struct jpeg_upsampler pub; /* public fields */
+
+ /* Pointer to routine to do actual upsampling/conversion of one row group */
+ void (*upmethod) (j_decompress_ptr cinfo, JSAMPIMAGE input_buf,
+ JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf);
+
+ /* Private state for YCC->RGB conversion */
+ int *Cr_r_tab; /* => table for Cr to R conversion */
+ int *Cb_b_tab; /* => table for Cb to B conversion */
+ JLONG *Cr_g_tab; /* => table for Cr to G conversion */
+ JLONG *Cb_g_tab; /* => table for Cb to G conversion */
+
+ /* For 2:1 vertical sampling, we produce two output rows at a time.
+ * We need a "spare" row buffer to hold the second output row if the
+ * application provides just a one-row buffer; we also use the spare
+ * to discard the dummy last row if the image height is odd.
+ */
+ JSAMPROW spare_row;
+ boolean spare_full; /* T if spare buffer is occupied */
+
+ JDIMENSION out_row_width; /* samples per output row */
+ JDIMENSION rows_to_go; /* counts rows remaining in image */
+} my_merged_upsampler;
+
+typedef my_merged_upsampler *my_merged_upsample_ptr;
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */
diff --git a/jdmrg565.c b/jdmrg565.c
index 1b87e37..980a4e2 100644
--- a/jdmrg565.c
+++ b/jdmrg565.c
@@ -5,7 +5,7 @@
* Copyright (C) 1994-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
* Copyright (C) 2013, Linaro Limited.
- * Copyright (C) 2014-2015, 2018, D. R. Commander.
+ * Copyright (C) 2014-2015, 2018, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -19,7 +19,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr;
@@ -43,20 +43,20 @@
/* Loop for each pair of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 2 Y values and emit 2 pixels */
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
@@ -68,12 +68,12 @@
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr0);
+ y = *inptr0;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
@@ -90,7 +90,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr;
@@ -115,21 +115,21 @@
/* Loop for each pair of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 2 Y values and emit 2 pixels */
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
d0 = DITHER_ROTATE(d0);
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
@@ -142,12 +142,12 @@
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr0);
+ y = *inptr0;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
@@ -163,7 +163,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr0, outptr1;
@@ -189,20 +189,20 @@
/* Loop for each group of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 4 Y values and emit 4 pixels */
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
@@ -211,13 +211,13 @@
WRITE_TWO_PIXELS(outptr0, rgb);
outptr0 += 4;
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
@@ -229,20 +229,20 @@
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr00);
+ y = *inptr00;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr0 = (INT16)rgb;
- y = GETJSAMPLE(*inptr01);
+ y = *inptr01;
r = range_limit[y + cred];
g = range_limit[y + cgreen];
b = range_limit[y + cblue];
@@ -259,7 +259,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr0, outptr1;
@@ -287,21 +287,21 @@
/* Loop for each group of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 4 Y values and emit 4 pixels */
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
d0 = DITHER_ROTATE(d0);
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
@@ -311,14 +311,14 @@
WRITE_TWO_PIXELS(outptr0, rgb);
outptr0 += 4;
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
r = range_limit[DITHER_565_R(y + cred, d1)];
g = range_limit[DITHER_565_G(y + cgreen, d1)];
b = range_limit[DITHER_565_B(y + cblue, d1)];
d1 = DITHER_ROTATE(d1);
rgb = PACK_SHORT_565(r, g, b);
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
r = range_limit[DITHER_565_R(y + cred, d1)];
g = range_limit[DITHER_565_G(y + cgreen, d1)];
b = range_limit[DITHER_565_B(y + cblue, d1)];
@@ -331,20 +331,20 @@
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr00);
+ y = *inptr00;
r = range_limit[DITHER_565_R(y + cred, d0)];
g = range_limit[DITHER_565_G(y + cgreen, d0)];
b = range_limit[DITHER_565_B(y + cblue, d0)];
rgb = PACK_SHORT_565(r, g, b);
*(INT16 *)outptr0 = (INT16)rgb;
- y = GETJSAMPLE(*inptr01);
+ y = *inptr01;
r = range_limit[DITHER_565_R(y + cred, d1)];
g = range_limit[DITHER_565_G(y + cgreen, d1)];
b = range_limit[DITHER_565_B(y + cblue, d1)];
diff --git a/jdmrgext.c b/jdmrgext.c
index b1c27df..9bf4f1a 100644
--- a/jdmrgext.c
+++ b/jdmrgext.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1994-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2011, 2015, D. R. Commander.
+ * Copyright (C) 2011, 2015, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -25,7 +25,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr;
@@ -46,13 +46,13 @@
/* Loop for each pair of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 2 Y values and emit 2 pixels */
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
outptr[RGB_RED] = range_limit[y + cred];
outptr[RGB_GREEN] = range_limit[y + cgreen];
outptr[RGB_BLUE] = range_limit[y + cblue];
@@ -60,7 +60,7 @@
outptr[RGB_ALPHA] = 0xFF;
#endif
outptr += RGB_PIXELSIZE;
- y = GETJSAMPLE(*inptr0++);
+ y = *inptr0++;
outptr[RGB_RED] = range_limit[y + cred];
outptr[RGB_GREEN] = range_limit[y + cgreen];
outptr[RGB_BLUE] = range_limit[y + cblue];
@@ -71,12 +71,12 @@
}
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr0);
+ y = *inptr0;
outptr[RGB_RED] = range_limit[y + cred];
outptr[RGB_GREEN] = range_limit[y + cgreen];
outptr[RGB_BLUE] = range_limit[y + cblue];
@@ -97,7 +97,7 @@
JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf)
{
- my_upsample_ptr upsample = (my_upsample_ptr)cinfo->upsample;
+ my_merged_upsample_ptr upsample = (my_merged_upsample_ptr)cinfo->upsample;
register int y, cred, cgreen, cblue;
int cb, cr;
register JSAMPROW outptr0, outptr1;
@@ -120,13 +120,13 @@
/* Loop for each group of output pixels */
for (col = cinfo->output_width >> 1; col > 0; col--) {
/* Do the chroma part of the calculation */
- cb = GETJSAMPLE(*inptr1++);
- cr = GETJSAMPLE(*inptr2++);
+ cb = *inptr1++;
+ cr = *inptr2++;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
/* Fetch 4 Y values and emit 4 pixels */
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
outptr0[RGB_RED] = range_limit[y + cred];
outptr0[RGB_GREEN] = range_limit[y + cgreen];
outptr0[RGB_BLUE] = range_limit[y + cblue];
@@ -134,7 +134,7 @@
outptr0[RGB_ALPHA] = 0xFF;
#endif
outptr0 += RGB_PIXELSIZE;
- y = GETJSAMPLE(*inptr00++);
+ y = *inptr00++;
outptr0[RGB_RED] = range_limit[y + cred];
outptr0[RGB_GREEN] = range_limit[y + cgreen];
outptr0[RGB_BLUE] = range_limit[y + cblue];
@@ -142,7 +142,7 @@
outptr0[RGB_ALPHA] = 0xFF;
#endif
outptr0 += RGB_PIXELSIZE;
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
outptr1[RGB_RED] = range_limit[y + cred];
outptr1[RGB_GREEN] = range_limit[y + cgreen];
outptr1[RGB_BLUE] = range_limit[y + cblue];
@@ -150,7 +150,7 @@
outptr1[RGB_ALPHA] = 0xFF;
#endif
outptr1 += RGB_PIXELSIZE;
- y = GETJSAMPLE(*inptr01++);
+ y = *inptr01++;
outptr1[RGB_RED] = range_limit[y + cred];
outptr1[RGB_GREEN] = range_limit[y + cgreen];
outptr1[RGB_BLUE] = range_limit[y + cblue];
@@ -161,19 +161,19 @@
}
/* If image width is odd, do the last output column separately */
if (cinfo->output_width & 1) {
- cb = GETJSAMPLE(*inptr1);
- cr = GETJSAMPLE(*inptr2);
+ cb = *inptr1;
+ cr = *inptr2;
cred = Crrtab[cr];
cgreen = (int)RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
cblue = Cbbtab[cb];
- y = GETJSAMPLE(*inptr00);
+ y = *inptr00;
outptr0[RGB_RED] = range_limit[y + cred];
outptr0[RGB_GREEN] = range_limit[y + cgreen];
outptr0[RGB_BLUE] = range_limit[y + cblue];
#ifdef RGB_ALPHA
outptr0[RGB_ALPHA] = 0xFF;
#endif
- y = GETJSAMPLE(*inptr01);
+ y = *inptr01;
outptr1[RGB_RED] = range_limit[y + cred];
outptr1[RGB_GREEN] = range_limit[y + cgreen];
outptr1[RGB_BLUE] = range_limit[y + cblue];
diff --git a/jdphuff.c b/jdphuff.c
index 9e82636..cac4582 100644
--- a/jdphuff.c
+++ b/jdphuff.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1995-1997, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2015-2016, 2018, D. R. Commander.
+ * Copyright (C) 2015-2016, 2018-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -41,25 +41,6 @@
int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
} savable_state;
-/* This macro is to work around compilers with missing or broken
- * structure assignment. You'll need to fix this code if you have
- * such a compiler and you change MAX_COMPS_IN_SCAN.
- */
-
-#ifndef NO_STRUCT_ASSIGN
-#define ASSIGN_STATE(dest, src) ((dest) = (src))
-#else
-#if MAX_COMPS_IN_SCAN == 4
-#define ASSIGN_STATE(dest, src) \
- ((dest).EOBRUN = (src).EOBRUN, \
- (dest).last_dc_val[0] = (src).last_dc_val[0], \
- (dest).last_dc_val[1] = (src).last_dc_val[1], \
- (dest).last_dc_val[2] = (src).last_dc_val[2], \
- (dest).last_dc_val[3] = (src).last_dc_val[3])
-#endif
-#endif
-
-
typedef struct {
struct jpeg_entropy_decoder pub; /* public fields */
@@ -102,7 +83,7 @@
boolean is_DC_band, bad;
int ci, coefi, tbl;
d_derived_tbl **pdtbl;
- int *coef_bit_ptr;
+ int *coef_bit_ptr, *prev_coef_bit_ptr;
jpeg_component_info *compptr;
is_DC_band = (cinfo->Ss == 0);
@@ -143,8 +124,13 @@
for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
int cindex = cinfo->cur_comp_info[ci]->component_index;
coef_bit_ptr = &cinfo->coef_bits[cindex][0];
+ prev_coef_bit_ptr = &cinfo->coef_bits[cindex + cinfo->num_components][0];
if (!is_DC_band && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+ for (coefi = MIN(cinfo->Ss, 1); coefi <= MAX(cinfo->Se, 9); coefi++) {
+ if (cinfo->input_scan_number > 1)
+ prev_coef_bit_ptr[coefi] = coef_bit_ptr[coefi];
+ }
for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
if (cinfo->Ah != expected)
@@ -323,7 +309,7 @@
/* Load up working state */
BITREAD_LOAD_STATE(cinfo, entropy->bitstate);
- ASSIGN_STATE(state, entropy->saved);
+ state = entropy->saved;
/* Outer loop handles each block in the MCU */
@@ -356,7 +342,7 @@
/* Completed MCU, so update state */
BITREAD_SAVE_STATE(cinfo, entropy->bitstate);
- ASSIGN_STATE(entropy->saved, state);
+ entropy->saved = state;
}
/* Account for restart interval (no-op if not using restarts) */
@@ -676,7 +662,7 @@
/* Create progression status table */
cinfo->coef_bits = (int (*)[DCTSIZE2])
(*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- cinfo->num_components * DCTSIZE2 *
+ cinfo->num_components * 2 * DCTSIZE2 *
sizeof(int));
coef_bit_ptr = &cinfo->coef_bits[0][0];
for (ci = 0; ci < cinfo->num_components; ci++)
diff --git a/jdsample.c b/jdsample.c
index 49016ce..eaad72a 100644
--- a/jdsample.c
+++ b/jdsample.c
@@ -8,7 +8,7 @@
* Copyright (C) 2010, 2015-2016, D. R. Commander.
* Copyright (C) 2014, MIPS Technologies, Inc., California.
* Copyright (C) 2015, Google, Inc.
- * Copyright (C) 2019, Arm Limited.
+ * Copyright (C) 2019-2020, Arm Limited.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -177,7 +177,7 @@
outptr = output_data[outrow];
outend = outptr + cinfo->output_width;
while (outptr < outend) {
- invalue = *inptr++; /* don't need GETJSAMPLE() here */
+ invalue = *inptr++;
for (h = h_expand; h > 0; h--) {
*outptr++ = invalue;
}
@@ -213,7 +213,7 @@
outptr = output_data[inrow];
outend = outptr + cinfo->output_width;
while (outptr < outend) {
- invalue = *inptr++; /* don't need GETJSAMPLE() here */
+ invalue = *inptr++;
*outptr++ = invalue;
*outptr++ = invalue;
}
@@ -242,7 +242,7 @@
outptr = output_data[outrow];
outend = outptr + cinfo->output_width;
while (outptr < outend) {
- invalue = *inptr++; /* don't need GETJSAMPLE() here */
+ invalue = *inptr++;
*outptr++ = invalue;
*outptr++ = invalue;
}
@@ -283,20 +283,20 @@
inptr = input_data[inrow];
outptr = output_data[inrow];
/* Special case for first column */
- invalue = GETJSAMPLE(*inptr++);
+ invalue = *inptr++;
*outptr++ = (JSAMPLE)invalue;
- *outptr++ = (JSAMPLE)((invalue * 3 + GETJSAMPLE(*inptr) + 2) >> 2);
+ *outptr++ = (JSAMPLE)((invalue * 3 + inptr[0] + 2) >> 2);
for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
/* General case: 3/4 * nearer pixel + 1/4 * further pixel */
- invalue = GETJSAMPLE(*inptr++) * 3;
- *outptr++ = (JSAMPLE)((invalue + GETJSAMPLE(inptr[-2]) + 1) >> 2);
- *outptr++ = (JSAMPLE)((invalue + GETJSAMPLE(*inptr) + 2) >> 2);
+ invalue = (*inptr++) * 3;
+ *outptr++ = (JSAMPLE)((invalue + inptr[-2] + 1) >> 2);
+ *outptr++ = (JSAMPLE)((invalue + inptr[0] + 2) >> 2);
}
/* Special case for last column */
- invalue = GETJSAMPLE(*inptr);
- *outptr++ = (JSAMPLE)((invalue * 3 + GETJSAMPLE(inptr[-1]) + 1) >> 2);
+ invalue = *inptr;
+ *outptr++ = (JSAMPLE)((invalue * 3 + inptr[-1] + 1) >> 2);
*outptr++ = (JSAMPLE)invalue;
}
}
@@ -338,7 +338,7 @@
outptr = output_data[outrow++];
for (colctr = 0; colctr < compptr->downsampled_width; colctr++) {
- thiscolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+ thiscolsum = (*inptr0++) * 3 + (*inptr1++);
*outptr++ = (JSAMPLE)((thiscolsum + bias) >> 2);
}
}
@@ -381,8 +381,8 @@
outptr = output_data[outrow++];
/* Special case for first column */
- thiscolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
- nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+ thiscolsum = (*inptr0++) * 3 + (*inptr1++);
+ nextcolsum = (*inptr0++) * 3 + (*inptr1++);
*outptr++ = (JSAMPLE)((thiscolsum * 4 + 8) >> 4);
*outptr++ = (JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4);
lastcolsum = thiscolsum; thiscolsum = nextcolsum;
@@ -390,7 +390,7 @@
for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
/* General case: 3/4 * nearer pixel + 1/4 * further pixel in each */
/* dimension, thus 9/16, 3/16, 3/16, 1/16 overall */
- nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+ nextcolsum = (*inptr0++) * 3 + (*inptr1++);
*outptr++ = (JSAMPLE)((thiscolsum * 3 + lastcolsum + 8) >> 4);
*outptr++ = (JSAMPLE)((thiscolsum * 3 + nextcolsum + 7) >> 4);
lastcolsum = thiscolsum; thiscolsum = nextcolsum;
@@ -477,9 +477,12 @@
} else if (h_in_group == h_out_group &&
v_in_group * 2 == v_out_group && do_fancy) {
/* Non-fancy upsampling is handled by the generic method */
+#if defined(__arm__) || defined(__aarch64__) || \
+ defined(_M_ARM) || defined(_M_ARM64)
if (jsimd_can_h1v2_fancy_upsample())
upsample->methods[ci] = jsimd_h1v2_fancy_upsample;
else
+#endif
upsample->methods[ci] = h1v2_fancy_upsample;
upsample->pub.need_context_rows = TRUE;
} else if (h_in_group * 2 == h_out_group &&
diff --git a/jdtrans.c b/jdtrans.c
index 56713ef..d7ec4b8 100644
--- a/jdtrans.c
+++ b/jdtrans.c
@@ -3,8 +3,8 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1995-1997, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code relevant
- * to libjpeg-turbo.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -16,6 +16,7 @@
#define JPEG_INTERNALS
#include "jinclude.h"
#include "jpeglib.h"
+#include "jpegcomp.h"
/* Forward declarations */
diff --git a/jerror.h b/jerror.h
index 933a369..4476df2 100644
--- a/jerror.h
+++ b/jerror.h
@@ -207,6 +207,10 @@
#endif
#endif
JMESSAGE(JWRN_BOGUS_ICC, "Corrupt JPEG data: bad ICC marker")
+#if JPEG_LIB_VERSION < 70
+JMESSAGE(JERR_BAD_DROP_SAMPLING,
+ "Component index %d: mismatching sampling ratio %d:%d, %d:%d, %c")
+#endif
#ifdef JMAKE_ENUM_LIST
@@ -252,6 +256,15 @@
(cinfo)->err->msg_parm.i[2] = (p3), \
(cinfo)->err->msg_parm.i[3] = (p4), \
(*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
+#define ERREXIT6(cinfo, code, p1, p2, p3, p4, p5, p6) \
+ ((cinfo)->err->msg_code = (code), \
+ (cinfo)->err->msg_parm.i[0] = (p1), \
+ (cinfo)->err->msg_parm.i[1] = (p2), \
+ (cinfo)->err->msg_parm.i[2] = (p3), \
+ (cinfo)->err->msg_parm.i[3] = (p4), \
+ (cinfo)->err->msg_parm.i[4] = (p5), \
+ (cinfo)->err->msg_parm.i[5] = (p6), \
+ (*(cinfo)->err->error_exit) ((j_common_ptr)(cinfo)))
#define ERREXITS(cinfo, code, str) \
((cinfo)->err->msg_code = (code), \
strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
diff --git a/jfdctint.c b/jfdctint.c
index b47c306..c95a3a7 100644
--- a/jfdctint.c
+++ b/jfdctint.c
@@ -4,11 +4,11 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2015, D. R. Commander.
+ * Copyright (C) 2015, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
- * This file contains a slow-but-accurate integer implementation of the
+ * This file contains a slower but more accurate integer implementation of the
* forward DCT (Discrete Cosine Transform).
*
* A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
diff --git a/jidctint.c b/jidctint.c
index 98425d5..bb08748 100644
--- a/jidctint.c
+++ b/jidctint.c
@@ -3,13 +3,13 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1998, Thomas G. Lane.
- * Modification developed 2002-2009 by Guido Vollbeding.
+ * Modification developed 2002-2018 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2015, D. R. Commander.
+ * Copyright (C) 2015, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
- * This file contains a slow-but-accurate integer implementation of the
+ * This file contains a slower but more accurate integer implementation of the
* inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
* must also perform dequantization of the input coefficients.
*
@@ -417,7 +417,7 @@
/*
* Perform dequantization and inverse DCT on one block of coefficients,
- * producing a 7x7 output block.
+ * producing a reduced-size 7x7 output block.
*
* Optimized algorithm with 12 multiplications in the 1-D kernel.
* cK represents sqrt(2) * cos(K*pi/14).
@@ -1258,7 +1258,7 @@
/*
* Perform dequantization and inverse DCT on one block of coefficients,
- * producing a 11x11 output block.
+ * producing an 11x11 output block.
*
* Optimized algorithm with 24 multiplications in the 1-D kernel.
* cK represents sqrt(2) * cos(K*pi/22).
@@ -2398,7 +2398,7 @@
tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0]);
tmp0 = LEFT_SHIFT(tmp0, CONST_BITS);
/* Add fudge factor here for final descale. */
- tmp0 += 1 << (CONST_BITS - PASS1_BITS - 1);
+ tmp0 += ONE << (CONST_BITS - PASS1_BITS - 1);
z1 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4]);
tmp1 = MULTIPLY(z1, FIX(1.306562965)); /* c4[16] = c2[8] */
diff --git a/jmemmgr.c b/jmemmgr.c
index deffc9c..508ca74 100644
--- a/jmemmgr.c
+++ b/jmemmgr.c
@@ -95,19 +95,6 @@
#endif
#endif
-#ifdef WITH_SIMD
-#if (ALIGN_SIZE % 16)
- #error "ALIGN_SIZE is not a multiple of 16 bytes - required for SIMD instructions."
-#endif
-#if defined(__AVX2__) && (ALIGN_SIZE % 32)
- #error "AVX2 requires 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes."
-#elif defined(__ARM_NEON) && (ALIGN_SIZE % 32)
- /* 32-byte alignment allows us to extract more performance from */
- /* fancy-upsampling algorithms when using NEON. */
- #error "NEON optimizations rely on 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes."
-#endif
-#endif
-
/*
* We allocate objects from "pools", where each pool is gotten with a single
* request to jpeg_get_small() or jpeg_get_large(). There is no per-object
diff --git a/jmorecfg.h b/jmorecfg.h
index d0b9300..fb3a9cf 100644
--- a/jmorecfg.h
+++ b/jmorecfg.h
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1997, Thomas G. Lane.
* Modified 1997-2009 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2009, 2011, 2014-2015, 2018, D. R. Commander.
+ * Copyright (C) 2009, 2011, 2014-2015, 2018, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -43,25 +43,11 @@
#if BITS_IN_JSAMPLE == 8
/* JSAMPLE should be the smallest type that will hold the values 0..255.
- * You can use a signed char by having GETJSAMPLE mask it with 0xFF.
*/
-#ifdef HAVE_UNSIGNED_CHAR
-
typedef unsigned char JSAMPLE;
#define GETJSAMPLE(value) ((int)(value))
-#else /* not HAVE_UNSIGNED_CHAR */
-
-typedef char JSAMPLE;
-#ifdef __CHAR_UNSIGNED__
-#define GETJSAMPLE(value) ((int)(value))
-#else
-#define GETJSAMPLE(value) ((int)(value) & 0xFF)
-#endif /* __CHAR_UNSIGNED__ */
-
-#endif /* HAVE_UNSIGNED_CHAR */
-
#define MAXJSAMPLE 255
#define CENTERJSAMPLE 128
@@ -97,22 +83,9 @@
* managers, this is also the data type passed to fread/fwrite.
*/
-#ifdef HAVE_UNSIGNED_CHAR
-
typedef unsigned char JOCTET;
#define GETJOCTET(value) (value)
-#else /* not HAVE_UNSIGNED_CHAR */
-
-typedef char JOCTET;
-#ifdef __CHAR_UNSIGNED__
-#define GETJOCTET(value) (value)
-#else
-#define GETJOCTET(value) ((value) & 0xFF)
-#endif /* __CHAR_UNSIGNED__ */
-
-#endif /* HAVE_UNSIGNED_CHAR */
-
/* These typedefs are used for various table entries and so forth.
* They must be at least as wide as specified; but making them too big
@@ -123,15 +96,7 @@
/* UINT8 must hold at least the values 0..255. */
-#ifdef HAVE_UNSIGNED_CHAR
typedef unsigned char UINT8;
-#else /* not HAVE_UNSIGNED_CHAR */
-#ifdef __CHAR_UNSIGNED__
-typedef char UINT8;
-#else /* not __CHAR_UNSIGNED__ */
-typedef short UINT8;
-#endif /* __CHAR_UNSIGNED__ */
-#endif /* HAVE_UNSIGNED_CHAR */
/* UINT16 must hold at least the values 0..65535. */
@@ -273,9 +238,9 @@
/* Capability options common to encoder and decoder: */
-#define DCT_ISLOW_SUPPORTED /* slow but accurate integer algorithm */
-#define DCT_IFAST_SUPPORTED /* faster, less accurate integer method */
-#define DCT_FLOAT_SUPPORTED /* floating-point: accurate, fast on fast HW */
+#define DCT_ISLOW_SUPPORTED /* accurate integer method */
+#define DCT_IFAST_SUPPORTED /* less accurate int method [legacy feature] */
+#define DCT_FLOAT_SUPPORTED /* floating-point method [legacy feature] */
/* Encoder capability options: */
diff --git a/jpegcomp.h b/jpegcomp.h
index b32d544..c4834ac 100644
--- a/jpegcomp.h
+++ b/jpegcomp.h
@@ -1,7 +1,7 @@
/*
* jpegcomp.h
*
- * Copyright (C) 2010, D. R. Commander.
+ * Copyright (C) 2010, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -19,6 +19,7 @@
#define _min_DCT_v_scaled_size min_DCT_v_scaled_size
#define _jpeg_width jpeg_width
#define _jpeg_height jpeg_height
+#define JERR_ARITH_NOTIMPL JERR_NOT_COMPILED
#else
#define _DCT_scaled_size DCT_scaled_size
#define _DCT_h_scaled_size DCT_scaled_size
diff --git a/jpegint.h b/jpegint.h
index ad36ca8..195fbcb 100644
--- a/jpegint.h
+++ b/jpegint.h
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1997, Thomas G. Lane.
* Modified 1997-2009 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2015-2016, D. R. Commander.
+ * Copyright (C) 2015-2016, 2019, D. R. Commander.
* Copyright (C) 2015, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
@@ -158,6 +158,9 @@
JDIMENSION first_MCU_col[MAX_COMPONENTS];
JDIMENSION last_MCU_col[MAX_COMPONENTS];
boolean jinit_upsampler_no_alloc;
+
+ /* Last iMCU row that was successfully decoded */
+ JDIMENSION last_good_iMCU_row;
};
/* Input control module */
diff --git a/jpeglib.h b/jpeglib.h
index dfa68fa..d7664f0 100644
--- a/jpeglib.h
+++ b/jpeglib.h
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1998, Thomas G. Lane.
* Modified 2002-2009 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2009-2011, 2013-2014, 2016-2017, D. R. Commander.
+ * Copyright (C) 2009-2011, 2013-2014, 2016-2017, 2020, D. R. Commander.
* Copyright (C) 2015, Google, Inc.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
@@ -18,10 +18,6 @@
#ifndef JPEGLIB_H
#define JPEGLIB_H
-/* Begin chromium edits */
-#include "jpeglibmangler.h"
-/* End chromium edits */
-
/*
* First we include the configuration files that record how this
* installation of the JPEG library is set up. jconfig.h can be
@@ -248,9 +244,9 @@
/* DCT/IDCT algorithm options. */
typedef enum {
- JDCT_ISLOW, /* slow but accurate integer algorithm */
- JDCT_IFAST, /* faster, less accurate integer method */
- JDCT_FLOAT /* floating-point: accurate, fast on fast HW */
+ JDCT_ISLOW, /* accurate integer method */
+ JDCT_IFAST, /* less accurate integer method [legacy feature] */
+ JDCT_FLOAT /* floating-point method [legacy feature] */
} J_DCT_METHOD;
#ifndef JDCT_DEFAULT /* may be overridden in jconfig.h */
diff --git a/jpegtran.1 b/jpegtran.1
index 2efb264..da7a266 100644
--- a/jpegtran.1
+++ b/jpegtran.1
@@ -1,4 +1,4 @@
-.TH JPEGTRAN 1 "18 March 2017"
+.TH JPEGTRAN 1 "26 October 2020"
.SH NAME
jpegtran \- lossless transformation of JPEG files
.SH SYNOPSIS
@@ -161,13 +161,13 @@
.PP
This version of \fBjpegtran\fR also offers a lossless crop option, which
discards data outside of a given image region but losslessly preserves what is
-inside. Like the rotate and flip transforms, lossless crop is restricted by the
-current JPEG format; the upper left corner of the selected region must fall on
-an iMCU boundary. If it doesn't, then it is silently moved up and/or left to
-the nearest iMCU boundary (the lower right corner is unchanged.) Thus, the
+inside. Like the rotate and flip transforms, lossless crop is restricted by
+the current JPEG format; the upper left corner of the selected region must fall
+on an iMCU boundary. If it doesn't, then it is silently moved up and/or left
+to the nearest iMCU boundary (the lower right corner is unchanged.) Thus, the
output image covers at least the requested region, but it may cover more. The
-adjustment of the region dimensions may be optionally disabled by attaching
-an 'f' character ("force") to the width or height number.
+adjustment of the region dimensions may be optionally disabled by attaching an
+'f' character ("force") to the width or height number.
The image can be losslessly cropped by giving the switch:
.TP
@@ -180,6 +180,47 @@
doesn't, then it is silently moved up and/or left to the nearest iMCU boundary
(the lower right corner is unchanged.)
.PP
+If W or H is larger than the width/height of the input image, then the output
+image is expanded in size, and the expanded region is filled in with zeros
+(neutral gray). Attaching an 'f' character ("flatten") to the width number
+will cause each block in the expanded region to be filled in with the DC
+coefficient of the nearest block in the input image rather than grayed out.
+Attaching an 'r' character ("reflect") to the width number will cause the
+expanded region to be filled in with repeated reflections of the input image
+rather than grayed out.
+.PP
+A complementary lossless wipe option is provided to discard (gray out) data
+inside a given image region while losslessly preserving what is outside:
+.TP
+.B \-wipe WxH+X+Y
+Wipe (gray out) a rectangular region of width W and height H from the input
+image, starting at point X,Y.
+.PP
+Attaching an 'f' character ("flatten") to the width number will cause the
+region to be filled with the average of adjacent blocks rather than grayed out.
+If the wipe region and the region outside the wipe region, when adjusted to the
+nearest iMCU boundary, form two horizontally adjacent rectangles, then
+attaching an 'r' character ("reflect") to the width number will cause the wipe
+region to be filled with repeated reflections of the outside region rather than
+grayed out.
+.PP
+A lossless drop option is also provided, which allows another JPEG image to be
+inserted ("dropped") into the input image data at a given position, replacing
+the existing image data at that position:
+.TP
+.B \-drop +X+Y filename
+Drop (insert) another image at point X,Y
+.PP
+Both the input image and the drop image must have the same subsampling level.
+It is best if they also have the same quantization (quality.) Otherwise, the
+quantization of the output image will be adapted to accommodate the higher of
+the input image quality and the drop image quality. The trim option can be
+used with the drop option to requantize the drop image to match the input
+image. Note that a grayscale image can be dropped into a full-color image or
+vice versa, as long as the full-color image has no vertical subsampling. If
+the input image is grayscale and the drop image is full-color, then the
+chrominance channels from the drop image will be discarded.
+.PP
Other not-strictly-lossless transformation switches are:
.TP
.B \-grayscale
@@ -229,9 +270,31 @@
.B \-max 4m
selects 4000000 bytes. If more space is needed, an error will occur.
.TP
+.BI \-maxscans " N"
+Abort if the input image contains more than
+.I N
+scans. This feature demonstrates a method by which applications can guard
+against denial-of-service attacks instigated by specially-crafted malformed
+JPEG images containing numerous scans with missing image data or image data
+consisting only of "EOB runs" (a feature of progressive JPEG images that allows
+potentially hundreds of thousands of adjoining zero-value pixels to be
+represented using only a few bytes.) Attempting to transform such malformed
+JPEG images can cause excessive CPU activity, since the decompressor must fully
+process each scan (even if the scan is corrupt) before it can proceed to the
+next scan.
+.TP
.BI \-outfile " name"
Send output image to the named file, not to standard output.
.TP
+.BI \-report
+Report transformation progress.
+.TP
+.BI \-strict
+Treat all warnings as fatal. This feature also demonstrates a method by which
+applications can guard against attacks instigated by specially-crafted
+malformed JPEG images. Enabling this option will cause the decompressor to
+abort if the input image contains incomplete or corrupt image data.
+.TP
.B \-verbose
Enable debug printout. More
.BR \-v 's
diff --git a/jpegtran.c b/jpegtran.c
index 1f7e521..90fda7d 100644
--- a/jpegtran.c
+++ b/jpegtran.c
@@ -2,9 +2,9 @@
* jpegtran.c
*
* This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1995-2010, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 1995-2019, Thomas G. Lane, Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2010, 2014, 2017, D. R. Commander.
+ * Copyright (C) 2010, 2014, 2017, 2019-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -41,7 +41,11 @@
static const char *progname; /* program name for error messages */
static char *icc_filename; /* for -icc switch */
+static JDIMENSION max_scans; /* for -maxscans switch */
static char *outfilename; /* for -outfile switch */
+static char *dropfilename; /* for -drop switch */
+static boolean report; /* for -report switch */
+static boolean strict; /* for -strict switch */
static JCOPY_OPTION copyoption; /* -copy switch */
static jpeg_transform_info transformoption; /* image transformation options */
@@ -69,9 +73,10 @@
#endif
fprintf(stderr, "Switches for modifying the image:\n");
#if TRANSFORMS_SUPPORTED
- fprintf(stderr, " -crop WxH+X+Y Crop to a rectangular subarea\n");
- fprintf(stderr, " -grayscale Reduce to grayscale (omit color data)\n");
+ fprintf(stderr, " -crop WxH+X+Y Crop to a rectangular region\n");
+ fprintf(stderr, " -drop +X+Y filename Drop (insert) another image\n");
fprintf(stderr, " -flip [horizontal|vertical] Mirror image (left-right or top-bottom)\n");
+ fprintf(stderr, " -grayscale Reduce to grayscale (omit color data)\n");
fprintf(stderr, " -perfect Fail if there is non-transformable edge blocks\n");
fprintf(stderr, " -rotate [90|180|270] Rotate image (degrees clockwise)\n");
#endif
@@ -79,6 +84,8 @@
fprintf(stderr, " -transpose Transpose image\n");
fprintf(stderr, " -transverse Transverse transpose image\n");
fprintf(stderr, " -trim Drop non-transformable edge blocks\n");
+ fprintf(stderr, " with -drop: Requantize drop file to match source file\n");
+ fprintf(stderr, " -wipe WxH+X+Y Wipe (gray out) a rectangular region\n");
#endif
fprintf(stderr, "Switches for advanced users:\n");
#ifdef C_ARITH_CODING_SUPPORTED
@@ -87,7 +94,10 @@
fprintf(stderr, " -icc FILE Embed ICC profile contained in FILE\n");
fprintf(stderr, " -restart N Set restart interval in rows, or in blocks with B\n");
fprintf(stderr, " -maxmemory N Maximum memory to use (in kbytes)\n");
+ fprintf(stderr, " -maxscans N Maximum number of scans to allow in input file\n");
fprintf(stderr, " -outfile name Specify name for output file\n");
+ fprintf(stderr, " -report Report transformation progress\n");
+ fprintf(stderr, " -strict Treat all warnings as fatal\n");
fprintf(stderr, " -verbose or -debug Emit debug output\n");
fprintf(stderr, " -version Print version information and exit\n");
fprintf(stderr, "Switches for wizards:\n");
@@ -141,7 +151,10 @@
/* Set up default JPEG parameters. */
simple_progressive = FALSE;
icc_filename = NULL;
+ max_scans = 0;
outfilename = NULL;
+ report = FALSE;
+ strict = FALSE;
copyoption = JCOPYOPT_DEFAULT;
transformoption.transform = JXFORM_NONE;
transformoption.perfect = FALSE;
@@ -193,7 +206,8 @@
#if TRANSFORMS_SUPPORTED
if (++argn >= argc) /* advance to next argument */
usage();
- if (!jtransform_parse_crop_spec(&transformoption, argv[argn])) {
+ if (transformoption.crop /* reject multiple crop/drop/wipe requests */ ||
+ !jtransform_parse_crop_spec(&transformoption, argv[argn])) {
fprintf(stderr, "%s: bogus -crop argument '%s'\n",
progname, argv[argn]);
exit(EXIT_FAILURE);
@@ -202,6 +216,26 @@
select_transform(JXFORM_NONE); /* force an error */
#endif
+ } else if (keymatch(arg, "drop", 2)) {
+#if TRANSFORMS_SUPPORTED
+ if (++argn >= argc) /* advance to next argument */
+ usage();
+ if (transformoption.crop /* reject multiple crop/drop/wipe requests */ ||
+ !jtransform_parse_crop_spec(&transformoption, argv[argn]) ||
+ transformoption.crop_width_set != JCROP_UNSET ||
+ transformoption.crop_height_set != JCROP_UNSET) {
+ fprintf(stderr, "%s: bogus -drop argument '%s'\n",
+ progname, argv[argn]);
+ exit(EXIT_FAILURE);
+ }
+ if (++argn >= argc) /* advance to next argument */
+ usage();
+ dropfilename = argv[argn];
+ select_transform(JXFORM_DROP);
+#else
+ select_transform(JXFORM_NONE); /* force an error */
+#endif
+
} else if (keymatch(arg, "debug", 1) || keymatch(arg, "verbose", 1)) {
/* Enable debug printouts. */
/* On first -d, print version identification */
@@ -261,6 +295,12 @@
lval *= 1000L;
cinfo->mem->max_memory_to_use = lval * 1000L;
+ } else if (keymatch(arg, "maxscans", 4)) {
+ if (++argn >= argc) /* advance to next argument */
+ usage();
+ if (sscanf(argv[argn], "%u", &max_scans) != 1)
+ usage();
+
} else if (keymatch(arg, "optimize", 1) || keymatch(arg, "optimise", 1)) {
/* Enable entropy parm optimization. */
#ifdef ENTROPY_OPT_SUPPORTED
@@ -293,6 +333,9 @@
exit(EXIT_FAILURE);
#endif
+ } else if (keymatch(arg, "report", 3)) {
+ report = TRUE;
+
} else if (keymatch(arg, "restart", 1)) {
/* Restart interval in MCU rows (or in MCUs with 'b'). */
long lval;
@@ -338,6 +381,9 @@
exit(EXIT_FAILURE);
#endif
+ } else if (keymatch(arg, "strict", 2)) {
+ strict = TRUE;
+
} else if (keymatch(arg, "transpose", 1)) {
/* Transpose (across UL-to-LR axis). */
select_transform(JXFORM_TRANSPOSE);
@@ -350,6 +396,21 @@
/* Trim off any partial edge MCUs that the transform can't handle. */
transformoption.trim = TRUE;
+ } else if (keymatch(arg, "wipe", 1)) {
+#if TRANSFORMS_SUPPORTED
+ if (++argn >= argc) /* advance to next argument */
+ usage();
+ if (transformoption.crop /* reject multiple crop/drop/wipe requests */ ||
+ !jtransform_parse_crop_spec(&transformoption, argv[argn])) {
+ fprintf(stderr, "%s: bogus -wipe argument '%s'\n",
+ progname, argv[argn]);
+ exit(EXIT_FAILURE);
+ }
+ select_transform(JXFORM_WIPE);
+#else
+ select_transform(JXFORM_NONE); /* force an error */
+#endif
+
} else {
usage(); /* bogus switch */
}
@@ -375,6 +436,19 @@
}
+METHODDEF(void)
+my_emit_message(j_common_ptr cinfo, int msg_level)
+{
+ if (msg_level < 0) {
+ /* Treat warning as fatal */
+ cinfo->err->error_exit(cinfo);
+ } else {
+ if (cinfo->err->trace_level >= msg_level)
+ cinfo->err->output_message(cinfo);
+ }
+}
+
+
/*
* The main program.
*/
@@ -387,11 +461,14 @@
#endif
{
struct jpeg_decompress_struct srcinfo;
+#if TRANSFORMS_SUPPORTED
+ struct jpeg_decompress_struct dropinfo;
+ struct jpeg_error_mgr jdroperr;
+ FILE *drop_file;
+#endif
struct jpeg_compress_struct dstinfo;
struct jpeg_error_mgr jsrcerr, jdsterr;
-#ifdef PROGRESS_REPORT
- struct cdjpeg_progress_mgr progress;
-#endif
+ struct cdjpeg_progress_mgr src_progress, dst_progress;
jvirt_barray_ptr *src_coef_arrays;
jvirt_barray_ptr *dst_coef_arrays;
int file_index;
@@ -424,13 +501,16 @@
* values read here are mostly ignored; we will rescan the switches after
* opening the input file. Also note that most of the switches affect the
* destination JPEG object, so we parse into that and then copy over what
- * needs to affects the source too.
+ * needs to affect the source too.
*/
file_index = parse_switches(&dstinfo, argc, argv, 0, FALSE);
jsrcerr.trace_level = jdsterr.trace_level;
srcinfo.mem->max_memory_to_use = dstinfo.mem->max_memory_to_use;
+ if (strict)
+ jsrcerr.emit_message = my_emit_message;
+
#ifdef TWO_FILE_COMMANDLINE
/* Must have either -outfile switch or explicit output file name */
if (outfilename == NULL) {
@@ -496,8 +576,29 @@
copyoption = JCOPYOPT_ALL_EXCEPT_ICC;
}
-#ifdef PROGRESS_REPORT
- start_progress_monitor((j_common_ptr)&dstinfo, &progress);
+ if (report) {
+ start_progress_monitor((j_common_ptr)&dstinfo, &dst_progress);
+ dst_progress.report = report;
+ }
+ if (report || max_scans != 0) {
+ start_progress_monitor((j_common_ptr)&srcinfo, &src_progress);
+ src_progress.report = report;
+ src_progress.max_scans = max_scans;
+ }
+#if TRANSFORMS_SUPPORTED
+ /* Open the drop file. */
+ if (dropfilename != NULL) {
+ if ((drop_file = fopen(dropfilename, READ_BINARY)) == NULL) {
+ fprintf(stderr, "%s: can't open %s for reading\n", progname,
+ dropfilename);
+ return EXIT_FAILURE;
+ }
+ dropinfo.err = jpeg_std_error(&jdroperr);
+ jpeg_create_decompress(&dropinfo);
+ jpeg_stdio_src(&dropinfo, drop_file);
+ } else {
+ drop_file = NULL;
+ }
#endif
/* Specify data source for decompression */
@@ -509,6 +610,17 @@
/* Read file header */
(void)jpeg_read_header(&srcinfo, TRUE);
+#if TRANSFORMS_SUPPORTED
+ if (dropfilename != NULL) {
+ (void)jpeg_read_header(&dropinfo, TRUE);
+ transformoption.crop_width = dropinfo.image_width;
+ transformoption.crop_width_set = JCROP_POS;
+ transformoption.crop_height = dropinfo.image_height;
+ transformoption.crop_height_set = JCROP_POS;
+ transformoption.drop_ptr = &dropinfo;
+ }
+#endif
+
/* Any space needed by a transform option must be requested before
* jpeg_read_coefficients so that memory allocation will be done right.
*/
@@ -524,6 +636,12 @@
/* Read source file as DCT coefficients */
src_coef_arrays = jpeg_read_coefficients(&srcinfo);
+#if TRANSFORMS_SUPPORTED
+ if (dropfilename != NULL) {
+ transformoption.drop_coef_arrays = jpeg_read_coefficients(&dropinfo);
+ }
+#endif
+
/* Initialize destination compression parameters from source values */
jpeg_copy_critical_parameters(&srcinfo, &dstinfo);
@@ -584,20 +702,36 @@
/* Finish compression and release memory */
jpeg_finish_compress(&dstinfo);
jpeg_destroy_compress(&dstinfo);
+#if TRANSFORMS_SUPPORTED
+ if (dropfilename != NULL) {
+ (void)jpeg_finish_decompress(&dropinfo);
+ jpeg_destroy_decompress(&dropinfo);
+ }
+#endif
(void)jpeg_finish_decompress(&srcinfo);
jpeg_destroy_decompress(&srcinfo);
/* Close output file, if we opened it */
if (fp != stdout)
fclose(fp);
-
-#ifdef PROGRESS_REPORT
- end_progress_monitor((j_common_ptr)&dstinfo);
+#if TRANSFORMS_SUPPORTED
+ if (drop_file != NULL)
+ fclose(drop_file);
#endif
+ if (report)
+ end_progress_monitor((j_common_ptr)&dstinfo);
+ if (report || max_scans != 0)
+ end_progress_monitor((j_common_ptr)&srcinfo);
+
free(icc_profile);
/* All done. */
+#if TRANSFORMS_SUPPORTED
+ if (dropfilename != NULL)
+ return (jsrcerr.num_warnings + jdroperr.num_warnings +
+ jdsterr.num_warnings ? EXIT_WARNING : EXIT_SUCCESS);
+#endif
return (jsrcerr.num_warnings + jdsterr.num_warnings ?
EXIT_WARNING : EXIT_SUCCESS);
}
diff --git a/jquant1.c b/jquant1.c
index 40bbb28..73b83e1 100644
--- a/jquant1.c
+++ b/jquant1.c
@@ -479,7 +479,7 @@
for (col = width; col > 0; col--) {
pixcode = 0;
for (ci = 0; ci < nc; ci++) {
- pixcode += GETJSAMPLE(colorindex[ci][GETJSAMPLE(*ptrin++)]);
+ pixcode += colorindex[ci][*ptrin++];
}
*ptrout++ = (JSAMPLE)pixcode;
}
@@ -506,9 +506,9 @@
ptrin = input_buf[row];
ptrout = output_buf[row];
for (col = width; col > 0; col--) {
- pixcode = GETJSAMPLE(colorindex0[GETJSAMPLE(*ptrin++)]);
- pixcode += GETJSAMPLE(colorindex1[GETJSAMPLE(*ptrin++)]);
- pixcode += GETJSAMPLE(colorindex2[GETJSAMPLE(*ptrin++)]);
+ pixcode = colorindex0[*ptrin++];
+ pixcode += colorindex1[*ptrin++];
+ pixcode += colorindex2[*ptrin++];
*ptrout++ = (JSAMPLE)pixcode;
}
}
@@ -552,7 +552,7 @@
* required amount of padding.
*/
*output_ptr +=
- colorindex_ci[GETJSAMPLE(*input_ptr) + dither[col_index]];
+ colorindex_ci[*input_ptr + dither[col_index]];
input_ptr += nc;
output_ptr++;
col_index = (col_index + 1) & ODITHER_MASK;
@@ -595,12 +595,9 @@
col_index = 0;
for (col = width; col > 0; col--) {
- pixcode =
- GETJSAMPLE(colorindex0[GETJSAMPLE(*input_ptr++) + dither0[col_index]]);
- pixcode +=
- GETJSAMPLE(colorindex1[GETJSAMPLE(*input_ptr++) + dither1[col_index]]);
- pixcode +=
- GETJSAMPLE(colorindex2[GETJSAMPLE(*input_ptr++) + dither2[col_index]]);
+ pixcode = colorindex0[(*input_ptr++) + dither0[col_index]];
+ pixcode += colorindex1[(*input_ptr++) + dither1[col_index]];
+ pixcode += colorindex2[(*input_ptr++) + dither2[col_index]];
*output_ptr++ = (JSAMPLE)pixcode;
col_index = (col_index + 1) & ODITHER_MASK;
}
@@ -677,15 +674,15 @@
* The maximum error is +- MAXJSAMPLE; this sets the required size
* of the range_limit array.
*/
- cur += GETJSAMPLE(*input_ptr);
- cur = GETJSAMPLE(range_limit[cur]);
+ cur += *input_ptr;
+ cur = range_limit[cur];
/* Select output value, accumulate into output code for this pixel */
- pixcode = GETJSAMPLE(colorindex_ci[cur]);
+ pixcode = colorindex_ci[cur];
*output_ptr += (JSAMPLE)pixcode;
/* Compute actual representation error at this pixel */
/* Note: we can do this even though we don't have the final */
/* pixel code, because the colormap is orthogonal. */
- cur -= GETJSAMPLE(colormap_ci[pixcode]);
+ cur -= colormap_ci[pixcode];
/* Compute error fractions to be propagated to adjacent pixels.
* Add these into the running sums, and simultaneously shift the
* next-line error sums left by 1 column.
diff --git a/jquant2.c b/jquant2.c
index 0ce0ca5..44efb18 100644
--- a/jquant2.c
+++ b/jquant2.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2009, 2014-2015, D. R. Commander.
+ * Copyright (C) 2009, 2014-2015, 2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -215,9 +215,9 @@
ptr = input_buf[row];
for (col = width; col > 0; col--) {
/* get pixel value and index into the histogram */
- histp = &histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
- [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
- [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
+ histp = &histogram[ptr[0] >> C0_SHIFT]
+ [ptr[1] >> C1_SHIFT]
+ [ptr[2] >> C2_SHIFT];
/* increment, check for overflow and undo increment if so. */
if (++(*histp) <= 0)
(*histp)--;
@@ -665,7 +665,7 @@
for (i = 0; i < numcolors; i++) {
/* We compute the squared-c0-distance term, then add in the other two. */
- x = GETJSAMPLE(cinfo->colormap[0][i]);
+ x = cinfo->colormap[0][i];
if (x < minc0) {
tdist = (x - minc0) * C0_SCALE;
min_dist = tdist * tdist;
@@ -688,7 +688,7 @@
}
}
- x = GETJSAMPLE(cinfo->colormap[1][i]);
+ x = cinfo->colormap[1][i];
if (x < minc1) {
tdist = (x - minc1) * C1_SCALE;
min_dist += tdist * tdist;
@@ -710,7 +710,7 @@
}
}
- x = GETJSAMPLE(cinfo->colormap[2][i]);
+ x = cinfo->colormap[2][i];
if (x < minc2) {
tdist = (x - minc2) * C2_SCALE;
min_dist += tdist * tdist;
@@ -788,13 +788,13 @@
#define STEP_C2 ((1 << C2_SHIFT) * C2_SCALE)
for (i = 0; i < numcolors; i++) {
- icolor = GETJSAMPLE(colorlist[i]);
+ icolor = colorlist[i];
/* Compute (square of) distance from minc0/c1/c2 to this color */
- inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
+ inc0 = (minc0 - cinfo->colormap[0][icolor]) * C0_SCALE;
dist0 = inc0 * inc0;
- inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
+ inc1 = (minc1 - cinfo->colormap[1][icolor]) * C1_SCALE;
dist0 += inc1 * inc1;
- inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
+ inc2 = (minc2 - cinfo->colormap[2][icolor]) * C2_SCALE;
dist0 += inc2 * inc2;
/* Form the initial difference increments */
inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
@@ -879,7 +879,7 @@
for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
cachep = &histogram[c0 + ic0][c1 + ic1][c2];
for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
- *cachep++ = (histcell)(GETJSAMPLE(*cptr++) + 1);
+ *cachep++ = (histcell)((*cptr++) + 1);
}
}
}
@@ -909,9 +909,9 @@
outptr = output_buf[row];
for (col = width; col > 0; col--) {
/* get pixel value and index into the cache */
- c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
- c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
- c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
+ c0 = (*inptr++) >> C0_SHIFT;
+ c1 = (*inptr++) >> C1_SHIFT;
+ c2 = (*inptr++) >> C2_SHIFT;
cachep = &histogram[c0][c1][c2];
/* If we have not seen this color before, find nearest colormap entry */
/* and update the cache */
@@ -996,12 +996,12 @@
* The maximum error is +- MAXJSAMPLE (or less with error limiting);
* this sets the required size of the range_limit array.
*/
- cur0 += GETJSAMPLE(inptr[0]);
- cur1 += GETJSAMPLE(inptr[1]);
- cur2 += GETJSAMPLE(inptr[2]);
- cur0 = GETJSAMPLE(range_limit[cur0]);
- cur1 = GETJSAMPLE(range_limit[cur1]);
- cur2 = GETJSAMPLE(range_limit[cur2]);
+ cur0 += inptr[0];
+ cur1 += inptr[1];
+ cur2 += inptr[2];
+ cur0 = range_limit[cur0];
+ cur1 = range_limit[cur1];
+ cur2 = range_limit[cur2];
/* Index into the cache with adjusted pixel value */
cachep =
&histogram[cur0 >> C0_SHIFT][cur1 >> C1_SHIFT][cur2 >> C2_SHIFT];
@@ -1015,9 +1015,9 @@
register int pixcode = *cachep - 1;
*outptr = (JSAMPLE)pixcode;
/* Compute representation error for this pixel */
- cur0 -= GETJSAMPLE(colormap0[pixcode]);
- cur1 -= GETJSAMPLE(colormap1[pixcode]);
- cur2 -= GETJSAMPLE(colormap2[pixcode]);
+ cur0 -= colormap0[pixcode];
+ cur1 -= colormap1[pixcode];
+ cur2 -= colormap2[pixcode];
}
/* Compute error fractions to be propagated to adjacent pixels.
* Add these into the running sums, and simultaneously shift the
@@ -1145,7 +1145,7 @@
int i;
/* Only F-S dithering or no dithering is supported. */
- /* If user asks for ordered dither, give him F-S. */
+ /* If user asks for ordered dither, give them F-S. */
if (cinfo->dither_mode != JDITHER_NONE)
cinfo->dither_mode = JDITHER_FS;
@@ -1263,7 +1263,7 @@
cquantize->sv_colormap = NULL;
/* Only F-S dithering or no dithering is supported. */
- /* If user asks for ordered dither, give him F-S. */
+ /* If user asks for ordered dither, give them F-S. */
if (cinfo->dither_mode != JDITHER_NONE)
cinfo->dither_mode = JDITHER_FS;
diff --git a/jsimd.h b/jsimd.h
index 33851dc..6c20365 100644
--- a/jsimd.h
+++ b/jsimd.h
@@ -4,6 +4,7 @@
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
* Copyright (C) 2011, 2014, D. R. Commander.
* Copyright (C) 2015-2016, 2018, Matthieu Darbois.
+ * Copyright (C) 2020, Arm Limited.
*
* Based on the x86 SIMD extension for IJG JPEG library,
* Copyright (C) 1999-2006, MIYASAKA Masaru.
diff --git a/jsimd_none.c b/jsimd_none.c
index 2087e69..5b38a9f 100644
--- a/jsimd_none.c
+++ b/jsimd_none.c
@@ -4,6 +4,7 @@
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
* Copyright (C) 2009-2011, 2014, D. R. Commander.
* Copyright (C) 2015-2016, 2018, Matthieu Darbois.
+ * Copyright (C) 2020, Arm Limited.
*
* Based on the x86 SIMD extension for IJG JPEG library,
* Copyright (C) 1999-2006, MIYASAKA Masaru.
diff --git a/jversion.h b/jversion.h
index ab4a2c5..d1d94f5 100644
--- a/jversion.h
+++ b/jversion.h
@@ -2,7 +2,7 @@
* jversion.h
*
* This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1991-2012, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 1991-2020, Thomas G. Lane, Guido Vollbeding.
* libjpeg-turbo Modifications:
* Copyright (C) 2010, 2012-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
@@ -30,23 +30,25 @@
* NOTE: It is our convention to place the authors in the following order:
* - libjpeg-turbo authors (2009-) in descending order of the date of their
* most recent contribution to the project, then in ascending order of the
- * date of their first contribution to the project
+ * date of their first contribution to the project, then in alphabetical
+ * order
* - Upstream authors in descending order of the date of the first inclusion of
* their code
*/
#define JCOPYRIGHT \
"Copyright (C) 2009-2020 D. R. Commander\n" \
- "Copyright (C) 2011-2016 Siarhei Siamashka\n" \
+ "Copyright (C) 2015, 2020 Google, Inc.\n" \
+ "Copyright (C) 2019-2020 Arm Limited\n" \
"Copyright (C) 2015-2016, 2018 Matthieu Darbois\n" \
+ "Copyright (C) 2011-2016 Siarhei Siamashka\n" \
"Copyright (C) 2015 Intel Corporation\n" \
- "Copyright (C) 2015 Google, Inc.\n" \
+ "Copyright (C) 2013-2014 Linaro Limited\n" \
"Copyright (C) 2013-2014 MIPS Technologies, Inc.\n" \
- "Copyright (C) 2013 Linaro Limited\n" \
+ "Copyright (C) 2009, 2012 Pierre Ossman for Cendio AB\n" \
"Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies)\n" \
- "Copyright (C) 2009 Pierre Ossman for Cendio AB\n" \
"Copyright (C) 1999-2006 MIYASAKA Masaru\n" \
- "Copyright (C) 1991-2016 Thomas G. Lane, Guido Vollbeding"
+ "Copyright (C) 1991-2020 Thomas G. Lane, Guido Vollbeding"
#define JCOPYRIGHT_SHORT \
"Copyright (C) 1991-2020 The libjpeg-turbo Project and many others"
diff --git a/libjpeg.txt b/libjpeg.txt
index c50cf90..3c680b5 100644
--- a/libjpeg.txt
+++ b/libjpeg.txt
@@ -3,7 +3,7 @@
This file was part of the Independent JPEG Group's software:
Copyright (C) 1994-2013, Thomas G. Lane, Guido Vollbeding.
libjpeg-turbo Modifications:
-Copyright (C) 2010, 2014-2018, D. R. Commander.
+Copyright (C) 2010, 2014-2018, 2020, D. R. Commander.
Copyright (C) 2015, Google, Inc.
For conditions of distribution and use, see the accompanying README.ijg file.
@@ -750,7 +750,9 @@
Suspending data sources are not supported by this function. Calling
jpeg_skip_scanlines() with a suspending data source will result in undefined
-behavior.
+behavior. Two-pass color quantization is also not supported by this function.
+Calling jpeg_skip_scanlines() with two-pass color quantization enabled will
+result in an error.
jpeg_skip_scanlines() will not allow skipping past the bottom of the image. If
the value of num_lines is large enough to skip past the bottom of the image,
@@ -967,30 +969,38 @@
J_DCT_METHOD dct_method
Selects the algorithm used for the DCT step. Choices are:
- JDCT_ISLOW: slow but accurate integer algorithm
- JDCT_IFAST: faster, less accurate integer method
- JDCT_FLOAT: floating-point method
+ JDCT_ISLOW: accurate integer method
+ JDCT_IFAST: less accurate integer method [legacy feature]
+ JDCT_FLOAT: floating-point method [legacy feature]
JDCT_DEFAULT: default method (normally JDCT_ISLOW)
JDCT_FASTEST: fastest method (normally JDCT_IFAST)
- In libjpeg-turbo, JDCT_IFAST is generally about 5-15% faster than
- JDCT_ISLOW when using the x86/x86-64 SIMD extensions (results may vary
- with other SIMD implementations, or when using libjpeg-turbo without
- SIMD extensions.) For quality levels of 90 and below, there should be
- little or no perceptible difference between the two algorithms. For
- quality levels above 90, however, the difference between JDCT_IFAST and
+ When the Independent JPEG Group's software was first released in 1991,
+ the compression time for a 1-megapixel JPEG image on a mainstream PC
+ was measured in minutes. Thus, JDCT_IFAST provided noticeable
+ performance benefits. On modern CPUs running libjpeg-turbo, however,
+ the compression time for a 1-megapixel JPEG image is measured in
+ milliseconds, and thus the performance benefits of JDCT_IFAST are much
+ less noticeable. On modern x86/x86-64 CPUs that support AVX2
+ instructions, JDCT_IFAST and JDCT_ISLOW have similar performance. On
+ other types of CPUs, JDCT_IFAST is generally about 5-15% faster than
+ JDCT_ISLOW.
+
+ For quality levels of 90 and below, there should be little or no
+ perceptible quality difference between the two algorithms. For quality
+ levels above 90, however, the difference between JDCT_IFAST and
JDCT_ISLOW becomes more pronounced. With quality=97, for instance,
- JDCT_IFAST incurs generally about a 1-3 dB loss (in PSNR) relative to
+ JDCT_IFAST incurs generally about a 1-3 dB loss in PSNR relative to
JDCT_ISLOW, but this can be larger for some images. Do not use
JDCT_IFAST with quality levels above 97. The algorithm often
degenerates at quality=98 and above and can actually produce a more
lossy image than if lower quality levels had been used. Also, in
libjpeg-turbo, JDCT_IFAST is not fully accelerated for quality levels
- above 97, so it will be slower than JDCT_ISLOW. JDCT_FLOAT is mainly a
- legacy feature. It does not produce significantly more accurate
- results than the ISLOW method, and it is much slower. The FLOAT method
- may also give different results on different machines due to varying
- roundoff behavior, whereas the integer methods should give the same
- results on all machines.
+ above 97, so it will be slower than JDCT_ISLOW.
+
+ JDCT_FLOAT does not produce significantly more accurate results than
+ JDCT_ISLOW, and it is much slower. JDCT_FLOAT may also give different
+ results on different machines due to varying roundoff behavior, whereas
+ the integer methods should give the same results on all machines.
J_COLOR_SPACE jpeg_color_space
int num_components
@@ -1268,31 +1278,39 @@
J_DCT_METHOD dct_method
Selects the algorithm used for the DCT step. Choices are:
- JDCT_ISLOW: slow but accurate integer algorithm
- JDCT_IFAST: faster, less accurate integer method
- JDCT_FLOAT: floating-point method
+ JDCT_ISLOW: accurate integer method
+ JDCT_IFAST: less accurate integer method [legacy feature]
+ JDCT_FLOAT: floating-point method [legacy feature]
JDCT_DEFAULT: default method (normally JDCT_ISLOW)
JDCT_FASTEST: fastest method (normally JDCT_IFAST)
- In libjpeg-turbo, JDCT_IFAST is generally about 5-15% faster than
- JDCT_ISLOW when using the x86/x86-64 SIMD extensions (results may vary
- with other SIMD implementations, or when using libjpeg-turbo without
- SIMD extensions.) If the JPEG image was compressed using a quality
- level of 85 or below, then there should be little or no perceptible
- difference between the two algorithms. When decompressing images that
- were compressed using quality levels above 85, however, the difference
+ When the Independent JPEG Group's software was first released in 1991,
+ the decompression time for a 1-megapixel JPEG image on a mainstream PC
+ was measured in minutes. Thus, JDCT_IFAST provided noticeable
+ performance benefits. On modern CPUs running libjpeg-turbo, however,
+ the decompression time for a 1-megapixel JPEG image is measured in
+ milliseconds, and thus the performance benefits of JDCT_IFAST are much
+ less noticeable. On modern x86/x86-64 CPUs that support AVX2
+ instructions, JDCT_IFAST and JDCT_ISLOW have similar performance. On
+ other types of CPUs, JDCT_IFAST is generally about 5-15% faster than
+ JDCT_ISLOW.
+
+ If the JPEG image was compressed using a quality level of 85 or below,
+ then there should be little or no perceptible quality difference
+ between the two algorithms. When decompressing images that were
+ compressed using quality levels above 85, however, the difference
between JDCT_IFAST and JDCT_ISLOW becomes more pronounced. With images
compressed using quality=97, for instance, JDCT_IFAST incurs generally
- about a 4-6 dB loss (in PSNR) relative to JDCT_ISLOW, but this can be
+ about a 4-6 dB loss in PSNR relative to JDCT_ISLOW, but this can be
larger for some images. If you can avoid it, do not use JDCT_IFAST
when decompressing images that were compressed using quality levels
above 97. The algorithm often degenerates for such images and can
actually produce a more lossy output image than if the JPEG image had
- been compressed using lower quality levels. JDCT_FLOAT is mainly a
- legacy feature. It does not produce significantly more accurate
- results than the ISLOW method, and it is much slower. The FLOAT method
- may also give different results on different machines due to varying
- roundoff behavior, whereas the integer methods should give the same
- results on all machines.
+ been compressed using lower quality levels.
+
+ JDCT_FLOAT does not produce significantly more accurate results than
+ JDCT_ISLOW, and it is much slower. JDCT_FLOAT may also give different
+ results on different machines due to varying roundoff behavior, whereas
+ the integer methods should give the same results on all machines.
boolean do_fancy_upsampling
If TRUE, do careful upsampling of chroma components. If FALSE,
diff --git a/rdbmp.c b/rdbmp.c
index 51af237..6ba584a 100644
--- a/rdbmp.c
+++ b/rdbmp.c
@@ -12,7 +12,7 @@
*
* This file contains routines to read input images in Microsoft "BMP"
* format (MS Windows 3.x, OS/2 1.x, and OS/2 2.x flavors).
- * Currently, only 8-bit and 24-bit images are supported, not 1-bit or
+ * Currently, only 8-, 24-, and 32-bit images are supported, not 1-bit or
* 4-bit (feeding such low-depth images into JPEG would be silly anyway).
* Also, we don't support RLE-compressed files.
*
@@ -34,18 +34,8 @@
/* Macros to deal with unsigned chars as efficiently as compiler allows */
-#ifdef HAVE_UNSIGNED_CHAR
typedef unsigned char U_CHAR;
#define UCH(x) ((int)(x))
-#else /* !HAVE_UNSIGNED_CHAR */
-#ifdef __CHAR_UNSIGNED__
-typedef char U_CHAR;
-#define UCH(x) ((int)(x))
-#else
-typedef char U_CHAR;
-#define UCH(x) ((int)(x) & 0xFF)
-#endif
-#endif /* HAVE_UNSIGNED_CHAR */
#define ReadOK(file, buffer, len) \
@@ -71,7 +61,7 @@
JDIMENSION source_row; /* Current source row number */
JDIMENSION row_width; /* Physical width of scanlines in file */
- int bits_per_pixel; /* remembers 8- or 24-bit format */
+ int bits_per_pixel; /* remembers 8-, 24-, or 32-bit format */
int cmap_length; /* colormap length */
boolean use_inversion_array; /* TRUE = preload the whole image, which is
@@ -179,14 +169,14 @@
outptr = source->pub.buffer[0];
if (cinfo->in_color_space == JCS_GRAYSCALE) {
for (col = cinfo->image_width; col > 0; col--) {
- t = GETJSAMPLE(*inptr++);
+ t = *inptr++;
if (t >= cmaplen)
ERREXIT(cinfo, JERR_BMP_OUTOFRANGE);
*outptr++ = colormap[0][t];
}
} else if (cinfo->in_color_space == JCS_CMYK) {
for (col = cinfo->image_width; col > 0; col--) {
- t = GETJSAMPLE(*inptr++);
+ t = *inptr++;
if (t >= cmaplen)
ERREXIT(cinfo, JERR_BMP_OUTOFRANGE);
rgb_to_cmyk(colormap[0][t], colormap[1][t], colormap[2][t], outptr,
@@ -202,7 +192,7 @@
if (aindex >= 0) {
for (col = cinfo->image_width; col > 0; col--) {
- t = GETJSAMPLE(*inptr++);
+ t = *inptr++;
if (t >= cmaplen)
ERREXIT(cinfo, JERR_BMP_OUTOFRANGE);
outptr[rindex] = colormap[0][t];
@@ -213,7 +203,7 @@
}
} else {
for (col = cinfo->image_width; col > 0; col--) {
- t = GETJSAMPLE(*inptr++);
+ t = *inptr++;
if (t >= cmaplen)
ERREXIT(cinfo, JERR_BMP_OUTOFRANGE);
outptr[rindex] = colormap[0][t];
@@ -258,7 +248,6 @@
MEMCOPY(outptr, inptr, source->row_width);
} else if (cinfo->in_color_space == JCS_CMYK) {
for (col = cinfo->image_width; col > 0; col--) {
- /* can omit GETJSAMPLE() safely */
JSAMPLE b = *inptr++, g = *inptr++, r = *inptr++;
rgb_to_cmyk(r, g, b, outptr, outptr + 1, outptr + 2, outptr + 3);
outptr += 4;
@@ -272,7 +261,7 @@
if (aindex >= 0) {
for (col = cinfo->image_width; col > 0; col--) {
- outptr[bindex] = *inptr++; /* can omit GETJSAMPLE() safely */
+ outptr[bindex] = *inptr++;
outptr[gindex] = *inptr++;
outptr[rindex] = *inptr++;
outptr[aindex] = 0xFF;
@@ -280,7 +269,7 @@
}
} else {
for (col = cinfo->image_width; col > 0; col--) {
- outptr[bindex] = *inptr++; /* can omit GETJSAMPLE() safely */
+ outptr[bindex] = *inptr++;
outptr[gindex] = *inptr++;
outptr[rindex] = *inptr++;
outptr += ps;
@@ -323,7 +312,6 @@
MEMCOPY(outptr, inptr, source->row_width);
} else if (cinfo->in_color_space == JCS_CMYK) {
for (col = cinfo->image_width; col > 0; col--) {
- /* can omit GETJSAMPLE() safely */
JSAMPLE b = *inptr++, g = *inptr++, r = *inptr++;
rgb_to_cmyk(r, g, b, outptr, outptr + 1, outptr + 2, outptr + 3);
inptr++; /* skip the 4th byte (Alpha channel) */
@@ -338,7 +326,7 @@
if (aindex >= 0) {
for (col = cinfo->image_width; col > 0; col--) {
- outptr[bindex] = *inptr++; /* can omit GETJSAMPLE() safely */
+ outptr[bindex] = *inptr++;
outptr[gindex] = *inptr++;
outptr[rindex] = *inptr++;
outptr[aindex] = *inptr++;
@@ -346,7 +334,7 @@
}
} else {
for (col = cinfo->image_width; col > 0; col--) {
- outptr[bindex] = *inptr++; /* can omit GETJSAMPLE() safely */
+ outptr[bindex] = *inptr++;
outptr[gindex] = *inptr++;
outptr[rindex] = *inptr++;
inptr++; /* skip the 4th byte (Alpha channel) */
@@ -481,7 +469,9 @@
TRACEMS2(cinfo, 1, JTRC_BMP_OS2_MAPPED, biWidth, biHeight);
break;
case 24: /* RGB image */
- TRACEMS2(cinfo, 1, JTRC_BMP_OS2, biWidth, biHeight);
+ case 32: /* RGB image + Alpha channel */
+ TRACEMS3(cinfo, 1, JTRC_BMP_OS2, biWidth, biHeight,
+ source->bits_per_pixel);
break;
default:
ERREXIT(cinfo, JERR_BMP_BADDEPTH);
@@ -508,10 +498,8 @@
TRACEMS2(cinfo, 1, JTRC_BMP_MAPPED, biWidth, biHeight);
break;
case 24: /* RGB image */
- TRACEMS2(cinfo, 1, JTRC_BMP, biWidth, biHeight);
- break;
case 32: /* RGB image + Alpha channel */
- TRACEMS2(cinfo, 1, JTRC_BMP, biWidth, biHeight);
+ TRACEMS3(cinfo, 1, JTRC_BMP, biWidth, biHeight, source->bits_per_pixel);
break;
default:
ERREXIT(cinfo, JERR_BMP_BADDEPTH);
diff --git a/rdcolmap.c b/rdcolmap.c
index cbbef59..d2ed95c 100644
--- a/rdcolmap.c
+++ b/rdcolmap.c
@@ -54,9 +54,8 @@
/* Check for duplicate color. */
for (index = 0; index < ncolors; index++) {
- if (GETJSAMPLE(colormap0[index]) == R &&
- GETJSAMPLE(colormap1[index]) == G &&
- GETJSAMPLE(colormap2[index]) == B)
+ if (colormap0[index] == R && colormap1[index] == G &&
+ colormap2[index] == B)
return; /* color is already in map */
}
diff --git a/rdgif.c b/rdgif.c
index ff9258d..e1ea56c 100644
--- a/rdgif.c
+++ b/rdgif.c
@@ -2,28 +2,656 @@
* rdgif.c
*
* Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2019 by Guido Vollbeding.
* This file is part of the Independent JPEG Group's software.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
* This file contains routines to read input images in GIF format.
*
- *****************************************************************************
- * NOTE: to avoid entanglements with Unisys' patent on LZW compression, *
- * the ability to read GIF files has been removed from the IJG distribution. *
- * Sorry about that. *
- *****************************************************************************
- *
- * We are required to state that
- * "The Graphics Interchange Format(c) is the Copyright property of
- * CompuServe Incorporated. GIF(sm) is a Service Mark property of
- * CompuServe Incorporated."
+ * These routines may need modification for non-Unix environments or
+ * specialized applications. As they stand, they assume input from
+ * an ordinary stdio stream. They further assume that reading begins
+ * at the start of the file; start_input may need work if the
+ * user interface has already read some data (e.g., to determine that
+ * the file is indeed GIF format).
+ */
+
+/*
+ * This code is loosely based on giftoppm from the PBMPLUS distribution
+ * of Feb. 1991. That file contains the following copyright notice:
+ * +-------------------------------------------------------------------+
+ * | Copyright 1990, David Koblas. |
+ * | Permission to use, copy, modify, and distribute this software |
+ * | and its documentation for any purpose and without fee is hereby |
+ * | granted, provided that the above copyright notice appear in all |
+ * | copies and that both that copyright notice and this permission |
+ * | notice appear in supporting documentation. This software is |
+ * | provided "as is" without express or implied warranty. |
+ * +-------------------------------------------------------------------+
*/
#include "cdjpeg.h" /* Common decls for cjpeg/djpeg applications */
#ifdef GIF_SUPPORTED
+
+/* Macros to deal with unsigned chars as efficiently as compiler allows */
+
+typedef unsigned char U_CHAR;
+#define UCH(x) ((int)(x))
+
+
+#define ReadOK(file, buffer, len) \
+ (JFREAD(file, buffer, len) == ((size_t)(len)))
+
+
+#define MAXCOLORMAPSIZE 256 /* max # of colors in a GIF colormap */
+#define NUMCOLORS 3 /* # of colors */
+#define CM_RED 0 /* color component numbers */
+#define CM_GREEN 1
+#define CM_BLUE 2
+
+#define MAX_LZW_BITS 12 /* maximum LZW code size */
+#define LZW_TABLE_SIZE (1 << MAX_LZW_BITS) /* # of possible LZW symbols */
+
+/* Macros for extracting header data --- note we assume chars may be signed */
+
+#define LM_to_uint(array, offset) \
+ ((unsigned int)UCH(array[offset]) + \
+ (((unsigned int)UCH(array[offset + 1])) << 8))
+
+#define BitSet(byte, bit) ((byte) & (bit))
+#define INTERLACE 0x40 /* mask for bit signifying interlaced image */
+#define COLORMAPFLAG 0x80 /* mask for bit signifying colormap presence */
+
+
+/*
+ * LZW decompression tables look like this:
+ * symbol_head[K] = prefix symbol of any LZW symbol K (0..LZW_TABLE_SIZE-1)
+ * symbol_tail[K] = suffix byte of any LZW symbol K (0..LZW_TABLE_SIZE-1)
+ * Note that entries 0..end_code of the above tables are not used,
+ * since those symbols represent raw bytes or special codes.
+ *
+ * The stack represents the not-yet-used expansion of the last LZW symbol.
+ * In the worst case, a symbol could expand to as many bytes as there are
+ * LZW symbols, so we allocate LZW_TABLE_SIZE bytes for the stack.
+ * (This is conservative since that number includes the raw-byte symbols.)
+ */
+
+
+/* Private version of data source object */
+
+typedef struct {
+ struct cjpeg_source_struct pub; /* public fields */
+
+ j_compress_ptr cinfo; /* back link saves passing separate parm */
+
+ JSAMPARRAY colormap; /* GIF colormap (converted to my format) */
+
+ /* State for GetCode and LZWReadByte */
+ U_CHAR code_buf[256 + 4]; /* current input data block */
+ int last_byte; /* # of bytes in code_buf */
+ int last_bit; /* # of bits in code_buf */
+ int cur_bit; /* next bit index to read */
+ boolean first_time; /* flags first call to GetCode */
+ boolean out_of_blocks; /* TRUE if hit terminator data block */
+
+ int input_code_size; /* codesize given in GIF file */
+ int clear_code, end_code; /* values for Clear and End codes */
+
+ int code_size; /* current actual code size */
+ int limit_code; /* 2^code_size */
+ int max_code; /* first unused code value */
+
+ /* Private state for LZWReadByte */
+ int oldcode; /* previous LZW symbol */
+ int firstcode; /* first byte of oldcode's expansion */
+
+ /* LZW symbol table and expansion stack */
+ UINT16 *symbol_head; /* => table of prefix symbols */
+ UINT8 *symbol_tail; /* => table of suffix bytes */
+ UINT8 *symbol_stack; /* => stack for symbol expansions */
+ UINT8 *sp; /* stack pointer */
+
+ /* State for interlaced image processing */
+ boolean is_interlaced; /* TRUE if have interlaced image */
+ jvirt_sarray_ptr interlaced_image; /* full image in interlaced order */
+ JDIMENSION cur_row_number; /* need to know actual row number */
+ JDIMENSION pass2_offset; /* # of pixel rows in pass 1 */
+ JDIMENSION pass3_offset; /* # of pixel rows in passes 1&2 */
+ JDIMENSION pass4_offset; /* # of pixel rows in passes 1,2,3 */
+} gif_source_struct;
+
+typedef gif_source_struct *gif_source_ptr;
+
+
+/* Forward declarations */
+METHODDEF(JDIMENSION) get_pixel_rows(j_compress_ptr cinfo,
+ cjpeg_source_ptr sinfo);
+METHODDEF(JDIMENSION) load_interlaced_image(j_compress_ptr cinfo,
+ cjpeg_source_ptr sinfo);
+METHODDEF(JDIMENSION) get_interlaced_row(j_compress_ptr cinfo,
+ cjpeg_source_ptr sinfo);
+
+
+LOCAL(int)
+ReadByte(gif_source_ptr sinfo)
+/* Read next byte from GIF file */
+{
+ register FILE *infile = sinfo->pub.input_file;
+ register int c;
+
+ if ((c = getc(infile)) == EOF)
+ ERREXIT(sinfo->cinfo, JERR_INPUT_EOF);
+ return c;
+}
+
+
+LOCAL(int)
+GetDataBlock(gif_source_ptr sinfo, U_CHAR *buf)
+/* Read a GIF data block, which has a leading count byte */
+/* A zero-length block marks the end of a data block sequence */
+{
+ int count;
+
+ count = ReadByte(sinfo);
+ if (count > 0) {
+ if (!ReadOK(sinfo->pub.input_file, buf, count))
+ ERREXIT(sinfo->cinfo, JERR_INPUT_EOF);
+ }
+ return count;
+}
+
+
+LOCAL(void)
+SkipDataBlocks(gif_source_ptr sinfo)
+/* Skip a series of data blocks, until a block terminator is found */
+{
+ U_CHAR buf[256];
+
+ while (GetDataBlock(sinfo, buf) > 0)
+ /* skip */;
+}
+
+
+LOCAL(void)
+ReInitLZW(gif_source_ptr sinfo)
+/* (Re)initialize LZW state; shared code for startup and Clear processing */
+{
+ sinfo->code_size = sinfo->input_code_size + 1;
+ sinfo->limit_code = sinfo->clear_code << 1; /* 2^code_size */
+ sinfo->max_code = sinfo->clear_code + 2; /* first unused code value */
+ sinfo->sp = sinfo->symbol_stack; /* init stack to empty */
+}
+
+
+LOCAL(void)
+InitLZWCode(gif_source_ptr sinfo)
+/* Initialize for a series of LZWReadByte (and hence GetCode) calls */
+{
+ /* GetCode initialization */
+ sinfo->last_byte = 2; /* make safe to "recopy last two bytes" */
+ sinfo->code_buf[0] = 0;
+ sinfo->code_buf[1] = 0;
+ sinfo->last_bit = 0; /* nothing in the buffer */
+ sinfo->cur_bit = 0; /* force buffer load on first call */
+ sinfo->first_time = TRUE;
+ sinfo->out_of_blocks = FALSE;
+
+ /* LZWReadByte initialization: */
+ /* compute special code values (note that these do not change later) */
+ sinfo->clear_code = 1 << sinfo->input_code_size;
+ sinfo->end_code = sinfo->clear_code + 1;
+ ReInitLZW(sinfo);
+}
+
+
+LOCAL(int)
+GetCode(gif_source_ptr sinfo)
+/* Fetch the next code_size bits from the GIF data */
+/* We assume code_size is less than 16 */
+{
+ register int accum;
+ int offs, count;
+
+ while (sinfo->cur_bit + sinfo->code_size > sinfo->last_bit) {
+ /* Time to reload the buffer */
+ /* First time, share code with Clear case */
+ if (sinfo->first_time) {
+ sinfo->first_time = FALSE;
+ return sinfo->clear_code;
+ }
+ if (sinfo->out_of_blocks) {
+ WARNMS(sinfo->cinfo, JWRN_GIF_NOMOREDATA);
+ return sinfo->end_code; /* fake something useful */
+ }
+ /* preserve last two bytes of what we have -- assume code_size <= 16 */
+ sinfo->code_buf[0] = sinfo->code_buf[sinfo->last_byte-2];
+ sinfo->code_buf[1] = sinfo->code_buf[sinfo->last_byte-1];
+ /* Load more bytes; set flag if we reach the terminator block */
+ if ((count = GetDataBlock(sinfo, &sinfo->code_buf[2])) == 0) {
+ sinfo->out_of_blocks = TRUE;
+ WARNMS(sinfo->cinfo, JWRN_GIF_NOMOREDATA);
+ return sinfo->end_code; /* fake something useful */
+ }
+ /* Reset counters */
+ sinfo->cur_bit = (sinfo->cur_bit - sinfo->last_bit) + 16;
+ sinfo->last_byte = 2 + count;
+ sinfo->last_bit = sinfo->last_byte * 8;
+ }
+
+ /* Form up next 24 bits in accum */
+ offs = sinfo->cur_bit >> 3; /* byte containing cur_bit */
+ accum = UCH(sinfo->code_buf[offs + 2]);
+ accum <<= 8;
+ accum |= UCH(sinfo->code_buf[offs + 1]);
+ accum <<= 8;
+ accum |= UCH(sinfo->code_buf[offs]);
+
+ /* Right-align cur_bit in accum, then mask off desired number of bits */
+ accum >>= (sinfo->cur_bit & 7);
+ sinfo->cur_bit += sinfo->code_size;
+ return accum & ((1 << sinfo->code_size) - 1);
+}
+
+
+LOCAL(int)
+LZWReadByte(gif_source_ptr sinfo)
+/* Read an LZW-compressed byte */
+{
+ register int code; /* current working code */
+ int incode; /* saves actual input code */
+
+ /* If any codes are stacked from a previously read symbol, return them */
+ if (sinfo->sp > sinfo->symbol_stack)
+ return (int)(*(--sinfo->sp));
+
+ /* Time to read a new symbol */
+ code = GetCode(sinfo);
+
+ if (code == sinfo->clear_code) {
+ /* Reinit state, swallow any extra Clear codes, and */
+ /* return next code, which is expected to be a raw byte. */
+ ReInitLZW(sinfo);
+ do {
+ code = GetCode(sinfo);
+ } while (code == sinfo->clear_code);
+ if (code > sinfo->clear_code) { /* make sure it is a raw byte */
+ WARNMS(sinfo->cinfo, JWRN_GIF_BADDATA);
+ code = 0; /* use something valid */
+ }
+ /* make firstcode, oldcode valid! */
+ sinfo->firstcode = sinfo->oldcode = code;
+ return code;
+ }
+
+ if (code == sinfo->end_code) {
+ /* Skip the rest of the image, unless GetCode already read terminator */
+ if (!sinfo->out_of_blocks) {
+ SkipDataBlocks(sinfo);
+ sinfo->out_of_blocks = TRUE;
+ }
+ /* Complain that there's not enough data */
+ WARNMS(sinfo->cinfo, JWRN_GIF_ENDCODE);
+ /* Pad data with 0's */
+ return 0; /* fake something usable */
+ }
+
+ /* Got normal raw byte or LZW symbol */
+ incode = code; /* save for a moment */
+
+ if (code >= sinfo->max_code) { /* special case for not-yet-defined symbol */
+ /* code == max_code is OK; anything bigger is bad data */
+ if (code > sinfo->max_code) {
+ WARNMS(sinfo->cinfo, JWRN_GIF_BADDATA);
+ incode = 0; /* prevent creation of loops in symbol table */
+ }
+ /* this symbol will be defined as oldcode/firstcode */
+ *(sinfo->sp++) = (UINT8)sinfo->firstcode;
+ code = sinfo->oldcode;
+ }
+
+ /* If it's a symbol, expand it into the stack */
+ while (code >= sinfo->clear_code) {
+ *(sinfo->sp++) = sinfo->symbol_tail[code]; /* tail is a byte value */
+ code = sinfo->symbol_head[code]; /* head is another LZW symbol */
+ }
+ /* At this point code just represents a raw byte */
+ sinfo->firstcode = code; /* save for possible future use */
+
+ /* If there's room in table... */
+ if ((code = sinfo->max_code) < LZW_TABLE_SIZE) {
+ /* Define a new symbol = prev sym + head of this sym's expansion */
+ sinfo->symbol_head[code] = (UINT16)sinfo->oldcode;
+ sinfo->symbol_tail[code] = (UINT8)sinfo->firstcode;
+ sinfo->max_code++;
+ /* Is it time to increase code_size? */
+ if (sinfo->max_code >= sinfo->limit_code &&
+ sinfo->code_size < MAX_LZW_BITS) {
+ sinfo->code_size++;
+ sinfo->limit_code <<= 1; /* keep equal to 2^code_size */
+ }
+ }
+
+ sinfo->oldcode = incode; /* save last input symbol for future use */
+ return sinfo->firstcode; /* return first byte of symbol's expansion */
+}
+
+
+LOCAL(void)
+ReadColorMap(gif_source_ptr sinfo, int cmaplen, JSAMPARRAY cmap)
+/* Read a GIF colormap */
+{
+ int i;
+
+ for (i = 0; i < cmaplen; i++) {
+#if BITS_IN_JSAMPLE == 8
+#define UPSCALE(x) (x)
+#else
+#define UPSCALE(x) ((x) << (BITS_IN_JSAMPLE - 8))
+#endif
+ cmap[CM_RED][i] = (JSAMPLE)UPSCALE(ReadByte(sinfo));
+ cmap[CM_GREEN][i] = (JSAMPLE)UPSCALE(ReadByte(sinfo));
+ cmap[CM_BLUE][i] = (JSAMPLE)UPSCALE(ReadByte(sinfo));
+ }
+}
+
+
+LOCAL(void)
+DoExtension(gif_source_ptr sinfo)
+/* Process an extension block */
+/* Currently we ignore 'em all */
+{
+ int extlabel;
+
+ /* Read extension label byte */
+ extlabel = ReadByte(sinfo);
+ TRACEMS1(sinfo->cinfo, 1, JTRC_GIF_EXTENSION, extlabel);
+ /* Skip the data block(s) associated with the extension */
+ SkipDataBlocks(sinfo);
+}
+
+
+/*
+ * Read the file header; return image size and component count.
+ */
+
+METHODDEF(void)
+start_input_gif(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+ gif_source_ptr source = (gif_source_ptr)sinfo;
+ U_CHAR hdrbuf[10]; /* workspace for reading control blocks */
+ unsigned int width, height; /* image dimensions */
+ int colormaplen, aspectRatio;
+ int c;
+
+ /* Read and verify GIF Header */
+ if (!ReadOK(source->pub.input_file, hdrbuf, 6))
+ ERREXIT(cinfo, JERR_GIF_NOT);
+ if (hdrbuf[0] != 'G' || hdrbuf[1] != 'I' || hdrbuf[2] != 'F')
+ ERREXIT(cinfo, JERR_GIF_NOT);
+ /* Check for expected version numbers.
+ * If unknown version, give warning and try to process anyway;
+ * this is per recommendation in GIF89a standard.
+ */
+ if ((hdrbuf[3] != '8' || hdrbuf[4] != '7' || hdrbuf[5] != 'a') &&
+ (hdrbuf[3] != '8' || hdrbuf[4] != '9' || hdrbuf[5] != 'a'))
+ TRACEMS3(cinfo, 1, JTRC_GIF_BADVERSION, hdrbuf[3], hdrbuf[4], hdrbuf[5]);
+
+ /* Read and decipher Logical Screen Descriptor */
+ if (!ReadOK(source->pub.input_file, hdrbuf, 7))
+ ERREXIT(cinfo, JERR_INPUT_EOF);
+ width = LM_to_uint(hdrbuf, 0);
+ height = LM_to_uint(hdrbuf, 2);
+ /* we ignore the color resolution, sort flag, and background color index */
+ aspectRatio = UCH(hdrbuf[6]);
+ if (aspectRatio != 0 && aspectRatio != 49)
+ TRACEMS(cinfo, 1, JTRC_GIF_NONSQUARE);
+
+ /* Allocate space to store the colormap */
+ source->colormap = (*cinfo->mem->alloc_sarray)
+ ((j_common_ptr)cinfo, JPOOL_IMAGE, (JDIMENSION)MAXCOLORMAPSIZE,
+ (JDIMENSION)NUMCOLORS);
+ colormaplen = 0; /* indicate initialization */
+
+ /* Read global colormap if header indicates it is present */
+ if (BitSet(hdrbuf[4], COLORMAPFLAG)) {
+ colormaplen = 2 << (hdrbuf[4] & 0x07);
+ ReadColorMap(source, colormaplen, source->colormap);
+ }
+
+ /* Scan until we reach start of desired image.
+ * We don't currently support skipping images, but could add it easily.
+ */
+ for (;;) {
+ c = ReadByte(source);
+
+ if (c == ';') /* GIF terminator?? */
+ ERREXIT(cinfo, JERR_GIF_IMAGENOTFOUND);
+
+ if (c == '!') { /* Extension */
+ DoExtension(source);
+ continue;
+ }
+
+ if (c != ',') { /* Not an image separator? */
+ WARNMS1(cinfo, JWRN_GIF_CHAR, c);
+ continue;
+ }
+
+ /* Read and decipher Local Image Descriptor */
+ if (!ReadOK(source->pub.input_file, hdrbuf, 9))
+ ERREXIT(cinfo, JERR_INPUT_EOF);
+ /* we ignore top/left position info, also sort flag */
+ width = LM_to_uint(hdrbuf, 4);
+ height = LM_to_uint(hdrbuf, 6);
+ source->is_interlaced = (BitSet(hdrbuf[8], INTERLACE) != 0);
+
+ /* Read local colormap if header indicates it is present */
+ /* Note: if we wanted to support skipping images, */
+ /* we'd need to skip rather than read colormap for ignored images */
+ if (BitSet(hdrbuf[8], COLORMAPFLAG)) {
+ colormaplen = 2 << (hdrbuf[8] & 0x07);
+ ReadColorMap(source, colormaplen, source->colormap);
+ }
+
+ source->input_code_size = ReadByte(source); /* get min-code-size byte */
+ if (source->input_code_size < 2 || source->input_code_size > 8)
+ ERREXIT1(cinfo, JERR_GIF_CODESIZE, source->input_code_size);
+
+ /* Reached desired image, so break out of loop */
+ /* If we wanted to skip this image, */
+ /* we'd call SkipDataBlocks and then continue the loop */
+ break;
+ }
+
+ /* Prepare to read selected image: first initialize LZW decompressor */
+ source->symbol_head = (UINT16 *)
+ (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ LZW_TABLE_SIZE * sizeof(UINT16));
+ source->symbol_tail = (UINT8 *)
+ (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ LZW_TABLE_SIZE * sizeof(UINT8));
+ source->symbol_stack = (UINT8 *)
+ (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ LZW_TABLE_SIZE * sizeof(UINT8));
+ InitLZWCode(source);
+
+ /*
+ * If image is interlaced, we read it into a full-size sample array,
+ * decompressing as we go; then get_interlaced_row selects rows from the
+ * sample array in the proper order.
+ */
+ if (source->is_interlaced) {
+ /* We request the virtual array now, but can't access it until virtual
+ * arrays have been allocated. Hence, the actual work of reading the
+ * image is postponed until the first call to get_pixel_rows.
+ */
+ source->interlaced_image = (*cinfo->mem->request_virt_sarray)
+ ((j_common_ptr)cinfo, JPOOL_IMAGE, FALSE,
+ (JDIMENSION)width, (JDIMENSION)height, (JDIMENSION)1);
+ if (cinfo->progress != NULL) {
+ cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
+ progress->total_extra_passes++; /* count file input as separate pass */
+ }
+ source->pub.get_pixel_rows = load_interlaced_image;
+ } else {
+ source->pub.get_pixel_rows = get_pixel_rows;
+ }
+
+ /* Create compressor input buffer. */
+ source->pub.buffer = (*cinfo->mem->alloc_sarray)
+ ((j_common_ptr)cinfo, JPOOL_IMAGE, (JDIMENSION)width * NUMCOLORS,
+ (JDIMENSION)1);
+ source->pub.buffer_height = 1;
+
+ /* Pad colormap for safety. */
+ for (c = colormaplen; c < source->clear_code; c++) {
+ source->colormap[CM_RED][c] =
+ source->colormap[CM_GREEN][c] =
+ source->colormap[CM_BLUE][c] = CENTERJSAMPLE;
+ }
+
+ /* Return info about the image. */
+ cinfo->in_color_space = JCS_RGB;
+ cinfo->input_components = NUMCOLORS;
+ cinfo->data_precision = BITS_IN_JSAMPLE; /* we always rescale data to this */
+ cinfo->image_width = width;
+ cinfo->image_height = height;
+
+ TRACEMS3(cinfo, 1, JTRC_GIF, width, height, colormaplen);
+}
+
+
+/*
+ * Read one row of pixels.
+ * This version is used for noninterlaced GIF images:
+ * we read directly from the GIF file.
+ */
+
+METHODDEF(JDIMENSION)
+get_pixel_rows(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+ gif_source_ptr source = (gif_source_ptr)sinfo;
+ register int c;
+ register JSAMPROW ptr;
+ register JDIMENSION col;
+ register JSAMPARRAY colormap = source->colormap;
+
+ ptr = source->pub.buffer[0];
+ for (col = cinfo->image_width; col > 0; col--) {
+ c = LZWReadByte(source);
+ *ptr++ = colormap[CM_RED][c];
+ *ptr++ = colormap[CM_GREEN][c];
+ *ptr++ = colormap[CM_BLUE][c];
+ }
+ return 1;
+}
+
+
+/*
+ * Read one row of pixels.
+ * This version is used for the first call on get_pixel_rows when
+ * reading an interlaced GIF file: we read the whole image into memory.
+ */
+
+METHODDEF(JDIMENSION)
+load_interlaced_image(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+ gif_source_ptr source = (gif_source_ptr)sinfo;
+ register JSAMPROW sptr;
+ register JDIMENSION col;
+ JDIMENSION row;
+ cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
+
+ /* Read the interlaced image into the virtual array we've created. */
+ for (row = 0; row < cinfo->image_height; row++) {
+ if (progress != NULL) {
+ progress->pub.pass_counter = (long)row;
+ progress->pub.pass_limit = (long)cinfo->image_height;
+ (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
+ }
+ sptr = *(*cinfo->mem->access_virt_sarray)
+ ((j_common_ptr)cinfo, source->interlaced_image, row, (JDIMENSION)1,
+ TRUE);
+ for (col = cinfo->image_width; col > 0; col--) {
+ *sptr++ = (JSAMPLE)LZWReadByte(source);
+ }
+ }
+ if (progress != NULL)
+ progress->completed_extra_passes++;
+
+ /* Replace method pointer so subsequent calls don't come here. */
+ source->pub.get_pixel_rows = get_interlaced_row;
+ /* Initialize for get_interlaced_row, and perform first call on it. */
+ source->cur_row_number = 0;
+ source->pass2_offset = (cinfo->image_height + 7) / 8;
+ source->pass3_offset = source->pass2_offset + (cinfo->image_height + 3) / 8;
+ source->pass4_offset = source->pass3_offset + (cinfo->image_height + 1) / 4;
+
+ return get_interlaced_row(cinfo, sinfo);
+}
+
+
+/*
+ * Read one row of pixels.
+ * This version is used for interlaced GIF images:
+ * we read from the virtual array.
+ */
+
+METHODDEF(JDIMENSION)
+get_interlaced_row(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+ gif_source_ptr source = (gif_source_ptr)sinfo;
+ register int c;
+ register JSAMPROW sptr, ptr;
+ register JDIMENSION col;
+ register JSAMPARRAY colormap = source->colormap;
+ JDIMENSION irow;
+
+ /* Figure out which row of interlaced image is needed, and access it. */
+ switch ((int)(source->cur_row_number & 7)) {
+ case 0: /* first-pass row */
+ irow = source->cur_row_number >> 3;
+ break;
+ case 4: /* second-pass row */
+ irow = (source->cur_row_number >> 3) + source->pass2_offset;
+ break;
+ case 2: /* third-pass row */
+ case 6:
+ irow = (source->cur_row_number >> 2) + source->pass3_offset;
+ break;
+ default: /* fourth-pass row */
+ irow = (source->cur_row_number >> 1) + source->pass4_offset;
+ }
+ sptr = *(*cinfo->mem->access_virt_sarray)
+ ((j_common_ptr)cinfo, source->interlaced_image, irow, (JDIMENSION)1,
+ FALSE);
+ /* Scan the row, expand colormap, and output */
+ ptr = source->pub.buffer[0];
+ for (col = cinfo->image_width; col > 0; col--) {
+ c = *sptr++;
+ *ptr++ = colormap[CM_RED][c];
+ *ptr++ = colormap[CM_GREEN][c];
+ *ptr++ = colormap[CM_BLUE][c];
+ }
+ source->cur_row_number++; /* for next time */
+ return 1;
+}
+
+
+/*
+ * Finish up at the end of the file.
+ */
+
+METHODDEF(void)
+finish_input_gif(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
+{
+ /* no work */
+}
+
+
/*
* The module selection routine for GIF format input.
*/
@@ -31,9 +659,18 @@
GLOBAL(cjpeg_source_ptr)
jinit_read_gif(j_compress_ptr cinfo)
{
- fprintf(stderr, "GIF input is unsupported for legal reasons. Sorry.\n");
- exit(EXIT_FAILURE);
- return NULL; /* keep compiler happy */
+ gif_source_ptr source;
+
+ /* Create module interface object */
+ source = (gif_source_ptr)
+ (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ sizeof(gif_source_struct));
+ source->cinfo = cinfo; /* make back link for subroutines */
+ /* Fill in method ptrs, except get_pixel_rows which start_input sets */
+ source->pub.start_input = start_input_gif;
+ source->pub.finish_input = finish_input_gif;
+
+ return (cjpeg_source_ptr)source;
}
#endif /* GIF_SUPPORTED */
diff --git a/rdppm.c b/rdppm.c
index a8507b9..c4c937e 100644
--- a/rdppm.c
+++ b/rdppm.c
@@ -43,18 +43,8 @@
/* Macros to deal with unsigned chars as efficiently as compiler allows */
-#ifdef HAVE_UNSIGNED_CHAR
typedef unsigned char U_CHAR;
#define UCH(x) ((int)(x))
-#else /* !HAVE_UNSIGNED_CHAR */
-#ifdef __CHAR_UNSIGNED__
-typedef char U_CHAR;
-#define UCH(x) ((int)(x))
-#else
-typedef char U_CHAR;
-#define UCH(x) ((int)(x) & 0xFF)
-#endif
-#endif /* HAVE_UNSIGNED_CHAR */
#define ReadOK(file, buffer, len) \
@@ -659,11 +649,12 @@
if (maxval > 255) {
source->pub.get_pixel_rows = get_word_rgb_row;
} else if (maxval == MAXJSAMPLE && sizeof(JSAMPLE) == sizeof(U_CHAR) &&
- (cinfo->in_color_space == JCS_EXT_RGB
#if RGB_RED == 0 && RGB_GREEN == 1 && RGB_BLUE == 2 && RGB_PIXELSIZE == 3
- || cinfo->in_color_space == JCS_RGB
+ (cinfo->in_color_space == JCS_EXT_RGB ||
+ cinfo->in_color_space == JCS_RGB)) {
+#else
+ cinfo->in_color_space == JCS_EXT_RGB) {
#endif
- )) {
source->pub.get_pixel_rows = get_raw_row;
use_raw_buffer = TRUE;
need_rescale = FALSE;
diff --git a/rdrle.c b/rdrle.c
deleted file mode 100644
index b694514..0000000
--- a/rdrle.c
+++ /dev/null
@@ -1,389 +0,0 @@
-/*
- * rdrle.c
- *
- * This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1991-1996, Thomas G. Lane.
- * It was modified by The libjpeg-turbo Project to include only code and
- * information relevant to libjpeg-turbo.
- * For conditions of distribution and use, see the accompanying README.ijg
- * file.
- *
- * This file contains routines to read input images in Utah RLE format.
- * The Utah Raster Toolkit library is required (version 3.1 or later).
- *
- * These routines may need modification for non-Unix environments or
- * specialized applications. As they stand, they assume input from
- * an ordinary stdio stream. They further assume that reading begins
- * at the start of the file; start_input may need work if the
- * user interface has already read some data (e.g., to determine that
- * the file is indeed RLE format).
- *
- * Based on code contributed by Mike Lijewski,
- * with updates from Robert Hutchinson.
- */
-
-#include "cdjpeg.h" /* Common decls for cjpeg/djpeg applications */
-
-#ifdef RLE_SUPPORTED
-
-/* rle.h is provided by the Utah Raster Toolkit. */
-
-#include <rle.h>
-
-/*
- * We assume that JSAMPLE has the same representation as rle_pixel,
- * to wit, "unsigned char". Hence we can't cope with 12- or 16-bit samples.
- */
-
-#if BITS_IN_JSAMPLE != 8
- Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
-#endif
-
-/*
- * We support the following types of RLE files:
- *
- * GRAYSCALE - 8 bits, no colormap
- * MAPPEDGRAY - 8 bits, 1 channel colomap
- * PSEUDOCOLOR - 8 bits, 3 channel colormap
- * TRUECOLOR - 24 bits, 3 channel colormap
- * DIRECTCOLOR - 24 bits, no colormap
- *
- * For now, we ignore any alpha channel in the image.
- */
-
-typedef enum
- { GRAYSCALE, MAPPEDGRAY, PSEUDOCOLOR, TRUECOLOR, DIRECTCOLOR } rle_kind;
-
-
-/*
- * Since RLE stores scanlines bottom-to-top, we have to invert the image
- * to conform to JPEG's top-to-bottom order. To do this, we read the
- * incoming image into a virtual array on the first get_pixel_rows call,
- * then fetch the required row from the virtual array on subsequent calls.
- */
-
-typedef struct _rle_source_struct *rle_source_ptr;
-
-typedef struct _rle_source_struct {
- struct cjpeg_source_struct pub; /* public fields */
-
- rle_kind visual; /* actual type of input file */
- jvirt_sarray_ptr image; /* virtual array to hold the image */
- JDIMENSION row; /* current row # in the virtual array */
- rle_hdr header; /* Input file information */
- rle_pixel **rle_row; /* holds a row returned by rle_getrow() */
-
-} rle_source_struct;
-
-
-/*
- * Read the file header; return image size and component count.
- */
-
-METHODDEF(void)
-start_input_rle(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
-{
- rle_source_ptr source = (rle_source_ptr)sinfo;
- JDIMENSION width, height;
-#ifdef PROGRESS_REPORT
- cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
-#endif
-
- /* Use RLE library routine to get the header info */
- source->header = *rle_hdr_init(NULL);
- source->header.rle_file = source->pub.input_file;
- switch (rle_get_setup(&(source->header))) {
- case RLE_SUCCESS:
- /* A-OK */
- break;
- case RLE_NOT_RLE:
- ERREXIT(cinfo, JERR_RLE_NOT);
- break;
- case RLE_NO_SPACE:
- ERREXIT(cinfo, JERR_RLE_MEM);
- break;
- case RLE_EMPTY:
- ERREXIT(cinfo, JERR_RLE_EMPTY);
- break;
- case RLE_EOF:
- ERREXIT(cinfo, JERR_RLE_EOF);
- break;
- default:
- ERREXIT(cinfo, JERR_RLE_BADERROR);
- break;
- }
-
- /* Figure out what we have, set private vars and return values accordingly */
-
- width = source->header.xmax - source->header.xmin + 1;
- height = source->header.ymax - source->header.ymin + 1;
- source->header.xmin = 0; /* realign horizontally */
- source->header.xmax = width - 1;
-
- cinfo->image_width = width;
- cinfo->image_height = height;
- cinfo->data_precision = 8; /* we can only handle 8 bit data */
-
- if (source->header.ncolors == 1 && source->header.ncmap == 0) {
- source->visual = GRAYSCALE;
- TRACEMS2(cinfo, 1, JTRC_RLE_GRAY, width, height);
- } else if (source->header.ncolors == 1 && source->header.ncmap == 1) {
- source->visual = MAPPEDGRAY;
- TRACEMS3(cinfo, 1, JTRC_RLE_MAPGRAY, width, height,
- 1 << source->header.cmaplen);
- } else if (source->header.ncolors == 1 && source->header.ncmap == 3) {
- source->visual = PSEUDOCOLOR;
- TRACEMS3(cinfo, 1, JTRC_RLE_MAPPED, width, height,
- 1 << source->header.cmaplen);
- } else if (source->header.ncolors == 3 && source->header.ncmap == 3) {
- source->visual = TRUECOLOR;
- TRACEMS3(cinfo, 1, JTRC_RLE_FULLMAP, width, height,
- 1 << source->header.cmaplen);
- } else if (source->header.ncolors == 3 && source->header.ncmap == 0) {
- source->visual = DIRECTCOLOR;
- TRACEMS2(cinfo, 1, JTRC_RLE, width, height);
- } else
- ERREXIT(cinfo, JERR_RLE_UNSUPPORTED);
-
- if (source->visual == GRAYSCALE || source->visual == MAPPEDGRAY) {
- cinfo->in_color_space = JCS_GRAYSCALE;
- cinfo->input_components = 1;
- } else {
- cinfo->in_color_space = JCS_RGB;
- cinfo->input_components = 3;
- }
-
- /*
- * A place to hold each scanline while it's converted.
- * (GRAYSCALE scanlines don't need converting)
- */
- if (source->visual != GRAYSCALE) {
- source->rle_row = (rle_pixel **)(*cinfo->mem->alloc_sarray)
- ((j_common_ptr)cinfo, JPOOL_IMAGE,
- (JDIMENSION)width, (JDIMENSION)cinfo->input_components);
- }
-
- /* request a virtual array to hold the image */
- source->image = (*cinfo->mem->request_virt_sarray)
- ((j_common_ptr)cinfo, JPOOL_IMAGE, FALSE,
- (JDIMENSION)(width * source->header.ncolors),
- (JDIMENSION)height, (JDIMENSION)1);
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- /* count file input as separate pass */
- progress->total_extra_passes++;
- }
-#endif
-
- source->pub.buffer_height = 1;
-}
-
-
-/*
- * Read one row of pixels.
- * Called only after load_image has read the image into the virtual array.
- * Used for GRAYSCALE, MAPPEDGRAY, TRUECOLOR, and DIRECTCOLOR images.
- */
-
-METHODDEF(JDIMENSION)
-get_rle_row(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
-{
- rle_source_ptr source = (rle_source_ptr)sinfo;
-
- source->row--;
- source->pub.buffer = (*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, source->image, source->row, (JDIMENSION)1, FALSE);
-
- return 1;
-}
-
-/*
- * Read one row of pixels.
- * Called only after load_image has read the image into the virtual array.
- * Used for PSEUDOCOLOR images.
- */
-
-METHODDEF(JDIMENSION)
-get_pseudocolor_row(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
-{
- rle_source_ptr source = (rle_source_ptr)sinfo;
- JSAMPROW src_row, dest_row;
- JDIMENSION col;
- rle_map *colormap;
- int val;
-
- colormap = source->header.cmap;
- dest_row = source->pub.buffer[0];
- source->row--;
- src_row = *(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, source->image, source->row, (JDIMENSION)1, FALSE);
-
- for (col = cinfo->image_width; col > 0; col--) {
- val = GETJSAMPLE(*src_row++);
- *dest_row++ = (JSAMPLE)(colormap[val ] >> 8);
- *dest_row++ = (JSAMPLE)(colormap[val + 256] >> 8);
- *dest_row++ = (JSAMPLE)(colormap[val + 512] >> 8);
- }
-
- return 1;
-}
-
-
-/*
- * Load the image into a virtual array. We have to do this because RLE
- * files start at the lower left while the JPEG standard has them starting
- * in the upper left. This is called the first time we want to get a row
- * of input. What we do is load the RLE data into the array and then call
- * the appropriate routine to read one row from the array. Before returning,
- * we set source->pub.get_pixel_rows so that subsequent calls go straight to
- * the appropriate row-reading routine.
- */
-
-METHODDEF(JDIMENSION)
-load_image(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
-{
- rle_source_ptr source = (rle_source_ptr)sinfo;
- JDIMENSION row, col;
- JSAMPROW scanline, red_ptr, green_ptr, blue_ptr;
- rle_pixel **rle_row;
- rle_map *colormap;
- char channel;
-#ifdef PROGRESS_REPORT
- cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
-#endif
-
- colormap = source->header.cmap;
- rle_row = source->rle_row;
-
- /* Read the RLE data into our virtual array.
- * We assume here that rle_pixel is represented the same as JSAMPLE.
- */
- RLE_CLR_BIT(source->header, RLE_ALPHA); /* don't read the alpha channel */
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_limit = cinfo->image_height;
- progress->pub.pass_counter = 0;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
-
- switch (source->visual) {
-
- case GRAYSCALE:
- case PSEUDOCOLOR:
- for (row = 0; row < cinfo->image_height; row++) {
- rle_row = (rle_pixel **)(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, source->image, row, (JDIMENSION)1, TRUE);
- rle_getrow(&source->header, rle_row);
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_counter++;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
- }
- break;
-
- case MAPPEDGRAY:
- case TRUECOLOR:
- for (row = 0; row < cinfo->image_height; row++) {
- scanline = *(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, source->image, row, (JDIMENSION)1, TRUE);
- rle_row = source->rle_row;
- rle_getrow(&source->header, rle_row);
-
- for (col = 0; col < cinfo->image_width; col++) {
- for (channel = 0; channel < source->header.ncolors; channel++) {
- *scanline++ = (JSAMPLE)
- (colormap[GETJSAMPLE(rle_row[channel][col]) + 256 * channel] >> 8);
- }
- }
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_counter++;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
- }
- break;
-
- case DIRECTCOLOR:
- for (row = 0; row < cinfo->image_height; row++) {
- scanline = *(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, source->image, row, (JDIMENSION)1, TRUE);
- rle_getrow(&source->header, rle_row);
-
- red_ptr = rle_row[0];
- green_ptr = rle_row[1];
- blue_ptr = rle_row[2];
-
- for (col = cinfo->image_width; col > 0; col--) {
- *scanline++ = *red_ptr++;
- *scanline++ = *green_ptr++;
- *scanline++ = *blue_ptr++;
- }
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_counter++;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
- }
- }
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL)
- progress->completed_extra_passes++;
-#endif
-
- /* Set up to call proper row-extraction routine in future */
- if (source->visual == PSEUDOCOLOR) {
- source->pub.buffer = source->rle_row;
- source->pub.get_pixel_rows = get_pseudocolor_row;
- } else {
- source->pub.get_pixel_rows = get_rle_row;
- }
- source->row = cinfo->image_height;
-
- /* And fetch the topmost (bottommost) row */
- return (*source->pub.get_pixel_rows) (cinfo, sinfo);
-}
-
-
-/*
- * Finish up at the end of the file.
- */
-
-METHODDEF(void)
-finish_input_rle(j_compress_ptr cinfo, cjpeg_source_ptr sinfo)
-{
- /* no work */
-}
-
-
-/*
- * The module selection routine for RLE format input.
- */
-
-GLOBAL(cjpeg_source_ptr)
-jinit_read_rle(j_compress_ptr cinfo)
-{
- rle_source_ptr source;
-
- /* Create module interface object */
- source = (rle_source_ptr)
- (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- sizeof(rle_source_struct));
- /* Fill in method ptrs */
- source->pub.start_input = start_input_rle;
- source->pub.finish_input = finish_input_rle;
- source->pub.get_pixel_rows = load_image;
-
- return (cjpeg_source_ptr)source;
-}
-
-#endif /* RLE_SUPPORTED */
diff --git a/rdtarga.c b/rdtarga.c
index 37bd286..c17073f 100644
--- a/rdtarga.c
+++ b/rdtarga.c
@@ -28,18 +28,8 @@
/* Macros to deal with unsigned chars as efficiently as compiler allows */
-#ifdef HAVE_UNSIGNED_CHAR
typedef unsigned char U_CHAR;
#define UCH(x) ((int)(x))
-#else /* !HAVE_UNSIGNED_CHAR */
-#ifdef __CHAR_UNSIGNED__
-typedef char U_CHAR;
-#define UCH(x) ((int)(x))
-#else
-typedef char U_CHAR;
-#define UCH(x) ((int)(x) & 0xFF)
-#endif
-#endif /* HAVE_UNSIGNED_CHAR */
#define ReadOK(file, buffer, len) \
@@ -344,8 +334,9 @@
unsigned int width, height, maplen;
boolean is_bottom_up;
-#define GET_2B(offset) ((unsigned int)UCH(targaheader[offset]) + \
- (((unsigned int)UCH(targaheader[offset + 1])) << 8))
+#define GET_2B(offset) \
+ ((unsigned int)UCH(targaheader[offset]) + \
+ (((unsigned int)UCH(targaheader[offset + 1])) << 8))
if (!ReadOK(source->pub.input_file, targaheader, 18))
ERREXIT(cinfo, JERR_INPUT_EOF);
diff --git a/simd/CMakeLists.txt b/simd/CMakeLists.txt
index 5c8009a..f3c24ef 100755
--- a/simd/CMakeLists.txt
+++ b/simd/CMakeLists.txt
@@ -30,6 +30,9 @@
if(CYGWIN)
set(CMAKE_ASM_NASM_OBJECT_FORMAT win64)
endif()
+ if(CMAKE_C_COMPILER_ABI MATCHES "ELF X32")
+ set(CMAKE_ASM_NASM_OBJECT_FORMAT elfx32)
+ endif()
elseif(CPU_TYPE STREQUAL "i386")
if(BORLAND)
set(CMAKE_ASM_NASM_OBJECT_FORMAT obj)
@@ -205,64 +208,76 @@
###############################################################################
-# ARM (GAS)
+# Arm (Intrinsics or GAS)
###############################################################################
elseif(CPU_TYPE STREQUAL "arm64" OR CPU_TYPE STREQUAL "arm")
-enable_language(ASM)
+include(CheckSymbolExists)
+if(BITS EQUAL 32)
+ set(CMAKE_REQUIRED_FLAGS -mfpu=neon)
+endif()
+check_symbol_exists(vld1_s16_x3 arm_neon.h HAVE_VLD1_S16_X3)
+check_symbol_exists(vld1_u16_x2 arm_neon.h HAVE_VLD1_U16_X2)
+check_symbol_exists(vld1q_u8_x4 arm_neon.h HAVE_VLD1Q_U8_X4)
+if(BITS EQUAL 32)
+ unset(CMAKE_REQUIRED_FLAGS)
+endif()
+configure_file(arm/neon-compat.h.in arm/neon-compat.h @ONLY)
+include_directories(${CMAKE_CURRENT_BINARY_DIR}/arm)
-set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} ${CMAKE_ASM_FLAGS}")
-
-string(TOUPPER ${CMAKE_BUILD_TYPE} CMAKE_BUILD_TYPE_UC)
-set(EFFECTIVE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CMAKE_ASM_FLAGS_${CMAKE_BUILD_TYPE_UC}}")
-message(STATUS "CMAKE_ASM_FLAGS = ${EFFECTIVE_ASM_FLAGS}")
-
-# Test whether we need gas-preprocessor.pl
-if(CPU_TYPE STREQUAL "arm")
- file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S "
- .text
- .fpu neon
- .arch armv7a
- .object_arch armv4
- .arm
- pld [r0]
- vmovn.u16 d0, q0")
+# GCC (as of this writing) and some older versions of Clang do not have a full
+# or optimal set of Neon intrinsics, so for performance reasons, when using
+# those compilers, we default to using the older GAS implementation of the Neon
+# SIMD extensions for certain algorithms. The presence or absence of the three
+# intrinsics we tested above is a reasonable proxy for this. We always default
+# to using the full Neon intrinsics implementation when building for macOS or
+# iOS, to avoid the need for gas-preprocessor.
+if((HAVE_VLD1_S16_X3 AND HAVE_VLD1_U16_X2 AND HAVE_VLD1Q_U8_X4) OR APPLE)
+ set(DEFAULT_NEON_INTRINSICS 1)
else()
- file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S "
- .text
- MYVAR .req x0
- movi v0.16b, #100
- mov MYVAR, #100
- .unreq MYVAR")
+ set(DEFAULT_NEON_INTRINSICS 0)
+endif()
+option(NEON_INTRINSICS
+ "Because GCC (as of this writing) and some older versions of Clang do not have a full or optimal set of Neon intrinsics, for performance reasons, the default when building libjpeg-turbo with those compilers is to continue using the older GAS implementation of the Neon SIMD extensions for certain algorithms. Setting this option forces the full Neon intrinsics implementation to be used with all compilers. Unsetting this option forces the hybrid GAS/intrinsics implementation to be used with all compilers."
+ ${DEFAULT_NEON_INTRINSICS})
+boolean_number(NEON_INTRINSICS PARENT_SCOPE)
+if(NEON_INTRINSICS)
+ add_definitions(-DNEON_INTRINSICS)
+ message(STATUS "Use full Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
+else()
+ message(STATUS "Use partial Neon SIMD intrinsics implementation (NEON_INTRINSICS = ${NEON_INTRINSICS})")
endif()
-separate_arguments(CMAKE_ASM_FLAGS_SEP UNIX_COMMAND "${CMAKE_ASM_FLAGS}")
+set(SIMD_SOURCES arm/jcgray-neon.c arm/jcphuff-neon.c arm/jcsample-neon.c
+ arm/jdmerge-neon.c arm/jdsample-neon.c arm/jfdctfst-neon.c
+ arm/jidctred-neon.c arm/jquanti-neon.c)
+if(NEON_INTRINSICS)
+ set(SIMD_SOURCES ${SIMD_SOURCES} arm/jccolor-neon.c arm/jidctint-neon.c)
+endif()
+if(NEON_INTRINSICS OR BITS EQUAL 64)
+ set(SIMD_SOURCES ${SIMD_SOURCES} arm/jidctfst-neon.c)
+endif()
+if(NEON_INTRINSICS OR BITS EQUAL 32)
+ set(SIMD_SOURCES ${SIMD_SOURCES} arm/aarch${BITS}/jchuff-neon.c
+ arm/jdcolor-neon.c arm/jfdctint-neon.c)
+endif()
+if(BITS EQUAL 32)
+ set_source_files_properties(${SIMD_SOURCES} COMPILE_FLAGS -mfpu=neon)
+endif()
+if(NOT NEON_INTRINSICS)
+ enable_language(ASM)
-execute_process(COMMAND ${CMAKE_ASM_COMPILER} ${CMAKE_ASM_FLAGS_SEP}
- -x assembler-with-cpp -c ${CMAKE_CURRENT_BINARY_DIR}/gastest.S
- RESULT_VARIABLE RESULT OUTPUT_VARIABLE OUTPUT ERROR_VARIABLE ERROR)
-if(NOT RESULT EQUAL 0)
- message(STATUS "GAS appears to be broken. Trying gas-preprocessor.pl ...")
- execute_process(COMMAND gas-preprocessor.pl ${CMAKE_ASM_COMPILER}
- ${CMAKE_ASM_FLAGS_SEP} -x assembler-with-cpp -c
- ${CMAKE_CURRENT_BINARY_DIR}/gastest.S
- RESULT_VARIABLE RESULT OUTPUT_VARIABLE OUTPUT ERROR_VARIABLE ERROR)
- if(NOT RESULT EQUAL 0)
- simd_fail("SIMD extensions disabled: GAS is not working properly")
- return()
- else()
- message(STATUS "Using gas-preprocessor.pl")
- configure_file(gas-preprocessor.in gas-preprocessor @ONLY)
- set(CMAKE_ASM_COMPILER ${CMAKE_CURRENT_BINARY_DIR}/gas-preprocessor)
- endif()
-else()
- message(STATUS "GAS is working properly")
+ set(CMAKE_ASM_FLAGS "${CMAKE_C_FLAGS} ${CMAKE_ASM_FLAGS}")
+
+ string(TOUPPER ${CMAKE_BUILD_TYPE} CMAKE_BUILD_TYPE_UC)
+ set(EFFECTIVE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CMAKE_ASM_FLAGS_${CMAKE_BUILD_TYPE_UC}}")
+ message(STATUS "CMAKE_ASM_FLAGS = ${EFFECTIVE_ASM_FLAGS}")
+
+ set(SIMD_SOURCES ${SIMD_SOURCES} arm/aarch${BITS}/jsimd_neon.S)
endif()
-file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/gastest.S)
-
-add_library(simd OBJECT ${CPU_TYPE}/jsimd_neon.S ${CPU_TYPE}/jsimd.c)
+add_library(simd OBJECT ${SIMD_SOURCES} arm/aarch${BITS}/jsimd.c)
if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
set_target_properties(simd PROPERTIES POSITION_INDEPENDENT_CODE 1)
@@ -311,14 +326,35 @@
endif()
###############################################################################
-# Loongson (Intrinsics)
+# MIPS64 (Intrinsics)
###############################################################################
-elseif(CPU_TYPE STREQUAL "loongson")
+elseif(CPU_TYPE STREQUAL "loongson" OR CPU_TYPE MATCHES "mips64*")
-set(SIMD_SOURCES loongson/jccolor-mmi.c loongson/jcsample-mmi.c
- loongson/jdcolor-mmi.c loongson/jdsample-mmi.c loongson/jfdctint-mmi.c
- loongson/jidctint-mmi.c loongson/jquanti-mmi.c)
+set(CMAKE_REQUIRED_FLAGS -Wa,-mloongson-mmi,-mloongson-ext)
+
+check_c_source_compiles("
+ int main(void) {
+ int c = 0, a = 0, b = 0;
+ asm (
+ \"paddb %0, %1, %2\"
+ : \"=f\" (c)
+ : \"f\" (a), \"f\" (b)
+ );
+ return c;
+ }" HAVE_MMI)
+
+unset(CMAKE_REQUIRED_FLAGS)
+
+if(NOT HAVE_MMI)
+ simd_fail("SIMD extensions not available for this CPU")
+ return()
+endif()
+
+set(SIMD_SOURCES mips64/jccolor-mmi.c mips64/jcgray-mmi.c mips64/jcsample-mmi.c
+ mips64/jdcolor-mmi.c mips64/jdmerge-mmi.c mips64/jdsample-mmi.c
+ mips64/jfdctfst-mmi.c mips64/jfdctint-mmi.c mips64/jidctfst-mmi.c
+ mips64/jidctint-mmi.c mips64/jquanti-mmi.c)
if(CMAKE_COMPILER_IS_GNUCC)
foreach(file ${SIMD_SOURCES})
@@ -326,8 +362,12 @@
" -fno-strict-aliasing")
endforeach()
endif()
+foreach(file ${SIMD_SOURCES})
+ set_property(SOURCE ${file} APPEND_STRING PROPERTY COMPILE_FLAGS
+ " -Wa,-mloongson-mmi,-mloongson-ext")
+endforeach()
-add_library(simd OBJECT ${SIMD_SOURCES} loongson/jsimd.c)
+add_library(simd OBJECT ${SIMD_SOURCES} mips64/jsimd.c)
if(CMAKE_POSITION_INDEPENDENT_CODE OR ENABLE_SHARED)
set_target_properties(simd PROPERTIES POSITION_INDEPENDENT_CODE 1)
diff --git a/simd/arm/arm/jccolext-neon.c b/simd/arm/aarch32/jccolext-neon.c
similarity index 81%
rename from simd/arm/arm/jccolext-neon.c
rename to simd/arm/aarch32/jccolext-neon.c
index 4f22e1f..362102d 100644
--- a/simd/arm/arm/jccolext-neon.c
+++ b/simd/arm/aarch32/jccolext-neon.c
@@ -1,7 +1,8 @@
/*
- * jccolext-neon.c - colorspace conversion (Arm NEON)
+ * jccolext-neon.c - colorspace conversion (32-bit Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -22,8 +23,8 @@
/* This file is included by jccolor-neon.c */
-/*
- * RGB -> YCbCr conversion is defined by the following equations:
+
+/* RGB -> YCbCr conversion is defined by the following equations:
* Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
* Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + 128
* Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + 128
@@ -39,28 +40,29 @@
* 0.08131409 = 5329 * 2^-16
* These constants are defined in jccolor-neon.c
*
- * To ensure rounding gives correct values, we add 0.5 to Cb and Cr.
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
*/
-void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width,
- JSAMPARRAY input_buf,
- JSAMPIMAGE output_buf,
- JDIMENSION output_row,
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+ JSAMPIMAGE output_buf, JDIMENSION output_row,
int num_rows)
{
- /* Pointer to RGB(X/A) input data. */
+ /* Pointer to RGB(X/A) input data */
JSAMPROW inptr;
- /* Pointers to Y, Cb and Cr output data. */
+ /* Pointers to Y, Cb, and Cr output data */
JSAMPROW outptr0, outptr1, outptr2;
+ /* Allocate temporary buffer for final (image_width % 8) pixels in row. */
+ ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
- /* Setup conversion constants. */
-#if defined(__clang__)
+ /* Set up conversion constants. */
+#ifdef HAVE_VLD1_U16_X2
const uint16x4x2_t consts = vld1_u16_x2(jsimd_rgb_ycc_neon_consts);
#else
/* GCC does not currently support the intrinsic vld1_<type>_x2(). */
const uint16x4_t consts1 = vld1_u16(jsimd_rgb_ycc_neon_consts);
const uint16x4_t consts2 = vld1_u16(jsimd_rgb_ycc_neon_consts + 4);
- const uint16x4x2_t consts = { consts1, consts2 };
+ const uint16x4x2_t consts = { { consts1, consts2 } };
#endif
const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
@@ -74,11 +76,11 @@
int cols_remaining = image_width;
for (; cols_remaining > 0; cols_remaining -= 8) {
- /* To prevent buffer overread by the vector load instructions, the */
- /* last (image_width % 8) columns of data are first memcopied to a */
- /* temporary buffer large enough to accommodate the vector load. */
+ /* To prevent buffer overread by the vector load instructions, the last
+ * (image_width % 8) columns of data are first memcopied to a temporary
+ * buffer large enough to accommodate the vector load.
+ */
if (cols_remaining < 8) {
- ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
inptr = tmp_buf;
}
@@ -129,8 +131,9 @@
/* Descale Cr values (right shift) and narrow to 16-bit. */
uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_low, 16),
vshrn_n_u32(cr_high, 16));
- /* Narrow Y, Cb and Cr values to 8-bit and store to memory. Buffer */
- /* overwrite is permitted up to the next multiple of ALIGN_SIZE bytes. */
+ /* Narrow Y, Cb, and Cr values to 8-bit and store to memory. Buffer
+ * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+ */
vst1_u8(outptr0, vmovn_u16(y_u16));
vst1_u8(outptr1, vmovn_u16(cb_u16));
vst1_u8(outptr2, vmovn_u16(cr_u16));
diff --git a/simd/arm/aarch32/jchuff-neon.c b/simd/arm/aarch32/jchuff-neon.c
new file mode 100644
index 0000000..19d94f7
--- /dev/null
+++ b/simd/arm/aarch32/jchuff-neon.c
@@ -0,0 +1,334 @@
+/*
+ * jchuff-neon.c - Huffman entropy encoding (32-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+#include "../jchuff.h"
+#include "neon-compat.h"
+
+#include <limits.h>
+
+#include <arm_neon.h>
+
+
+JOCTET *jsimd_huff_encode_one_block_neon(void *state, JOCTET *buffer,
+ JCOEFPTR block, int last_dc_val,
+ c_derived_tbl *dctbl,
+ c_derived_tbl *actbl)
+{
+ uint8_t block_nbits[DCTSIZE2];
+ uint16_t block_diff[DCTSIZE2];
+
+ /* Load rows of coefficients from DCT block in zig-zag order. */
+
+ /* Compute DC coefficient difference value. (F.1.1.5.1) */
+ int16x8_t row0 = vdupq_n_s16(block[0] - last_dc_val);
+ row0 = vld1q_lane_s16(block + 1, row0, 1);
+ row0 = vld1q_lane_s16(block + 8, row0, 2);
+ row0 = vld1q_lane_s16(block + 16, row0, 3);
+ row0 = vld1q_lane_s16(block + 9, row0, 4);
+ row0 = vld1q_lane_s16(block + 2, row0, 5);
+ row0 = vld1q_lane_s16(block + 3, row0, 6);
+ row0 = vld1q_lane_s16(block + 10, row0, 7);
+
+ int16x8_t row1 = vld1q_dup_s16(block + 17);
+ row1 = vld1q_lane_s16(block + 24, row1, 1);
+ row1 = vld1q_lane_s16(block + 32, row1, 2);
+ row1 = vld1q_lane_s16(block + 25, row1, 3);
+ row1 = vld1q_lane_s16(block + 18, row1, 4);
+ row1 = vld1q_lane_s16(block + 11, row1, 5);
+ row1 = vld1q_lane_s16(block + 4, row1, 6);
+ row1 = vld1q_lane_s16(block + 5, row1, 7);
+
+ int16x8_t row2 = vld1q_dup_s16(block + 12);
+ row2 = vld1q_lane_s16(block + 19, row2, 1);
+ row2 = vld1q_lane_s16(block + 26, row2, 2);
+ row2 = vld1q_lane_s16(block + 33, row2, 3);
+ row2 = vld1q_lane_s16(block + 40, row2, 4);
+ row2 = vld1q_lane_s16(block + 48, row2, 5);
+ row2 = vld1q_lane_s16(block + 41, row2, 6);
+ row2 = vld1q_lane_s16(block + 34, row2, 7);
+
+ int16x8_t row3 = vld1q_dup_s16(block + 27);
+ row3 = vld1q_lane_s16(block + 20, row3, 1);
+ row3 = vld1q_lane_s16(block + 13, row3, 2);
+ row3 = vld1q_lane_s16(block + 6, row3, 3);
+ row3 = vld1q_lane_s16(block + 7, row3, 4);
+ row3 = vld1q_lane_s16(block + 14, row3, 5);
+ row3 = vld1q_lane_s16(block + 21, row3, 6);
+ row3 = vld1q_lane_s16(block + 28, row3, 7);
+
+ int16x8_t abs_row0 = vabsq_s16(row0);
+ int16x8_t abs_row1 = vabsq_s16(row1);
+ int16x8_t abs_row2 = vabsq_s16(row2);
+ int16x8_t abs_row3 = vabsq_s16(row3);
+
+ int16x8_t row0_lz = vclzq_s16(abs_row0);
+ int16x8_t row1_lz = vclzq_s16(abs_row1);
+ int16x8_t row2_lz = vclzq_s16(abs_row2);
+ int16x8_t row3_lz = vclzq_s16(abs_row3);
+
+ /* Compute number of bits required to represent each coefficient. */
+ uint8x8_t row0_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row0_lz)));
+ uint8x8_t row1_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row1_lz)));
+ uint8x8_t row2_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row2_lz)));
+ uint8x8_t row3_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row3_lz)));
+
+ vst1_u8(block_nbits + 0 * DCTSIZE, row0_nbits);
+ vst1_u8(block_nbits + 1 * DCTSIZE, row1_nbits);
+ vst1_u8(block_nbits + 2 * DCTSIZE, row2_nbits);
+ vst1_u8(block_nbits + 3 * DCTSIZE, row3_nbits);
+
+ uint16x8_t row0_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row0, 15)),
+ vnegq_s16(row0_lz));
+ uint16x8_t row1_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row1, 15)),
+ vnegq_s16(row1_lz));
+ uint16x8_t row2_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row2, 15)),
+ vnegq_s16(row2_lz));
+ uint16x8_t row3_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row3, 15)),
+ vnegq_s16(row3_lz));
+
+ uint16x8_t row0_diff = veorq_u16(vreinterpretq_u16_s16(abs_row0), row0_mask);
+ uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1), row1_mask);
+ uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2), row2_mask);
+ uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3), row3_mask);
+
+ /* Store diff values for rows 0, 1, 2, and 3. */
+ vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+ vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+ vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+ vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+
+ /* Load last four rows of coefficients from DCT block in zig-zag order. */
+ int16x8_t row4 = vld1q_dup_s16(block + 35);
+ row4 = vld1q_lane_s16(block + 42, row4, 1);
+ row4 = vld1q_lane_s16(block + 49, row4, 2);
+ row4 = vld1q_lane_s16(block + 56, row4, 3);
+ row4 = vld1q_lane_s16(block + 57, row4, 4);
+ row4 = vld1q_lane_s16(block + 50, row4, 5);
+ row4 = vld1q_lane_s16(block + 43, row4, 6);
+ row4 = vld1q_lane_s16(block + 36, row4, 7);
+
+ int16x8_t row5 = vld1q_dup_s16(block + 29);
+ row5 = vld1q_lane_s16(block + 22, row5, 1);
+ row5 = vld1q_lane_s16(block + 15, row5, 2);
+ row5 = vld1q_lane_s16(block + 23, row5, 3);
+ row5 = vld1q_lane_s16(block + 30, row5, 4);
+ row5 = vld1q_lane_s16(block + 37, row5, 5);
+ row5 = vld1q_lane_s16(block + 44, row5, 6);
+ row5 = vld1q_lane_s16(block + 51, row5, 7);
+
+ int16x8_t row6 = vld1q_dup_s16(block + 58);
+ row6 = vld1q_lane_s16(block + 59, row6, 1);
+ row6 = vld1q_lane_s16(block + 52, row6, 2);
+ row6 = vld1q_lane_s16(block + 45, row6, 3);
+ row6 = vld1q_lane_s16(block + 38, row6, 4);
+ row6 = vld1q_lane_s16(block + 31, row6, 5);
+ row6 = vld1q_lane_s16(block + 39, row6, 6);
+ row6 = vld1q_lane_s16(block + 46, row6, 7);
+
+ int16x8_t row7 = vld1q_dup_s16(block + 53);
+ row7 = vld1q_lane_s16(block + 60, row7, 1);
+ row7 = vld1q_lane_s16(block + 61, row7, 2);
+ row7 = vld1q_lane_s16(block + 54, row7, 3);
+ row7 = vld1q_lane_s16(block + 47, row7, 4);
+ row7 = vld1q_lane_s16(block + 55, row7, 5);
+ row7 = vld1q_lane_s16(block + 62, row7, 6);
+ row7 = vld1q_lane_s16(block + 63, row7, 7);
+
+ int16x8_t abs_row4 = vabsq_s16(row4);
+ int16x8_t abs_row5 = vabsq_s16(row5);
+ int16x8_t abs_row6 = vabsq_s16(row6);
+ int16x8_t abs_row7 = vabsq_s16(row7);
+
+ int16x8_t row4_lz = vclzq_s16(abs_row4);
+ int16x8_t row5_lz = vclzq_s16(abs_row5);
+ int16x8_t row6_lz = vclzq_s16(abs_row6);
+ int16x8_t row7_lz = vclzq_s16(abs_row7);
+
+ /* Compute number of bits required to represent each coefficient. */
+ uint8x8_t row4_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row4_lz)));
+ uint8x8_t row5_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row5_lz)));
+ uint8x8_t row6_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row6_lz)));
+ uint8x8_t row7_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row7_lz)));
+
+ vst1_u8(block_nbits + 4 * DCTSIZE, row4_nbits);
+ vst1_u8(block_nbits + 5 * DCTSIZE, row5_nbits);
+ vst1_u8(block_nbits + 6 * DCTSIZE, row6_nbits);
+ vst1_u8(block_nbits + 7 * DCTSIZE, row7_nbits);
+
+ uint16x8_t row4_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row4, 15)),
+ vnegq_s16(row4_lz));
+ uint16x8_t row5_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row5, 15)),
+ vnegq_s16(row5_lz));
+ uint16x8_t row6_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row6, 15)),
+ vnegq_s16(row6_lz));
+ uint16x8_t row7_mask =
+ vshlq_u16(vreinterpretq_u16_s16(vshrq_n_s16(row7, 15)),
+ vnegq_s16(row7_lz));
+
+ uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4), row4_mask);
+ uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5), row5_mask);
+ uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6), row6_mask);
+ uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7), row7_mask);
+
+ /* Store diff values for rows 4, 5, 6, and 7. */
+ vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+ vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+ vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+ vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+ /* Construct bitmap to accelerate encoding of AC coefficients. A set bit
+ * means that the corresponding coefficient != 0.
+ */
+ uint8x8_t row0_nbits_gt0 = vcgt_u8(row0_nbits, vdup_n_u8(0));
+ uint8x8_t row1_nbits_gt0 = vcgt_u8(row1_nbits, vdup_n_u8(0));
+ uint8x8_t row2_nbits_gt0 = vcgt_u8(row2_nbits, vdup_n_u8(0));
+ uint8x8_t row3_nbits_gt0 = vcgt_u8(row3_nbits, vdup_n_u8(0));
+ uint8x8_t row4_nbits_gt0 = vcgt_u8(row4_nbits, vdup_n_u8(0));
+ uint8x8_t row5_nbits_gt0 = vcgt_u8(row5_nbits, vdup_n_u8(0));
+ uint8x8_t row6_nbits_gt0 = vcgt_u8(row6_nbits, vdup_n_u8(0));
+ uint8x8_t row7_nbits_gt0 = vcgt_u8(row7_nbits, vdup_n_u8(0));
+
+ /* { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 } */
+ const uint8x8_t bitmap_mask =
+ vreinterpret_u8_u64(vmov_n_u64(0x0102040810204080));
+
+ row0_nbits_gt0 = vand_u8(row0_nbits_gt0, bitmap_mask);
+ row1_nbits_gt0 = vand_u8(row1_nbits_gt0, bitmap_mask);
+ row2_nbits_gt0 = vand_u8(row2_nbits_gt0, bitmap_mask);
+ row3_nbits_gt0 = vand_u8(row3_nbits_gt0, bitmap_mask);
+ row4_nbits_gt0 = vand_u8(row4_nbits_gt0, bitmap_mask);
+ row5_nbits_gt0 = vand_u8(row5_nbits_gt0, bitmap_mask);
+ row6_nbits_gt0 = vand_u8(row6_nbits_gt0, bitmap_mask);
+ row7_nbits_gt0 = vand_u8(row7_nbits_gt0, bitmap_mask);
+
+ uint8x8_t bitmap_rows_10 = vpadd_u8(row1_nbits_gt0, row0_nbits_gt0);
+ uint8x8_t bitmap_rows_32 = vpadd_u8(row3_nbits_gt0, row2_nbits_gt0);
+ uint8x8_t bitmap_rows_54 = vpadd_u8(row5_nbits_gt0, row4_nbits_gt0);
+ uint8x8_t bitmap_rows_76 = vpadd_u8(row7_nbits_gt0, row6_nbits_gt0);
+ uint8x8_t bitmap_rows_3210 = vpadd_u8(bitmap_rows_32, bitmap_rows_10);
+ uint8x8_t bitmap_rows_7654 = vpadd_u8(bitmap_rows_76, bitmap_rows_54);
+ uint8x8_t bitmap = vpadd_u8(bitmap_rows_7654, bitmap_rows_3210);
+
+ /* Shift left to remove DC bit. */
+ bitmap = vreinterpret_u8_u64(vshl_n_u64(vreinterpret_u64_u8(bitmap), 1));
+ /* Move bitmap to 32-bit scalar registers. */
+ uint32_t bitmap_1_32 = vget_lane_u32(vreinterpret_u32_u8(bitmap), 1);
+ uint32_t bitmap_33_63 = vget_lane_u32(vreinterpret_u32_u8(bitmap), 0);
+
+ /* Set up state and bit buffer for output bitstream. */
+ working_state *state_ptr = (working_state *)state;
+ int free_bits = state_ptr->cur.free_bits;
+ size_t put_buffer = state_ptr->cur.put_buffer;
+
+ /* Encode DC coefficient. */
+
+ unsigned int nbits = block_nbits[0];
+ /* Emit Huffman-coded symbol and additional diff bits. */
+ unsigned int diff = block_diff[0];
+ PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits], diff)
+
+ /* Encode AC coefficients. */
+
+ unsigned int r = 0; /* r = run length of zeros */
+ unsigned int i = 1; /* i = number of coefficients encoded */
+ /* Code and size information for a run length of 16 zero coefficients */
+ const unsigned int code_0xf0 = actbl->ehufco[0xf0];
+ const unsigned int size_0xf0 = actbl->ehufsi[0xf0];
+
+ while (bitmap_1_32 != 0) {
+ r = BUILTIN_CLZ(bitmap_1_32);
+ i += r;
+ bitmap_1_32 <<= r;
+ nbits = block_nbits[i];
+ diff = block_diff[i];
+ while (r > 15) {
+ /* If run length > 15, emit special run-length-16 codes. */
+ PUT_BITS(code_0xf0, size_0xf0)
+ r -= 16;
+ }
+ /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+ unsigned int rs = (r << 4) + nbits;
+ PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+ i++;
+ bitmap_1_32 <<= 1;
+ }
+
+ r = 33 - i;
+ i = 33;
+
+ while (bitmap_33_63 != 0) {
+ unsigned int leading_zeros = BUILTIN_CLZ(bitmap_33_63);
+ r += leading_zeros;
+ i += leading_zeros;
+ bitmap_33_63 <<= leading_zeros;
+ nbits = block_nbits[i];
+ diff = block_diff[i];
+ while (r > 15) {
+ /* If run length > 15, emit special run-length-16 codes. */
+ PUT_BITS(code_0xf0, size_0xf0)
+ r -= 16;
+ }
+ /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+ unsigned int rs = (r << 4) + nbits;
+ PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+ r = 0;
+ i++;
+ bitmap_33_63 <<= 1;
+ }
+
+ /* If the last coefficient(s) were zero, emit an end-of-block (EOB) code.
+ * The value of RS for the EOB code is 0.
+ */
+ if (i != 64) {
+ PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+ }
+
+ state_ptr->cur.put_buffer = put_buffer;
+ state_ptr->cur.free_bits = free_bits;
+
+ return buffer;
+}
diff --git a/simd/arm/arm/jsimd.c b/simd/arm/aarch32/jsimd.c
similarity index 96%
rename from simd/arm/arm/jsimd.c
rename to simd/arm/aarch32/jsimd.c
index c0d5d90..fac55df 100644
--- a/simd/arm/arm/jsimd.c
+++ b/simd/arm/aarch32/jsimd.c
@@ -6,6 +6,7 @@
* Copyright (C) 2009-2011, 2013-2014, 2016, 2018, D. R. Commander.
* Copyright (C) 2015-2016, 2018, Matthieu Darbois.
* Copyright (C) 2019, Google LLC.
+ * Copyright (C) 2020, Arm Limited.
*
* Based on the x86 SIMD extension for IJG JPEG library,
* Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -13,7 +14,7 @@
*
* This file contains the interface between the "normal" portions
* of the library and the SIMD implementations when running on a
- * 32-bit ARM architecture.
+ * 32-bit Arm architecture.
*/
#define JPEG_INTERNALS
@@ -118,7 +119,7 @@
#if defined(__ARM_NEON__)
simd_support |= JSIMD_NEON;
#elif defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
- /* We still have a chance to use NEON regardless of globally used
+ /* We still have a chance to use Neon regardless of globally used
* -mcpu/-mfpu options passed to gcc by performing runtime detection via
* /proc/cpuinfo parsing on linux/android */
while (!parse_proc_cpuinfo(bufsize)) {
@@ -422,7 +423,6 @@
return 0;
if (sizeof(JDIMENSION) != 4)
return 0;
-
if (simd_support & JSIMD_NEON)
return 1;
@@ -934,6 +934,16 @@
GLOBAL(int)
jsimd_can_encode_mcu_AC_first_prepare(void)
{
+ init_simd();
+
+ if (DCTSIZE != 8)
+ return 0;
+ if (sizeof(JCOEF) != 2)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -942,11 +952,23 @@
const int *jpeg_natural_order_start, int Sl,
int Al, JCOEF *values, size_t *zerobits)
{
+ jsimd_encode_mcu_AC_first_prepare_neon(block, jpeg_natural_order_start,
+ Sl, Al, values, zerobits);
}
GLOBAL(int)
jsimd_can_encode_mcu_AC_refine_prepare(void)
{
+ init_simd();
+
+ if (DCTSIZE != 8)
+ return 0;
+ if (sizeof(JCOEF) != 2)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -955,5 +977,7 @@
const int *jpeg_natural_order_start, int Sl,
int Al, JCOEF *absvalues, size_t *bits)
{
- return 0;
+ return jsimd_encode_mcu_AC_refine_prepare_neon(block,
+ jpeg_natural_order_start, Sl,
+ Al, absvalues, bits);
}
diff --git a/simd/arm/arm64/jccolext-neon.c b/simd/arm/aarch64/jccolext-neon.c
similarity index 71%
rename from simd/arm/arm64/jccolext-neon.c
rename to simd/arm/aarch64/jccolext-neon.c
index 89f520a..37130c2 100644
--- a/simd/arm/arm64/jccolext-neon.c
+++ b/simd/arm/aarch64/jccolext-neon.c
@@ -1,7 +1,7 @@
/*
- * jccolext-neon.c - colorspace conversion (Arm NEON)
+ * jccolext-neon.c - colorspace conversion (64-bit Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -22,8 +22,8 @@
/* This file is included by jccolor-neon.c */
-/*
- * RGB -> YCbCr conversion is defined by the following equations:
+
+/* RGB -> YCbCr conversion is defined by the following equations:
* Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
* Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + 128
* Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + 128
@@ -39,21 +39,22 @@
* 0.08131409 = 5329 * 2^-16
* These constants are defined in jccolor-neon.c
*
- * To ensure rounding gives correct values, we add 0.5 to Cb and Cr.
+ * We add the fixed-point equivalent of 0.5 to Cb and Cr, which effectively
+ * rounds up or down the result via integer truncation.
*/
-void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width,
- JSAMPARRAY input_buf,
- JSAMPIMAGE output_buf,
- JDIMENSION output_row,
+void jsimd_rgb_ycc_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+ JSAMPIMAGE output_buf, JDIMENSION output_row,
int num_rows)
{
- /* Pointer to RGB(X/A) input data. */
+ /* Pointer to RGB(X/A) input data */
JSAMPROW inptr;
- /* Pointers to Y, Cb and Cr output data. */
+ /* Pointers to Y, Cb, and Cr output data */
JSAMPROW outptr0, outptr1, outptr2;
+ /* Allocate temporary buffer for final (image_width % 16) pixels in row. */
+ ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
- /* Setup conversion constants. */
+ /* Set up conversion constants. */
const uint16x8_t consts = vld1q_u16(jsimd_rgb_ycc_neon_consts);
const uint32x4_t scaled_128_5 = vdupq_n_u32((128 << 16) + 32767);
@@ -83,15 +84,15 @@
uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
- uint32x4_t y_lh = vmull_high_laneq_u16(r_l, consts, 0);
- y_lh = vmlal_high_laneq_u16(y_lh, g_l, consts, 1);
- y_lh = vmlal_high_laneq_u16(y_lh, b_l, consts, 2);
+ uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+ y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+ y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
- uint32x4_t y_hh = vmull_high_laneq_u16(r_h, consts, 0);
- y_hh = vmlal_high_laneq_u16(y_hh, g_h, consts, 1);
- y_hh = vmlal_high_laneq_u16(y_hh, b_h, consts, 2);
+ uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+ y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+ y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
/* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + 128 */
uint32x4_t cb_ll = scaled_128_5;
@@ -99,17 +100,17 @@
cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
uint32x4_t cb_lh = scaled_128_5;
- cb_lh = vmlsl_high_laneq_u16(cb_lh, r_l, consts, 3);
- cb_lh = vmlsl_high_laneq_u16(cb_lh, g_l, consts, 4);
- cb_lh = vmlal_high_laneq_u16(cb_lh, b_l, consts, 5);
+ cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+ cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+ cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
uint32x4_t cb_hl = scaled_128_5;
cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
uint32x4_t cb_hh = scaled_128_5;
- cb_hh = vmlsl_high_laneq_u16(cb_hh, r_h, consts, 3);
- cb_hh = vmlsl_high_laneq_u16(cb_hh, g_h, consts, 4);
- cb_hh = vmlal_high_laneq_u16(cb_hh, b_h, consts, 5);
+ cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+ cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+ cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
/* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + 128 */
uint32x4_t cr_ll = scaled_128_5;
@@ -117,17 +118,17 @@
cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
uint32x4_t cr_lh = scaled_128_5;
- cr_lh = vmlal_high_laneq_u16(cr_lh, r_l, consts, 5);
- cr_lh = vmlsl_high_laneq_u16(cr_lh, g_l, consts, 6);
- cr_lh = vmlsl_high_laneq_u16(cr_lh, b_l, consts, 7);
+ cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+ cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+ cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
uint32x4_t cr_hl = scaled_128_5;
cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
uint32x4_t cr_hh = scaled_128_5;
- cr_hh = vmlal_high_laneq_u16(cr_hh, r_h, consts, 5);
- cr_hh = vmlsl_high_laneq_u16(cr_hh, g_h, consts, 6);
- cr_hh = vmlsl_high_laneq_u16(cr_hh, b_h, consts, 7);
+ cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+ cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+ cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
/* Descale Y values (rounding right shift) and narrow to 16-bit. */
uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
@@ -144,8 +145,9 @@
vshrn_n_u32(cr_lh, 16));
uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
vshrn_n_u32(cr_hh, 16));
- /* Narrow Y, Cb and Cr values to 8-bit and store to memory. Buffer */
- /* overwrite is permitted up to the next multiple of ALIGN_SIZE bytes. */
+ /* Narrow Y, Cb, and Cr values to 8-bit and store to memory. Buffer
+ * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+ */
vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
@@ -158,10 +160,10 @@
}
if (cols_remaining > 8) {
- /* To prevent buffer overread by the vector load instructions, the */
- /* last (image_width % 16) columns of data are first memcopied to a */
- /* temporary buffer large enough to accommodate the vector load. */
- ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
+ /* To prevent buffer overread by the vector load instructions, the last
+ * (image_width % 16) columns of data are first memcopied to a temporary
+ * buffer large enough to accommodate the vector load.
+ */
memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
inptr = tmp_buf;
@@ -181,15 +183,15 @@
uint32x4_t y_ll = vmull_laneq_u16(vget_low_u16(r_l), consts, 0);
y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(g_l), consts, 1);
y_ll = vmlal_laneq_u16(y_ll, vget_low_u16(b_l), consts, 2);
- uint32x4_t y_lh = vmull_high_laneq_u16(r_l, consts, 0);
- y_lh = vmlal_high_laneq_u16(y_lh, g_l, consts, 1);
- y_lh = vmlal_high_laneq_u16(y_lh, b_l, consts, 2);
+ uint32x4_t y_lh = vmull_laneq_u16(vget_high_u16(r_l), consts, 0);
+ y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(g_l), consts, 1);
+ y_lh = vmlal_laneq_u16(y_lh, vget_high_u16(b_l), consts, 2);
uint32x4_t y_hl = vmull_laneq_u16(vget_low_u16(r_h), consts, 0);
y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(g_h), consts, 1);
y_hl = vmlal_laneq_u16(y_hl, vget_low_u16(b_h), consts, 2);
- uint32x4_t y_hh = vmull_high_laneq_u16(r_h, consts, 0);
- y_hh = vmlal_high_laneq_u16(y_hh, g_h, consts, 1);
- y_hh = vmlal_high_laneq_u16(y_hh, b_h, consts, 2);
+ uint32x4_t y_hh = vmull_laneq_u16(vget_high_u16(r_h), consts, 0);
+ y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(g_h), consts, 1);
+ y_hh = vmlal_laneq_u16(y_hh, vget_high_u16(b_h), consts, 2);
/* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + 128 */
uint32x4_t cb_ll = scaled_128_5;
@@ -197,17 +199,17 @@
cb_ll = vmlsl_laneq_u16(cb_ll, vget_low_u16(g_l), consts, 4);
cb_ll = vmlal_laneq_u16(cb_ll, vget_low_u16(b_l), consts, 5);
uint32x4_t cb_lh = scaled_128_5;
- cb_lh = vmlsl_high_laneq_u16(cb_lh, r_l, consts, 3);
- cb_lh = vmlsl_high_laneq_u16(cb_lh, g_l, consts, 4);
- cb_lh = vmlal_high_laneq_u16(cb_lh, b_l, consts, 5);
+ cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(r_l), consts, 3);
+ cb_lh = vmlsl_laneq_u16(cb_lh, vget_high_u16(g_l), consts, 4);
+ cb_lh = vmlal_laneq_u16(cb_lh, vget_high_u16(b_l), consts, 5);
uint32x4_t cb_hl = scaled_128_5;
cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(r_h), consts, 3);
cb_hl = vmlsl_laneq_u16(cb_hl, vget_low_u16(g_h), consts, 4);
cb_hl = vmlal_laneq_u16(cb_hl, vget_low_u16(b_h), consts, 5);
uint32x4_t cb_hh = scaled_128_5;
- cb_hh = vmlsl_high_laneq_u16(cb_hh, r_h, consts, 3);
- cb_hh = vmlsl_high_laneq_u16(cb_hh, g_h, consts, 4);
- cb_hh = vmlal_high_laneq_u16(cb_hh, b_h, consts, 5);
+ cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(r_h), consts, 3);
+ cb_hh = vmlsl_laneq_u16(cb_hh, vget_high_u16(g_h), consts, 4);
+ cb_hh = vmlal_laneq_u16(cb_hh, vget_high_u16(b_h), consts, 5);
/* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + 128 */
uint32x4_t cr_ll = scaled_128_5;
@@ -215,17 +217,17 @@
cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(g_l), consts, 6);
cr_ll = vmlsl_laneq_u16(cr_ll, vget_low_u16(b_l), consts, 7);
uint32x4_t cr_lh = scaled_128_5;
- cr_lh = vmlal_high_laneq_u16(cr_lh, r_l, consts, 5);
- cr_lh = vmlsl_high_laneq_u16(cr_lh, g_l, consts, 6);
- cr_lh = vmlsl_high_laneq_u16(cr_lh, b_l, consts, 7);
+ cr_lh = vmlal_laneq_u16(cr_lh, vget_high_u16(r_l), consts, 5);
+ cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(g_l), consts, 6);
+ cr_lh = vmlsl_laneq_u16(cr_lh, vget_high_u16(b_l), consts, 7);
uint32x4_t cr_hl = scaled_128_5;
cr_hl = vmlal_laneq_u16(cr_hl, vget_low_u16(r_h), consts, 5);
cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(g_h), consts, 6);
cr_hl = vmlsl_laneq_u16(cr_hl, vget_low_u16(b_h), consts, 7);
uint32x4_t cr_hh = scaled_128_5;
- cr_hh = vmlal_high_laneq_u16(cr_hh, r_h, consts, 5);
- cr_hh = vmlsl_high_laneq_u16(cr_hh, g_h, consts, 6);
- cr_hh = vmlsl_high_laneq_u16(cr_hh, b_h, consts, 7);
+ cr_hh = vmlal_laneq_u16(cr_hh, vget_high_u16(r_h), consts, 5);
+ cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(g_h), consts, 6);
+ cr_hh = vmlsl_laneq_u16(cr_hh, vget_high_u16(b_h), consts, 7);
/* Descale Y values (rounding right shift) and narrow to 16-bit. */
uint16x8_t y_l = vcombine_u16(vrshrn_n_u32(y_ll, 16),
@@ -242,17 +244,18 @@
vshrn_n_u32(cr_lh, 16));
uint16x8_t cr_h = vcombine_u16(vshrn_n_u32(cr_hl, 16),
vshrn_n_u32(cr_hh, 16));
- /* Narrow Y, Cb and Cr values to 8-bit and store to memory. Buffer */
- /* overwrite is permitted up to the next multiple of ALIGN_SIZE bytes. */
+ /* Narrow Y, Cb, and Cr values to 8-bit and store to memory. Buffer
+ * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+ */
vst1q_u8(outptr0, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
vst1q_u8(outptr1, vcombine_u8(vmovn_u16(cb_l), vmovn_u16(cb_h)));
vst1q_u8(outptr2, vcombine_u8(vmovn_u16(cr_l), vmovn_u16(cr_h)));
} else if (cols_remaining > 0) {
- /* To prevent buffer overread by the vector load instructions, the */
- /* last (image_width % 8) columns of data are first memcopied to a */
- /* temporary buffer large enough to accommodate the vector load. */
- ALIGN(16) uint8_t tmp_buf[8 * RGB_PIXELSIZE];
+ /* To prevent buffer overread by the vector load instructions, the last
+ * (image_width % 8) columns of data are first memcopied to a temporary
+ * buffer large enough to accommodate the vector load.
+ */
memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
inptr = tmp_buf;
@@ -269,9 +272,9 @@
uint32x4_t y_l = vmull_laneq_u16(vget_low_u16(r), consts, 0);
y_l = vmlal_laneq_u16(y_l, vget_low_u16(g), consts, 1);
y_l = vmlal_laneq_u16(y_l, vget_low_u16(b), consts, 2);
- uint32x4_t y_h = vmull_high_laneq_u16(r, consts, 0);
- y_h = vmlal_high_laneq_u16(y_h, g, consts, 1);
- y_h = vmlal_high_laneq_u16(y_h, b, consts, 2);
+ uint32x4_t y_h = vmull_laneq_u16(vget_high_u16(r), consts, 0);
+ y_h = vmlal_laneq_u16(y_h, vget_high_u16(g), consts, 1);
+ y_h = vmlal_laneq_u16(y_h, vget_high_u16(b), consts, 2);
/* Compute Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + 128 */
uint32x4_t cb_l = scaled_128_5;
@@ -279,9 +282,9 @@
cb_l = vmlsl_laneq_u16(cb_l, vget_low_u16(g), consts, 4);
cb_l = vmlal_laneq_u16(cb_l, vget_low_u16(b), consts, 5);
uint32x4_t cb_h = scaled_128_5;
- cb_h = vmlsl_high_laneq_u16(cb_h, r, consts, 3);
- cb_h = vmlsl_high_laneq_u16(cb_h, g, consts, 4);
- cb_h = vmlal_high_laneq_u16(cb_h, b, consts, 5);
+ cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(r), consts, 3);
+ cb_h = vmlsl_laneq_u16(cb_h, vget_high_u16(g), consts, 4);
+ cb_h = vmlal_laneq_u16(cb_h, vget_high_u16(b), consts, 5);
/* Compute Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + 128 */
uint32x4_t cr_l = scaled_128_5;
@@ -289,9 +292,9 @@
cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(g), consts, 6);
cr_l = vmlsl_laneq_u16(cr_l, vget_low_u16(b), consts, 7);
uint32x4_t cr_h = scaled_128_5;
- cr_h = vmlal_high_laneq_u16(cr_h, r, consts, 5);
- cr_h = vmlsl_high_laneq_u16(cr_h, g, consts, 6);
- cr_h = vmlsl_high_laneq_u16(cr_h, b, consts, 7);
+ cr_h = vmlal_laneq_u16(cr_h, vget_high_u16(r), consts, 5);
+ cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(g), consts, 6);
+ cr_h = vmlsl_laneq_u16(cr_h, vget_high_u16(b), consts, 7);
/* Descale Y values (rounding right shift) and narrow to 16-bit. */
uint16x8_t y_u16 = vcombine_u16(vrshrn_n_u32(y_l, 16),
@@ -302,8 +305,9 @@
/* Descale Cr values (right shift) and narrow to 16-bit. */
uint16x8_t cr_u16 = vcombine_u16(vshrn_n_u32(cr_l, 16),
vshrn_n_u32(cr_h, 16));
- /* Narrow Y, Cb and Cr values to 8-bit and store to memory. Buffer */
- /* overwrite is permitted up to the next multiple of ALIGN_SIZE bytes. */
+ /* Narrow Y, Cb, and Cr values to 8-bit and store to memory. Buffer
+ * overwrite is permitted up to the next multiple of ALIGN_SIZE bytes.
+ */
vst1_u8(outptr0, vmovn_u16(y_u16));
vst1_u8(outptr1, vmovn_u16(cb_u16));
vst1_u8(outptr2, vmovn_u16(cr_u16));
diff --git a/simd/arm/aarch64/jchuff-neon.c b/simd/arm/aarch64/jchuff-neon.c
new file mode 100644
index 0000000..a0a57a6
--- /dev/null
+++ b/simd/arm/aarch64/jchuff-neon.c
@@ -0,0 +1,403 @@
+/*
+ * jchuff-neon.c - Huffman entropy encoding (64-bit Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ *
+ * NOTE: All referenced figures are from
+ * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
+ */
+
+#define JPEG_INTERNALS
+#include "../../../jinclude.h"
+#include "../../../jpeglib.h"
+#include "../../../jsimd.h"
+#include "../../../jdct.h"
+#include "../../../jsimddct.h"
+#include "../../jsimd.h"
+#include "../align.h"
+#include "../jchuff.h"
+#include "neon-compat.h"
+
+#include <limits.h>
+
+#include <arm_neon.h>
+
+
+ALIGN(16) static const uint8_t jsimd_huff_encode_one_block_consts[] = {
+ 0, 1, 2, 3, 16, 17, 32, 33,
+ 18, 19, 4, 5, 6, 7, 20, 21,
+ 34, 35, 48, 49, 255, 255, 50, 51,
+ 36, 37, 22, 23, 8, 9, 10, 11,
+ 255, 255, 6, 7, 20, 21, 34, 35,
+ 48, 49, 255, 255, 50, 51, 36, 37,
+ 54, 55, 40, 41, 26, 27, 12, 13,
+ 14, 15, 28, 29, 42, 43, 56, 57,
+ 6, 7, 20, 21, 34, 35, 48, 49,
+ 50, 51, 36, 37, 22, 23, 8, 9,
+ 26, 27, 12, 13, 255, 255, 14, 15,
+ 28, 29, 42, 43, 56, 57, 255, 255,
+ 52, 53, 54, 55, 40, 41, 26, 27,
+ 12, 13, 255, 255, 14, 15, 28, 29,
+ 26, 27, 40, 41, 42, 43, 28, 29,
+ 14, 15, 30, 31, 44, 45, 46, 47
+};
+
+JOCTET *jsimd_huff_encode_one_block_neon(void *state, JOCTET *buffer,
+ JCOEFPTR block, int last_dc_val,
+ c_derived_tbl *dctbl,
+ c_derived_tbl *actbl)
+{
+ uint16_t block_diff[DCTSIZE2];
+
+ /* Load lookup table indices for rows of zig-zag ordering. */
+#ifdef HAVE_VLD1Q_U8_X4
+ const uint8x16x4_t idx_rows_0123 =
+ vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE);
+ const uint8x16x4_t idx_rows_4567 =
+ vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE);
+#else
+ /* GCC does not currently support intrinsics vl1dq_<type>_x4(). */
+ const uint8x16x4_t idx_rows_0123 = { {
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 2 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 4 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 6 * DCTSIZE)
+ } };
+ const uint8x16x4_t idx_rows_4567 = { {
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 10 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 12 * DCTSIZE),
+ vld1q_u8(jsimd_huff_encode_one_block_consts + 14 * DCTSIZE)
+ } };
+#endif
+
+ /* Load 8x8 block of DCT coefficients. */
+#ifdef HAVE_VLD1Q_U8_X4
+ const int8x16x4_t tbl_rows_0123 =
+ vld1q_s8_x4((int8_t *)(block + 0 * DCTSIZE));
+ const int8x16x4_t tbl_rows_4567 =
+ vld1q_s8_x4((int8_t *)(block + 4 * DCTSIZE));
+#else
+ const int8x16x4_t tbl_rows_0123 = { {
+ vld1q_s8((int8_t *)(block + 0 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 1 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 2 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 3 * DCTSIZE))
+ } };
+ const int8x16x4_t tbl_rows_4567 = { {
+ vld1q_s8((int8_t *)(block + 4 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 5 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 6 * DCTSIZE)),
+ vld1q_s8((int8_t *)(block + 7 * DCTSIZE))
+ } };
+#endif
+
+ /* Initialise extra lookup tables. */
+ const int8x16x4_t tbl_rows_2345 = { {
+ tbl_rows_0123.val[2], tbl_rows_0123.val[3],
+ tbl_rows_4567.val[0], tbl_rows_4567.val[1]
+ } };
+ const int8x16x3_t tbl_rows_567 =
+ { { tbl_rows_4567.val[1], tbl_rows_4567.val[2], tbl_rows_4567.val[3] } };
+
+ /* Shuffle coefficients into zig-zag order. */
+ int16x8_t row0 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[0]));
+ int16x8_t row1 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[1]));
+ int16x8_t row2 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_0123.val[2]));
+ int16x8_t row3 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[3]));
+ int16x8_t row4 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[0]));
+ int16x8_t row5 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_4567.val[1]));
+ int16x8_t row6 =
+ vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[2]));
+ int16x8_t row7 =
+ vreinterpretq_s16_s8(vqtbl3q_s8(tbl_rows_567, idx_rows_4567.val[3]));
+
+ /* Compute DC coefficient difference value (F.1.1.5.1). */
+ row0 = vsetq_lane_s16(block[0] - last_dc_val, row0, 0);
+ /* Initialize AC coefficient lanes not reachable by lookup tables. */
+ row1 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[0]),
+ 0), row1, 2);
+ row2 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
+ 4), row2, 0);
+ row2 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
+ 0), row2, 5);
+ row5 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
+ 7), row5, 2);
+ row5 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
+ 3), row5, 7);
+ row6 =
+ vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[3]),
+ 7), row6, 5);
+
+ /* DCT block is now in zig-zag order; start Huffman encoding process. */
+ int16x8_t abs_row0 = vabsq_s16(row0);
+ int16x8_t abs_row1 = vabsq_s16(row1);
+ int16x8_t abs_row2 = vabsq_s16(row2);
+ int16x8_t abs_row3 = vabsq_s16(row3);
+ int16x8_t abs_row4 = vabsq_s16(row4);
+ int16x8_t abs_row5 = vabsq_s16(row5);
+ int16x8_t abs_row6 = vabsq_s16(row6);
+ int16x8_t abs_row7 = vabsq_s16(row7);
+
+ /* For negative coeffs: diff = abs(coeff) -1 = ~abs(coeff) */
+ uint16x8_t row0_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row0, vshrq_n_s16(row0, 15)));
+ uint16x8_t row1_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row1, vshrq_n_s16(row1, 15)));
+ uint16x8_t row2_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row2, vshrq_n_s16(row2, 15)));
+ uint16x8_t row3_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row3, vshrq_n_s16(row3, 15)));
+ uint16x8_t row4_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row4, vshrq_n_s16(row4, 15)));
+ uint16x8_t row5_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row5, vshrq_n_s16(row5, 15)));
+ uint16x8_t row6_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row6, vshrq_n_s16(row6, 15)));
+ uint16x8_t row7_diff =
+ vreinterpretq_u16_s16(veorq_s16(abs_row7, vshrq_n_s16(row7, 15)));
+
+ /* Construct bitmap to accelerate encoding of AC coefficients. A set bit
+ * means that the corresponding coefficient != 0.
+ */
+ uint8x8_t abs_row0_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row0),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row1_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row1),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row2_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row2),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row3_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row3),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row4_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row4),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row5_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row5),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row6_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row6),
+ vdupq_n_u16(0)));
+ uint8x8_t abs_row7_gt0 = vmovn_u16(vcgtq_u16(vreinterpretq_u16_s16(abs_row7),
+ vdupq_n_u16(0)));
+
+ /* { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 } */
+ const uint8x8_t bitmap_mask =
+ vreinterpret_u8_u64(vmov_n_u64(0x0102040810204080));
+
+ abs_row0_gt0 = vand_u8(abs_row0_gt0, bitmap_mask);
+ abs_row1_gt0 = vand_u8(abs_row1_gt0, bitmap_mask);
+ abs_row2_gt0 = vand_u8(abs_row2_gt0, bitmap_mask);
+ abs_row3_gt0 = vand_u8(abs_row3_gt0, bitmap_mask);
+ abs_row4_gt0 = vand_u8(abs_row4_gt0, bitmap_mask);
+ abs_row5_gt0 = vand_u8(abs_row5_gt0, bitmap_mask);
+ abs_row6_gt0 = vand_u8(abs_row6_gt0, bitmap_mask);
+ abs_row7_gt0 = vand_u8(abs_row7_gt0, bitmap_mask);
+
+ uint8x8_t bitmap_rows_10 = vpadd_u8(abs_row1_gt0, abs_row0_gt0);
+ uint8x8_t bitmap_rows_32 = vpadd_u8(abs_row3_gt0, abs_row2_gt0);
+ uint8x8_t bitmap_rows_54 = vpadd_u8(abs_row5_gt0, abs_row4_gt0);
+ uint8x8_t bitmap_rows_76 = vpadd_u8(abs_row7_gt0, abs_row6_gt0);
+ uint8x8_t bitmap_rows_3210 = vpadd_u8(bitmap_rows_32, bitmap_rows_10);
+ uint8x8_t bitmap_rows_7654 = vpadd_u8(bitmap_rows_76, bitmap_rows_54);
+ uint8x8_t bitmap_all = vpadd_u8(bitmap_rows_7654, bitmap_rows_3210);
+
+ /* Shift left to remove DC bit. */
+ bitmap_all =
+ vreinterpret_u8_u64(vshl_n_u64(vreinterpret_u64_u8(bitmap_all), 1));
+ /* Count bits set (number of non-zero coefficients) in bitmap. */
+ unsigned int non_zero_coefficients = vaddv_u8(vcnt_u8(bitmap_all));
+ /* Move bitmap to 64-bit scalar register. */
+ uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+
+ /* Set up state and bit buffer for output bitstream. */
+ working_state *state_ptr = (working_state *)state;
+ int free_bits = state_ptr->cur.free_bits;
+ size_t put_buffer = state_ptr->cur.put_buffer;
+
+ /* Encode DC coefficient. */
+
+ /* Find nbits required to specify sign and amplitude of coefficient. */
+#if defined(_MSC_VER) && !defined(__clang__)
+ unsigned int lz = BUILTIN_CLZ(vgetq_lane_s16(abs_row0, 0));
+#else
+ unsigned int lz;
+ __asm__("clz %w0, %w1" : "=r"(lz) : "r"(vgetq_lane_s16(abs_row0, 0)));
+#endif
+ unsigned int nbits = 32 - lz;
+ /* Emit Huffman-coded symbol and additional diff bits. */
+ unsigned int diff = (unsigned int)(vgetq_lane_u16(row0_diff, 0) << lz) >> lz;
+ PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits], diff)
+
+ /* Encode AC coefficients. */
+
+ unsigned int r = 0; /* r = run length of zeros */
+ unsigned int i = 1; /* i = number of coefficients encoded */
+ /* Code and size information for a run length of 16 zero coefficients */
+ const unsigned int code_0xf0 = actbl->ehufco[0xf0];
+ const unsigned int size_0xf0 = actbl->ehufsi[0xf0];
+
+ /* The most efficient method of computing nbits and diff depends on the
+ * number of non-zero coefficients. If the bitmap is not too sparse (> 8
+ * non-zero AC coefficients), it is beneficial to use Neon; else we compute
+ * nbits and diff on demand using scalar code.
+ */
+ if (non_zero_coefficients > 8) {
+ uint8_t block_nbits[DCTSIZE2];
+
+ int16x8_t row0_lz = vclzq_s16(abs_row0);
+ int16x8_t row1_lz = vclzq_s16(abs_row1);
+ int16x8_t row2_lz = vclzq_s16(abs_row2);
+ int16x8_t row3_lz = vclzq_s16(abs_row3);
+ int16x8_t row4_lz = vclzq_s16(abs_row4);
+ int16x8_t row5_lz = vclzq_s16(abs_row5);
+ int16x8_t row6_lz = vclzq_s16(abs_row6);
+ int16x8_t row7_lz = vclzq_s16(abs_row7);
+ /* Compute nbits needed to specify magnitude of each coefficient. */
+ uint8x8_t row0_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row0_lz)));
+ uint8x8_t row1_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row1_lz)));
+ uint8x8_t row2_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row2_lz)));
+ uint8x8_t row3_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row3_lz)));
+ uint8x8_t row4_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row4_lz)));
+ uint8x8_t row5_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row5_lz)));
+ uint8x8_t row6_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row6_lz)));
+ uint8x8_t row7_nbits = vsub_u8(vdup_n_u8(16),
+ vmovn_u16(vreinterpretq_u16_s16(row7_lz)));
+ /* Store nbits. */
+ vst1_u8(block_nbits + 0 * DCTSIZE, row0_nbits);
+ vst1_u8(block_nbits + 1 * DCTSIZE, row1_nbits);
+ vst1_u8(block_nbits + 2 * DCTSIZE, row2_nbits);
+ vst1_u8(block_nbits + 3 * DCTSIZE, row3_nbits);
+ vst1_u8(block_nbits + 4 * DCTSIZE, row4_nbits);
+ vst1_u8(block_nbits + 5 * DCTSIZE, row5_nbits);
+ vst1_u8(block_nbits + 6 * DCTSIZE, row6_nbits);
+ vst1_u8(block_nbits + 7 * DCTSIZE, row7_nbits);
+ /* Mask bits not required to specify sign and amplitude of diff. */
+ row0_diff = vshlq_u16(row0_diff, row0_lz);
+ row1_diff = vshlq_u16(row1_diff, row1_lz);
+ row2_diff = vshlq_u16(row2_diff, row2_lz);
+ row3_diff = vshlq_u16(row3_diff, row3_lz);
+ row4_diff = vshlq_u16(row4_diff, row4_lz);
+ row5_diff = vshlq_u16(row5_diff, row5_lz);
+ row6_diff = vshlq_u16(row6_diff, row6_lz);
+ row7_diff = vshlq_u16(row7_diff, row7_lz);
+ row0_diff = vshlq_u16(row0_diff, vnegq_s16(row0_lz));
+ row1_diff = vshlq_u16(row1_diff, vnegq_s16(row1_lz));
+ row2_diff = vshlq_u16(row2_diff, vnegq_s16(row2_lz));
+ row3_diff = vshlq_u16(row3_diff, vnegq_s16(row3_lz));
+ row4_diff = vshlq_u16(row4_diff, vnegq_s16(row4_lz));
+ row5_diff = vshlq_u16(row5_diff, vnegq_s16(row5_lz));
+ row6_diff = vshlq_u16(row6_diff, vnegq_s16(row6_lz));
+ row7_diff = vshlq_u16(row7_diff, vnegq_s16(row7_lz));
+ /* Store diff bits. */
+ vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+ vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+ vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+ vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+ vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+ vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+ vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+ vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+ while (bitmap != 0) {
+ r = BUILTIN_CLZL(bitmap);
+ i += r;
+ bitmap <<= r;
+ nbits = block_nbits[i];
+ diff = block_diff[i];
+ while (r > 15) {
+ /* If run length > 15, emit special run-length-16 codes. */
+ PUT_BITS(code_0xf0, size_0xf0)
+ r -= 16;
+ }
+ /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+ unsigned int rs = (r << 4) + nbits;
+ PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+ i++;
+ bitmap <<= 1;
+ }
+ } else if (bitmap != 0) {
+ uint16_t block_abs[DCTSIZE2];
+ /* Store absolute value of coefficients. */
+ vst1q_u16(block_abs + 0 * DCTSIZE, vreinterpretq_u16_s16(abs_row0));
+ vst1q_u16(block_abs + 1 * DCTSIZE, vreinterpretq_u16_s16(abs_row1));
+ vst1q_u16(block_abs + 2 * DCTSIZE, vreinterpretq_u16_s16(abs_row2));
+ vst1q_u16(block_abs + 3 * DCTSIZE, vreinterpretq_u16_s16(abs_row3));
+ vst1q_u16(block_abs + 4 * DCTSIZE, vreinterpretq_u16_s16(abs_row4));
+ vst1q_u16(block_abs + 5 * DCTSIZE, vreinterpretq_u16_s16(abs_row5));
+ vst1q_u16(block_abs + 6 * DCTSIZE, vreinterpretq_u16_s16(abs_row6));
+ vst1q_u16(block_abs + 7 * DCTSIZE, vreinterpretq_u16_s16(abs_row7));
+ /* Store diff bits. */
+ vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
+ vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
+ vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
+ vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
+ vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
+ vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
+ vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
+ vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);
+
+ /* Same as above but must mask diff bits and compute nbits on demand. */
+ while (bitmap != 0) {
+ r = BUILTIN_CLZL(bitmap);
+ i += r;
+ bitmap <<= r;
+ lz = BUILTIN_CLZ(block_abs[i]);
+ nbits = 32 - lz;
+ diff = (unsigned int)(block_diff[i] << lz) >> lz;
+ while (r > 15) {
+ /* If run length > 15, emit special run-length-16 codes. */
+ PUT_BITS(code_0xf0, size_0xf0)
+ r -= 16;
+ }
+ /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
+ unsigned int rs = (r << 4) + nbits;
+ PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
+ i++;
+ bitmap <<= 1;
+ }
+ }
+
+ /* If the last coefficient(s) were zero, emit an end-of-block (EOB) code.
+ * The value of RS for the EOB code is 0.
+ */
+ if (i != 64) {
+ PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
+ }
+
+ state_ptr->cur.put_buffer = put_buffer;
+ state_ptr->cur.free_bits = free_bits;
+
+ return buffer;
+}
diff --git a/simd/arm/arm64/jsimd.c b/simd/arm/aarch64/jsimd.c
similarity index 87%
rename from simd/arm/arm64/jsimd.c
rename to simd/arm/aarch64/jsimd.c
index ca29cd6..8570b82 100644
--- a/simd/arm/arm64/jsimd.c
+++ b/simd/arm/aarch64/jsimd.c
@@ -3,8 +3,9 @@
*
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
* Copyright (C) 2011, Nokia Corporation and/or its subsidiary(-ies).
- * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, D. R. Commander.
+ * Copyright (C) 2009-2011, 2013-2014, 2016, 2018, 2020, D. R. Commander.
* Copyright (C) 2015-2016, 2018, Matthieu Darbois.
+ * Copyright (C) 2020, Arm Limited.
*
* Based on the x86 SIMD extension for IJG JPEG library,
* Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -12,7 +13,7 @@
*
* This file contains the interface between the "normal" portions
* of the library and the SIMD implementations when running on a
- * 64-bit ARM architecture.
+ * 64-bit Arm architecture.
*/
#define JPEG_INTERNALS
@@ -22,16 +23,20 @@
#include "../../../jdct.h"
#include "../../../jsimddct.h"
#include "../../jsimd.h"
+#include "jconfigint.h"
#include <stdio.h>
#include <string.h>
#include <ctype.h>
+#define JSIMD_FASTLD3 1
+#define JSIMD_FASTST3 2
#define JSIMD_FASTTBL 4
static unsigned int simd_support = ~0;
static unsigned int simd_huffman = 1;
-static unsigned int simd_features = JSIMD_FASTTBL;
+static unsigned int simd_features = JSIMD_FASTLD3 | JSIMD_FASTST3 |
+ JSIMD_FASTTBL;
#if defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
@@ -111,8 +116,8 @@
*/
/*
- * ARMv8 architectures support NEON extensions by default.
- * It is no longer optional as it was with ARMv7.
+ * Armv8 architectures support Neon extensions by default.
+ * It is no longer optional as it was with Armv7.
*/
@@ -151,6 +156,16 @@
env = getenv("JSIMD_NOHUFFENC");
if ((env != NULL) && (strcmp(env, "1") == 0))
simd_huffman = 0;
+ env = getenv("JSIMD_FASTLD3");
+ if ((env != NULL) && (strcmp(env, "1") == 0))
+ simd_features |= JSIMD_FASTLD3;
+ if ((env != NULL) && (strcmp(env, "0") == 0))
+ simd_features &= ~JSIMD_FASTLD3;
+ env = getenv("JSIMD_FASTST3");
+ if ((env != NULL) && (strcmp(env, "1") == 0))
+ simd_features |= JSIMD_FASTST3;
+ if ((env != NULL) && (strcmp(env, "0") == 0))
+ simd_features &= ~JSIMD_FASTST3;
#endif
}
@@ -237,14 +252,28 @@
switch (cinfo->in_color_space) {
case JCS_EXT_RGB:
- neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTLD3)
+#endif
+ neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_extrgb_ycc_convert_neon_slowld3;
+#endif
break;
case JCS_EXT_RGBX:
case JCS_EXT_RGBA:
neonfct = jsimd_extrgbx_ycc_convert_neon;
break;
case JCS_EXT_BGR:
- neonfct = jsimd_extbgr_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTLD3)
+#endif
+ neonfct = jsimd_extbgr_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_extbgr_ycc_convert_neon_slowld3;
+#endif
break;
case JCS_EXT_BGRX:
case JCS_EXT_BGRA:
@@ -259,7 +288,14 @@
neonfct = jsimd_extxrgb_ycc_convert_neon;
break;
default:
- neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTLD3)
+#endif
+ neonfct = jsimd_extrgb_ycc_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_extrgb_ycc_convert_neon_slowld3;
+#endif
break;
}
@@ -313,14 +349,28 @@
switch (cinfo->out_color_space) {
case JCS_EXT_RGB:
- neonfct = jsimd_ycc_extrgb_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTST3)
+#endif
+ neonfct = jsimd_ycc_extrgb_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_ycc_extrgb_convert_neon_slowst3;
+#endif
break;
case JCS_EXT_RGBX:
case JCS_EXT_RGBA:
neonfct = jsimd_ycc_extrgbx_convert_neon;
break;
case JCS_EXT_BGR:
- neonfct = jsimd_ycc_extbgr_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTST3)
+#endif
+ neonfct = jsimd_ycc_extbgr_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_ycc_extbgr_convert_neon_slowst3;
+#endif
break;
case JCS_EXT_BGRX:
case JCS_EXT_BGRA:
@@ -335,7 +385,15 @@
neonfct = jsimd_ycc_extxrgb_convert_neon;
break;
default:
- neonfct = jsimd_ycc_extrgb_convert_neon;
+#ifndef NEON_INTRINSICS
+ if (simd_features & JSIMD_FASTST3)
+#endif
+ neonfct = jsimd_ycc_extrgb_convert_neon;
+#ifndef NEON_INTRINSICS
+ else
+ neonfct = jsimd_ycc_extrgb_convert_neon_slowst3;
+#endif
+ break;
}
neonfct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
@@ -433,7 +491,6 @@
return 0;
if (sizeof(JDIMENSION) != 4)
return 0;
-
if (simd_support & JSIMD_NEON)
return 1;
@@ -938,17 +995,33 @@
int last_dc_val, c_derived_tbl *dctbl,
c_derived_tbl *actbl)
{
+#ifndef NEON_INTRINSICS
if (simd_features & JSIMD_FASTTBL)
+#endif
return jsimd_huff_encode_one_block_neon(state, buffer, block, last_dc_val,
dctbl, actbl);
+#ifndef NEON_INTRINSICS
else
return jsimd_huff_encode_one_block_neon_slowtbl(state, buffer, block,
last_dc_val, dctbl, actbl);
+#endif
}
GLOBAL(int)
jsimd_can_encode_mcu_AC_first_prepare(void)
{
+ init_simd();
+
+ if (DCTSIZE != 8)
+ return 0;
+ if (sizeof(JCOEF) != 2)
+ return 0;
+ if (SIZEOF_SIZE_T != 8)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -957,11 +1030,25 @@
const int *jpeg_natural_order_start, int Sl,
int Al, JCOEF *values, size_t *zerobits)
{
+ jsimd_encode_mcu_AC_first_prepare_neon(block, jpeg_natural_order_start,
+ Sl, Al, values, zerobits);
}
GLOBAL(int)
jsimd_can_encode_mcu_AC_refine_prepare(void)
{
+ init_simd();
+
+ if (DCTSIZE != 8)
+ return 0;
+ if (sizeof(JCOEF) != 2)
+ return 0;
+ if (SIZEOF_SIZE_T != 8)
+ return 0;
+
+ if (simd_support & JSIMD_NEON)
+ return 1;
+
return 0;
}
@@ -970,5 +1057,7 @@
const int *jpeg_natural_order_start, int Sl,
int Al, JCOEF *absvalues, size_t *bits)
{
- return 0;
+ return jsimd_encode_mcu_AC_refine_prepare_neon(block,
+ jpeg_natural_order_start,
+ Sl, Al, absvalues, bits);
}
diff --git a/simd/arm/align.h b/simd/arm/align.h
new file mode 100644
index 0000000..cff4241
--- /dev/null
+++ b/simd/arm/align.h
@@ -0,0 +1,28 @@
+/*
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* How to obtain memory alignment for structures and variables */
+#if defined(_MSC_VER)
+#define ALIGN(alignment) __declspec(align(alignment))
+#elif defined(__clang__) || defined(__GNUC__)
+#define ALIGN(alignment) __attribute__((aligned(alignment)))
+#else
+#error "Unknown compiler"
+#endif
diff --git a/simd/arm/arm/jsimd_neon.S b/simd/arm/arm/jsimd_neon.S
deleted file mode 100644
index 2c45324..0000000
--- a/simd/arm/arm/jsimd_neon.S
+++ /dev/null
@@ -1,499 +0,0 @@
-/*
- * ARMv7 NEON optimizations for libjpeg-turbo
- *
- * Copyright (C) 2009-2011, Nokia Corporation and/or its subsidiary(-ies).
- * All Rights Reserved.
- * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com>
- * Copyright (C) 2014, Siarhei Siamashka. All Rights Reserved.
- * Copyright (C) 2014, Linaro Limited. All Rights Reserved.
- * Copyright (C) 2015, D. R. Commander. All Rights Reserved.
- * Copyright (C) 2015-2016, 2018, Matthieu Darbois. All Rights Reserved.
- *
- * This software is provided 'as-is', without any express or implied
- * warranty. In no event will the authors be held liable for any damages
- * arising from the use of this software.
- *
- * Permission is granted to anyone to use this software for any purpose,
- * including commercial applications, and to alter it and redistribute it
- * freely, subject to the following restrictions:
- *
- * 1. The origin of this software must not be misrepresented; you must not
- * claim that you wrote the original software. If you use this software
- * in a product, an acknowledgment in the product documentation would be
- * appreciated but is not required.
- * 2. Altered source versions must be plainly marked as such, and must not be
- * misrepresented as being the original software.
- * 3. This notice may not be removed or altered from any source distribution.
- */
-
-#if defined(__linux__) && defined(__ELF__)
-.section .note.GNU-stack, "", %progbits /* mark stack as non-executable */
-#endif
-
-.text
-.fpu neon
-.arch armv7a
-.object_arch armv4
-.arm
-.syntax unified
-
-
-#define RESPECT_STRICT_ALIGNMENT 1
-
-
-/*****************************************************************************/
-
-/* Supplementary macro for setting function attributes */
-.macro asm_function fname
-#ifdef __APPLE__
- .private_extern _\fname
- .globl _\fname
-_\fname:
-#else
- .global \fname
-#ifdef __ELF__
- .hidden \fname
- .type \fname, %function
-#endif
-\fname:
-#endif
-.endm
-
-
-#define CENTERJSAMPLE 128
-
-/*****************************************************************************/
-
-/*
- * GLOBAL(JOCTET*)
- * jsimd_huff_encode_one_block(working_state *state, JOCTET *buffer,
- * JCOEFPTR block, int last_dc_val,
- * c_derived_tbl *dctbl, c_derived_tbl *actbl)
- *
- */
-
-.macro emit_byte BUFFER, PUT_BUFFER, PUT_BITS, ZERO, TMP
- sub \PUT_BITS, \PUT_BITS, #0x8
- lsr \TMP, \PUT_BUFFER, \PUT_BITS
- uxtb \TMP, \TMP
- strb \TMP, [\BUFFER, #1]!
- cmp \TMP, #0xff
- /*it eq*/
- strbeq \ZERO, [\BUFFER, #1]!
-.endm
-
-.macro put_bits PUT_BUFFER, PUT_BITS, CODE, SIZE
- /*lsl \PUT_BUFFER, \PUT_BUFFER, \SIZE*/
- add \PUT_BITS, \SIZE
- /*orr \PUT_BUFFER, \PUT_BUFFER, \CODE*/
- orr \PUT_BUFFER, \CODE, \PUT_BUFFER, lsl \SIZE
-.endm
-
-.macro checkbuf15 BUFFER, PUT_BUFFER, PUT_BITS, ZERO, TMP
- cmp \PUT_BITS, #0x10
- blt 15f
- eor \ZERO, \ZERO, \ZERO
- emit_byte \BUFFER, \PUT_BUFFER, \PUT_BITS, \ZERO, \TMP
- emit_byte \BUFFER, \PUT_BUFFER, \PUT_BITS, \ZERO, \TMP
-15:
-.endm
-
-.balign 16
-jsimd_huff_encode_one_block_neon_consts:
- .byte 0x01
- .byte 0x02
- .byte 0x04
- .byte 0x08
- .byte 0x10
- .byte 0x20
- .byte 0x40
- .byte 0x80
-
-asm_function jsimd_huff_encode_one_block_neon
- push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
- add r7, sp, #0x1c
- sub r4, sp, #0x40
- bfc r4, #0, #5
- mov sp, r4 /* align sp on 32 bytes */
- vst1.64 {d8, d9, d10, d11}, [r4, :128]!
- vst1.64 {d12, d13, d14, d15}, [r4, :128]
- sub sp, #0x140 /* reserve 320 bytes */
- str r0, [sp, #0x18] /* working state > sp + Ox18 */
- add r4, sp, #0x20 /* r4 = t1 */
- ldr lr, [r7, #0x8] /* lr = dctbl */
- sub r10, r1, #0x1 /* r10=buffer-- */
- ldrsh r1, [r2]
- mov r9, #0x10
- mov r8, #0x1
- adr r5, jsimd_huff_encode_one_block_neon_consts
- /* prepare data */
- vld1.8 {d26}, [r5, :64]
- veor q8, q8, q8
- veor q9, q9, q9
- vdup.16 q14, r9
- vdup.16 q15, r8
- veor q10, q10, q10
- veor q11, q11, q11
- sub r1, r1, r3
- add r9, r2, #0x22
- add r8, r2, #0x18
- add r3, r2, #0x36
- vmov.16 d0[0], r1
- vld1.16 {d2[0]}, [r9, :16]
- vld1.16 {d4[0]}, [r8, :16]
- vld1.16 {d6[0]}, [r3, :16]
- add r1, r2, #0x2
- add r9, r2, #0x30
- add r8, r2, #0x26
- add r3, r2, #0x28
- vld1.16 {d0[1]}, [r1, :16]
- vld1.16 {d2[1]}, [r9, :16]
- vld1.16 {d4[1]}, [r8, :16]
- vld1.16 {d6[1]}, [r3, :16]
- add r1, r2, #0x10
- add r9, r2, #0x40
- add r8, r2, #0x34
- add r3, r2, #0x1a
- vld1.16 {d0[2]}, [r1, :16]
- vld1.16 {d2[2]}, [r9, :16]
- vld1.16 {d4[2]}, [r8, :16]
- vld1.16 {d6[2]}, [r3, :16]
- add r1, r2, #0x20
- add r9, r2, #0x32
- add r8, r2, #0x42
- add r3, r2, #0xc
- vld1.16 {d0[3]}, [r1, :16]
- vld1.16 {d2[3]}, [r9, :16]
- vld1.16 {d4[3]}, [r8, :16]
- vld1.16 {d6[3]}, [r3, :16]
- add r1, r2, #0x12
- add r9, r2, #0x24
- add r8, r2, #0x50
- add r3, r2, #0xe
- vld1.16 {d1[0]}, [r1, :16]
- vld1.16 {d3[0]}, [r9, :16]
- vld1.16 {d5[0]}, [r8, :16]
- vld1.16 {d7[0]}, [r3, :16]
- add r1, r2, #0x4
- add r9, r2, #0x16
- add r8, r2, #0x60
- add r3, r2, #0x1c
- vld1.16 {d1[1]}, [r1, :16]
- vld1.16 {d3[1]}, [r9, :16]
- vld1.16 {d5[1]}, [r8, :16]
- vld1.16 {d7[1]}, [r3, :16]
- add r1, r2, #0x6
- add r9, r2, #0x8
- add r8, r2, #0x52
- add r3, r2, #0x2a
- vld1.16 {d1[2]}, [r1, :16]
- vld1.16 {d3[2]}, [r9, :16]
- vld1.16 {d5[2]}, [r8, :16]
- vld1.16 {d7[2]}, [r3, :16]
- add r1, r2, #0x14
- add r9, r2, #0xa
- add r8, r2, #0x44
- add r3, r2, #0x38
- vld1.16 {d1[3]}, [r1, :16]
- vld1.16 {d3[3]}, [r9, :16]
- vld1.16 {d5[3]}, [r8, :16]
- vld1.16 {d7[3]}, [r3, :16]
- vcgt.s16 q8, q8, q0
- vcgt.s16 q9, q9, q1
- vcgt.s16 q10, q10, q2
- vcgt.s16 q11, q11, q3
- vabs.s16 q0, q0
- vabs.s16 q1, q1
- vabs.s16 q2, q2
- vabs.s16 q3, q3
- veor q8, q8, q0
- veor q9, q9, q1
- veor q10, q10, q2
- veor q11, q11, q3
- add r9, r4, #0x20
- add r8, r4, #0x80
- add r3, r4, #0xa0
- vclz.i16 q0, q0
- vclz.i16 q1, q1
- vclz.i16 q2, q2
- vclz.i16 q3, q3
- vsub.i16 q0, q14, q0
- vsub.i16 q1, q14, q1
- vsub.i16 q2, q14, q2
- vsub.i16 q3, q14, q3
- vst1.16 {d0, d1, d2, d3}, [r4, :256]
- vst1.16 {d4, d5, d6, d7}, [r9, :256]
- vshl.s16 q0, q15, q0
- vshl.s16 q1, q15, q1
- vshl.s16 q2, q15, q2
- vshl.s16 q3, q15, q3
- vsub.i16 q0, q0, q15
- vsub.i16 q1, q1, q15
- vsub.i16 q2, q2, q15
- vsub.i16 q3, q3, q15
- vand q8, q8, q0
- vand q9, q9, q1
- vand q10, q10, q2
- vand q11, q11, q3
- vst1.16 {d16, d17, d18, d19}, [r8, :256]
- vst1.16 {d20, d21, d22, d23}, [r3, :256]
- add r1, r2, #0x46
- add r9, r2, #0x3a
- add r8, r2, #0x74
- add r3, r2, #0x6a
- vld1.16 {d8[0]}, [r1, :16]
- vld1.16 {d10[0]}, [r9, :16]
- vld1.16 {d12[0]}, [r8, :16]
- vld1.16 {d14[0]}, [r3, :16]
- veor q8, q8, q8
- veor q9, q9, q9
- veor q10, q10, q10
- veor q11, q11, q11
- add r1, r2, #0x54
- add r9, r2, #0x2c
- add r8, r2, #0x76
- add r3, r2, #0x78
- vld1.16 {d8[1]}, [r1, :16]
- vld1.16 {d10[1]}, [r9, :16]
- vld1.16 {d12[1]}, [r8, :16]
- vld1.16 {d14[1]}, [r3, :16]
- add r1, r2, #0x62
- add r9, r2, #0x1e
- add r8, r2, #0x68
- add r3, r2, #0x7a
- vld1.16 {d8[2]}, [r1, :16]
- vld1.16 {d10[2]}, [r9, :16]
- vld1.16 {d12[2]}, [r8, :16]
- vld1.16 {d14[2]}, [r3, :16]
- add r1, r2, #0x70
- add r9, r2, #0x2e
- add r8, r2, #0x5a
- add r3, r2, #0x6c
- vld1.16 {d8[3]}, [r1, :16]
- vld1.16 {d10[3]}, [r9, :16]
- vld1.16 {d12[3]}, [r8, :16]
- vld1.16 {d14[3]}, [r3, :16]
- add r1, r2, #0x72
- add r9, r2, #0x3c
- add r8, r2, #0x4c
- add r3, r2, #0x5e
- vld1.16 {d9[0]}, [r1, :16]
- vld1.16 {d11[0]}, [r9, :16]
- vld1.16 {d13[0]}, [r8, :16]
- vld1.16 {d15[0]}, [r3, :16]
- add r1, r2, #0x64
- add r9, r2, #0x4a
- add r8, r2, #0x3e
- add r3, r2, #0x6e
- vld1.16 {d9[1]}, [r1, :16]
- vld1.16 {d11[1]}, [r9, :16]
- vld1.16 {d13[1]}, [r8, :16]
- vld1.16 {d15[1]}, [r3, :16]
- add r1, r2, #0x56
- add r9, r2, #0x58
- add r8, r2, #0x4e
- add r3, r2, #0x7c
- vld1.16 {d9[2]}, [r1, :16]
- vld1.16 {d11[2]}, [r9, :16]
- vld1.16 {d13[2]}, [r8, :16]
- vld1.16 {d15[2]}, [r3, :16]
- add r1, r2, #0x48
- add r9, r2, #0x66
- add r8, r2, #0x5c
- add r3, r2, #0x7e
- vld1.16 {d9[3]}, [r1, :16]
- vld1.16 {d11[3]}, [r9, :16]
- vld1.16 {d13[3]}, [r8, :16]
- vld1.16 {d15[3]}, [r3, :16]
- vcgt.s16 q8, q8, q4
- vcgt.s16 q9, q9, q5
- vcgt.s16 q10, q10, q6
- vcgt.s16 q11, q11, q7
- vabs.s16 q4, q4
- vabs.s16 q5, q5
- vabs.s16 q6, q6
- vabs.s16 q7, q7
- veor q8, q8, q4
- veor q9, q9, q5
- veor q10, q10, q6
- veor q11, q11, q7
- add r1, r4, #0x40
- add r9, r4, #0x60
- add r8, r4, #0xc0
- add r3, r4, #0xe0
- vclz.i16 q4, q4
- vclz.i16 q5, q5
- vclz.i16 q6, q6
- vclz.i16 q7, q7
- vsub.i16 q4, q14, q4
- vsub.i16 q5, q14, q5
- vsub.i16 q6, q14, q6
- vsub.i16 q7, q14, q7
- vst1.16 {d8, d9, d10, d11}, [r1, :256]
- vst1.16 {d12, d13, d14, d15}, [r9, :256]
- vshl.s16 q4, q15, q4
- vshl.s16 q5, q15, q5
- vshl.s16 q6, q15, q6
- vshl.s16 q7, q15, q7
- vsub.i16 q4, q4, q15
- vsub.i16 q5, q5, q15
- vsub.i16 q6, q6, q15
- vsub.i16 q7, q7, q15
- vand q8, q8, q4
- vand q9, q9, q5
- vand q10, q10, q6
- vand q11, q11, q7
- vst1.16 {d16, d17, d18, d19}, [r8, :256]
- vst1.16 {d20, d21, d22, d23}, [r3, :256]
- ldr r12, [r7, #0xc] /* r12 = actbl */
- add r1, lr, #0x400 /* r1 = dctbl->ehufsi */
- mov r9, r12 /* r9 = actbl */
- add r6, r4, #0x80 /* r6 = t2 */
- ldr r11, [r0, #0x8] /* r11 = put_buffer */
- ldr r4, [r0, #0xc] /* r4 = put_bits */
- ldrh r2, [r6, #-128] /* r2 = nbits */
- ldrh r3, [r6] /* r3 = temp2 & (((JLONG)1)<<nbits) - 1; */
- ldr r0, [lr, r2, lsl #2]
- ldrb r5, [r1, r2]
- put_bits r11, r4, r0, r5
- checkbuf15 r10, r11, r4, r5, r0
- put_bits r11, r4, r3, r2
- checkbuf15 r10, r11, r4, r5, r0
- mov lr, r6 /* lr = t2 */
- add r5, r9, #0x400 /* r5 = actbl->ehufsi */
- ldrsb r6, [r5, #0xf0] /* r6 = actbl->ehufsi[0xf0] */
- veor q8, q8, q8
- vceq.i16 q0, q0, q8
- vceq.i16 q1, q1, q8
- vceq.i16 q2, q2, q8
- vceq.i16 q3, q3, q8
- vceq.i16 q4, q4, q8
- vceq.i16 q5, q5, q8
- vceq.i16 q6, q6, q8
- vceq.i16 q7, q7, q8
- vmovn.i16 d0, q0
- vmovn.i16 d2, q1
- vmovn.i16 d4, q2
- vmovn.i16 d6, q3
- vmovn.i16 d8, q4
- vmovn.i16 d10, q5
- vmovn.i16 d12, q6
- vmovn.i16 d14, q7
- vand d0, d0, d26
- vand d2, d2, d26
- vand d4, d4, d26
- vand d6, d6, d26
- vand d8, d8, d26
- vand d10, d10, d26
- vand d12, d12, d26
- vand d14, d14, d26
- vpadd.i8 d0, d0, d2
- vpadd.i8 d4, d4, d6
- vpadd.i8 d8, d8, d10
- vpadd.i8 d12, d12, d14
- vpadd.i8 d0, d0, d4
- vpadd.i8 d8, d8, d12
- vpadd.i8 d0, d0, d8
- vmov.32 r1, d0[1]
- vmov.32 r8, d0[0]
- mvn r1, r1
- mvn r8, r8
- lsrs r1, r1, #0x1
- rrx r8, r8 /* shift in last r1 bit while shifting out DC bit */
- rbit r1, r1 /* r1 = index1 */
- rbit r8, r8 /* r8 = index0 */
- ldr r0, [r9, #0x3c0] /* r0 = actbl->ehufco[0xf0] */
- str r1, [sp, #0x14] /* index1 > sp + 0x14 */
- cmp r8, #0x0
- beq 6f
-1:
- clz r2, r8
- add lr, lr, r2, lsl #1
- lsl r8, r8, r2
- ldrh r1, [lr, #-126]
-2:
- cmp r2, #0x10
- blt 3f
- sub r2, r2, #0x10
- put_bits r11, r4, r0, r6
- cmp r4, #0x10
- blt 2b
- eor r3, r3, r3
- emit_byte r10, r11, r4, r3, r12
- emit_byte r10, r11, r4, r3, r12
- b 2b
-3:
- add r2, r1, r2, lsl #4
- ldrh r3, [lr, #2]!
- ldr r12, [r9, r2, lsl #2]
- ldrb r2, [r5, r2]
- put_bits r11, r4, r12, r2
- checkbuf15 r10, r11, r4, r2, r12
- put_bits r11, r4, r3, r1
- checkbuf15 r10, r11, r4, r2, r12
- lsls r8, r8, #0x1
- bne 1b
-6:
- add r12, sp, #0x20 /* r12 = t1 */
- ldr r8, [sp, #0x14] /* r8 = index1 */
- adds r12, #0xc0 /* r12 = t2 + (DCTSIZE2/2) */
- cmp r8, #0x0
- beq 6f
- clz r2, r8
- sub r12, r12, lr
- lsl r8, r8, r2
- add r2, r2, r12, lsr #1
- add lr, lr, r2, lsl #1
- b 7f
-1:
- clz r2, r8
- add lr, lr, r2, lsl #1
- lsl r8, r8, r2
-7:
- ldrh r1, [lr, #-126]
-2:
- cmp r2, #0x10
- blt 3f
- sub r2, r2, #0x10
- put_bits r11, r4, r0, r6
- cmp r4, #0x10
- blt 2b
- eor r3, r3, r3
- emit_byte r10, r11, r4, r3, r12
- emit_byte r10, r11, r4, r3, r12
- b 2b
-3:
- add r2, r1, r2, lsl #4
- ldrh r3, [lr, #2]!
- ldr r12, [r9, r2, lsl #2]
- ldrb r2, [r5, r2]
- put_bits r11, r4, r12, r2
- checkbuf15 r10, r11, r4, r2, r12
- put_bits r11, r4, r3, r1
- checkbuf15 r10, r11, r4, r2, r12
- lsls r8, r8, #0x1
- bne 1b
-6:
- add r0, sp, #0x20
- add r0, #0xfe
- cmp lr, r0
- bhs 1f
- ldr r1, [r9]
- ldrb r0, [r5]
- put_bits r11, r4, r1, r0
- checkbuf15 r10, r11, r4, r0, r1
-1:
- ldr r12, [sp, #0x18]
- str r11, [r12, #0x8]
- str r4, [r12, #0xc]
- add r0, r10, #0x1
- add r4, sp, #0x140
- vld1.64 {d8, d9, d10, d11}, [r4, :128]!
- vld1.64 {d12, d13, d14, d15}, [r4, :128]
- sub r4, r7, #0x1c
- mov sp, r4
- pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
-
-.purgem emit_byte
-.purgem put_bits
-.purgem checkbuf15
diff --git a/simd/arm/arm64/jsimd_neon.S b/simd/arm/arm64/jsimd_neon.S
deleted file mode 100644
index 898cf2c..0000000
--- a/simd/arm/arm64/jsimd_neon.S
+++ /dev/null
@@ -1,538 +0,0 @@
-/*
- * ARMv8 NEON optimizations for libjpeg-turbo
- *
- * Copyright (C) 2009-2011, Nokia Corporation and/or its subsidiary(-ies).
- * All Rights Reserved.
- * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com>
- * Copyright (C) 2013-2014, Linaro Limited. All Rights Reserved.
- * Author: Ragesh Radhakrishnan <ragesh.r@linaro.org>
- * Copyright (C) 2014-2016, D. R. Commander. All Rights Reserved.
- * Copyright (C) 2015-2016, 2018, Matthieu Darbois. All Rights Reserved.
- * Copyright (C) 2016, Siarhei Siamashka. All Rights Reserved.
- *
- * This software is provided 'as-is', without any express or implied
- * warranty. In no event will the authors be held liable for any damages
- * arising from the use of this software.
- *
- * Permission is granted to anyone to use this software for any purpose,
- * including commercial applications, and to alter it and redistribute it
- * freely, subject to the following restrictions:
- *
- * 1. The origin of this software must not be misrepresented; you must not
- * claim that you wrote the original software. If you use this software
- * in a product, an acknowledgment in the product documentation would be
- * appreciated but is not required.
- * 2. Altered source versions must be plainly marked as such, and must not be
- * misrepresented as being the original software.
- * 3. This notice may not be removed or altered from any source distribution.
- */
-
-#if defined(__linux__) && defined(__ELF__)
-.section .note.GNU-stack, "", %progbits /* mark stack as non-executable */
-#endif
-
-#if defined(__APPLE__)
-.section __DATA, __const
-#elif defined(_WIN32)
-.section .rdata
-#else
-.section .rodata, "a", %progbits
-#endif
-
-/* Constants for jsimd_huff_encode_one_block_neon() */
-
-.balign 16
-Ljsimd_huff_encode_one_block_neon_consts:
- .byte 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, \
- 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80
- .byte 0, 1, 2, 3, 16, 17, 32, 33, \
- 18, 19, 4, 5, 6, 7, 20, 21 /* L0 => L3 : 4 lines OK */
- .byte 34, 35, 48, 49, 255, 255, 50, 51, \
- 36, 37, 22, 23, 8, 9, 10, 11 /* L0 => L3 : 4 lines OK */
- .byte 8, 9, 22, 23, 36, 37, 50, 51, \
- 255, 255, 255, 255, 255, 255, 52, 53 /* L1 => L4 : 4 lines OK */
- .byte 54, 55, 40, 41, 26, 27, 12, 13, \
- 14, 15, 28, 29, 42, 43, 56, 57 /* L0 => L3 : 4 lines OK */
- .byte 6, 7, 20, 21, 34, 35, 48, 49, \
- 50, 51, 36, 37, 22, 23, 8, 9 /* L4 => L7 : 4 lines OK */
- .byte 42, 43, 28, 29, 14, 15, 30, 31, \
- 44, 45, 58, 59, 255, 255, 255, 255 /* L1 => L4 : 4 lines OK */
- .byte 255, 255, 255, 255, 56, 57, 42, 43, \
- 28, 29, 14, 15, 30, 31, 44, 45 /* L3 => L6 : 4 lines OK */
- .byte 26, 27, 40, 41, 42, 43, 28, 29, \
- 14, 15, 30, 31, 44, 45, 46, 47 /* L5 => L7 : 3 lines OK */
- .byte 255, 255, 255, 255, 0, 1, 255, 255, \
- 255, 255, 255, 255, 255, 255, 255, 255 /* L4 : 1 lines OK */
- .byte 255, 255, 255, 255, 255, 255, 255, 255, \
- 0, 1, 16, 17, 2, 3, 255, 255 /* L5 => L6 : 2 lines OK */
- .byte 255, 255, 255, 255, 255, 255, 255, 255, \
- 255, 255, 255, 255, 8, 9, 22, 23 /* L5 => L6 : 2 lines OK */
- .byte 4, 5, 6, 7, 255, 255, 255, 255, \
- 255, 255, 255, 255, 255, 255, 255, 255 /* L7 : 1 line OK */
-
-.text
-
-
-#define RESPECT_STRICT_ALIGNMENT 1
-
-
-/*****************************************************************************/
-
-/* Supplementary macro for setting function attributes */
-.macro asm_function fname
-#ifdef __APPLE__
- .private_extern _\fname
- .globl _\fname
-_\fname:
-#else
- .global \fname
-#ifdef __ELF__
- .hidden \fname
- .type \fname, %function
-#endif
-\fname:
-#endif
-.endm
-
-/* Get symbol location */
-.macro get_symbol_loc reg, symbol
-#ifdef __APPLE__
- adrp \reg, \symbol@PAGE
- add \reg, \reg, \symbol@PAGEOFF
-#else
- adrp \reg, \symbol
- add \reg, \reg, :lo12:\symbol
-#endif
-.endm
-
-
-#define CENTERJSAMPLE 128
-
-/*****************************************************************************/
-
-/*
- * GLOBAL(JOCTET *)
- * jsimd_huff_encode_one_block(working_state *state, JOCTET *buffer,
- * JCOEFPTR block, int last_dc_val,
- * c_derived_tbl *dctbl, c_derived_tbl *actbl)
- *
- */
-
- BUFFER .req x1
- PUT_BUFFER .req x6
- PUT_BITS .req x7
- PUT_BITSw .req w7
-
-.macro emit_byte
- sub PUT_BITS, PUT_BITS, #0x8
- lsr x19, PUT_BUFFER, PUT_BITS
- uxtb w19, w19
- strb w19, [BUFFER, #1]!
- cmp w19, #0xff
- b.ne 14f
- strb wzr, [BUFFER, #1]!
-14:
-.endm
-.macro put_bits CODE, SIZE
- lsl PUT_BUFFER, PUT_BUFFER, \SIZE
- add PUT_BITS, PUT_BITS, \SIZE
- orr PUT_BUFFER, PUT_BUFFER, \CODE
-.endm
-.macro checkbuf31
- cmp PUT_BITS, #0x20
- b.lt 31f
- emit_byte
- emit_byte
- emit_byte
- emit_byte
-31:
-.endm
-.macro checkbuf47
- cmp PUT_BITS, #0x30
- b.lt 47f
- emit_byte
- emit_byte
- emit_byte
- emit_byte
- emit_byte
- emit_byte
-47:
-.endm
-
-.macro generate_jsimd_huff_encode_one_block fast_tbl
-
-.balign 16
-.if \fast_tbl == 1
-asm_function jsimd_huff_encode_one_block_neon
-.else
-asm_function jsimd_huff_encode_one_block_neon_slowtbl
-.endif
- sub sp, sp, 272
- sub BUFFER, BUFFER, #0x1 /* BUFFER=buffer-- */
- /* Save ARM registers */
- stp x19, x20, [sp]
- get_symbol_loc x15, Ljsimd_huff_encode_one_block_neon_consts
- ldr PUT_BUFFER, [x0, #0x10]
- ldr PUT_BITSw, [x0, #0x18]
- ldrsh w12, [x2] /* load DC coeff in w12 */
- /* prepare data */
-.if \fast_tbl == 1
- ld1 {v23.16b}, [x15], #16
- ld1 {v0.16b, v1.16b, v2.16b, v3.16b}, [x15], #64
- ld1 {v4.16b, v5.16b, v6.16b, v7.16b}, [x15], #64
- ld1 {v16.16b, v17.16b, v18.16b, v19.16b}, [x15], #64
- ld1 {v24.16b, v25.16b, v26.16b, v27.16b}, [x2], #64
- ld1 {v28.16b, v29.16b, v30.16b, v31.16b}, [x2], #64
- sub w12, w12, w3 /* last_dc_val, not used afterwards */
- /* ZigZag 8x8 */
- tbl v0.16b, {v24.16b, v25.16b, v26.16b, v27.16b}, v0.16b
- tbl v1.16b, {v24.16b, v25.16b, v26.16b, v27.16b}, v1.16b
- tbl v2.16b, {v25.16b, v26.16b, v27.16b, v28.16b}, v2.16b
- tbl v3.16b, {v24.16b, v25.16b, v26.16b, v27.16b}, v3.16b
- tbl v4.16b, {v28.16b, v29.16b, v30.16b, v31.16b}, v4.16b
- tbl v5.16b, {v25.16b, v26.16b, v27.16b, v28.16b}, v5.16b
- tbl v6.16b, {v27.16b, v28.16b, v29.16b, v30.16b}, v6.16b
- tbl v7.16b, {v29.16b, v30.16b, v31.16b}, v7.16b
- ins v0.h[0], w12
- tbx v1.16b, {v28.16b}, v16.16b
- tbx v2.16b, {v29.16b, v30.16b}, v17.16b
- tbx v5.16b, {v29.16b, v30.16b}, v18.16b
- tbx v6.16b, {v31.16b}, v19.16b
-.else
- add x13, x2, #0x22
- sub w12, w12, w3 /* last_dc_val, not used afterwards */
- ld1 {v23.16b}, [x15]
- add x14, x2, #0x18
- add x3, x2, #0x36
- ins v0.h[0], w12
- add x9, x2, #0x2
- ld1 {v1.h}[0], [x13]
- add x15, x2, #0x30
- ld1 {v2.h}[0], [x14]
- add x19, x2, #0x26
- ld1 {v3.h}[0], [x3]
- add x20, x2, #0x28
- ld1 {v0.h}[1], [x9]
- add x12, x2, #0x10
- ld1 {v1.h}[1], [x15]
- add x13, x2, #0x40
- ld1 {v2.h}[1], [x19]
- add x14, x2, #0x34
- ld1 {v3.h}[1], [x20]
- add x3, x2, #0x1a
- ld1 {v0.h}[2], [x12]
- add x9, x2, #0x20
- ld1 {v1.h}[2], [x13]
- add x15, x2, #0x32
- ld1 {v2.h}[2], [x14]
- add x19, x2, #0x42
- ld1 {v3.h}[2], [x3]
- add x20, x2, #0xc
- ld1 {v0.h}[3], [x9]
- add x12, x2, #0x12
- ld1 {v1.h}[3], [x15]
- add x13, x2, #0x24
- ld1 {v2.h}[3], [x19]
- add x14, x2, #0x50
- ld1 {v3.h}[3], [x20]
- add x3, x2, #0xe
- ld1 {v0.h}[4], [x12]
- add x9, x2, #0x4
- ld1 {v1.h}[4], [x13]
- add x15, x2, #0x16
- ld1 {v2.h}[4], [x14]
- add x19, x2, #0x60
- ld1 {v3.h}[4], [x3]
- add x20, x2, #0x1c
- ld1 {v0.h}[5], [x9]
- add x12, x2, #0x6
- ld1 {v1.h}[5], [x15]
- add x13, x2, #0x8
- ld1 {v2.h}[5], [x19]
- add x14, x2, #0x52
- ld1 {v3.h}[5], [x20]
- add x3, x2, #0x2a
- ld1 {v0.h}[6], [x12]
- add x9, x2, #0x14
- ld1 {v1.h}[6], [x13]
- add x15, x2, #0xa
- ld1 {v2.h}[6], [x14]
- add x19, x2, #0x44
- ld1 {v3.h}[6], [x3]
- add x20, x2, #0x38
- ld1 {v0.h}[7], [x9]
- add x12, x2, #0x46
- ld1 {v1.h}[7], [x15]
- add x13, x2, #0x3a
- ld1 {v2.h}[7], [x19]
- add x14, x2, #0x74
- ld1 {v3.h}[7], [x20]
- add x3, x2, #0x6a
- ld1 {v4.h}[0], [x12]
- add x9, x2, #0x54
- ld1 {v5.h}[0], [x13]
- add x15, x2, #0x2c
- ld1 {v6.h}[0], [x14]
- add x19, x2, #0x76
- ld1 {v7.h}[0], [x3]
- add x20, x2, #0x78
- ld1 {v4.h}[1], [x9]
- add x12, x2, #0x62
- ld1 {v5.h}[1], [x15]
- add x13, x2, #0x1e
- ld1 {v6.h}[1], [x19]
- add x14, x2, #0x68
- ld1 {v7.h}[1], [x20]
- add x3, x2, #0x7a
- ld1 {v4.h}[2], [x12]
- add x9, x2, #0x70
- ld1 {v5.h}[2], [x13]
- add x15, x2, #0x2e
- ld1 {v6.h}[2], [x14]
- add x19, x2, #0x5a
- ld1 {v7.h}[2], [x3]
- add x20, x2, #0x6c
- ld1 {v4.h}[3], [x9]
- add x12, x2, #0x72
- ld1 {v5.h}[3], [x15]
- add x13, x2, #0x3c
- ld1 {v6.h}[3], [x19]
- add x14, x2, #0x4c
- ld1 {v7.h}[3], [x20]
- add x3, x2, #0x5e
- ld1 {v4.h}[4], [x12]
- add x9, x2, #0x64
- ld1 {v5.h}[4], [x13]
- add x15, x2, #0x4a
- ld1 {v6.h}[4], [x14]
- add x19, x2, #0x3e
- ld1 {v7.h}[4], [x3]
- add x20, x2, #0x6e
- ld1 {v4.h}[5], [x9]
- add x12, x2, #0x56
- ld1 {v5.h}[5], [x15]
- add x13, x2, #0x58
- ld1 {v6.h}[5], [x19]
- add x14, x2, #0x4e
- ld1 {v7.h}[5], [x20]
- add x3, x2, #0x7c
- ld1 {v4.h}[6], [x12]
- add x9, x2, #0x48
- ld1 {v5.h}[6], [x13]
- add x15, x2, #0x66
- ld1 {v6.h}[6], [x14]
- add x19, x2, #0x5c
- ld1 {v7.h}[6], [x3]
- add x20, x2, #0x7e
- ld1 {v4.h}[7], [x9]
- ld1 {v5.h}[7], [x15]
- ld1 {v6.h}[7], [x19]
- ld1 {v7.h}[7], [x20]
-.endif
- cmlt v24.8h, v0.8h, #0
- cmlt v25.8h, v1.8h, #0
- cmlt v26.8h, v2.8h, #0
- cmlt v27.8h, v3.8h, #0
- cmlt v28.8h, v4.8h, #0
- cmlt v29.8h, v5.8h, #0
- cmlt v30.8h, v6.8h, #0
- cmlt v31.8h, v7.8h, #0
- abs v0.8h, v0.8h
- abs v1.8h, v1.8h
- abs v2.8h, v2.8h
- abs v3.8h, v3.8h
- abs v4.8h, v4.8h
- abs v5.8h, v5.8h
- abs v6.8h, v6.8h
- abs v7.8h, v7.8h
- eor v24.16b, v24.16b, v0.16b
- eor v25.16b, v25.16b, v1.16b
- eor v26.16b, v26.16b, v2.16b
- eor v27.16b, v27.16b, v3.16b
- eor v28.16b, v28.16b, v4.16b
- eor v29.16b, v29.16b, v5.16b
- eor v30.16b, v30.16b, v6.16b
- eor v31.16b, v31.16b, v7.16b
- cmeq v16.8h, v0.8h, #0
- cmeq v17.8h, v1.8h, #0
- cmeq v18.8h, v2.8h, #0
- cmeq v19.8h, v3.8h, #0
- cmeq v20.8h, v4.8h, #0
- cmeq v21.8h, v5.8h, #0
- cmeq v22.8h, v6.8h, #0
- xtn v16.8b, v16.8h
- xtn v18.8b, v18.8h
- xtn v20.8b, v20.8h
- xtn v22.8b, v22.8h
- umov w14, v0.h[0]
- xtn2 v16.16b, v17.8h
- umov w13, v24.h[0]
- xtn2 v18.16b, v19.8h
- clz w14, w14
- xtn2 v20.16b, v21.8h
- lsl w13, w13, w14
- cmeq v17.8h, v7.8h, #0
- sub w12, w14, #32
- xtn2 v22.16b, v17.8h
- lsr w13, w13, w14
- and v16.16b, v16.16b, v23.16b
- neg w12, w12
- and v18.16b, v18.16b, v23.16b
- add x3, x4, #0x400 /* r1 = dctbl->ehufsi */
- and v20.16b, v20.16b, v23.16b
- add x15, sp, #0x90 /* x15 = t2 */
- and v22.16b, v22.16b, v23.16b
- ldr w10, [x4, x12, lsl #2]
- addp v16.16b, v16.16b, v18.16b
- ldrb w11, [x3, x12]
- addp v20.16b, v20.16b, v22.16b
- checkbuf47
- addp v16.16b, v16.16b, v20.16b
- put_bits x10, x11
- addp v16.16b, v16.16b, v18.16b
- checkbuf47
- umov x9, v16.D[0]
- put_bits x13, x12
- cnt v17.8b, v16.8b
- mvn x9, x9
- addv B18, v17.8b
- add x4, x5, #0x400 /* x4 = actbl->ehufsi */
- umov w12, v18.b[0]
- lsr x9, x9, #0x1 /* clear AC coeff */
- ldr w13, [x5, #0x3c0] /* x13 = actbl->ehufco[0xf0] */
- rbit x9, x9 /* x9 = index0 */
- ldrb w14, [x4, #0xf0] /* x14 = actbl->ehufsi[0xf0] */
- cmp w12, #(64-8)
- add x11, sp, #16
- b.lt 4f
- cbz x9, 6f
- st1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x11], #64
- st1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x11], #64
- st1 {v24.8h, v25.8h, v26.8h, v27.8h}, [x11], #64
- st1 {v28.8h, v29.8h, v30.8h, v31.8h}, [x11], #64
-1:
- clz x2, x9
- add x15, x15, x2, lsl #1
- lsl x9, x9, x2
- ldrh w20, [x15, #-126]
-2:
- cmp x2, #0x10
- b.lt 3f
- sub x2, x2, #0x10
- checkbuf47
- put_bits x13, x14
- b 2b
-3:
- clz w20, w20
- ldrh w3, [x15, #2]!
- sub w11, w20, #32
- lsl w3, w3, w20
- neg w11, w11
- lsr w3, w3, w20
- add x2, x11, x2, lsl #4
- lsl x9, x9, #0x1
- ldr w12, [x5, x2, lsl #2]
- ldrb w10, [x4, x2]
- checkbuf31
- put_bits x12, x10
- put_bits x3, x11
- cbnz x9, 1b
- b 6f
-4:
- movi v21.8h, #0x0010
- clz v0.8h, v0.8h
- clz v1.8h, v1.8h
- clz v2.8h, v2.8h
- clz v3.8h, v3.8h
- clz v4.8h, v4.8h
- clz v5.8h, v5.8h
- clz v6.8h, v6.8h
- clz v7.8h, v7.8h
- ushl v24.8h, v24.8h, v0.8h
- ushl v25.8h, v25.8h, v1.8h
- ushl v26.8h, v26.8h, v2.8h
- ushl v27.8h, v27.8h, v3.8h
- ushl v28.8h, v28.8h, v4.8h
- ushl v29.8h, v29.8h, v5.8h
- ushl v30.8h, v30.8h, v6.8h
- ushl v31.8h, v31.8h, v7.8h
- neg v0.8h, v0.8h
- neg v1.8h, v1.8h
- neg v2.8h, v2.8h
- neg v3.8h, v3.8h
- neg v4.8h, v4.8h
- neg v5.8h, v5.8h
- neg v6.8h, v6.8h
- neg v7.8h, v7.8h
- ushl v24.8h, v24.8h, v0.8h
- ushl v25.8h, v25.8h, v1.8h
- ushl v26.8h, v26.8h, v2.8h
- ushl v27.8h, v27.8h, v3.8h
- ushl v28.8h, v28.8h, v4.8h
- ushl v29.8h, v29.8h, v5.8h
- ushl v30.8h, v30.8h, v6.8h
- ushl v31.8h, v31.8h, v7.8h
- add v0.8h, v21.8h, v0.8h
- add v1.8h, v21.8h, v1.8h
- add v2.8h, v21.8h, v2.8h
- add v3.8h, v21.8h, v3.8h
- add v4.8h, v21.8h, v4.8h
- add v5.8h, v21.8h, v5.8h
- add v6.8h, v21.8h, v6.8h
- add v7.8h, v21.8h, v7.8h
- st1 {v0.8h, v1.8h, v2.8h, v3.8h}, [x11], #64
- st1 {v4.8h, v5.8h, v6.8h, v7.8h}, [x11], #64
- st1 {v24.8h, v25.8h, v26.8h, v27.8h}, [x11], #64
- st1 {v28.8h, v29.8h, v30.8h, v31.8h}, [x11], #64
-1:
- clz x2, x9
- add x15, x15, x2, lsl #1
- lsl x9, x9, x2
- ldrh w11, [x15, #-126]
-2:
- cmp x2, #0x10
- b.lt 3f
- sub x2, x2, #0x10
- checkbuf47
- put_bits x13, x14
- b 2b
-3:
- ldrh w3, [x15, #2]!
- add x2, x11, x2, lsl #4
- lsl x9, x9, #0x1
- ldr w12, [x5, x2, lsl #2]
- ldrb w10, [x4, x2]
- checkbuf31
- put_bits x12, x10
- put_bits x3, x11
- cbnz x9, 1b
-6:
- add x13, sp, #0x10e
- cmp x15, x13
- b.hs 1f
- ldr w12, [x5]
- ldrb w14, [x4]
- checkbuf47
- put_bits x12, x14
-1:
- str PUT_BUFFER, [x0, #0x10]
- str PUT_BITSw, [x0, #0x18]
- ldp x19, x20, [sp], 16
- add x0, BUFFER, #0x1
- add sp, sp, 256
- br x30
-
-.endm
-
-generate_jsimd_huff_encode_one_block 1
-generate_jsimd_huff_encode_one_block 0
-
- .unreq BUFFER
- .unreq PUT_BUFFER
- .unreq PUT_BITS
- .unreq PUT_BITSw
-
-.purgem emit_byte
-.purgem put_bits
-.purgem checkbuf31
-.purgem checkbuf47
diff --git a/simd/arm/common/jdcolext-neon.c b/simd/arm/common/jdcolext-neon.c
deleted file mode 100644
index b201792..0000000
--- a/simd/arm/common/jdcolext-neon.c
+++ /dev/null
@@ -1,330 +0,0 @@
-/*
- * jdcolext-neon.c - colorspace conversion (Arm NEON)
- *
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
- *
- * This software is provided 'as-is', without any express or implied
- * warranty. In no event will the authors be held liable for any damages
- * arising from the use of this software.
- *
- * Permission is granted to anyone to use this software for any purpose,
- * including commercial applications, and to alter it and redistribute it
- * freely, subject to the following restrictions:
- *
- * 1. The origin of this software must not be misrepresented; you must not
- * claim that you wrote the original software. If you use this software
- * in a product, an acknowledgment in the product documentation would be
- * appreciated but is not required.
- * 2. Altered source versions must be plainly marked as such, and must not be
- * misrepresented as being the original software.
- * 3. This notice may not be removed or altered from any source distribution.
- */
-
-/* This file is included by jdcolor-neon.c. */
-
-/*
- * YCbCr -> RGB conversion is defined by the following equations:
- * R = Y + 1.40200 * (Cr - 128)
- * G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
- * B = Y + 1.77200 * (Cb - 128)
- *
- * Scaled integer constants are used to avoid floating-point arithmetic:
- * 0.3441467 = 11277 * 2^-15
- * 0.7141418 = 23401 * 2^-15
- * 1.4020386 = 22971 * 2^-14
- * 1.7720337 = 29033 * 2^-14
- * These constants are defined in jdcolor-neon.c.
- *
- * Rounding is used when descaling to ensure correct results.
- */
-
-/*
- * Notes on safe memory access for YCbCr -> RGB conversion routines:
- *
- * Input memory buffers can be safely overread up to the next multiple of
- * ALIGN_SIZE bytes since they are always allocated by alloc_sarray() in
- * jmemmgr.c.
- *
- * The output buffer cannot safely be written beyond output_width since the
- * TurboJPEG API permits it to be allocated with or without padding up to the
- * next multiple of ALIGN_SIZE bytes.
- */
-
-void jsimd_ycc_rgb_convert_neon(JDIMENSION output_width,
- JSAMPIMAGE input_buf,
- JDIMENSION input_row,
- JSAMPARRAY output_buf,
- int num_rows)
-{
- JSAMPROW outptr;
- /* Pointers to Y, Cb and Cr data. */
- JSAMPROW inptr0, inptr1, inptr2;
-
- const int16x8_t neg_128 = vdupq_n_s16(-128);
-
- while (--num_rows >= 0) {
- inptr0 = input_buf[0][input_row];
- inptr1 = input_buf[1][input_row];
- inptr2 = input_buf[2][input_row];
- input_row++;
- outptr = *output_buf++;
- int cols_remaining = output_width;
- for (; cols_remaining >= 16; cols_remaining -= 16) {
- uint8x16_t y = vld1q_u8(inptr0);
- uint8x16_t cb = vld1q_u8(inptr1);
- uint8x16_t cr = vld1q_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128_l = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), vget_low_u8(cr)));
- int16x8_t cr_128_h = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), vget_high_u8(cr)));
- int16x8_t cb_128_l = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), vget_low_u8(cb)));
- int16x8_t cb_128_h = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), vget_high_u8(cb)));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_ll = vmull_n_s16(vget_low_s16(cb_128_l), -F_0_344);
- int32x4_t g_sub_y_lh = vmull_n_s16(vget_high_s16(cb_128_l), -F_0_344);
- int32x4_t g_sub_y_hl = vmull_n_s16(vget_low_s16(cb_128_h), -F_0_344);
- int32x4_t g_sub_y_hh = vmull_n_s16(vget_high_s16(cb_128_h), -F_0_344);
- g_sub_y_ll = vmlsl_n_s16(g_sub_y_ll, vget_low_s16(cr_128_l), F_0_714);
- g_sub_y_lh = vmlsl_n_s16(g_sub_y_lh, vget_high_s16(cr_128_l), F_0_714);
- g_sub_y_hl = vmlsl_n_s16(g_sub_y_hl, vget_low_s16(cr_128_h), F_0_714);
- g_sub_y_hh = vmlsl_n_s16(g_sub_y_hh, vget_high_s16(cr_128_h), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y_l = vcombine_s16(vrshrn_n_s32(g_sub_y_ll, 15),
- vrshrn_n_s32(g_sub_y_lh, 15));
- int16x8_t g_sub_y_h = vcombine_s16(vrshrn_n_s32(g_sub_y_hl, 15),
- vrshrn_n_s32(g_sub_y_hh, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y_l = vqrdmulhq_n_s16(vshlq_n_s16(cr_128_l, 1), F_1_402);
- int16x8_t r_sub_y_h = vqrdmulhq_n_s16(vshlq_n_s16(cr_128_h, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y_l = vqrdmulhq_n_s16(vshlq_n_s16(cb_128_l, 1), F_1_772);
- int16x8_t b_sub_y_h = vqrdmulhq_n_s16(vshlq_n_s16(cb_128_h, 1), F_1_772);
- /* Add Y. */
- int16x8_t r_l = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y_l), vget_low_u8(y)));
- int16x8_t r_h = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y_h), vget_high_u8(y)));
- int16x8_t b_l = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y_l), vget_low_u8(y)));
- int16x8_t b_h = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y_h), vget_high_u8(y)));
- int16x8_t g_l = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y_l), vget_low_u8(y)));
- int16x8_t g_h = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y_h), vget_high_u8(y)));
-
-#if RGB_PIXELSIZE == 4
- uint8x16x4_t rgba;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgba.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
- rgba.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
- rgba.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
- /* Set alpha channel to opaque (0xFF). */
- rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- vst4q_u8(outptr, rgba);
-#elif RGB_PIXELSIZE == 3
- uint8x16x3_t rgb;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgb.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
- rgb.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
- rgb.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
- /* Store RGB pixel data to memory. */
- vst3q_u8(outptr, rgb);
-#else /* RGB565 */
- /* Pack R, G and B values in ratio 5:6:5. */
- uint16x8_t rgb565_l = vqshluq_n_s16(r_l, 8);
- rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(g_l, 8), 5);
- rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(b_l, 8), 11);
- uint16x8_t rgb565_h = vqshluq_n_s16(r_h, 8);
- rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(g_h, 8), 5);
- rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(b_h, 8), 11);
- /* Store RGB pixel data to memory. */
- vst1q_u16((uint16_t *)outptr, rgb565_l);
- vst1q_u16(((uint16_t *)outptr) + 8, rgb565_h);
-#endif /* RGB565 */
-
- /* Increment pointers. */
- inptr0 += 16;
- inptr1 += 16;
- inptr2 += 16;
- outptr += (RGB_PIXELSIZE * 16);
- }
-
- if (cols_remaining >= 8) {
- uint8x8_t y = vld1_u8(inptr0);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y. */
- int16x8_t r = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
- int16x8_t b = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
- int16x8_t g = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
-
-#if RGB_PIXELSIZE == 4
- uint8x8x4_t rgba;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgba.val[RGB_RED] = vqmovun_s16(r);
- rgba.val[RGB_GREEN] = vqmovun_s16(g);
- rgba.val[RGB_BLUE] = vqmovun_s16(b);
- /* Set alpha channel to opaque (0xFF). */
- rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- vst4_u8(outptr, rgba);
-#elif RGB_PIXELSIZE == 3
- uint8x8x3_t rgb;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgb.val[RGB_RED] = vqmovun_s16(r);
- rgb.val[RGB_GREEN] = vqmovun_s16(g);
- rgb.val[RGB_BLUE] = vqmovun_s16(b);
- /* Store RGB pixel data to memory. */
- vst3_u8(outptr, rgb);
-#else /* RGB565 */
- /* Pack R, G and B values in ratio 5:6:5. */
- uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
- rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
- rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
- /* Store RGB pixel data to memory. */
- vst1q_u16((uint16_t *)outptr, rgb565);
-#endif /* RGB565 */
-
- /* Increment pointers. */
- inptr0 += 8;
- inptr1 += 8;
- inptr2 += 8;
- outptr += (RGB_PIXELSIZE * 8);
- cols_remaining -= 8;
- }
-
- /* Handle the tail elements. */
- if (cols_remaining > 0) {
- uint8x8_t y = vld1_u8(inptr0);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y. */
- int16x8_t r = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
- int16x8_t b = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
- int16x8_t g = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
-
-#if RGB_PIXELSIZE == 4
- uint8x8x4_t rgba;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgba.val[RGB_RED] = vqmovun_s16(r);
- rgba.val[RGB_GREEN] = vqmovun_s16(g);
- rgba.val[RGB_BLUE] = vqmovun_s16(b);
- /* Set alpha channel to opaque (0xFF). */
- rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- switch (cols_remaining) {
- case 7 :
- vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba, 6);
- case 6 :
- vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba, 5);
- case 5 :
- vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba, 4);
- case 4 :
- vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba, 3);
- case 3 :
- vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba, 2);
- case 2 :
- vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba, 1);
- case 1 :
- vst4_lane_u8(outptr, rgba, 0);
- default:
- break;
- }
-#elif RGB_PIXELSIZE == 3
- uint8x8x3_t rgb;
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- rgb.val[RGB_RED] = vqmovun_s16(r);
- rgb.val[RGB_GREEN] = vqmovun_s16(g);
- rgb.val[RGB_BLUE] = vqmovun_s16(b);
- /* Store RGB pixel data to memory. */
- switch (cols_remaining) {
- case 7 :
- vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb, 6);
- case 6 :
- vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb, 5);
- case 5 :
- vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb, 4);
- case 4 :
- vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb, 3);
- case 3 :
- vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb, 2);
- case 2 :
- vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb, 1);
- case 1 :
- vst3_lane_u8(outptr, rgb, 0);
- default:
- break;
- }
-#else /* RGB565 */
- /* Pack R, G and B values in ratio 5:6:5. */
- uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
- rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
- rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
- /* Store RGB565 pixel data to memory. */
- switch (cols_remaining) {
- case 7 :
- vst1q_lane_u16(outptr + 6 * RGB_PIXELSIZE, rgb565, 6);
- case 6 :
- vst1q_lane_u16(outptr + 5 * RGB_PIXELSIZE, rgb565, 5);
- case 5 :
- vst1q_lane_u16(outptr + 4 * RGB_PIXELSIZE, rgb565, 4);
- case 4 :
- vst1q_lane_u16(outptr + 3 * RGB_PIXELSIZE, rgb565, 3);
- case 3 :
- vst1q_lane_u16(outptr + 2 * RGB_PIXELSIZE, rgb565, 2);
- case 2 :
- vst1q_lane_u16(outptr + RGB_PIXELSIZE, rgb565, 1);
- case 1 :
- vst1q_lane_u16(outptr, rgb565, 0);
- default:
- break;
- }
-#endif /* RGB565 */
- }
- }
-}
diff --git a/simd/arm/common/jdmrgext-neon.c b/simd/arm/common/jdmrgext-neon.c
deleted file mode 100644
index 8533d71..0000000
--- a/simd/arm/common/jdmrgext-neon.c
+++ /dev/null
@@ -1,607 +0,0 @@
-/*
- * jdmrgext-neon.c - merged upsampling/color conversion (Arm NEON)
- *
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
- *
- * This software is provided 'as-is', without any express or implied
- * warranty. In no event will the authors be held liable for any damages
- * arising from the use of this software.
- *
- * Permission is granted to anyone to use this software for any purpose,
- * including commercial applications, and to alter it and redistribute it
- * freely, subject to the following restrictions:
- *
- * 1. The origin of this software must not be misrepresented; you must not
- * claim that you wrote the original software. If you use this software
- * in a product, an acknowledgment in the product documentation would be
- * appreciated but is not required.
- * 2. Altered source versions must be plainly marked as such, and must not be
- * misrepresented as being the original software.
- * 3. This notice may not be removed or altered from any source distribution.
- */
-
-/* This file is included by jdmerge-neon.c. */
-
-/*
- * These routines perform simple chroma upsampling - h2v1 or h2v2 - followed by
- * YCbCr -> RGB color conversion all in the same function.
- *
- * As with the standalone functions, YCbCr -> RGB conversion is defined by the
- * following equations:
- * R = Y + 1.40200 * (Cr - 128)
- * G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
- * B = Y + 1.77200 * (Cb - 128)
- *
- * Scaled integer constants are used to avoid floating-point arithmetic:
- * 0.3441467 = 11277 * 2^-15
- * 0.7141418 = 23401 * 2^-15
- * 1.4020386 = 22971 * 2^-14
- * 1.7720337 = 29033 * 2^-14
- * These constants are defined in jdmerge-neon.c.
- *
- * Rounding is used when descaling to ensure correct results.
- */
-
-/*
- * Notes on safe memory access for merged upsampling/YCbCr -> RGB conversion
- * routines:
- *
- * Input memory buffers can be safely overread up to the next multiple of
- * ALIGN_SIZE bytes since they are always allocated by alloc_sarray() in
- * jmemmgr.c.
- *
- * The output buffer cannot safely be written beyond output_width since the
- * TurboJPEG API permits it to be allocated with or without padding up to the
- * next multiple of ALIGN_SIZE bytes.
- */
-
-/*
- * Upsample and color convert from YCbCr -> RGB for the case of 2:1 horizontal.
- */
-
-void jsimd_h2v1_merged_upsample_neon(JDIMENSION output_width,
- JSAMPIMAGE input_buf,
- JDIMENSION in_row_group_ctr,
- JSAMPARRAY output_buf)
-{
- JSAMPROW outptr;
- /* Pointers to Y, Cb and Cr data. */
- JSAMPROW inptr0, inptr1, inptr2;
-
- int16x8_t neg_128 = vdupq_n_s16(-128);
-
- inptr0 = input_buf[0][in_row_group_ctr];
- inptr1 = input_buf[1][in_row_group_ctr];
- inptr2 = input_buf[2][in_row_group_ctr];
- outptr = output_buf[0];
-
- int cols_remaining = output_width;
- for (; cols_remaining >= 16; cols_remaining -= 16) {
- /* Load Y-values such that even pixel indices are in one vector and odd */
- /* pixel indices are in another vector. */
- uint8x8x2_t y = vld2_u8(inptr0);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y and duplicate chroma components; upsampling horizontally. */
- int16x8_t g_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y.val[0]));
- int16x8_t r_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y.val[0]));
- int16x8_t b_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y.val[0]));
- int16x8_t g_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y.val[1]));
- int16x8_t r_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y.val[1]));
- int16x8_t b_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y.val[1]));
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- /* Interleave pixel channel values having odd and even pixel indices. */
- uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
- uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
- uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
-
-#ifdef RGB_ALPHA
- uint8x16x4_t rgba;
- rgba.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
- rgba.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
- rgba.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
- /* Set alpha channel to opaque (0xFF). */
- rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- vst4q_u8(outptr, rgba);
-#else
- uint8x16x3_t rgb;
- rgb.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
- rgb.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
- rgb.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
- /* Store RGB pixel data to memory. */
- vst3q_u8(outptr, rgb);
-#endif
-
- /* Increment pointers. */
- inptr0 += 16;
- inptr1 += 8;
- inptr2 += 8;
- outptr += (RGB_PIXELSIZE * 16);
- }
-
- if (cols_remaining > 0) {
- /* Load y-values such that even pixel indices are in one vector and odd */
- /* pixel indices are in another vector. */
- uint8x8x2_t y = vld2_u8(inptr0);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y and duplicate chroma components - upsample horizontally. */
- int16x8_t g_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y.val[0]));
- int16x8_t r_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y.val[0]));
- int16x8_t b_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y.val[0]));
- int16x8_t g_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y.val[1]));
- int16x8_t r_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y.val[1]));
- int16x8_t b_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y.val[1]));
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- /* Interleave pixel channel values having odd and even pixel indices. */
- uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
- uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
- uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
-
-#ifdef RGB_ALPHA
- uint8x8x4_t rgba_h;
- rgba_h.val[RGB_RED] = r.val[1];
- rgba_h.val[RGB_GREEN] = g.val[1];
- rgba_h.val[RGB_BLUE] = b.val[1];
- /* Set alpha channel to opaque (0xFF). */
- rgba_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- uint8x8x4_t rgba_l;
- rgba_l.val[RGB_RED] = r.val[0];
- rgba_l.val[RGB_GREEN] = g.val[0];
- rgba_l.val[RGB_BLUE] = b.val[0];
- /* Set alpha channel to opaque (0xFF). */
- rgba_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- switch (cols_remaining) {
- case 15 :
- vst4_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgba_h, 6);
- case 14 :
- vst4_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgba_h, 5);
- case 13 :
- vst4_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgba_h, 4);
- case 12 :
- vst4_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgba_h, 3);
- case 11 :
- vst4_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgba_h, 2);
- case 10 :
- vst4_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgba_h, 1);
- case 9 :
- vst4_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgba_h, 0);
- case 8 :
- vst4_u8(outptr, rgba_l);
- break;
- case 7 :
- vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba_l, 6);
- case 6 :
- vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba_l, 5);
- case 5 :
- vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba_l, 4);
- case 4 :
- vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba_l, 3);
- case 3 :
- vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba_l, 2);
- case 2 :
- vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba_l, 1);
- case 1 :
- vst4_lane_u8(outptr, rgba_l, 0);
- default :
- break;
- }
-#else
- uint8x8x3_t rgb_h;
- rgb_h.val[RGB_RED] = r.val[1];
- rgb_h.val[RGB_GREEN] = g.val[1];
- rgb_h.val[RGB_BLUE] = b.val[1];
- uint8x8x3_t rgb_l;
- rgb_l.val[RGB_RED] = r.val[0];
- rgb_l.val[RGB_GREEN] = g.val[0];
- rgb_l.val[RGB_BLUE] = b.val[0];
- /* Store RGB pixel data to memory. */
- switch (cols_remaining) {
- case 15 :
- vst3_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgb_h, 6);
- case 14 :
- vst3_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgb_h, 5);
- case 13 :
- vst3_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgb_h, 4);
- case 12 :
- vst3_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgb_h, 3);
- case 11 :
- vst3_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgb_h, 2);
- case 10 :
- vst3_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgb_h, 1);
- case 9 :
- vst3_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgb_h, 0);
- case 8 :
- vst3_u8(outptr, rgb_l);
- break;
- case 7 :
- vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb_l, 6);
- case 6 :
- vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb_l, 5);
- case 5 :
- vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb_l, 4);
- case 4 :
- vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb_l, 3);
- case 3 :
- vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb_l, 2);
- case 2 :
- vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb_l, 1);
- case 1 :
- vst3_lane_u8(outptr, rgb_l, 0);
- default :
- break;
- }
-#endif
- }
-}
-
-
-/*
- * Upsample and color convert from YCbCr -> RGB for the case of 2:1 horizontal
- * and 2:1 vertical.
- *
- * See above for details of color conversion and safe memory buffer access.
- */
-
-void jsimd_h2v2_merged_upsample_neon(JDIMENSION output_width,
- JSAMPIMAGE input_buf,
- JDIMENSION in_row_group_ctr,
- JSAMPARRAY output_buf)
-{
- JSAMPROW outptr0, outptr1;
- /* Pointers to Y (both rows), Cb and Cr data. */
- JSAMPROW inptr0_0, inptr0_1, inptr1, inptr2;
-
- int16x8_t neg_128 = vdupq_n_s16(-128);
-
- inptr0_0 = input_buf[0][in_row_group_ctr * 2];
- inptr0_1 = input_buf[0][in_row_group_ctr * 2 + 1];
- inptr1 = input_buf[1][in_row_group_ctr];
- inptr2 = input_buf[2][in_row_group_ctr];
- outptr0 = output_buf[0];
- outptr1 = output_buf[1];
-
- int cols_remaining = output_width;
- for (; cols_remaining >= 16; cols_remaining -= 16) {
- /* Load Y-values such that even pixel indices are in one vector and odd */
- /* pixel indices are in another vector. */
- uint8x8x2_t y0 = vld2_u8(inptr0_0);
- uint8x8x2_t y1 = vld2_u8(inptr0_1);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y and duplicate chroma components - upsample horizontally. */
- int16x8_t g0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y0.val[0]));
- int16x8_t r0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y0.val[0]));
- int16x8_t b0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y0.val[0]));
- int16x8_t g0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y0.val[1]));
- int16x8_t r0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y0.val[1]));
- int16x8_t b0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y0.val[1]));
- int16x8_t g1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y1.val[0]));
- int16x8_t r1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y1.val[0]));
- int16x8_t b1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y1.val[0]));
- int16x8_t g1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y1.val[1]));
- int16x8_t r1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y1.val[1]));
- int16x8_t b1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y1.val[1]));
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- /* Interleave pixel channel values having odd and even pixel indices. */
- uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
- uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
- uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
- uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
- uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
- uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
-
-#ifdef RGB_ALPHA
- uint8x16x4_t rgba0, rgba1;
- rgba0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
- rgba1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
- rgba0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
- rgba1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
- rgba0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
- rgba1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
- /* Set alpha channel to opaque (0xFF). */
- rgba0.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
- rgba1.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- vst4q_u8(outptr0, rgba0);
- vst4q_u8(outptr1, rgba1);
-#else
- uint8x16x3_t rgb0, rgb1;
- rgb0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
- rgb1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
- rgb0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
- rgb1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
- rgb0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
- rgb1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
- /* Store RGB pixel data to memory. */
- vst3q_u8(outptr0, rgb0);
- vst3q_u8(outptr1, rgb1);
-#endif
-
- /* Increment pointers. */
- inptr0_0 += 16;
- inptr0_1 += 16;
- inptr1 += 8;
- inptr2 += 8;
- outptr0 += (RGB_PIXELSIZE * 16);
- outptr1 += (RGB_PIXELSIZE * 16);
- }
-
- if (cols_remaining > 0) {
- /* Load Y-values such that even pixel indices are in one vector and */
- /* odd pixel indices are in another vector. */
- uint8x8x2_t y0 = vld2_u8(inptr0_0);
- uint8x8x2_t y1 = vld2_u8(inptr0_1);
- uint8x8_t cb = vld1_u8(inptr1);
- uint8x8_t cr = vld1_u8(inptr2);
- /* Subtract 128 from Cb and Cr. */
- int16x8_t cr_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
- int16x8_t cb_128 = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
- /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
- int32x4_t g_sub_y_l = vmull_n_s16(vget_low_s16(cb_128), -F_0_344);
- int32x4_t g_sub_y_h = vmull_n_s16(vget_high_s16(cb_128), -F_0_344);
- g_sub_y_l = vmlsl_n_s16(g_sub_y_l, vget_low_s16(cr_128), F_0_714);
- g_sub_y_h = vmlsl_n_s16(g_sub_y_h, vget_high_s16(cr_128), F_0_714);
- /* Descale G components: shift right 15, round and narrow to 16-bit. */
- int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
- vrshrn_n_s32(g_sub_y_h, 15));
- /* Compute R-Y: 1.40200 * (Cr - 128) */
- int16x8_t r_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cr_128, 1), F_1_402);
- /* Compute B-Y: 1.77200 * (Cb - 128) */
- int16x8_t b_sub_y = vqrdmulhq_n_s16(vshlq_n_s16(cb_128, 1), F_1_772);
- /* Add Y and duplicate chroma components - upsample horizontally. */
- int16x8_t g0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y0.val[0]));
- int16x8_t r0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y0.val[0]));
- int16x8_t b0_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y0.val[0]));
- int16x8_t g0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y0.val[1]));
- int16x8_t r0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y0.val[1]));
- int16x8_t b0_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y0.val[1]));
- int16x8_t g1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y1.val[0]));
- int16x8_t r1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y1.val[0]));
- int16x8_t b1_even = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y1.val[0]));
- int16x8_t g1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y1.val[1]));
- int16x8_t r1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y1.val[1]));
- int16x8_t b1_odd = vreinterpretq_s16_u16(
- vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y1.val[1]));
- /* Convert each component to unsigned and narrow, clamping to [0-255]. */
- /* Interleave pixel channel values having odd and even pixel indices. */
- uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
- uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
- uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
- uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
- uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
- uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
-
-#ifdef RGB_ALPHA
- uint8x8x4_t rgba0_h, rgba1_h;
- rgba0_h.val[RGB_RED] = r0.val[1];
- rgba1_h.val[RGB_RED] = r1.val[1];
- rgba0_h.val[RGB_GREEN] = g0.val[1];
- rgba1_h.val[RGB_GREEN] = g1.val[1];
- rgba0_h.val[RGB_BLUE] = b0.val[1];
- rgba1_h.val[RGB_BLUE] = b1.val[1];
- /* Set alpha channel to opaque (0xFF). */
- rgba0_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- rgba1_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
-
- uint8x8x4_t rgba0_l, rgba1_l;
- rgba0_l.val[RGB_RED] = r0.val[0];
- rgba1_l.val[RGB_RED] = r1.val[0];
- rgba0_l.val[RGB_GREEN] = g0.val[0];
- rgba1_l.val[RGB_GREEN] = g1.val[0];
- rgba0_l.val[RGB_BLUE] = b0.val[0];
- rgba1_l.val[RGB_BLUE] = b1.val[0];
- /* Set alpha channel to opaque (0xFF). */
- rgba0_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- rgba1_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
- /* Store RGBA pixel data to memory. */
- switch (cols_remaining) {
- case 15 :
- vst4_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgba0_h, 6);
- vst4_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgba1_h, 6);
- case 14 :
- vst4_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgba0_h, 5);
- vst4_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgba1_h, 5);
- case 13 :
- vst4_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgba0_h, 4);
- vst4_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgba1_h, 4);
- case 12 :
- vst4_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgba0_h, 3);
- vst4_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgba1_h, 3);
- case 11 :
- vst4_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgba0_h, 2);
- vst4_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgba1_h, 2);
- case 10 :
- vst4_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgba0_h, 1);
- vst4_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgba1_h, 1);
- case 9 :
- vst4_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgba0_h, 0);
- vst4_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgba1_h, 0);
- case 8 :
- vst4_u8(outptr0, rgba0_l);
- vst4_u8(outptr1, rgba1_l);
- break;
- case 7 :
- vst4_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgba0_l, 6);
- vst4_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgba1_l, 6);
- case 6 :
- vst4_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgba0_l, 5);
- vst4_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgba1_l, 5);
- case 5 :
- vst4_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgba0_l, 4);
- vst4_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgba1_l, 4);
- case 4 :
- vst4_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgba0_l, 3);
- vst4_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgba1_l, 3);
- case 3 :
- vst4_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgba0_l, 2);
- vst4_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgba1_l, 2);
- case 2 :
- vst4_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgba0_l, 1);
- vst4_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgba1_l, 1);
- case 1 :
- vst4_lane_u8(outptr0, rgba0_l, 0);
- vst4_lane_u8(outptr1, rgba1_l, 0);
- default :
- break;
- }
-#else
- uint8x8x3_t rgb0_h, rgb1_h;
- rgb0_h.val[RGB_RED] = r0.val[1];
- rgb1_h.val[RGB_RED] = r1.val[1];
- rgb0_h.val[RGB_GREEN] = g0.val[1];
- rgb1_h.val[RGB_GREEN] = g1.val[1];
- rgb0_h.val[RGB_BLUE] = b0.val[1];
- rgb1_h.val[RGB_BLUE] = b1.val[1];
-
- uint8x8x3_t rgb0_l, rgb1_l;
- rgb0_l.val[RGB_RED] = r0.val[0];
- rgb1_l.val[RGB_RED] = r1.val[0];
- rgb0_l.val[RGB_GREEN] = g0.val[0];
- rgb1_l.val[RGB_GREEN] = g1.val[0];
- rgb0_l.val[RGB_BLUE] = b0.val[0];
- rgb1_l.val[RGB_BLUE] = b1.val[0];
- /* Store RGB pixel data to memory. */
- switch (cols_remaining) {
- case 15 :
- vst3_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgb0_h, 6);
- vst3_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgb1_h, 6);
- case 14 :
- vst3_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgb0_h, 5);
- vst3_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgb1_h, 5);
- case 13 :
- vst3_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgb0_h, 4);
- vst3_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgb1_h, 4);
- case 12 :
- vst3_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgb0_h, 3);
- vst3_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgb1_h, 3);
- case 11 :
- vst3_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgb0_h, 2);
- vst3_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgb1_h, 2);
- case 10 :
- vst3_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgb0_h, 1);
- vst3_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgb1_h, 1);
- case 9 :
- vst3_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgb0_h, 0);
- vst3_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgb1_h, 0);
- case 8 :
- vst3_u8(outptr0, rgb0_l);
- vst3_u8(outptr1, rgb1_l);
- break;
- case 7 :
- vst3_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgb0_l, 6);
- vst3_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgb1_l, 6);
- case 6 :
- vst3_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgb0_l, 5);
- vst3_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgb1_l, 5);
- case 5 :
- vst3_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgb0_l, 4);
- vst3_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgb1_l, 4);
- case 4 :
- vst3_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgb0_l, 3);
- vst3_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgb1_l, 3);
- case 3 :
- vst3_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgb0_l, 2);
- vst3_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgb1_l, 2);
- case 2 :
- vst3_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgb0_l, 1);
- vst3_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgb1_l, 1);
- case 1 :
- vst3_lane_u8(outptr0, rgb0_l, 0);
- vst3_lane_u8(outptr1, rgb1_l, 0);
- default :
- break;
- }
-#endif
- }
-}
diff --git a/simd/arm/common/jdsample-neon.c b/simd/arm/common/jdsample-neon.c
deleted file mode 100644
index e4f5129..0000000
--- a/simd/arm/common/jdsample-neon.c
+++ /dev/null
@@ -1,557 +0,0 @@
-/*
- * jdsample-neon.c - upsampling (Arm NEON)
- *
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
- *
- * This software is provided 'as-is', without any express or implied
- * warranty. In no event will the authors be held liable for any damages
- * arising from the use of this software.
- *
- * Permission is granted to anyone to use this software for any purpose,
- * including commercial applications, and to alter it and redistribute it
- * freely, subject to the following restrictions:
- *
- * 1. The origin of this software must not be misrepresented; you must not
- * claim that you wrote the original software. If you use this software
- * in a product, an acknowledgment in the product documentation would be
- * appreciated but is not required.
- * 2. Altered source versions must be plainly marked as such, and must not be
- * misrepresented as being the original software.
- * 3. This notice may not be removed or altered from any source distribution.
- */
-
-#define JPEG_INTERNALS
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
-#include "../../jsimd.h"
-
-#include <arm_neon.h>
-
-/*
- * The diagram below shows a row of samples (luma or chroma) produced by h2v1
- * downsampling.
- *
- * s0 s1 s2
- * +---------+---------+---------+
- * | | | |
- * | p0 p1 | p2 p3 | p4 p5 |
- * | | | |
- * +---------+---------+---------+
- *
- * Each sample contains two of the original pixel channel values. These pixel
- * channel values are centred at positions p0, p1, p2, p3, p4 and p5 above. To
- * compute the channel values of the original image, we proportionally blend
- * the adjacent samples in each row.
- *
- * There are three cases to consider:
- *
- * 1) The first pixel in the original image.
- * Pixel channel value p0 contains only a component from sample s0, so we
- * set p0 = s0.
- * 2) The last pixel in the original image.
- * Pixel channel value p5 contains only a component from sample s2, so we
- * set p5 = s2.
- * 3) General case (all other pixels in the row).
- * Apart from the first and last pixels, every other pixel channel value is
- * computed by blending the containing sample and the nearest neigbouring
- * sample in the ratio 3:1.
- * For example, the pixel channel value centred at p1 would be computed as
- * follows:
- * 3/4 * s0 + 1/4 * s1
- * while the pixel channel value centred at p2 would be:
- * 3/4 * s1 + 1/4 * s0
- */
-
-void jsimd_h2v1_fancy_upsample_neon(int max_v_samp_factor,
- JDIMENSION downsampled_width,
- JSAMPARRAY input_data,
- JSAMPARRAY *output_data_ptr)
-{
- JSAMPARRAY output_data = *output_data_ptr;
- JSAMPROW inptr, outptr;
- /* Setup constants. */
- const uint16x8_t one_u16 = vdupq_n_u16(1);
- const uint8x8_t three_u8 = vdup_n_u8(3);
-
- for (int inrow = 0; inrow < max_v_samp_factor; inrow++) {
- inptr = input_data[inrow];
- outptr = output_data[inrow];
- /* Case 1: first pixel channel value in this row of the original image. */
- *outptr = (JSAMPLE)GETJSAMPLE(*inptr);
-
- /* General case: */
- /* 3/4 * containing sample + 1/4 * nearest neighbouring sample */
- /* For p1: containing sample = s0, nearest neighbouring sample = s1. */
- /* For p2: containing sample = s1, nearest neighbouring sample = s0. */
- uint8x16_t s0 = vld1q_u8(inptr);
- uint8x16_t s1 = vld1q_u8(inptr + 1);
- /* Multiplication makes vectors twice as wide: '_l' and '_h' suffixes */
- /* denote low half and high half respectively. */
- uint16x8_t s1_add_3s0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1)),
- vget_low_u8(s0), three_u8);
- uint16x8_t s1_add_3s0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1)),
- vget_high_u8(s0), three_u8);
- uint16x8_t s0_add_3s1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0)),
- vget_low_u8(s1), three_u8);
- uint16x8_t s0_add_3s1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0)),
- vget_high_u8(s1), three_u8);
- /* Add ordered dithering bias to odd pixel values. */
- s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
- s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
-
- /* Initially 1 - due to having already stored the first pixel of the */
- /* image. However, in subsequent iterations of the SIMD loop this offset */
- /* is (2 * colctr - 1) to stay within the bounds of the sample buffers */
- /* without having to resort to a slow scalar tail case for the last */
- /* (downsampled_width % 16) samples. See "Creation of 2-D sample arrays" */
- /* in jmemmgr.c for details. */
- unsigned outptr_offset = 1;
- uint8x16x2_t output_pixels;
-
-#if defined(__aarch64__) && defined(__clang__) && !defined(__OPTIMIZE_SIZE__)
- /* Unrolling by four is beneficial on AArch64 as there are 16 additional */
- /* 128-bit SIMD registers to accommodate the extra data in flight. */
- #pragma clang loop unroll_count(4)
-#endif
- /* We use software pipelining to maximise performance. The code indented */
- /* an extra 6 spaces begins the next iteration of the loop. */
- for (unsigned colctr = 16; colctr < downsampled_width; colctr += 16) {
- s0 = vld1q_u8(inptr + colctr - 1);
- s1 = vld1q_u8(inptr + colctr);
- /* Right-shift by 2 (divide by 4), narrow to 8-bit and combine. */
- output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
- vrshrn_n_u16(s1_add_3s0_h, 2));
- output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
- vshrn_n_u16(s0_add_3s1_h, 2));
- /* Multiplication makes vectors twice as wide: '_l' and '_h' */
- /* suffixes denote low half and high half respectively. */
- s1_add_3s0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1)),
- vget_low_u8(s0), three_u8);
- s1_add_3s0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1)),
- vget_high_u8(s0), three_u8);
- s0_add_3s1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0)),
- vget_low_u8(s1), three_u8);
- s0_add_3s1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0)),
- vget_high_u8(s1), three_u8);
- /* Add ordered dithering bias to odd pixel values. */
- s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
- s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
- /* Store pixel channel values to memory. */
- vst2q_u8(outptr + outptr_offset, output_pixels);
- outptr_offset = 2 * colctr - 1;
- }
-
- /* Complete the last iteration of the loop. */
- /* Right-shift by 2 (divide by 4), narrow to 8-bit and combine. */
- output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
- vrshrn_n_u16(s1_add_3s0_h, 2));
- output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
- vshrn_n_u16(s0_add_3s1_h, 2));
- /* Store pixel channel values to memory. */
- vst2q_u8(outptr + outptr_offset, output_pixels);
-
- /* Case 2: last pixel channel value in this row of the original image. */
- outptr[2 * downsampled_width - 1] =
- GETJSAMPLE(inptr[downsampled_width - 1]);
- }
-}
-
-
-/*
- * The diagram below shows a grid-window of samples (luma or chroma) produced
- * by h2v2 downsampling.
- *
- * s0 s1
- * +---------+---------+
- * | p0 p1 | p2 p3 |
- * r0 | | |
- * | p4 p5 | p6 p7 |
- * +---------+---------+
- * | p8 p9 | p10 p11|
- * r1 | | |
- * | p12 p13| p14 p15|
- * +---------+---------+
- * | p16 p17| p18 p19|
- * r2 | | |
- * | p20 p21| p22 p23|
- * +---------+---------+
- *
- * Every sample contains four of the original pixel channel values. The pixels'
- * channel values are centred at positions p0, p1, p2,..., p23 above. For a
- * given grid-window position, r1 is always used to denote the row of samples
- * containing the pixel channel values we are computing. For the top row of
- * pixel channel values in r1 (p8-p11), the nearest neighbouring samples are in
- * the row above - denoted by r0. Likewise, for the bottom row of pixels in r1
- * (p12-p15), the nearest neighbouring samples are in the row below - denoted
- * by r2.
- *
- * To compute the pixel channel values of the original image, we proportionally
- * blend the sample containing the pixel centre with the nearest neighbouring
- * samples in each row, column and diagonal.
- *
- * There are three cases to consider:
- *
- * 1) The first pixel in this row of the original image.
- * Pixel channel value p8 only contains components from sample column s0.
- * Its value is computed by blending samples s0r1 and s0r0 in the ratio 3:1.
- * 2) The last pixel in this row of the original image.
- * Pixel channel value p11 only contains components from sample column s1.
- * Its value is computed by blending samples s1r1 and s1r0 in the ratio 3:1.
- * 3) General case (all other pixels in the row).
- * Apart from the first and last pixels, every other pixel channel value in
- * the row contains components from samples in adjacent columns.
- *
- * For example, the pixel centred at p9 would be computed as follows:
- * (9/16 * s0r1) + (3/16 * s0r0) + (3/16 * s1r1) + (1/16 * s1r0)
- *
- * This can be broken down into two steps:
- * 1) Blend samples vertically in columns s0 and s1 in the ratio 3:1:
- * s0colsum = 3/4 * s0r1 + 1/4 * s0r0
- * s1colsum = 3/4 * s1r1 + 1/4 * s1r0
- * 2) Blend the already-blended columns in the ratio 3:1:
- * p9 = 3/4 * s0colsum + 1/4 * s1colsum
- *
- * The bottom row of pixel channel values in row r1 can be computed in the same
- * way for each of the three cases, only using samples in row r2 instead of row
- * r0 - as r2 is the nearest neighbouring row.
- */
-
-void jsimd_h2v2_fancy_upsample_neon(int max_v_samp_factor,
- JDIMENSION downsampled_width,
- JSAMPARRAY input_data,
- JSAMPARRAY *output_data_ptr)
-{
- JSAMPARRAY output_data = *output_data_ptr;
- JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
- int inrow, outrow;
- /* Setup constants. */
- const uint16x8_t seven_u16 = vdupq_n_u16(7);
- const uint8x8_t three_u8 = vdup_n_u8(3);
- const uint16x8_t three_u16 = vdupq_n_u16(3);
-
- inrow = outrow = 0;
- while (outrow < max_v_samp_factor) {
- inptr0 = input_data[inrow - 1];
- inptr1 = input_data[inrow];
- inptr2 = input_data[inrow + 1];
- /* Suffixes 0 and 1 denote the top and bottom rows of output pixels */
- /* respectively. */
- outptr0 = output_data[outrow++];
- outptr1 = output_data[outrow++];
-
- /* Case 1: first pixel channel value in this row of original image. */
- int s0colsum0 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr0);
- *outptr0 = (JSAMPLE)((s0colsum0 * 4 + 8) >> 4);
- int s0colsum1 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr2);
- *outptr1 = (JSAMPLE)((s0colsum1 * 4 + 8) >> 4);
-
- /* General case as described above. */
- /* Step 1: Blend samples vertically in columns s0 and s1. */
- /* Leave the divide by 4 to the end when it can be done for both */
- /* dimensions at once, right-shifting by 4. */
-
- /* Load and compute s0colsum0 and s0colsum1. */
- uint8x16_t s0r0 = vld1q_u8(inptr0);
- uint8x16_t s0r1 = vld1q_u8(inptr1);
- uint8x16_t s0r2 = vld1q_u8(inptr2);
- /* Multiplication makes vectors twice as wide: '_l' and '_h' suffixes */
- /* denote low half and high half respectively. */
- uint16x8_t s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0r0)),
- vget_low_u8(s0r1), three_u8);
- uint16x8_t s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0r0)),
- vget_high_u8(s0r1), three_u8);
- uint16x8_t s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0r2)),
- vget_low_u8(s0r1), three_u8);
- uint16x8_t s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0r2)),
- vget_high_u8(s0r1), three_u8);
- /* Load and compute s1colsum0 and s1colsum1. */
- uint8x16_t s1r0 = vld1q_u8(inptr0 + 1);
- uint8x16_t s1r1 = vld1q_u8(inptr1 + 1);
- uint8x16_t s1r2 = vld1q_u8(inptr2 + 1);
- uint16x8_t s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1r0)),
- vget_low_u8(s1r1), three_u8);
- uint16x8_t s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1r0)),
- vget_high_u8(s1r1), three_u8);
- uint16x8_t s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1r2)),
- vget_low_u8(s1r1), three_u8);
- uint16x8_t s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1r2)),
- vget_high_u8(s1r1), three_u8);
- /* Step 2: Blend the already-blended columns. */
- uint16x8_t output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
- uint16x8_t output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
- uint16x8_t output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
- uint16x8_t output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
- uint16x8_t output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
- uint16x8_t output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
- uint16x8_t output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
- uint16x8_t output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
- /* Add ordered dithering bias to odd pixel values. */
- output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
- output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
- output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
- output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
- /* Right-shift by 4 (divide by 16), narrow to 8-bit and combine. */
- uint8x16x2_t output_pixels0 = { vcombine_u8(vshrn_n_u16(output0_p1_l, 4),
- vshrn_n_u16(output0_p1_h, 4)),
- vcombine_u8(vrshrn_n_u16(output0_p2_l, 4),
- vrshrn_n_u16(output0_p2_h, 4))
- };
- uint8x16x2_t output_pixels1 = { vcombine_u8(vshrn_n_u16(output1_p1_l, 4),
- vshrn_n_u16(output1_p1_h, 4)),
- vcombine_u8(vrshrn_n_u16(output1_p2_l, 4),
- vrshrn_n_u16(output1_p2_h, 4))
- };
- /* Store pixel channel values to memory. */
- /* The minimum size of the output buffer for each row is 64 bytes => no */
- /* need to worry about buffer overflow here. See "Creation of 2-D sample */
- /* arrays" in jmemmgr.c for details. */
- vst2q_u8(outptr0 + 1, output_pixels0);
- vst2q_u8(outptr1 + 1, output_pixels1);
-
- /* The first pixel of the image shifted our loads and stores by one */
- /* byte. We have to re-align on a 32-byte boundary at some point before */
- /* the end of the row (we do it now on the 32/33 pixel boundary) to stay */
- /* within the bounds of the sample buffers without having to resort to a */
- /* slow scalar tail case for the last (downsampled_width % 16) samples. */
- /* See "Creation of 2-D sample arrays" in jmemmgr.c for details.*/
- for (unsigned colctr = 16; colctr < downsampled_width; colctr += 16) {
- /* Step 1: Blend samples vertically in columns s0 and s1. */
- /* Load and compute s0colsum0 and s0colsum1. */
- s0r0 = vld1q_u8(inptr0 + colctr - 1);
- s0r1 = vld1q_u8(inptr1 + colctr - 1);
- s0r2 = vld1q_u8(inptr2 + colctr - 1);
- s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0r0)),
- vget_low_u8(s0r1), three_u8);
- s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0r0)),
- vget_high_u8(s0r1), three_u8);
- s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0r2)),
- vget_low_u8(s0r1), three_u8);
- s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0r2)),
- vget_high_u8(s0r1), three_u8);
- /* Load and compute s1colsum0 and s1colsum1. */
- s1r0 = vld1q_u8(inptr0 + colctr);
- s1r1 = vld1q_u8(inptr1 + colctr);
- s1r2 = vld1q_u8(inptr2 + colctr);
- s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1r0)),
- vget_low_u8(s1r1), three_u8);
- s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1r0)),
- vget_high_u8(s1r1), three_u8);
- s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1r2)),
- vget_low_u8(s1r1), three_u8);
- s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1r2)),
- vget_high_u8(s1r1), three_u8);
- /* Step 2: Blend the already-blended columns. */
- output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
- output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
- output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
- output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
- output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
- output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
- output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
- output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
- /* Add ordered dithering bias to odd pixel values. */
- output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
- output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
- output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
- output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
- /* Right-shift by 4 (divide by 16), narrow to 8-bit and combine. */
- output_pixels0.val[0] = vcombine_u8(vshrn_n_u16(output0_p1_l, 4),
- vshrn_n_u16(output0_p1_h, 4));
- output_pixels0.val[1] = vcombine_u8(vrshrn_n_u16(output0_p2_l, 4),
- vrshrn_n_u16(output0_p2_h, 4));
- output_pixels1.val[0] = vcombine_u8(vshrn_n_u16(output1_p1_l, 4),
- vshrn_n_u16(output1_p1_h, 4));
- output_pixels1.val[1] = vcombine_u8(vrshrn_n_u16(output1_p2_l, 4),
- vrshrn_n_u16(output1_p2_h, 4));
- /* Store pixel channel values to memory. */
- vst2q_u8(outptr0 + 2 * colctr - 1, output_pixels0);
- vst2q_u8(outptr1 + 2 * colctr - 1, output_pixels1);
- }
-
- /* Case 2: last pixel channel value in this row of the original image. */
- int s1colsum0 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
- GETJSAMPLE(inptr0[downsampled_width - 1]);
- outptr0[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum0 * 4 + 7) >> 4);
- int s1colsum1 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
- GETJSAMPLE(inptr2[downsampled_width - 1]);
- outptr1[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum1 * 4 + 7) >> 4);
- inrow++;
- }
-}
-
-
-/*
- * The diagram below shows a grid-window of samples (luma or chroma) produced
- * by h2v1 downsampling; which has been subsequently rotated 90 degrees. (The
- * usual use of h1v2 upsampling is upsampling rotated or transposed h2v1
- * downsampled images.)
- *
- * s0 s1
- * +---------+---------+
- * | p0 | p1 |
- * r0 | | |
- * | p2 | p3 |
- * +---------+---------+
- * | p4 | p5 |
- * r1 | | |
- * | p6 | p7 |
- * +---------+---------+
- * | p8 | p9 |
- * r2 | | |
- * | p10 | p11 |
- * +---------+---------+
- *
- * Every sample contains two of the original pixel channel values. The pixels'
- * channel values are centred at positions p0, p1, p2,..., p11 above. For a
- * given grid-window position, r1 is always used to denote the row of samples
- * containing the pixel channel values we are computing. For the top row of
- * pixel channel values in r1 (p4 and p5), the nearest neighbouring samples are
- * in the row above - denoted by r0. Likewise, for the bottom row of pixels in
- * r1 (p6 and p7), the nearest neighbouring samples are in the row below -
- * denoted by r2.
- *
- * To compute the pixel channel values of the original image, we proportionally
- * blend the adjacent samples in each column.
- *
- * For example, the pixel channel value centred at p4 would be computed as
- * follows:
- * 3/4 * s0r1 + 1/4 * s0r0
- * while the pixel channel value centred at p6 would be:
- * 3/4 * s0r1 + 1/4 * s0r2
- */
-
-void jsimd_h1v2_fancy_upsample_neon(int max_v_samp_factor,
- JDIMENSION downsampled_width,
- JSAMPARRAY input_data,
- JSAMPARRAY *output_data_ptr)
-{
- JSAMPARRAY output_data = *output_data_ptr;
- JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
- int inrow, outrow;
- /* Setup constants. */
- const uint16x8_t one_u16 = vdupq_n_u16(1);
- const uint8x8_t three_u8 = vdup_n_u8(3);
-
- inrow = outrow = 0;
- while (outrow < max_v_samp_factor) {
- inptr0 = input_data[inrow - 1];
- inptr1 = input_data[inrow];
- inptr2 = input_data[inrow + 1];
- /* Suffixes 0 and 1 denote the top and bottom rows of output pixels */
- /* respectively. */
- outptr0 = output_data[outrow++];
- outptr1 = output_data[outrow++];
- inrow++;
-
- /* The size of the input and output buffers is always a multiple of 32 */
- /* bytes => no need to worry about buffer overflow when reading/writing */
- /* memory. See "Creation of 2-D sample arrays" in jmemmgr.c for details. */
- for (unsigned colctr = 0; colctr < downsampled_width; colctr += 16) {
- /* Load samples. */
- uint8x16_t r0 = vld1q_u8(inptr0 + colctr);
- uint8x16_t r1 = vld1q_u8(inptr1 + colctr);
- uint8x16_t r2 = vld1q_u8(inptr2 + colctr);
- /* Blend samples vertically. */
- uint16x8_t colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(r0)),
- vget_low_u8(r1), three_u8);
- uint16x8_t colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(r0)),
- vget_high_u8(r1), three_u8);
- uint16x8_t colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(r2)),
- vget_low_u8(r1), three_u8);
- uint16x8_t colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(r2)),
- vget_high_u8(r1), three_u8);
- /* Add ordered dithering bias to pixel values in even output rows. */
- colsum0_l = vaddq_u16(colsum0_l, one_u16);
- colsum0_h = vaddq_u16(colsum0_h, one_u16);
- /* Right-shift by 2 (divide by 4), narrow to 8-bit and combine. */
- uint8x16_t output_pixels0 = vcombine_u8(vshrn_n_u16(colsum0_l, 2),
- vshrn_n_u16(colsum0_h, 2));
- uint8x16_t output_pixels1 = vcombine_u8(vrshrn_n_u16(colsum1_l, 2),
- vrshrn_n_u16(colsum1_h, 2));
- /* Store pixel channel values to memory. */
- vst1q_u8(outptr0 + colctr, output_pixels0);
- vst1q_u8(outptr1 + colctr, output_pixels1);
- }
- }
-}
-
-
-/*
- * The diagram below shows the operation of h2v1 (simple) upsampling. Each
- * sample in the row is duplicated to form two output pixel channel values.
- *
- * p0 p1 p2 p3
- * +----+----+ +----+----+----+----+
- * | s0 | s1 | -> | s0 | s0 | s1 | s1 |
- * +----+----+ +----+----+----+----+
- */
-
-void jsimd_h2v1_upsample_neon(int max_v_samp_factor,
- JDIMENSION output_width,
- JSAMPARRAY input_data,
- JSAMPARRAY *output_data_ptr)
-{
- JSAMPARRAY output_data = *output_data_ptr;
- JSAMPROW inptr, outptr;
-
- for (int inrow = 0; inrow < max_v_samp_factor; inrow++) {
- inptr = input_data[inrow];
- outptr = output_data[inrow];
- for (unsigned colctr = 0; 2 * colctr < output_width; colctr += 16) {
- uint8x16_t samples = vld1q_u8(inptr + colctr);
- /* Duplicate the samples - the store interleaves them to produce the */
- /* pattern in the diagram above. */
- uint8x16x2_t output_pixels = { samples, samples };
- /* Store pixel values to memory. */
- /* Due to the way sample buffers are allocated, we don't need to worry */
- /* about tail cases when output_width is not a multiple of 32. */
- /* See "Creation of 2-D sample arrays" in jmemmgr.c for details. */
- vst2q_u8(outptr + 2 * colctr, output_pixels);
- }
- }
-}
-
-
-/*
- * The diagram below shows the operation of h2v2 (simple) upsampling. Each
- * sample in the row is duplicated to form two output pixel channel values.
- * This horizontally-upsampled row is then also duplicated.
- *
- * p0 p1 p2 p3
- * +-----+-----+ +-----+-----+-----+-----+
- * | s0 | s1 | -> | s0 | s0 | s1 | s1 |
- * +-----+-----+ +-----+-----+-----+-----+
- * | s0 | s0 | s1 | s1 |
- * +-----+-----+-----+-----+
- */
-
-void jsimd_h2v2_upsample_neon(int max_v_samp_factor,
- JDIMENSION output_width,
- JSAMPARRAY input_data,
- JSAMPARRAY *output_data_ptr)
-{
- JSAMPARRAY output_data = *output_data_ptr;
- JSAMPROW inptr, outptr0, outptr1;
-
- for (int inrow = 0, outrow = 0; outrow < max_v_samp_factor; inrow++) {
- inptr = input_data[inrow];
- outptr0 = output_data[outrow++];
- outptr1 = output_data[outrow++];
-
- for (unsigned colctr = 0; 2 * colctr < output_width; colctr += 16) {
- uint8x16_t samples = vld1q_u8(inptr + colctr);
- /* Duplicate the samples - the store interleaves them to produce the */
- /* pattern in the diagram above. */
- uint8x16x2_t output_pixels = { samples, samples };
- /* Store pixel values to memory for both output rows. */
- /* Due to the way sample buffers are allocated, we don't need to worry */
- /* about tail cases when output_width is not a multiple of 32. */
- /* See "Creation of 2-D sample arrays" in jmemmgr.c for details. */
- vst2q_u8(outptr0 + 2 * colctr, output_pixels);
- vst2q_u8(outptr1 + 2 * colctr, output_pixels);
- }
- }
-}
diff --git a/simd/arm/common/jccolor-neon.c b/simd/arm/jccolor-neon.c
similarity index 72%
rename from simd/arm/common/jccolor-neon.c
rename to simd/arm/jccolor-neon.c
index f87c8d9..9fcc62d 100644
--- a/simd/arm/common/jccolor-neon.c
+++ b/simd/arm/jccolor-neon.c
@@ -1,7 +1,8 @@
/*
* jccolor-neon.c - colorspace conversion (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,17 +22,19 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
#include <arm_neon.h>
-/* RGB -> YCbCr conversion constants. */
+
+/* RGB -> YCbCr conversion constants */
#define F_0_298 19595
#define F_0_587 38470
@@ -43,18 +46,17 @@
#define F_0_081 5329
ALIGN(16) static const uint16_t jsimd_rgb_ycc_neon_consts[] = {
- F_0_298, F_0_587,
- F_0_113, F_0_168,
- F_0_331, F_0_500,
- F_0_418, F_0_081
- };
+ F_0_298, F_0_587, F_0_113, F_0_168,
+ F_0_331, F_0_500, F_0_418, F_0_081
+};
+
/* Include inline routines for colorspace extensions. */
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -66,10 +68,10 @@
#define RGB_BLUE EXT_RGB_BLUE
#define RGB_PIXELSIZE EXT_RGB_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extrgb_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -82,10 +84,10 @@
#define RGB_BLUE EXT_RGBX_BLUE
#define RGB_PIXELSIZE EXT_RGBX_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extrgbx_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -98,10 +100,10 @@
#define RGB_BLUE EXT_BGR_BLUE
#define RGB_PIXELSIZE EXT_BGR_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extbgr_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -114,10 +116,10 @@
#define RGB_BLUE EXT_BGRX_BLUE
#define RGB_PIXELSIZE EXT_BGRX_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extbgrx_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -130,10 +132,10 @@
#define RGB_BLUE EXT_XBGR_BLUE
#define RGB_PIXELSIZE EXT_XBGR_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extxbgr_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
@@ -146,10 +148,10 @@
#define RGB_BLUE EXT_XRGB_BLUE
#define RGB_PIXELSIZE EXT_XRGB_PIXELSIZE
#define jsimd_rgb_ycc_convert_neon jsimd_extxrgb_ycc_convert_neon
-#if defined(__aarch64__)
-#include "../arm64/jccolext-neon.c"
+#if defined(__aarch64__) || defined(_M_ARM64)
+#include "aarch64/jccolext-neon.c"
#else
-#include "../arm/jccolext-neon.c"
+#include "aarch32/jccolext-neon.c"
#endif
#undef RGB_RED
#undef RGB_GREEN
diff --git a/simd/arm/common/jcgray-neon.c b/simd/arm/jcgray-neon.c
similarity index 89%
rename from simd/arm/common/jcgray-neon.c
rename to simd/arm/jcgray-neon.c
index 39d903f..71c7b2d 100644
--- a/simd/arm/common/jcgray-neon.c
+++ b/simd/arm/jcgray-neon.c
@@ -1,7 +1,7 @@
/*
- * jcgray-neon.c - grayscale colorspace conversion (Arm NEON)
+ * jcgray-neon.c - grayscale colorspace conversion (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,22 +21,24 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
-/* RGB -> Grayscale conversion constants. */
+
+/* RGB -> Grayscale conversion constants */
#define F_0_298 19595
#define F_0_587 38470
#define F_0_113 7471
+
/* Include inline routines for colorspace extensions. */
#include "jcgryext-neon.c"
diff --git a/simd/arm/common/jcgryext-neon.c b/simd/arm/jcgryext-neon.c
similarity index 81%
rename from simd/arm/common/jcgryext-neon.c
rename to simd/arm/jcgryext-neon.c
index 69ea67f..416a738 100644
--- a/simd/arm/common/jcgryext-neon.c
+++ b/simd/arm/jcgryext-neon.c
@@ -1,7 +1,7 @@
/*
- * jcgryext-neon.c - grayscale colorspace conversion (Arm NEON)
+ * jcgryext-neon.c - grayscale colorspace conversion (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -22,8 +22,8 @@
/* This file is included by jcgray-neon.c */
-/*
- * RGB -> Grayscale conversion is defined by the following equation:
+
+/* RGB -> Grayscale conversion is defined by the following equation:
* Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
*
* Avoid floating point arithmetic by using shifted integer constants:
@@ -32,19 +32,17 @@
* 0.11399841 = 7471 * 2^-16
* These constants are defined in jcgray-neon.c
*
- * We use rounding later to get correct values.
- *
* This is the same computation as the RGB -> Y portion of RGB -> YCbCr.
*/
-void jsimd_rgb_gray_convert_neon(JDIMENSION image_width,
- JSAMPARRAY input_buf,
- JSAMPIMAGE output_buf,
- JDIMENSION output_row,
+void jsimd_rgb_gray_convert_neon(JDIMENSION image_width, JSAMPARRAY input_buf,
+ JSAMPIMAGE output_buf, JDIMENSION output_row,
int num_rows)
{
JSAMPROW inptr;
JSAMPROW outptr;
+ /* Allocate temporary buffer for final (image_width % 16) pixels in row. */
+ ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
while (--num_rows >= 0) {
inptr = *input_buf++;
@@ -54,11 +52,11 @@
int cols_remaining = image_width;
for (; cols_remaining > 0; cols_remaining -= 16) {
- /* To prevent buffer overread by the vector load instructions, the */
- /* last (image_width % 16) columns of data are first memcopied to a */
- /* temporary buffer large enough to accommodate the vector load. */
+ /* To prevent buffer overread by the vector load instructions, the last
+ * (image_width % 16) columns of data are first memcopied to a temporary
+ * buffer large enough to accommodate the vector load.
+ */
if (cols_remaining < 16) {
- ALIGN(16) uint8_t tmp_buf[16 * RGB_PIXELSIZE];
memcpy(tmp_buf, inptr, cols_remaining * RGB_PIXELSIZE);
inptr = tmp_buf;
}
@@ -95,8 +93,9 @@
uint16x8_t y_h = vcombine_u16(vrshrn_n_u32(y_hl, 16),
vrshrn_n_u32(y_hh, 16));
- /* Narrow Y values to 8-bit and store to memory. Buffer overwrite is */
- /* permitted up to the next multiple of ALIGN_SIZE bytes. */
+ /* Narrow Y values to 8-bit and store to memory. Buffer overwrite is
+ * permitted up to the next multiple of ALIGN_SIZE bytes.
+ */
vst1q_u8(outptr, vcombine_u8(vmovn_u16(y_l), vmovn_u16(y_h)));
/* Increment pointers. */
diff --git a/simd/arm/jchuff.h b/simd/arm/jchuff.h
new file mode 100644
index 0000000..d30759f
--- /dev/null
+++ b/simd/arm/jchuff.h
@@ -0,0 +1,149 @@
+/*
+ * jchuff.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * libjpeg-turbo Modifications:
+ * Copyright (C) 2009, 2018, D. R. Commander.
+ * Copyright (C) 2018, Matthias Räncker.
+ * Copyright (C) 2020, Arm Limited.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define BIT_BUF_SIZE 64
+#else
+#define BIT_BUF_SIZE 32
+#endif
+
+typedef struct {
+ size_t put_buffer; /* current bit accumulation buffer */
+ int free_bits; /* # of bits available in it */
+ int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+typedef struct {
+ JOCTET *next_output_byte; /* => next byte to write in buffer */
+ size_t free_in_buffer; /* # of byte spaces remaining in buffer */
+ savable_state cur; /* Current bit buffer & DC state */
+ j_compress_ptr cinfo; /* dump_buffer needs access to this */
+ int simd;
+} working_state;
+
+/* Outputting bits to the file */
+
+/* Output byte b and, speculatively, an additional 0 byte. 0xFF must be encoded
+ * as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the byte is
+ * 0xFF. Otherwise, the output buffer pointer is advanced by 1, and the
+ * speculative 0 byte will be overwritten by the next byte.
+ */
+#define EMIT_BYTE(b) { \
+ buffer[0] = (JOCTET)(b); \
+ buffer[1] = 0; \
+ buffer -= -2 + ((JOCTET)(b) < 0xFF); \
+}
+
+/* Output the entire bit buffer. If there are no 0xFF bytes in it, then write
+ * directly to the output buffer. Otherwise, use the EMIT_BYTE() macro to
+ * encode 0xFF as 0xFF 0x00.
+ */
+#if defined(__aarch64__) || defined(_M_ARM64)
+
+#if defined(_MSC_VER) && !defined(__clang__)
+#define SPLAT() { \
+ buffer[0] = (JOCTET)(put_buffer >> 56); \
+ buffer[1] = (JOCTET)(put_buffer >> 48); \
+ buffer[2] = (JOCTET)(put_buffer >> 40); \
+ buffer[3] = (JOCTET)(put_buffer >> 32); \
+ buffer[4] = (JOCTET)(put_buffer >> 24); \
+ buffer[5] = (JOCTET)(put_buffer >> 16); \
+ buffer[6] = (JOCTET)(put_buffer >> 8); \
+ buffer[7] = (JOCTET)(put_buffer ); \
+}
+#else
+#define SPLAT() { \
+ __asm__("rev %x0, %x1" : "=r"(put_buffer) : "r"(put_buffer)); \
+ *((uint64_t *)buffer) = put_buffer; \
+}
+#endif
+
+#define FLUSH() { \
+ if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \
+ EMIT_BYTE(put_buffer >> 56) \
+ EMIT_BYTE(put_buffer >> 48) \
+ EMIT_BYTE(put_buffer >> 40) \
+ EMIT_BYTE(put_buffer >> 32) \
+ EMIT_BYTE(put_buffer >> 24) \
+ EMIT_BYTE(put_buffer >> 16) \
+ EMIT_BYTE(put_buffer >> 8) \
+ EMIT_BYTE(put_buffer ) \
+ } else { \
+ SPLAT() \
+ buffer += 8; \
+ } \
+}
+
+#else
+
+#if defined(_MSC_VER) && !defined(__clang__)
+#define SPLAT() { \
+ buffer[0] = (JOCTET)(put_buffer >> 24); \
+ buffer[1] = (JOCTET)(put_buffer >> 16); \
+ buffer[2] = (JOCTET)(put_buffer >> 8); \
+ buffer[3] = (JOCTET)(put_buffer ); \
+}
+#else
+#define SPLAT() { \
+ __asm__("rev %0, %1" : "=r"(put_buffer) : "r"(put_buffer)); \
+ *((uint32_t *)buffer) = put_buffer; \
+}
+#endif
+
+#define FLUSH() { \
+ if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \
+ EMIT_BYTE(put_buffer >> 24) \
+ EMIT_BYTE(put_buffer >> 16) \
+ EMIT_BYTE(put_buffer >> 8) \
+ EMIT_BYTE(put_buffer ) \
+ } else { \
+ SPLAT() \
+ buffer += 4; \
+ } \
+}
+
+#endif
+
+/* Fill the bit buffer to capacity with the leading bits from code, then output
+ * the bit buffer and put the remaining bits from code into the bit buffer.
+ */
+#define PUT_AND_FLUSH(code, size) { \
+ put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \
+ FLUSH() \
+ free_bits += BIT_BUF_SIZE; \
+ put_buffer = code; \
+}
+
+/* Insert code into the bit buffer and output the bit buffer if needed.
+ * NOTE: We can't flush with free_bits == 0, since the left shift in
+ * PUT_AND_FLUSH() would have undefined behavior.
+ */
+#define PUT_BITS(code, size) { \
+ free_bits -= size; \
+ if (free_bits < 0) \
+ PUT_AND_FLUSH(code, size) \
+ else \
+ put_buffer = (put_buffer << size) | code; \
+}
+
+#define PUT_CODE(code, size, diff) { \
+ diff |= code << nbits; \
+ nbits += size; \
+ PUT_BITS(diff, nbits) \
+}
diff --git a/simd/arm/jcphuff-neon.c b/simd/arm/jcphuff-neon.c
new file mode 100644
index 0000000..8b6d53b
--- /dev/null
+++ b/simd/arm/jcphuff-neon.c
@@ -0,0 +1,591 @@
+/*
+ * jcphuff-neon.c - prepare data for progressive Huffman encoding (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "neon-compat.h"
+
+#include <arm_neon.h>
+
+
+/* Data preparation for encode_mcu_AC_first().
+ *
+ * The equivalent scalar C function (encode_mcu_AC_first_prepare()) can be
+ * found in jcphuff.c.
+ */
+
+void jsimd_encode_mcu_AC_first_prepare_neon
+ (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+ JCOEF *values, size_t *zerobits)
+{
+ JCOEF *values_ptr = values;
+ JCOEF *diff_values_ptr = values + DCTSIZE2;
+
+ /* Rows of coefficients to zero (since they haven't been processed) */
+ int i, rows_to_zero = 8;
+
+ for (i = 0; i < Sl / 16; i++) {
+ int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+ int16x8_t coefs2 = vld1q_dup_s16(block + jpeg_natural_order_start[8]);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[15], coefs2, 7);
+
+ /* Isolate sign of coefficients. */
+ int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);
+ int16x8_t sign_coefs2 = vshrq_n_s16(coefs2, 15);
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+ int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+ coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+ coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+
+ /* Compute diff values. */
+ int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);
+ int16x8_t diff2 = veorq_s16(coefs2, sign_coefs2);
+
+ /* Store transformed coefficients and diff values. */
+ vst1q_s16(values_ptr, coefs1);
+ vst1q_s16(values_ptr + DCTSIZE, coefs2);
+ vst1q_s16(diff_values_ptr, diff1);
+ vst1q_s16(diff_values_ptr + DCTSIZE, diff2);
+ values_ptr += 16;
+ diff_values_ptr += 16;
+ jpeg_natural_order_start += 16;
+ rows_to_zero -= 2;
+ }
+
+ /* Same operation but for remaining partial vector */
+ int remaining_coefs = Sl % 16;
+ if (remaining_coefs > 8) {
+ int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+ int16x8_t coefs2 = vdupq_n_s16(0);
+ switch (remaining_coefs) {
+ case 15:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+ case 14:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+ case 13:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+ case 12:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+ case 11:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+ case 10:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+ case 9:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[8], coefs2, 0);
+ default:
+ break;
+ }
+
+ /* Isolate sign of coefficients. */
+ int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);
+ int16x8_t sign_coefs2 = vshrq_n_s16(coefs2, 15);
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+ int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+ coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+ coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+
+ /* Compute diff values. */
+ int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);
+ int16x8_t diff2 = veorq_s16(coefs2, sign_coefs2);
+
+ /* Store transformed coefficients and diff values. */
+ vst1q_s16(values_ptr, coefs1);
+ vst1q_s16(values_ptr + DCTSIZE, coefs2);
+ vst1q_s16(diff_values_ptr, diff1);
+ vst1q_s16(diff_values_ptr + DCTSIZE, diff2);
+ values_ptr += 16;
+ diff_values_ptr += 16;
+ rows_to_zero -= 2;
+
+ } else if (remaining_coefs > 0) {
+ int16x8_t coefs = vdupq_n_s16(0);
+
+ switch (remaining_coefs) {
+ case 8:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs, 7);
+ case 7:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs, 6);
+ case 6:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs, 5);
+ case 5:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs, 4);
+ case 4:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs, 3);
+ case 3:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs, 2);
+ case 2:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs, 1);
+ case 1:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[0], coefs, 0);
+ default:
+ break;
+ }
+
+ /* Isolate sign of coefficients. */
+ int16x8_t sign_coefs = vshrq_n_s16(coefs, 15);
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs = vabsq_s16(coefs);
+ coefs = vshlq_s16(abs_coefs, vdupq_n_s16(-Al));
+
+ /* Compute diff values. */
+ int16x8_t diff = veorq_s16(coefs, sign_coefs);
+
+ /* Store transformed coefficients and diff values. */
+ vst1q_s16(values_ptr, coefs);
+ vst1q_s16(diff_values_ptr, diff);
+ values_ptr += 8;
+ diff_values_ptr += 8;
+ rows_to_zero--;
+ }
+
+ /* Zero remaining memory in the values and diff_values blocks. */
+ for (i = 0; i < rows_to_zero; i++) {
+ vst1q_s16(values_ptr, vdupq_n_s16(0));
+ vst1q_s16(diff_values_ptr, vdupq_n_s16(0));
+ values_ptr += 8;
+ diff_values_ptr += 8;
+ }
+
+ /* Construct zerobits bitmap. A set bit means that the corresponding
+ * coefficient != 0.
+ */
+ int16x8_t row0 = vld1q_s16(values + 0 * DCTSIZE);
+ int16x8_t row1 = vld1q_s16(values + 1 * DCTSIZE);
+ int16x8_t row2 = vld1q_s16(values + 2 * DCTSIZE);
+ int16x8_t row3 = vld1q_s16(values + 3 * DCTSIZE);
+ int16x8_t row4 = vld1q_s16(values + 4 * DCTSIZE);
+ int16x8_t row5 = vld1q_s16(values + 5 * DCTSIZE);
+ int16x8_t row6 = vld1q_s16(values + 6 * DCTSIZE);
+ int16x8_t row7 = vld1q_s16(values + 7 * DCTSIZE);
+
+ uint8x8_t row0_eq0 = vmovn_u16(vceqq_s16(row0, vdupq_n_s16(0)));
+ uint8x8_t row1_eq0 = vmovn_u16(vceqq_s16(row1, vdupq_n_s16(0)));
+ uint8x8_t row2_eq0 = vmovn_u16(vceqq_s16(row2, vdupq_n_s16(0)));
+ uint8x8_t row3_eq0 = vmovn_u16(vceqq_s16(row3, vdupq_n_s16(0)));
+ uint8x8_t row4_eq0 = vmovn_u16(vceqq_s16(row4, vdupq_n_s16(0)));
+ uint8x8_t row5_eq0 = vmovn_u16(vceqq_s16(row5, vdupq_n_s16(0)));
+ uint8x8_t row6_eq0 = vmovn_u16(vceqq_s16(row6, vdupq_n_s16(0)));
+ uint8x8_t row7_eq0 = vmovn_u16(vceqq_s16(row7, vdupq_n_s16(0)));
+
+ /* { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } */
+ const uint8x8_t bitmap_mask =
+ vreinterpret_u8_u64(vmov_n_u64(0x8040201008040201));
+
+ row0_eq0 = vand_u8(row0_eq0, bitmap_mask);
+ row1_eq0 = vand_u8(row1_eq0, bitmap_mask);
+ row2_eq0 = vand_u8(row2_eq0, bitmap_mask);
+ row3_eq0 = vand_u8(row3_eq0, bitmap_mask);
+ row4_eq0 = vand_u8(row4_eq0, bitmap_mask);
+ row5_eq0 = vand_u8(row5_eq0, bitmap_mask);
+ row6_eq0 = vand_u8(row6_eq0, bitmap_mask);
+ row7_eq0 = vand_u8(row7_eq0, bitmap_mask);
+
+ uint8x8_t bitmap_rows_01 = vpadd_u8(row0_eq0, row1_eq0);
+ uint8x8_t bitmap_rows_23 = vpadd_u8(row2_eq0, row3_eq0);
+ uint8x8_t bitmap_rows_45 = vpadd_u8(row4_eq0, row5_eq0);
+ uint8x8_t bitmap_rows_67 = vpadd_u8(row6_eq0, row7_eq0);
+ uint8x8_t bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+ uint8x8_t bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+ uint8x8_t bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+ /* Move bitmap to a 64-bit scalar register. */
+ uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+ /* Store zerobits bitmap. */
+ *zerobits = ~bitmap;
+#else
+ /* Move bitmap to two 32-bit scalar registers. */
+ uint32_t bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+ uint32_t bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+ /* Store zerobits bitmap. */
+ zerobits[0] = ~bitmap0;
+ zerobits[1] = ~bitmap1;
+#endif
+}
+
+
+/* Data preparation for encode_mcu_AC_refine().
+ *
+ * The equivalent scalar C function (encode_mcu_AC_refine_prepare()) can be
+ * found in jcphuff.c.
+ */
+
+int jsimd_encode_mcu_AC_refine_prepare_neon
+ (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+ JCOEF *absvalues, size_t *bits)
+{
+ /* Temporary storage buffers for data used to compute the signbits bitmap and
+ * the end-of-block (EOB) position
+ */
+ uint8_t coef_sign_bits[64];
+ uint8_t coef_eq1_bits[64];
+
+ JCOEF *absvalues_ptr = absvalues;
+ uint8_t *coef_sign_bits_ptr = coef_sign_bits;
+ uint8_t *eq1_bits_ptr = coef_eq1_bits;
+
+ /* Rows of coefficients to zero (since they haven't been processed) */
+ int i, rows_to_zero = 8;
+
+ for (i = 0; i < Sl / 16; i++) {
+ int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+ int16x8_t coefs2 = vld1q_dup_s16(block + jpeg_natural_order_start[8]);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[15], coefs2, 7);
+
+ /* Compute and store data for signbits bitmap. */
+ uint8x8_t sign_coefs1 =
+ vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));
+ uint8x8_t sign_coefs2 =
+ vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs2, 15)));
+ vst1_u8(coef_sign_bits_ptr, sign_coefs1);
+ vst1_u8(coef_sign_bits_ptr + DCTSIZE, sign_coefs2);
+
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+ int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+ coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+ coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+ vst1q_s16(absvalues_ptr, coefs1);
+ vst1q_s16(absvalues_ptr + DCTSIZE, coefs2);
+
+ /* Test whether transformed coefficient values == 1 (used to find EOB
+ * position.)
+ */
+ uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));
+ uint8x8_t coefs_eq12 = vmovn_u16(vceqq_s16(coefs2, vdupq_n_s16(1)));
+ vst1_u8(eq1_bits_ptr, coefs_eq11);
+ vst1_u8(eq1_bits_ptr + DCTSIZE, coefs_eq12);
+
+ absvalues_ptr += 16;
+ coef_sign_bits_ptr += 16;
+ eq1_bits_ptr += 16;
+ jpeg_natural_order_start += 16;
+ rows_to_zero -= 2;
+ }
+
+ /* Same operation but for remaining partial vector */
+ int remaining_coefs = Sl % 16;
+ if (remaining_coefs > 8) {
+ int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);
+ coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);
+ int16x8_t coefs2 = vdupq_n_s16(0);
+ switch (remaining_coefs) {
+ case 15:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[14], coefs2, 6);
+ case 14:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[13], coefs2, 5);
+ case 13:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[12], coefs2, 4);
+ case 12:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[11], coefs2, 3);
+ case 11:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[10], coefs2, 2);
+ case 10:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[9], coefs2, 1);
+ case 9:
+ coefs2 = vld1q_lane_s16(block + jpeg_natural_order_start[8], coefs2, 0);
+ default:
+ break;
+ }
+
+ /* Compute and store data for signbits bitmap. */
+ uint8x8_t sign_coefs1 =
+ vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));
+ uint8x8_t sign_coefs2 =
+ vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs2, 15)));
+ vst1_u8(coef_sign_bits_ptr, sign_coefs1);
+ vst1_u8(coef_sign_bits_ptr + DCTSIZE, sign_coefs2);
+
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs1 = vabsq_s16(coefs1);
+ int16x8_t abs_coefs2 = vabsq_s16(coefs2);
+ coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));
+ coefs2 = vshlq_s16(abs_coefs2, vdupq_n_s16(-Al));
+ vst1q_s16(absvalues_ptr, coefs1);
+ vst1q_s16(absvalues_ptr + DCTSIZE, coefs2);
+
+ /* Test whether transformed coefficient values == 1 (used to find EOB
+ * position.)
+ */
+ uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));
+ uint8x8_t coefs_eq12 = vmovn_u16(vceqq_s16(coefs2, vdupq_n_s16(1)));
+ vst1_u8(eq1_bits_ptr, coefs_eq11);
+ vst1_u8(eq1_bits_ptr + DCTSIZE, coefs_eq12);
+
+ absvalues_ptr += 16;
+ coef_sign_bits_ptr += 16;
+ eq1_bits_ptr += 16;
+ jpeg_natural_order_start += 16;
+ rows_to_zero -= 2;
+
+ } else if (remaining_coefs > 0) {
+ int16x8_t coefs = vdupq_n_s16(0);
+
+ switch (remaining_coefs) {
+ case 8:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs, 7);
+ case 7:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs, 6);
+ case 6:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs, 5);
+ case 5:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs, 4);
+ case 4:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs, 3);
+ case 3:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs, 2);
+ case 2:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs, 1);
+ case 1:
+ coefs = vld1q_lane_s16(block + jpeg_natural_order_start[0], coefs, 0);
+ default:
+ break;
+ }
+
+ /* Compute and store data for signbits bitmap. */
+ uint8x8_t sign_coefs =
+ vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs, 15)));
+ vst1_u8(coef_sign_bits_ptr, sign_coefs);
+
+ /* Compute absolute value of coefficients and apply point transform Al. */
+ int16x8_t abs_coefs = vabsq_s16(coefs);
+ coefs = vshlq_s16(abs_coefs, vdupq_n_s16(-Al));
+ vst1q_s16(absvalues_ptr, coefs);
+
+ /* Test whether transformed coefficient values == 1 (used to find EOB
+ * position.)
+ */
+ uint8x8_t coefs_eq1 = vmovn_u16(vceqq_s16(coefs, vdupq_n_s16(1)));
+ vst1_u8(eq1_bits_ptr, coefs_eq1);
+
+ absvalues_ptr += 8;
+ coef_sign_bits_ptr += 8;
+ eq1_bits_ptr += 8;
+ rows_to_zero--;
+ }
+
+ /* Zero remaining memory in blocks. */
+ for (i = 0; i < rows_to_zero; i++) {
+ vst1q_s16(absvalues_ptr, vdupq_n_s16(0));
+ vst1_u8(coef_sign_bits_ptr, vdup_n_u8(0));
+ vst1_u8(eq1_bits_ptr, vdup_n_u8(0));
+ absvalues_ptr += 8;
+ coef_sign_bits_ptr += 8;
+ eq1_bits_ptr += 8;
+ }
+
+ /* Construct zerobits bitmap. */
+ int16x8_t abs_row0 = vld1q_s16(absvalues + 0 * DCTSIZE);
+ int16x8_t abs_row1 = vld1q_s16(absvalues + 1 * DCTSIZE);
+ int16x8_t abs_row2 = vld1q_s16(absvalues + 2 * DCTSIZE);
+ int16x8_t abs_row3 = vld1q_s16(absvalues + 3 * DCTSIZE);
+ int16x8_t abs_row4 = vld1q_s16(absvalues + 4 * DCTSIZE);
+ int16x8_t abs_row5 = vld1q_s16(absvalues + 5 * DCTSIZE);
+ int16x8_t abs_row6 = vld1q_s16(absvalues + 6 * DCTSIZE);
+ int16x8_t abs_row7 = vld1q_s16(absvalues + 7 * DCTSIZE);
+
+ uint8x8_t abs_row0_eq0 = vmovn_u16(vceqq_s16(abs_row0, vdupq_n_s16(0)));
+ uint8x8_t abs_row1_eq0 = vmovn_u16(vceqq_s16(abs_row1, vdupq_n_s16(0)));
+ uint8x8_t abs_row2_eq0 = vmovn_u16(vceqq_s16(abs_row2, vdupq_n_s16(0)));
+ uint8x8_t abs_row3_eq0 = vmovn_u16(vceqq_s16(abs_row3, vdupq_n_s16(0)));
+ uint8x8_t abs_row4_eq0 = vmovn_u16(vceqq_s16(abs_row4, vdupq_n_s16(0)));
+ uint8x8_t abs_row5_eq0 = vmovn_u16(vceqq_s16(abs_row5, vdupq_n_s16(0)));
+ uint8x8_t abs_row6_eq0 = vmovn_u16(vceqq_s16(abs_row6, vdupq_n_s16(0)));
+ uint8x8_t abs_row7_eq0 = vmovn_u16(vceqq_s16(abs_row7, vdupq_n_s16(0)));
+
+ /* { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } */
+ const uint8x8_t bitmap_mask =
+ vreinterpret_u8_u64(vmov_n_u64(0x8040201008040201));
+
+ abs_row0_eq0 = vand_u8(abs_row0_eq0, bitmap_mask);
+ abs_row1_eq0 = vand_u8(abs_row1_eq0, bitmap_mask);
+ abs_row2_eq0 = vand_u8(abs_row2_eq0, bitmap_mask);
+ abs_row3_eq0 = vand_u8(abs_row3_eq0, bitmap_mask);
+ abs_row4_eq0 = vand_u8(abs_row4_eq0, bitmap_mask);
+ abs_row5_eq0 = vand_u8(abs_row5_eq0, bitmap_mask);
+ abs_row6_eq0 = vand_u8(abs_row6_eq0, bitmap_mask);
+ abs_row7_eq0 = vand_u8(abs_row7_eq0, bitmap_mask);
+
+ uint8x8_t bitmap_rows_01 = vpadd_u8(abs_row0_eq0, abs_row1_eq0);
+ uint8x8_t bitmap_rows_23 = vpadd_u8(abs_row2_eq0, abs_row3_eq0);
+ uint8x8_t bitmap_rows_45 = vpadd_u8(abs_row4_eq0, abs_row5_eq0);
+ uint8x8_t bitmap_rows_67 = vpadd_u8(abs_row6_eq0, abs_row7_eq0);
+ uint8x8_t bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+ uint8x8_t bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+ uint8x8_t bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+ /* Move bitmap to a 64-bit scalar register. */
+ uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+ /* Store zerobits bitmap. */
+ bits[0] = ~bitmap;
+#else
+ /* Move bitmap to two 32-bit scalar registers. */
+ uint32_t bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+ uint32_t bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+ /* Store zerobits bitmap. */
+ bits[0] = ~bitmap0;
+ bits[1] = ~bitmap1;
+#endif
+
+ /* Construct signbits bitmap. */
+ uint8x8_t signbits_row0 = vld1_u8(coef_sign_bits + 0 * DCTSIZE);
+ uint8x8_t signbits_row1 = vld1_u8(coef_sign_bits + 1 * DCTSIZE);
+ uint8x8_t signbits_row2 = vld1_u8(coef_sign_bits + 2 * DCTSIZE);
+ uint8x8_t signbits_row3 = vld1_u8(coef_sign_bits + 3 * DCTSIZE);
+ uint8x8_t signbits_row4 = vld1_u8(coef_sign_bits + 4 * DCTSIZE);
+ uint8x8_t signbits_row5 = vld1_u8(coef_sign_bits + 5 * DCTSIZE);
+ uint8x8_t signbits_row6 = vld1_u8(coef_sign_bits + 6 * DCTSIZE);
+ uint8x8_t signbits_row7 = vld1_u8(coef_sign_bits + 7 * DCTSIZE);
+
+ signbits_row0 = vand_u8(signbits_row0, bitmap_mask);
+ signbits_row1 = vand_u8(signbits_row1, bitmap_mask);
+ signbits_row2 = vand_u8(signbits_row2, bitmap_mask);
+ signbits_row3 = vand_u8(signbits_row3, bitmap_mask);
+ signbits_row4 = vand_u8(signbits_row4, bitmap_mask);
+ signbits_row5 = vand_u8(signbits_row5, bitmap_mask);
+ signbits_row6 = vand_u8(signbits_row6, bitmap_mask);
+ signbits_row7 = vand_u8(signbits_row7, bitmap_mask);
+
+ bitmap_rows_01 = vpadd_u8(signbits_row0, signbits_row1);
+ bitmap_rows_23 = vpadd_u8(signbits_row2, signbits_row3);
+ bitmap_rows_45 = vpadd_u8(signbits_row4, signbits_row5);
+ bitmap_rows_67 = vpadd_u8(signbits_row6, signbits_row7);
+ bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+ bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+ bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+ /* Move bitmap to a 64-bit scalar register. */
+ bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+ /* Store signbits bitmap. */
+ bits[1] = ~bitmap;
+#else
+ /* Move bitmap to two 32-bit scalar registers. */
+ bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+ bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+ /* Store signbits bitmap. */
+ bits[2] = ~bitmap0;
+ bits[3] = ~bitmap1;
+#endif
+
+ /* Construct bitmap to find EOB position (the index of the last coefficient
+ * equal to 1.)
+ */
+ uint8x8_t row0_eq1 = vld1_u8(coef_eq1_bits + 0 * DCTSIZE);
+ uint8x8_t row1_eq1 = vld1_u8(coef_eq1_bits + 1 * DCTSIZE);
+ uint8x8_t row2_eq1 = vld1_u8(coef_eq1_bits + 2 * DCTSIZE);
+ uint8x8_t row3_eq1 = vld1_u8(coef_eq1_bits + 3 * DCTSIZE);
+ uint8x8_t row4_eq1 = vld1_u8(coef_eq1_bits + 4 * DCTSIZE);
+ uint8x8_t row5_eq1 = vld1_u8(coef_eq1_bits + 5 * DCTSIZE);
+ uint8x8_t row6_eq1 = vld1_u8(coef_eq1_bits + 6 * DCTSIZE);
+ uint8x8_t row7_eq1 = vld1_u8(coef_eq1_bits + 7 * DCTSIZE);
+
+ row0_eq1 = vand_u8(row0_eq1, bitmap_mask);
+ row1_eq1 = vand_u8(row1_eq1, bitmap_mask);
+ row2_eq1 = vand_u8(row2_eq1, bitmap_mask);
+ row3_eq1 = vand_u8(row3_eq1, bitmap_mask);
+ row4_eq1 = vand_u8(row4_eq1, bitmap_mask);
+ row5_eq1 = vand_u8(row5_eq1, bitmap_mask);
+ row6_eq1 = vand_u8(row6_eq1, bitmap_mask);
+ row7_eq1 = vand_u8(row7_eq1, bitmap_mask);
+
+ bitmap_rows_01 = vpadd_u8(row0_eq1, row1_eq1);
+ bitmap_rows_23 = vpadd_u8(row2_eq1, row3_eq1);
+ bitmap_rows_45 = vpadd_u8(row4_eq1, row5_eq1);
+ bitmap_rows_67 = vpadd_u8(row6_eq1, row7_eq1);
+ bitmap_rows_0123 = vpadd_u8(bitmap_rows_01, bitmap_rows_23);
+ bitmap_rows_4567 = vpadd_u8(bitmap_rows_45, bitmap_rows_67);
+ bitmap_all = vpadd_u8(bitmap_rows_0123, bitmap_rows_4567);
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+ /* Move bitmap to a 64-bit scalar register. */
+ bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);
+
+ /* Return EOB position. */
+ if (bitmap == 0) {
+ /* EOB position is defined to be 0 if all coefficients != 1. */
+ return 0;
+ } else {
+ return 63 - BUILTIN_CLZL(bitmap);
+ }
+#else
+ /* Move bitmap to two 32-bit scalar registers. */
+ bitmap0 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 0);
+ bitmap1 = vget_lane_u32(vreinterpret_u32_u8(bitmap_all), 1);
+
+ /* Return EOB position. */
+ if (bitmap0 == 0 && bitmap1 == 0) {
+ return 0;
+ } else if (bitmap1 != 0) {
+ return 63 - BUILTIN_CLZ(bitmap1);
+ } else {
+ return 31 - BUILTIN_CLZ(bitmap0);
+ }
+#endif
+}
diff --git a/simd/arm/common/jcsample-neon.c b/simd/arm/jcsample-neon.c
similarity index 69%
rename from simd/arm/common/jcsample-neon.c
rename to simd/arm/jcsample-neon.c
index a5ddf16..8a3e237 100644
--- a/simd/arm/common/jcsample-neon.c
+++ b/simd/arm/jcsample-neon.c
@@ -1,7 +1,7 @@
/*
- * jcsample-neon.c - downsampling (Arm NEON)
+ * jcsample-neon.c - downsampling (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,13 +21,13 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
@@ -68,35 +68,34 @@
};
-/*
- * Downsample pixel values of a single chroma component i.e. Cb, Cr.
+/* Downsample pixel values of a single component.
* This version handles the common case of 2:1 horizontal and 1:1 vertical,
* without smoothing.
*/
-void jsimd_h2v1_downsample_neon(JDIMENSION image_width,
- int max_v_samp_factor,
+void jsimd_h2v1_downsample_neon(JDIMENSION image_width, int max_v_samp_factor,
JDIMENSION v_samp_factor,
JDIMENSION width_in_blocks,
- JSAMPARRAY input_data,
- JSAMPARRAY output_data)
+ JSAMPARRAY input_data, JSAMPARRAY output_data)
{
JSAMPROW inptr, outptr;
/* Load expansion mask to pad remaining elements of last DCT block. */
const int mask_offset = 16 * ((width_in_blocks * 2 * DCTSIZE) - image_width);
- const uint8x16_t expand_mask = vld1q_u8(
- &jsimd_h2_downsample_consts[mask_offset]);
- /* Load bias pattern alternating every pixel. */
- const uint16x8_t bias = { 0, 1, 0, 1, 0, 1, 0, 1 };
+ const uint8x16_t expand_mask =
+ vld1q_u8(&jsimd_h2_downsample_consts[mask_offset]);
+ /* Load bias pattern (alternating every pixel.) */
+ /* { 0, 1, 0, 1, 0, 1, 0, 1 } */
+ const uint16x8_t bias = vreinterpretq_u16_u32(vdupq_n_u32(0x00010000));
+ unsigned i, outrow;
- for (unsigned outrow = 0; outrow < v_samp_factor; outrow++) {
+ for (outrow = 0; outrow < v_samp_factor; outrow++) {
outptr = output_data[outrow];
inptr = input_data[outrow];
/* Downsample all but the last DCT block of pixels. */
- for (unsigned i = 0; i < width_in_blocks - 1; i++) {
+ for (i = 0; i < width_in_blocks - 1; i++) {
uint8x16_t pixels = vld1q_u8(inptr + i * 2 * DCTSIZE);
- /* Add adjacent pixel values, widen to 16-bit and add bias. */
+ /* Add adjacent pixel values, widen to 16-bit, and add bias. */
uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
/* Divide total by 2 and narrow to 8-bit. */
uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
@@ -106,56 +105,56 @@
/* Load pixels in last DCT block into a table. */
uint8x16_t pixels = vld1q_u8(inptr + (width_in_blocks - 1) * 2 * DCTSIZE);
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(_M_ARM64)
/* Pad the empty elements with the value of the last pixel. */
pixels = vqtbl1q_u8(pixels, expand_mask);
#else
- uint8x8x2_t table = { vget_low_u8(pixels), vget_high_u8(pixels) };
+ uint8x8x2_t table = { { vget_low_u8(pixels), vget_high_u8(pixels) } };
pixels = vcombine_u8(vtbl2_u8(table, vget_low_u8(expand_mask)),
vtbl2_u8(table, vget_high_u8(expand_mask)));
#endif
- /* Add adjacent pixel values, widen to 16-bit and add bias. */
+ /* Add adjacent pixel values, widen to 16-bit, and add bias. */
uint16x8_t samples_u16 = vpadalq_u8(bias, pixels);
- /* Divide total by 2, narrow to 8-bit and store. */
+ /* Divide total by 2, narrow to 8-bit, and store. */
uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 1);
vst1_u8(outptr + (width_in_blocks - 1) * DCTSIZE, samples_u8);
}
}
-/*
- * Downsample pixel values of a single chroma component i.e. Cb, Cr.
+/* Downsample pixel values of a single component.
* This version handles the standard case of 2:1 horizontal and 2:1 vertical,
* without smoothing.
*/
-void jsimd_h2v2_downsample_neon(JDIMENSION image_width,
- int max_v_samp_factor,
+void jsimd_h2v2_downsample_neon(JDIMENSION image_width, int max_v_samp_factor,
JDIMENSION v_samp_factor,
JDIMENSION width_in_blocks,
- JSAMPARRAY input_data,
- JSAMPARRAY output_data)
+ JSAMPARRAY input_data, JSAMPARRAY output_data)
{
JSAMPROW inptr0, inptr1, outptr;
/* Load expansion mask to pad remaining elements of last DCT block. */
const int mask_offset = 16 * ((width_in_blocks * 2 * DCTSIZE) - image_width);
- const uint8x16_t expand_mask = vld1q_u8(
- &jsimd_h2_downsample_consts[mask_offset]);
- /* Load bias pattern alternating every pixel. */
- const uint16x8_t bias = { 1, 2, 1, 2, 1, 2, 1, 2 };
+ const uint8x16_t expand_mask =
+ vld1q_u8(&jsimd_h2_downsample_consts[mask_offset]);
+ /* Load bias pattern (alternating every pixel.) */
+ /* { 1, 2, 1, 2, 1, 2, 1, 2 } */
+ const uint16x8_t bias = vreinterpretq_u16_u32(vdupq_n_u32(0x00020001));
+ unsigned i, outrow;
- for (unsigned outrow = 0; outrow < v_samp_factor; outrow++) {
+ for (outrow = 0; outrow < v_samp_factor; outrow++) {
outptr = output_data[outrow];
inptr0 = input_data[outrow];
inptr1 = input_data[outrow + 1];
/* Downsample all but the last DCT block of pixels. */
- for (unsigned i = 0; i < width_in_blocks - 1; i++) {
+ for (i = 0; i < width_in_blocks - 1; i++) {
uint8x16_t pixels_r0 = vld1q_u8(inptr0 + i * 2 * DCTSIZE);
uint8x16_t pixels_r1 = vld1q_u8(inptr1 + i * 2 * DCTSIZE);
- /* Add adjacent pixel values in row 0, widen to 16-bit and add bias. */
+ /* Add adjacent pixel values in row 0, widen to 16-bit, and add bias. */
uint16x8_t samples_u16 = vpadalq_u8(bias, pixels_r0);
- /* Add adjacent pixel values in row 1, widen to 16-bit and accumulate. */
+ /* Add adjacent pixel values in row 1, widen to 16-bit, and accumulate.
+ */
samples_u16 = vpadalq_u8(samples_u16, pixels_r1);
/* Divide total by 4 and narrow to 8-bit. */
uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 2);
@@ -164,27 +163,29 @@
}
/* Load pixels in last DCT block into a table. */
- uint8x16_t pixels_r0 = vld1q_u8(
- inptr0 + (width_in_blocks - 1) * 2 * DCTSIZE);
- uint8x16_t pixels_r1 = vld1q_u8(
- inptr1 + (width_in_blocks - 1) * 2 * DCTSIZE);
-#if defined(__aarch64__)
+ uint8x16_t pixels_r0 =
+ vld1q_u8(inptr0 + (width_in_blocks - 1) * 2 * DCTSIZE);
+ uint8x16_t pixels_r1 =
+ vld1q_u8(inptr1 + (width_in_blocks - 1) * 2 * DCTSIZE);
+#if defined(__aarch64__) || defined(_M_ARM64)
/* Pad the empty elements with the value of the last pixel. */
pixels_r0 = vqtbl1q_u8(pixels_r0, expand_mask);
pixels_r1 = vqtbl1q_u8(pixels_r1, expand_mask);
#else
- uint8x8x2_t table_r0 = { vget_low_u8(pixels_r0), vget_high_u8(pixels_r0) };
- uint8x8x2_t table_r1 = { vget_low_u8(pixels_r1), vget_high_u8(pixels_r1) };
+ uint8x8x2_t table_r0 =
+ { { vget_low_u8(pixels_r0), vget_high_u8(pixels_r0) } };
+ uint8x8x2_t table_r1 =
+ { { vget_low_u8(pixels_r1), vget_high_u8(pixels_r1) } };
pixels_r0 = vcombine_u8(vtbl2_u8(table_r0, vget_low_u8(expand_mask)),
vtbl2_u8(table_r0, vget_high_u8(expand_mask)));
pixels_r1 = vcombine_u8(vtbl2_u8(table_r1, vget_low_u8(expand_mask)),
vtbl2_u8(table_r1, vget_high_u8(expand_mask)));
#endif
- /* Add adjacent pixel values in row 0, widen to 16-bit and add bias. */
+ /* Add adjacent pixel values in row 0, widen to 16-bit, and add bias. */
uint16x8_t samples_u16 = vpadalq_u8(bias, pixels_r0);
- /* Add adjacent pixel values in row 1, widen to 16-bit and accumulate. */
+ /* Add adjacent pixel values in row 1, widen to 16-bit, and accumulate. */
samples_u16 = vpadalq_u8(samples_u16, pixels_r1);
- /* Divide total by 4, narrow to 8-bit and store. */
+ /* Divide total by 4, narrow to 8-bit, and store. */
uint8x8_t samples_u8 = vshrn_n_u16(samples_u16, 2);
vst1_u8(outptr + (width_in_blocks - 1) * DCTSIZE, samples_u8);
}
diff --git a/simd/arm/jdcolext-neon.c b/simd/arm/jdcolext-neon.c
new file mode 100644
index 0000000..ae440f4
--- /dev/null
+++ b/simd/arm/jdcolext-neon.c
@@ -0,0 +1,353 @@
+/*
+ * jdcolext-neon.c - colorspace conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jdcolor-neon.c. */
+
+
+/* YCbCr -> RGB conversion is defined by the following equations:
+ * R = Y + 1.40200 * (Cr - 128)
+ * G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
+ * B = Y + 1.77200 * (Cb - 128)
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ * 0.3441467 = 11277 * 2^-15
+ * 0.7141418 = 23401 * 2^-15
+ * 1.4020386 = 22971 * 2^-14
+ * 1.7720337 = 29033 * 2^-14
+ * These constants are defined in jdcolor-neon.c.
+ *
+ * To ensure correct results, rounding is used when descaling.
+ */
+
+/* Notes on safe memory access for YCbCr -> RGB conversion routines:
+ *
+ * Input memory buffers can be safely overread up to the next multiple of
+ * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
+ * jmemmgr.c.
+ *
+ * The output buffer cannot safely be written beyond output_width, since
+ * output_buf points to a possibly unpadded row in the decompressed image
+ * buffer allocated by the calling program.
+ */
+
+void jsimd_ycc_rgb_convert_neon(JDIMENSION output_width, JSAMPIMAGE input_buf,
+ JDIMENSION input_row, JSAMPARRAY output_buf,
+ int num_rows)
+{
+ JSAMPROW outptr;
+ /* Pointers to Y, Cb, and Cr data */
+ JSAMPROW inptr0, inptr1, inptr2;
+
+ const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+ const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+ while (--num_rows >= 0) {
+ inptr0 = input_buf[0][input_row];
+ inptr1 = input_buf[1][input_row];
+ inptr2 = input_buf[2][input_row];
+ input_row++;
+ outptr = *output_buf++;
+ int cols_remaining = output_width;
+ for (; cols_remaining >= 16; cols_remaining -= 16) {
+ uint8x16_t y = vld1q_u8(inptr0);
+ uint8x16_t cb = vld1q_u8(inptr1);
+ uint8x16_t cr = vld1q_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128_l =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+ vget_low_u8(cr)));
+ int16x8_t cr_128_h =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+ vget_high_u8(cr)));
+ int16x8_t cb_128_l =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+ vget_low_u8(cb)));
+ int16x8_t cb_128_h =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128),
+ vget_high_u8(cb)));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_ll = vmull_lane_s16(vget_low_s16(cb_128_l), consts, 0);
+ int32x4_t g_sub_y_lh = vmull_lane_s16(vget_high_s16(cb_128_l),
+ consts, 0);
+ int32x4_t g_sub_y_hl = vmull_lane_s16(vget_low_s16(cb_128_h), consts, 0);
+ int32x4_t g_sub_y_hh = vmull_lane_s16(vget_high_s16(cb_128_h),
+ consts, 0);
+ g_sub_y_ll = vmlsl_lane_s16(g_sub_y_ll, vget_low_s16(cr_128_l),
+ consts, 1);
+ g_sub_y_lh = vmlsl_lane_s16(g_sub_y_lh, vget_high_s16(cr_128_l),
+ consts, 1);
+ g_sub_y_hl = vmlsl_lane_s16(g_sub_y_hl, vget_low_s16(cr_128_h),
+ consts, 1);
+ g_sub_y_hh = vmlsl_lane_s16(g_sub_y_hh, vget_high_s16(cr_128_h),
+ consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y_l = vcombine_s16(vrshrn_n_s32(g_sub_y_ll, 15),
+ vrshrn_n_s32(g_sub_y_lh, 15));
+ int16x8_t g_sub_y_h = vcombine_s16(vrshrn_n_s32(g_sub_y_hl, 15),
+ vrshrn_n_s32(g_sub_y_hh, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y_l = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128_l, 1),
+ consts, 2);
+ int16x8_t r_sub_y_h = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128_h, 1),
+ consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y_l = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128_l, 1),
+ consts, 3);
+ int16x8_t b_sub_y_h = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128_h, 1),
+ consts, 3);
+ /* Add Y. */
+ int16x8_t r_l =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y_l),
+ vget_low_u8(y)));
+ int16x8_t r_h =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y_h),
+ vget_high_u8(y)));
+ int16x8_t b_l =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y_l),
+ vget_low_u8(y)));
+ int16x8_t b_h =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y_h),
+ vget_high_u8(y)));
+ int16x8_t g_l =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y_l),
+ vget_low_u8(y)));
+ int16x8_t g_h =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y_h),
+ vget_high_u8(y)));
+
+#if RGB_PIXELSIZE == 4
+ uint8x16x4_t rgba;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgba.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
+ rgba.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
+ rgba.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
+ /* Set alpha channel to opaque (0xFF). */
+ rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ vst4q_u8(outptr, rgba);
+#elif RGB_PIXELSIZE == 3
+ uint8x16x3_t rgb;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgb.val[RGB_RED] = vcombine_u8(vqmovun_s16(r_l), vqmovun_s16(r_h));
+ rgb.val[RGB_GREEN] = vcombine_u8(vqmovun_s16(g_l), vqmovun_s16(g_h));
+ rgb.val[RGB_BLUE] = vcombine_u8(vqmovun_s16(b_l), vqmovun_s16(b_h));
+ /* Store RGB pixel data to memory. */
+ vst3q_u8(outptr, rgb);
+#else
+ /* Pack R, G, and B values in ratio 5:6:5. */
+ uint16x8_t rgb565_l = vqshluq_n_s16(r_l, 8);
+ rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(g_l, 8), 5);
+ rgb565_l = vsriq_n_u16(rgb565_l, vqshluq_n_s16(b_l, 8), 11);
+ uint16x8_t rgb565_h = vqshluq_n_s16(r_h, 8);
+ rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(g_h, 8), 5);
+ rgb565_h = vsriq_n_u16(rgb565_h, vqshluq_n_s16(b_h, 8), 11);
+ /* Store RGB pixel data to memory. */
+ vst1q_u16((uint16_t *)outptr, rgb565_l);
+ vst1q_u16(((uint16_t *)outptr) + 8, rgb565_h);
+#endif
+
+ /* Increment pointers. */
+ inptr0 += 16;
+ inptr1 += 16;
+ inptr2 += 16;
+ outptr += (RGB_PIXELSIZE * 16);
+ }
+
+ if (cols_remaining >= 8) {
+ uint8x8_t y = vld1_u8(inptr0);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1),
+ consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1),
+ consts, 3);
+ /* Add Y. */
+ int16x8_t r =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
+ int16x8_t b =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
+ int16x8_t g =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
+
+#if RGB_PIXELSIZE == 4
+ uint8x8x4_t rgba;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgba.val[RGB_RED] = vqmovun_s16(r);
+ rgba.val[RGB_GREEN] = vqmovun_s16(g);
+ rgba.val[RGB_BLUE] = vqmovun_s16(b);
+ /* Set alpha channel to opaque (0xFF). */
+ rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ vst4_u8(outptr, rgba);
+#elif RGB_PIXELSIZE == 3
+ uint8x8x3_t rgb;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgb.val[RGB_RED] = vqmovun_s16(r);
+ rgb.val[RGB_GREEN] = vqmovun_s16(g);
+ rgb.val[RGB_BLUE] = vqmovun_s16(b);
+ /* Store RGB pixel data to memory. */
+ vst3_u8(outptr, rgb);
+#else
+ /* Pack R, G, and B values in ratio 5:6:5. */
+ uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
+ rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
+ rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
+ /* Store RGB pixel data to memory. */
+ vst1q_u16((uint16_t *)outptr, rgb565);
+#endif
+
+ /* Increment pointers. */
+ inptr0 += 8;
+ inptr1 += 8;
+ inptr2 += 8;
+ outptr += (RGB_PIXELSIZE * 8);
+ cols_remaining -= 8;
+ }
+
+ /* Handle the tail elements. */
+ if (cols_remaining > 0) {
+ uint8x8_t y = vld1_u8(inptr0);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1),
+ consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1),
+ consts, 3);
+ /* Add Y. */
+ int16x8_t r =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y), y));
+ int16x8_t b =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y), y));
+ int16x8_t g =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y), y));
+
+#if RGB_PIXELSIZE == 4
+ uint8x8x4_t rgba;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgba.val[RGB_RED] = vqmovun_s16(r);
+ rgba.val[RGB_GREEN] = vqmovun_s16(g);
+ rgba.val[RGB_BLUE] = vqmovun_s16(b);
+ /* Set alpha channel to opaque (0xFF). */
+ rgba.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ switch (cols_remaining) {
+ case 7:
+ vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba, 6);
+ case 6:
+ vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba, 5);
+ case 5:
+ vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba, 4);
+ case 4:
+ vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba, 3);
+ case 3:
+ vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba, 2);
+ case 2:
+ vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba, 1);
+ case 1:
+ vst4_lane_u8(outptr, rgba, 0);
+ default:
+ break;
+ }
+#elif RGB_PIXELSIZE == 3
+ uint8x8x3_t rgb;
+ /* Convert each component to unsigned and narrow, clamping to [0-255]. */
+ rgb.val[RGB_RED] = vqmovun_s16(r);
+ rgb.val[RGB_GREEN] = vqmovun_s16(g);
+ rgb.val[RGB_BLUE] = vqmovun_s16(b);
+ /* Store RGB pixel data to memory. */
+ switch (cols_remaining) {
+ case 7:
+ vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb, 6);
+ case 6:
+ vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb, 5);
+ case 5:
+ vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb, 4);
+ case 4:
+ vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb, 3);
+ case 3:
+ vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb, 2);
+ case 2:
+ vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb, 1);
+ case 1:
+ vst3_lane_u8(outptr, rgb, 0);
+ default:
+ break;
+ }
+#else
+ /* Pack R, G, and B values in ratio 5:6:5. */
+ uint16x8_t rgb565 = vqshluq_n_s16(r, 8);
+ rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(g, 8), 5);
+ rgb565 = vsriq_n_u16(rgb565, vqshluq_n_s16(b, 8), 11);
+ /* Store RGB565 pixel data to memory. */
+ switch (cols_remaining) {
+ case 7:
+ vst1q_lane_u16((uint16_t *)(outptr + 6 * RGB_PIXELSIZE), rgb565, 6);
+ case 6:
+ vst1q_lane_u16((uint16_t *)(outptr + 5 * RGB_PIXELSIZE), rgb565, 5);
+ case 5:
+ vst1q_lane_u16((uint16_t *)(outptr + 4 * RGB_PIXELSIZE), rgb565, 4);
+ case 4:
+ vst1q_lane_u16((uint16_t *)(outptr + 3 * RGB_PIXELSIZE), rgb565, 3);
+ case 3:
+ vst1q_lane_u16((uint16_t *)(outptr + 2 * RGB_PIXELSIZE), rgb565, 2);
+ case 2:
+ vst1q_lane_u16((uint16_t *)(outptr + RGB_PIXELSIZE), rgb565, 1);
+ case 1:
+ vst1q_lane_u16((uint16_t *)outptr, rgb565, 0);
+ default:
+ break;
+ }
+#endif
+ }
+ }
+}
diff --git a/simd/arm/common/jdcolor-neon.c b/simd/arm/jdcolor-neon.c
similarity index 88%
rename from simd/arm/common/jdcolor-neon.c
rename to simd/arm/jdcolor-neon.c
index 52dab1e..28dbc57 100644
--- a/simd/arm/common/jdcolor-neon.c
+++ b/simd/arm/jdcolor-neon.c
@@ -1,7 +1,7 @@
/*
- * jdcolor-neon.c - colorspace conversion (Arm NEON)
+ * jdcolor-neon.c - colorspace conversion (Arm Neon)
*
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,22 +21,29 @@
*/
#define JPEG_INTERNALS
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
-/* YCbCr -> RGB conversion constants. */
+
+/* YCbCr -> RGB conversion constants */
#define F_0_344 11277 /* 0.3441467 = 11277 * 2^-15 */
#define F_0_714 23401 /* 0.7141418 = 23401 * 2^-15 */
#define F_1_402 22971 /* 1.4020386 = 22971 * 2^-14 */
#define F_1_772 29033 /* 1.7720337 = 29033 * 2^-14 */
+ALIGN(16) static const int16_t jsimd_ycc_rgb_convert_neon_consts[] = {
+ -F_0_344, F_0_714, F_1_402, F_1_772
+};
+
+
/* Include inline routines for colorspace extensions. */
#include "jdcolext-neon.c"
@@ -125,7 +132,7 @@
#undef RGB_PIXELSIZE
#undef jsimd_ycc_rgb_convert_neon
-/* YCbCr -> RGB565 Conversion. */
+/* YCbCr -> RGB565 Conversion */
#define RGB_PIXELSIZE 2
#define jsimd_ycc_rgb_convert_neon jsimd_ycc_rgb565_convert_neon
diff --git a/simd/arm/common/jdmerge-neon.c b/simd/arm/jdmerge-neon.c
similarity index 89%
rename from simd/arm/common/jdmerge-neon.c
rename to simd/arm/jdmerge-neon.c
index 71798c7..18fb9d8 100644
--- a/simd/arm/common/jdmerge-neon.c
+++ b/simd/arm/jdmerge-neon.c
@@ -1,7 +1,7 @@
/*
- * jdmerge-neon.c - merged upsampling/color conversion (Arm NEON)
+ * jdmerge-neon.c - merged upsampling/color conversion (Arm Neon)
*
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,23 +21,30 @@
*/
#define JPEG_INTERNALS
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
-/* YCbCr -> RGB conversion constants. */
+
+/* YCbCr -> RGB conversion constants */
#define F_0_344 11277 /* 0.3441467 = 11277 * 2^-15 */
#define F_0_714 23401 /* 0.7141418 = 23401 * 2^-15 */
#define F_1_402 22971 /* 1.4020386 = 22971 * 2^-14 */
#define F_1_772 29033 /* 1.7720337 = 29033 * 2^-14 */
-/* Include inline routines for colorspace extensions */
+ALIGN(16) static const int16_t jsimd_ycc_rgb_convert_neon_consts[] = {
+ -F_0_344, F_0_714, F_1_402, F_1_772
+};
+
+
+/* Include inline routines for colorspace extensions. */
#include "jdmrgext-neon.c"
#undef RGB_RED
@@ -135,4 +142,3 @@
#undef RGB_ALPHA
#undef RGB_PIXELSIZE
#undef jsimd_h2v1_merged_upsample_neon
-#undef jsimd_h2v2_merged_upsample_neon
diff --git a/simd/arm/jdmrgext-neon.c b/simd/arm/jdmrgext-neon.c
new file mode 100644
index 0000000..fa2ec05
--- /dev/null
+++ b/simd/arm/jdmrgext-neon.c
@@ -0,0 +1,667 @@
+/*
+ * jdmrgext-neon.c - merged upsampling/color conversion (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jdmerge-neon.c. */
+
+
+/* These routines combine simple (non-fancy, i.e. non-smooth) h2v1 or h2v2
+ * chroma upsampling and YCbCr -> RGB color conversion into a single function.
+ *
+ * As with the standalone functions, YCbCr -> RGB conversion is defined by the
+ * following equations:
+ * R = Y + 1.40200 * (Cr - 128)
+ * G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
+ * B = Y + 1.77200 * (Cb - 128)
+ *
+ * Scaled integer constants are used to avoid floating-point arithmetic:
+ * 0.3441467 = 11277 * 2^-15
+ * 0.7141418 = 23401 * 2^-15
+ * 1.4020386 = 22971 * 2^-14
+ * 1.7720337 = 29033 * 2^-14
+ * These constants are defined in jdmerge-neon.c.
+ *
+ * To ensure correct results, rounding is used when descaling.
+ */
+
+/* Notes on safe memory access for merged upsampling/YCbCr -> RGB conversion
+ * routines:
+ *
+ * Input memory buffers can be safely overread up to the next multiple of
+ * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
+ * jmemmgr.c.
+ *
+ * The output buffer cannot safely be written beyond output_width, since
+ * output_buf points to a possibly unpadded row in the decompressed image
+ * buffer allocated by the calling program.
+ */
+
+/* Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+void jsimd_h2v1_merged_upsample_neon(JDIMENSION output_width,
+ JSAMPIMAGE input_buf,
+ JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf)
+{
+ JSAMPROW outptr;
+ /* Pointers to Y, Cb, and Cr data */
+ JSAMPROW inptr0, inptr1, inptr2;
+
+ const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+ const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+ inptr0 = input_buf[0][in_row_group_ctr];
+ inptr1 = input_buf[1][in_row_group_ctr];
+ inptr2 = input_buf[2][in_row_group_ctr];
+ outptr = output_buf[0];
+
+ int cols_remaining = output_width;
+ for (; cols_remaining >= 16; cols_remaining -= 16) {
+ /* De-interleave Y component values into two separate vectors, one
+ * containing the component values with even-numbered indices and one
+ * containing the component values with odd-numbered indices.
+ */
+ uint8x8x2_t y = vld2_u8(inptr0);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+ /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+ * "odd" Y component values. This effectively upsamples the chroma
+ * components horizontally.
+ */
+ int16x8_t g_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y.val[0]));
+ int16x8_t r_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y.val[0]));
+ int16x8_t b_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y.val[0]));
+ int16x8_t g_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y.val[1]));
+ int16x8_t r_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y.val[1]));
+ int16x8_t b_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y.val[1]));
+ /* Convert each component to unsigned and narrow, clamping to [0-255].
+ * Re-interleave the "even" and "odd" component values.
+ */
+ uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+ uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+ uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+ uint8x16x4_t rgba;
+ rgba.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+ rgba.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+ rgba.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+ /* Set alpha channel to opaque (0xFF). */
+ rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ vst4q_u8(outptr, rgba);
+#else
+ uint8x16x3_t rgb;
+ rgb.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
+ rgb.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
+ rgb.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
+ /* Store RGB pixel data to memory. */
+ vst3q_u8(outptr, rgb);
+#endif
+
+ /* Increment pointers. */
+ inptr0 += 16;
+ inptr1 += 8;
+ inptr2 += 8;
+ outptr += (RGB_PIXELSIZE * 16);
+ }
+
+ if (cols_remaining > 0) {
+ /* De-interleave Y component values into two separate vectors, one
+ * containing the component values with even-numbered indices and one
+ * containing the component values with odd-numbered indices.
+ */
+ uint8x8x2_t y = vld2_u8(inptr0);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+ /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
+ * "odd" Y component values. This effectively upsamples the chroma
+ * components horizontally.
+ */
+ int16x8_t g_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y.val[0]));
+ int16x8_t r_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y.val[0]));
+ int16x8_t b_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y.val[0]));
+ int16x8_t g_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y.val[1]));
+ int16x8_t r_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y.val[1]));
+ int16x8_t b_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y.val[1]));
+ /* Convert each component to unsigned and narrow, clamping to [0-255].
+ * Re-interleave the "even" and "odd" component values.
+ */
+ uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
+ uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
+ uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));
+
+#ifdef RGB_ALPHA
+ uint8x8x4_t rgba_h;
+ rgba_h.val[RGB_RED] = r.val[1];
+ rgba_h.val[RGB_GREEN] = g.val[1];
+ rgba_h.val[RGB_BLUE] = b.val[1];
+ /* Set alpha channel to opaque (0xFF). */
+ rgba_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ uint8x8x4_t rgba_l;
+ rgba_l.val[RGB_RED] = r.val[0];
+ rgba_l.val[RGB_GREEN] = g.val[0];
+ rgba_l.val[RGB_BLUE] = b.val[0];
+ /* Set alpha channel to opaque (0xFF). */
+ rgba_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ switch (cols_remaining) {
+ case 15:
+ vst4_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgba_h, 6);
+ case 14:
+ vst4_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgba_h, 5);
+ case 13:
+ vst4_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgba_h, 4);
+ case 12:
+ vst4_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgba_h, 3);
+ case 11:
+ vst4_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgba_h, 2);
+ case 10:
+ vst4_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgba_h, 1);
+ case 9:
+ vst4_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgba_h, 0);
+ case 8:
+ vst4_u8(outptr, rgba_l);
+ break;
+ case 7:
+ vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba_l, 6);
+ case 6:
+ vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba_l, 5);
+ case 5:
+ vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba_l, 4);
+ case 4:
+ vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba_l, 3);
+ case 3:
+ vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba_l, 2);
+ case 2:
+ vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba_l, 1);
+ case 1:
+ vst4_lane_u8(outptr, rgba_l, 0);
+ default:
+ break;
+ }
+#else
+ uint8x8x3_t rgb_h;
+ rgb_h.val[RGB_RED] = r.val[1];
+ rgb_h.val[RGB_GREEN] = g.val[1];
+ rgb_h.val[RGB_BLUE] = b.val[1];
+ uint8x8x3_t rgb_l;
+ rgb_l.val[RGB_RED] = r.val[0];
+ rgb_l.val[RGB_GREEN] = g.val[0];
+ rgb_l.val[RGB_BLUE] = b.val[0];
+ /* Store RGB pixel data to memory. */
+ switch (cols_remaining) {
+ case 15:
+ vst3_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgb_h, 6);
+ case 14:
+ vst3_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgb_h, 5);
+ case 13:
+ vst3_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgb_h, 4);
+ case 12:
+ vst3_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgb_h, 3);
+ case 11:
+ vst3_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgb_h, 2);
+ case 10:
+ vst3_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgb_h, 1);
+ case 9:
+ vst3_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgb_h, 0);
+ case 8:
+ vst3_u8(outptr, rgb_l);
+ break;
+ case 7:
+ vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb_l, 6);
+ case 6:
+ vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb_l, 5);
+ case 5:
+ vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb_l, 4);
+ case 4:
+ vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb_l, 3);
+ case 3:
+ vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb_l, 2);
+ case 2:
+ vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb_l, 1);
+ case 1:
+ vst3_lane_u8(outptr, rgb_l, 0);
+ default:
+ break;
+ }
+#endif
+ }
+}
+
+
+/* Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ *
+ * See comments above for details regarding color conversion and safe memory
+ * access.
+ */
+
+void jsimd_h2v2_merged_upsample_neon(JDIMENSION output_width,
+ JSAMPIMAGE input_buf,
+ JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf)
+{
+ JSAMPROW outptr0, outptr1;
+ /* Pointers to Y (both rows), Cb, and Cr data */
+ JSAMPROW inptr0_0, inptr0_1, inptr1, inptr2;
+
+ const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
+ const int16x8_t neg_128 = vdupq_n_s16(-128);
+
+ inptr0_0 = input_buf[0][in_row_group_ctr * 2];
+ inptr0_1 = input_buf[0][in_row_group_ctr * 2 + 1];
+ inptr1 = input_buf[1][in_row_group_ctr];
+ inptr2 = input_buf[2][in_row_group_ctr];
+ outptr0 = output_buf[0];
+ outptr1 = output_buf[1];
+
+ int cols_remaining = output_width;
+ for (; cols_remaining >= 16; cols_remaining -= 16) {
+ /* For each row, de-interleave Y component values into two separate
+ * vectors, one containing the component values with even-numbered indices
+ * and one containing the component values with odd-numbered indices.
+ */
+ uint8x8x2_t y0 = vld2_u8(inptr0_0);
+ uint8x8x2_t y1 = vld2_u8(inptr0_1);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+ /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+ * the "even" and "odd" Y component values. This effectively upsamples the
+ * chroma components both horizontally and vertically.
+ */
+ int16x8_t g0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y0.val[0]));
+ int16x8_t r0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y0.val[0]));
+ int16x8_t b0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y0.val[0]));
+ int16x8_t g0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y0.val[1]));
+ int16x8_t r0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y0.val[1]));
+ int16x8_t b0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y0.val[1]));
+ int16x8_t g1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y1.val[0]));
+ int16x8_t r1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y1.val[0]));
+ int16x8_t b1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y1.val[0]));
+ int16x8_t g1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y1.val[1]));
+ int16x8_t r1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y1.val[1]));
+ int16x8_t b1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y1.val[1]));
+ /* Convert each component to unsigned and narrow, clamping to [0-255].
+ * Re-interleave the "even" and "odd" component values.
+ */
+ uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+ uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+ uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+ uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+ uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+ uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+ uint8x16x4_t rgba0, rgba1;
+ rgba0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+ rgba1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+ rgba0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+ rgba1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+ rgba0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+ rgba1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+ /* Set alpha channel to opaque (0xFF). */
+ rgba0.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+ rgba1.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ vst4q_u8(outptr0, rgba0);
+ vst4q_u8(outptr1, rgba1);
+#else
+ uint8x16x3_t rgb0, rgb1;
+ rgb0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
+ rgb1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
+ rgb0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
+ rgb1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
+ rgb0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
+ rgb1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
+ /* Store RGB pixel data to memory. */
+ vst3q_u8(outptr0, rgb0);
+ vst3q_u8(outptr1, rgb1);
+#endif
+
+ /* Increment pointers. */
+ inptr0_0 += 16;
+ inptr0_1 += 16;
+ inptr1 += 8;
+ inptr2 += 8;
+ outptr0 += (RGB_PIXELSIZE * 16);
+ outptr1 += (RGB_PIXELSIZE * 16);
+ }
+
+ if (cols_remaining > 0) {
+ /* For each row, de-interleave Y component values into two separate
+ * vectors, one containing the component values with even-numbered indices
+ * and one containing the component values with odd-numbered indices.
+ */
+ uint8x8x2_t y0 = vld2_u8(inptr0_0);
+ uint8x8x2_t y1 = vld2_u8(inptr0_1);
+ uint8x8_t cb = vld1_u8(inptr1);
+ uint8x8_t cr = vld1_u8(inptr2);
+ /* Subtract 128 from Cb and Cr. */
+ int16x8_t cr_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
+ int16x8_t cb_128 =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
+ /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
+ int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
+ int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
+ g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
+ g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
+ /* Descale G components: shift right 15, round, and narrow to 16-bit. */
+ int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
+ vrshrn_n_s32(g_sub_y_h, 15));
+ /* Compute R-Y: 1.40200 * (Cr - 128) */
+ int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
+ /* Compute B-Y: 1.77200 * (Cb - 128) */
+ int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
+ /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
+ * the "even" and "odd" Y component values. This effectively upsamples the
+ * chroma components both horizontally and vertically.
+ */
+ int16x8_t g0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y0.val[0]));
+ int16x8_t r0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y0.val[0]));
+ int16x8_t b0_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y0.val[0]));
+ int16x8_t g0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y0.val[1]));
+ int16x8_t r0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y0.val[1]));
+ int16x8_t b0_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y0.val[1]));
+ int16x8_t g1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y1.val[0]));
+ int16x8_t r1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y1.val[0]));
+ int16x8_t b1_even =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y1.val[0]));
+ int16x8_t g1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
+ y1.val[1]));
+ int16x8_t r1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
+ y1.val[1]));
+ int16x8_t b1_odd =
+ vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
+ y1.val[1]));
+ /* Convert each component to unsigned and narrow, clamping to [0-255].
+ * Re-interleave the "even" and "odd" component values.
+ */
+ uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
+ uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
+ uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
+ uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
+ uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
+ uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));
+
+#ifdef RGB_ALPHA
+ uint8x8x4_t rgba0_h, rgba1_h;
+ rgba0_h.val[RGB_RED] = r0.val[1];
+ rgba1_h.val[RGB_RED] = r1.val[1];
+ rgba0_h.val[RGB_GREEN] = g0.val[1];
+ rgba1_h.val[RGB_GREEN] = g1.val[1];
+ rgba0_h.val[RGB_BLUE] = b0.val[1];
+ rgba1_h.val[RGB_BLUE] = b1.val[1];
+ /* Set alpha channel to opaque (0xFF). */
+ rgba0_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ rgba1_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+
+ uint8x8x4_t rgba0_l, rgba1_l;
+ rgba0_l.val[RGB_RED] = r0.val[0];
+ rgba1_l.val[RGB_RED] = r1.val[0];
+ rgba0_l.val[RGB_GREEN] = g0.val[0];
+ rgba1_l.val[RGB_GREEN] = g1.val[0];
+ rgba0_l.val[RGB_BLUE] = b0.val[0];
+ rgba1_l.val[RGB_BLUE] = b1.val[0];
+ /* Set alpha channel to opaque (0xFF). */
+ rgba0_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ rgba1_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
+ /* Store RGBA pixel data to memory. */
+ switch (cols_remaining) {
+ case 15:
+ vst4_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgba0_h, 6);
+ vst4_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgba1_h, 6);
+ case 14:
+ vst4_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgba0_h, 5);
+ vst4_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgba1_h, 5);
+ case 13:
+ vst4_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgba0_h, 4);
+ vst4_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgba1_h, 4);
+ case 12:
+ vst4_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgba0_h, 3);
+ vst4_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgba1_h, 3);
+ case 11:
+ vst4_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgba0_h, 2);
+ vst4_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgba1_h, 2);
+ case 10:
+ vst4_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgba0_h, 1);
+ vst4_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgba1_h, 1);
+ case 9:
+ vst4_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgba0_h, 0);
+ vst4_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgba1_h, 0);
+ case 8:
+ vst4_u8(outptr0, rgba0_l);
+ vst4_u8(outptr1, rgba1_l);
+ break;
+ case 7:
+ vst4_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgba0_l, 6);
+ vst4_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgba1_l, 6);
+ case 6:
+ vst4_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgba0_l, 5);
+ vst4_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgba1_l, 5);
+ case 5:
+ vst4_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgba0_l, 4);
+ vst4_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgba1_l, 4);
+ case 4:
+ vst4_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgba0_l, 3);
+ vst4_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgba1_l, 3);
+ case 3:
+ vst4_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgba0_l, 2);
+ vst4_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgba1_l, 2);
+ case 2:
+ vst4_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgba0_l, 1);
+ vst4_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgba1_l, 1);
+ case 1:
+ vst4_lane_u8(outptr0, rgba0_l, 0);
+ vst4_lane_u8(outptr1, rgba1_l, 0);
+ default:
+ break;
+ }
+#else
+ uint8x8x3_t rgb0_h, rgb1_h;
+ rgb0_h.val[RGB_RED] = r0.val[1];
+ rgb1_h.val[RGB_RED] = r1.val[1];
+ rgb0_h.val[RGB_GREEN] = g0.val[1];
+ rgb1_h.val[RGB_GREEN] = g1.val[1];
+ rgb0_h.val[RGB_BLUE] = b0.val[1];
+ rgb1_h.val[RGB_BLUE] = b1.val[1];
+
+ uint8x8x3_t rgb0_l, rgb1_l;
+ rgb0_l.val[RGB_RED] = r0.val[0];
+ rgb1_l.val[RGB_RED] = r1.val[0];
+ rgb0_l.val[RGB_GREEN] = g0.val[0];
+ rgb1_l.val[RGB_GREEN] = g1.val[0];
+ rgb0_l.val[RGB_BLUE] = b0.val[0];
+ rgb1_l.val[RGB_BLUE] = b1.val[0];
+ /* Store RGB pixel data to memory. */
+ switch (cols_remaining) {
+ case 15:
+ vst3_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgb0_h, 6);
+ vst3_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgb1_h, 6);
+ case 14:
+ vst3_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgb0_h, 5);
+ vst3_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgb1_h, 5);
+ case 13:
+ vst3_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgb0_h, 4);
+ vst3_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgb1_h, 4);
+ case 12:
+ vst3_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgb0_h, 3);
+ vst3_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgb1_h, 3);
+ case 11:
+ vst3_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgb0_h, 2);
+ vst3_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgb1_h, 2);
+ case 10:
+ vst3_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgb0_h, 1);
+ vst3_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgb1_h, 1);
+ case 9:
+ vst3_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgb0_h, 0);
+ vst3_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgb1_h, 0);
+ case 8:
+ vst3_u8(outptr0, rgb0_l);
+ vst3_u8(outptr1, rgb1_l);
+ break;
+ case 7:
+ vst3_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgb0_l, 6);
+ vst3_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgb1_l, 6);
+ case 6:
+ vst3_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgb0_l, 5);
+ vst3_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgb1_l, 5);
+ case 5:
+ vst3_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgb0_l, 4);
+ vst3_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgb1_l, 4);
+ case 4:
+ vst3_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgb0_l, 3);
+ vst3_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgb1_l, 3);
+ case 3:
+ vst3_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgb0_l, 2);
+ vst3_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgb1_l, 2);
+ case 2:
+ vst3_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgb0_l, 1);
+ vst3_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgb1_l, 1);
+ case 1:
+ vst3_lane_u8(outptr0, rgb0_l, 0);
+ vst3_lane_u8(outptr1, rgb1_l, 0);
+ default:
+ break;
+ }
+#endif
+ }
+}
diff --git a/simd/arm/jdsample-neon.c b/simd/arm/jdsample-neon.c
new file mode 100644
index 0000000..90ec678
--- /dev/null
+++ b/simd/arm/jdsample-neon.c
@@ -0,0 +1,569 @@
+/*
+ * jdsample-neon.c - upsampling (Arm Neon)
+ *
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#define JPEG_INTERNALS
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
+#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+
+#include <arm_neon.h>
+
+
+/* The diagram below shows a row of samples produced by h2v1 downsampling.
+ *
+ * s0 s1 s2
+ * +---------+---------+---------+
+ * | | | |
+ * | p0 p1 | p2 p3 | p4 p5 |
+ * | | | |
+ * +---------+---------+---------+
+ *
+ * Samples s0-s2 were created by averaging the original pixel component values
+ * centered at positions p0-p5 above. To approximate those original pixel
+ * component values, we proportionally blend the adjacent samples in each row.
+ *
+ * An upsampled pixel component value is computed by blending the sample
+ * containing the pixel center with the nearest neighboring sample, in the
+ * ratio 3:1. For example:
+ * p1(upsampled) = 3/4 * s0 + 1/4 * s1
+ * p2(upsampled) = 3/4 * s1 + 1/4 * s0
+ * When computing the first and last pixel component values in the row, there
+ * is no adjacent sample to blend, so:
+ * p0(upsampled) = s0
+ * p5(upsampled) = s2
+ */
+
+void jsimd_h2v1_fancy_upsample_neon(int max_v_samp_factor,
+ JDIMENSION downsampled_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr, outptr;
+ int inrow;
+ unsigned colctr;
+ /* Set up constants. */
+ const uint16x8_t one_u16 = vdupq_n_u16(1);
+ const uint8x8_t three_u8 = vdup_n_u8(3);
+
+ for (inrow = 0; inrow < max_v_samp_factor; inrow++) {
+ inptr = input_data[inrow];
+ outptr = output_data[inrow];
+ /* First pixel component value in this row of the original image */
+ *outptr = (JSAMPLE)GETJSAMPLE(*inptr);
+
+ /* 3/4 * containing sample + 1/4 * nearest neighboring sample
+ * For p1: containing sample = s0, nearest neighboring sample = s1
+ * For p2: containing sample = s1, nearest neighboring sample = s0
+ */
+ uint8x16_t s0 = vld1q_u8(inptr);
+ uint8x16_t s1 = vld1q_u8(inptr + 1);
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ uint16x8_t s1_add_3s0_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+ uint16x8_t s1_add_3s0_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+ uint16x8_t s0_add_3s1_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+ uint16x8_t s0_add_3s1_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+ /* Add ordered dithering bias to odd pixel values. */
+ s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+ s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+ /* The offset is initially 1, because the first pixel component has already
+ * been stored. However, in subsequent iterations of the SIMD loop, this
+ * offset is (2 * colctr - 1) to stay within the bounds of the sample
+ * buffers without having to resort to a slow scalar tail case for the last
+ * (downsampled_width % 16) samples. See "Creation of 2-D sample arrays"
+ * in jmemmgr.c for more details.
+ */
+ unsigned outptr_offset = 1;
+ uint8x16x2_t output_pixels;
+
+ /* We use software pipelining to maximise performance. The code indented
+ * an extra two spaces begins the next iteration of the loop.
+ */
+ for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+
+ s0 = vld1q_u8(inptr + colctr - 1);
+ s1 = vld1q_u8(inptr + colctr);
+
+ /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+ output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+ vrshrn_n_u16(s1_add_3s0_h, 2));
+ output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+ vshrn_n_u16(s0_add_3s1_h, 2));
+
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ s1_add_3s0_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);
+ s1_add_3s0_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);
+ s0_add_3s1_l =
+ vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);
+ s0_add_3s1_h =
+ vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);
+ /* Add ordered dithering bias to odd pixel values. */
+ s0_add_3s1_l = vaddq_u16(s0_add_3s1_l, one_u16);
+ s0_add_3s1_h = vaddq_u16(s0_add_3s1_h, one_u16);
+
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr + outptr_offset, output_pixels);
+ outptr_offset = 2 * colctr - 1;
+ }
+
+ /* Complete the last iteration of the loop. */
+
+ /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+ output_pixels.val[0] = vcombine_u8(vrshrn_n_u16(s1_add_3s0_l, 2),
+ vrshrn_n_u16(s1_add_3s0_h, 2));
+ output_pixels.val[1] = vcombine_u8(vshrn_n_u16(s0_add_3s1_l, 2),
+ vshrn_n_u16(s0_add_3s1_h, 2));
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr + outptr_offset, output_pixels);
+
+ /* Last pixel component value in this row of the original image */
+ outptr[2 * downsampled_width - 1] =
+ GETJSAMPLE(inptr[downsampled_width - 1]);
+ }
+}
+
+
+/* The diagram below shows an array of samples produced by h2v2 downsampling.
+ *
+ * s0 s1 s2
+ * +---------+---------+---------+
+ * | p0 p1 | p2 p3 | p4 p5 |
+ * sA | | | |
+ * | p6 p7 | p8 p9 | p10 p11|
+ * +---------+---------+---------+
+ * | p12 p13| p14 p15| p16 p17|
+ * sB | | | |
+ * | p18 p19| p20 p21| p22 p23|
+ * +---------+---------+---------+
+ * | p24 p25| p26 p27| p28 p29|
+ * sC | | | |
+ * | p30 p31| p32 p33| p34 p35|
+ * +---------+---------+---------+
+ *
+ * Samples s0A-s2C were created by averaging the original pixel component
+ * values centered at positions p0-p35 above. To approximate one of those
+ * original pixel component values, we proportionally blend the sample
+ * containing the pixel center with the nearest neighboring samples in each
+ * row, column, and diagonal.
+ *
+ * An upsampled pixel component value is computed by first blending the sample
+ * containing the pixel center with the nearest neighboring samples in the
+ * same column, in the ratio 3:1, and then blending each column sum with the
+ * nearest neighboring column sum, in the ratio 3:1. For example:
+ * p14(upsampled) = 3/4 * (3/4 * s1B + 1/4 * s1A) +
+ * 1/4 * (3/4 * s0B + 1/4 * s0A)
+ * = 9/16 * s1B + 3/16 * s1A + 3/16 * s0B + 1/16 * s0A
+ * When computing the first and last pixel component values in the row, there
+ * is no horizontally adjacent sample to blend, so:
+ * p12(upsampled) = 3/4 * s0B + 1/4 * s0A
+ * p23(upsampled) = 3/4 * s2B + 1/4 * s2C
+ * When computing the first and last pixel component values in the column,
+ * there is no vertically adjacent sample to blend, so:
+ * p2(upsampled) = 3/4 * s1A + 1/4 * s0A
+ * p33(upsampled) = 3/4 * s1C + 1/4 * s2C
+ * When computing the corner pixel component values, there is no adjacent
+ * sample to blend, so:
+ * p0(upsampled) = s0A
+ * p35(upsampled) = s2C
+ */
+
+void jsimd_h2v2_fancy_upsample_neon(int max_v_samp_factor,
+ JDIMENSION downsampled_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
+ int inrow, outrow;
+ unsigned colctr;
+ /* Set up constants. */
+ const uint16x8_t seven_u16 = vdupq_n_u16(7);
+ const uint8x8_t three_u8 = vdup_n_u8(3);
+ const uint16x8_t three_u16 = vdupq_n_u16(3);
+
+ inrow = outrow = 0;
+ while (outrow < max_v_samp_factor) {
+ inptr0 = input_data[inrow - 1];
+ inptr1 = input_data[inrow];
+ inptr2 = input_data[inrow + 1];
+ /* Suffixes 0 and 1 denote the upper and lower rows of output pixels,
+ * respectively.
+ */
+ outptr0 = output_data[outrow++];
+ outptr1 = output_data[outrow++];
+
+ /* First pixel component value in this row of the original image */
+ int s0colsum0 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr0);
+ *outptr0 = (JSAMPLE)((s0colsum0 * 4 + 8) >> 4);
+ int s0colsum1 = GETJSAMPLE(*inptr1) * 3 + GETJSAMPLE(*inptr2);
+ *outptr1 = (JSAMPLE)((s0colsum1 * 4 + 8) >> 4);
+
+ /* Step 1: Blend samples vertically in columns s0 and s1.
+ * Leave the divide by 4 until the end, when it can be done for both
+ * dimensions at once, right-shifting by 4.
+ */
+
+ /* Load and compute s0colsum0 and s0colsum1. */
+ uint8x16_t s0A = vld1q_u8(inptr0);
+ uint8x16_t s0B = vld1q_u8(inptr1);
+ uint8x16_t s0C = vld1q_u8(inptr2);
+ /* Multiplication makes vectors twice as wide. '_l' and '_h' suffixes
+ * denote low half and high half respectively.
+ */
+ uint16x8_t s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)),
+ vget_low_u8(s0B), three_u8);
+ uint16x8_t s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)),
+ vget_high_u8(s0B), three_u8);
+ uint16x8_t s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)),
+ vget_low_u8(s0B), three_u8);
+ uint16x8_t s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)),
+ vget_high_u8(s0B), three_u8);
+ /* Load and compute s1colsum0 and s1colsum1. */
+ uint8x16_t s1A = vld1q_u8(inptr0 + 1);
+ uint8x16_t s1B = vld1q_u8(inptr1 + 1);
+ uint8x16_t s1C = vld1q_u8(inptr2 + 1);
+ uint16x8_t s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)),
+ vget_low_u8(s1B), three_u8);
+ uint16x8_t s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)),
+ vget_high_u8(s1B), three_u8);
+ uint16x8_t s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)),
+ vget_low_u8(s1B), three_u8);
+ uint16x8_t s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)),
+ vget_high_u8(s1B), three_u8);
+
+ /* Step 2: Blend the already-blended columns. */
+
+ uint16x8_t output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+ uint16x8_t output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+ uint16x8_t output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+ uint16x8_t output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+ uint16x8_t output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+ uint16x8_t output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+ uint16x8_t output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+ uint16x8_t output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+ /* Add ordered dithering bias to odd pixel values. */
+ output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+ output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+ output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+ output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+ /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+ uint8x16x2_t output_pixels0 = { {
+ vcombine_u8(vshrn_n_u16(output0_p1_l, 4), vshrn_n_u16(output0_p1_h, 4)),
+ vcombine_u8(vrshrn_n_u16(output0_p2_l, 4), vrshrn_n_u16(output0_p2_h, 4))
+ } };
+ uint8x16x2_t output_pixels1 = { {
+ vcombine_u8(vshrn_n_u16(output1_p1_l, 4), vshrn_n_u16(output1_p1_h, 4)),
+ vcombine_u8(vrshrn_n_u16(output1_p2_l, 4), vrshrn_n_u16(output1_p2_h, 4))
+ } };
+
+ /* Store pixel component values to memory.
+ * The minimum size of the output buffer for each row is 64 bytes => no
+ * need to worry about buffer overflow here. See "Creation of 2-D sample
+ * arrays" in jmemmgr.c for more details.
+ */
+ vst2q_u8(outptr0 + 1, output_pixels0);
+ vst2q_u8(outptr1 + 1, output_pixels1);
+
+ /* The first pixel of the image shifted our loads and stores by one byte.
+ * We have to re-align on a 32-byte boundary at some point before the end
+ * of the row (we do it now on the 32/33 pixel boundary) to stay within the
+ * bounds of the sample buffers without having to resort to a slow scalar
+ * tail case for the last (downsampled_width % 16) samples. See "Creation
+ * of 2-D sample arrays" in jmemmgr.c for more details.
+ */
+ for (colctr = 16; colctr < downsampled_width; colctr += 16) {
+ /* Step 1: Blend samples vertically in columns s0 and s1. */
+
+ /* Load and compute s0colsum0 and s0colsum1. */
+ s0A = vld1q_u8(inptr0 + colctr - 1);
+ s0B = vld1q_u8(inptr1 + colctr - 1);
+ s0C = vld1q_u8(inptr2 + colctr - 1);
+ s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)), vget_low_u8(s0B),
+ three_u8);
+ s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)), vget_high_u8(s0B),
+ three_u8);
+ s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)), vget_low_u8(s0B),
+ three_u8);
+ s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)), vget_high_u8(s0B),
+ three_u8);
+ /* Load and compute s1colsum0 and s1colsum1. */
+ s1A = vld1q_u8(inptr0 + colctr);
+ s1B = vld1q_u8(inptr1 + colctr);
+ s1C = vld1q_u8(inptr2 + colctr);
+ s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)), vget_low_u8(s1B),
+ three_u8);
+ s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)), vget_high_u8(s1B),
+ three_u8);
+ s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)), vget_low_u8(s1B),
+ three_u8);
+ s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)), vget_high_u8(s1B),
+ three_u8);
+
+ /* Step 2: Blend the already-blended columns. */
+
+ output0_p1_l = vmlaq_u16(s1colsum0_l, s0colsum0_l, three_u16);
+ output0_p1_h = vmlaq_u16(s1colsum0_h, s0colsum0_h, three_u16);
+ output0_p2_l = vmlaq_u16(s0colsum0_l, s1colsum0_l, three_u16);
+ output0_p2_h = vmlaq_u16(s0colsum0_h, s1colsum0_h, three_u16);
+ output1_p1_l = vmlaq_u16(s1colsum1_l, s0colsum1_l, three_u16);
+ output1_p1_h = vmlaq_u16(s1colsum1_h, s0colsum1_h, three_u16);
+ output1_p2_l = vmlaq_u16(s0colsum1_l, s1colsum1_l, three_u16);
+ output1_p2_h = vmlaq_u16(s0colsum1_h, s1colsum1_h, three_u16);
+ /* Add ordered dithering bias to odd pixel values. */
+ output0_p1_l = vaddq_u16(output0_p1_l, seven_u16);
+ output0_p1_h = vaddq_u16(output0_p1_h, seven_u16);
+ output1_p1_l = vaddq_u16(output1_p1_l, seven_u16);
+ output1_p1_h = vaddq_u16(output1_p1_h, seven_u16);
+ /* Right-shift by 4 (divide by 16), narrow to 8-bit, and combine. */
+ output_pixels0.val[0] = vcombine_u8(vshrn_n_u16(output0_p1_l, 4),
+ vshrn_n_u16(output0_p1_h, 4));
+ output_pixels0.val[1] = vcombine_u8(vrshrn_n_u16(output0_p2_l, 4),
+ vrshrn_n_u16(output0_p2_h, 4));
+ output_pixels1.val[0] = vcombine_u8(vshrn_n_u16(output1_p1_l, 4),
+ vshrn_n_u16(output1_p1_h, 4));
+ output_pixels1.val[1] = vcombine_u8(vrshrn_n_u16(output1_p2_l, 4),
+ vrshrn_n_u16(output1_p2_h, 4));
+ /* Store pixel component values to memory. */
+ vst2q_u8(outptr0 + 2 * colctr - 1, output_pixels0);
+ vst2q_u8(outptr1 + 2 * colctr - 1, output_pixels1);
+ }
+
+ /* Last pixel component value in this row of the original image */
+ int s1colsum0 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+ GETJSAMPLE(inptr0[downsampled_width - 1]);
+ outptr0[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum0 * 4 + 7) >> 4);
+ int s1colsum1 = GETJSAMPLE(inptr1[downsampled_width - 1]) * 3 +
+ GETJSAMPLE(inptr2[downsampled_width - 1]);
+ outptr1[2 * downsampled_width - 1] = (JSAMPLE)((s1colsum1 * 4 + 7) >> 4);
+ inrow++;
+ }
+}
+
+
+/* The diagram below shows a column of samples produced by h1v2 downsampling
+ * (or by losslessly rotating or transposing an h2v1-downsampled image.)
+ *
+ * +---------+
+ * | p0 |
+ * sA | |
+ * | p1 |
+ * +---------+
+ * | p2 |
+ * sB | |
+ * | p3 |
+ * +---------+
+ * | p4 |
+ * sC | |
+ * | p5 |
+ * +---------+
+ *
+ * Samples sA-sC were created by averaging the original pixel component values
+ * centered at positions p0-p5 above. To approximate those original pixel
+ * component values, we proportionally blend the adjacent samples in each
+ * column.
+ *
+ * An upsampled pixel component value is computed by blending the sample
+ * containing the pixel center with the nearest neighboring sample, in the
+ * ratio 3:1. For example:
+ * p1(upsampled) = 3/4 * sA + 1/4 * sB
+ * p2(upsampled) = 3/4 * sB + 1/4 * sA
+ * When computing the first and last pixel component values in the column,
+ * there is no adjacent sample to blend, so:
+ * p0(upsampled) = sA
+ * p5(upsampled) = sC
+ */
+
+void jsimd_h1v2_fancy_upsample_neon(int max_v_samp_factor,
+ JDIMENSION downsampled_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr0, inptr1, inptr2, outptr0, outptr1;
+ int inrow, outrow;
+ unsigned colctr;
+ /* Set up constants. */
+ const uint16x8_t one_u16 = vdupq_n_u16(1);
+ const uint8x8_t three_u8 = vdup_n_u8(3);
+
+ inrow = outrow = 0;
+ while (outrow < max_v_samp_factor) {
+ inptr0 = input_data[inrow - 1];
+ inptr1 = input_data[inrow];
+ inptr2 = input_data[inrow + 1];
+ /* Suffixes 0 and 1 denote the upper and lower rows of output pixels,
+ * respectively.
+ */
+ outptr0 = output_data[outrow++];
+ outptr1 = output_data[outrow++];
+ inrow++;
+
+ /* The size of the input and output buffers is always a multiple of 32
+ * bytes => no need to worry about buffer overflow when reading/writing
+ * memory. See "Creation of 2-D sample arrays" in jmemmgr.c for more
+ * details.
+ */
+ for (colctr = 0; colctr < downsampled_width; colctr += 16) {
+ /* Load samples. */
+ uint8x16_t sA = vld1q_u8(inptr0 + colctr);
+ uint8x16_t sB = vld1q_u8(inptr1 + colctr);
+ uint8x16_t sC = vld1q_u8(inptr2 + colctr);
+ /* Blend samples vertically. */
+ uint16x8_t colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(sA)),
+ vget_low_u8(sB), three_u8);
+ uint16x8_t colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(sA)),
+ vget_high_u8(sB), three_u8);
+ uint16x8_t colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(sC)),
+ vget_low_u8(sB), three_u8);
+ uint16x8_t colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(sC)),
+ vget_high_u8(sB), three_u8);
+ /* Add ordered dithering bias to pixel values in even output rows. */
+ colsum0_l = vaddq_u16(colsum0_l, one_u16);
+ colsum0_h = vaddq_u16(colsum0_h, one_u16);
+ /* Right-shift by 2 (divide by 4), narrow to 8-bit, and combine. */
+ uint8x16_t output_pixels0 = vcombine_u8(vshrn_n_u16(colsum0_l, 2),
+ vshrn_n_u16(colsum0_h, 2));
+ uint8x16_t output_pixels1 = vcombine_u8(vrshrn_n_u16(colsum1_l, 2),
+ vrshrn_n_u16(colsum1_h, 2));
+ /* Store pixel component values to memory. */
+ vst1q_u8(outptr0 + colctr, output_pixels0);
+ vst1q_u8(outptr1 + colctr, output_pixels1);
+ }
+ }
+}
+
+
+/* The diagram below shows a row of samples produced by h2v1 downsampling.
+ *
+ * s0 s1
+ * +---------+---------+
+ * | | |
+ * | p0 p1 | p2 p3 |
+ * | | |
+ * +---------+---------+
+ *
+ * Samples s0 and s1 were created by averaging the original pixel component
+ * values centered at positions p0-p3 above. To approximate those original
+ * pixel component values, we duplicate the samples horizontally:
+ * p0(upsampled) = p1(upsampled) = s0
+ * p2(upsampled) = p3(upsampled) = s1
+ */
+
+void jsimd_h2v1_upsample_neon(int max_v_samp_factor, JDIMENSION output_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr, outptr;
+ int inrow;
+ unsigned colctr;
+
+ for (inrow = 0; inrow < max_v_samp_factor; inrow++) {
+ inptr = input_data[inrow];
+ outptr = output_data[inrow];
+ for (colctr = 0; 2 * colctr < output_width; colctr += 16) {
+ uint8x16_t samples = vld1q_u8(inptr + colctr);
+ /* Duplicate the samples. The store operation below interleaves them so
+ * that adjacent pixel component values take on the same sample value,
+ * per above.
+ */
+ uint8x16x2_t output_pixels = { { samples, samples } };
+ /* Store pixel component values to memory.
+ * Due to the way sample buffers are allocated, we don't need to worry
+ * about tail cases when output_width is not a multiple of 32. See
+ * "Creation of 2-D sample arrays" in jmemmgr.c for details.
+ */
+ vst2q_u8(outptr + 2 * colctr, output_pixels);
+ }
+ }
+}
+
+
+/* The diagram below shows an array of samples produced by h2v2 downsampling.
+ *
+ * s0 s1
+ * +---------+---------+
+ * | p0 p1 | p2 p3 |
+ * sA | | |
+ * | p4 p5 | p6 p7 |
+ * +---------+---------+
+ * | p8 p9 | p10 p11|
+ * sB | | |
+ * | p12 p13| p14 p15|
+ * +---------+---------+
+ *
+ * Samples s0A-s1B were created by averaging the original pixel component
+ * values centered at positions p0-p15 above. To approximate those original
+ * pixel component values, we duplicate the samples both horizontally and
+ * vertically:
+ * p0(upsampled) = p1(upsampled) = p4(upsampled) = p5(upsampled) = s0A
+ * p2(upsampled) = p3(upsampled) = p6(upsampled) = p7(upsampled) = s1A
+ * p8(upsampled) = p9(upsampled) = p12(upsampled) = p13(upsampled) = s0B
+ * p10(upsampled) = p11(upsampled) = p14(upsampled) = p15(upsampled) = s1B
+ */
+
+void jsimd_h2v2_upsample_neon(int max_v_samp_factor, JDIMENSION output_width,
+ JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr)
+{
+ JSAMPARRAY output_data = *output_data_ptr;
+ JSAMPROW inptr, outptr0, outptr1;
+ int inrow, outrow;
+ unsigned colctr;
+
+ for (inrow = 0, outrow = 0; outrow < max_v_samp_factor; inrow++) {
+ inptr = input_data[inrow];
+ outptr0 = output_data[outrow++];
+ outptr1 = output_data[outrow++];
+
+ for (colctr = 0; 2 * colctr < output_width; colctr += 16) {
+ uint8x16_t samples = vld1q_u8(inptr + colctr);
+ /* Duplicate the samples. The store operation below interleaves them so
+ * that adjacent pixel component values take on the same sample value,
+ * per above.
+ */
+ uint8x16x2_t output_pixels = { { samples, samples } };
+ /* Store pixel component values for both output rows to memory.
+ * Due to the way sample buffers are allocated, we don't need to worry
+ * about tail cases when output_width is not a multiple of 32. See
+ * "Creation of 2-D sample arrays" in jmemmgr.c for details.
+ */
+ vst2q_u8(outptr0 + 2 * colctr, output_pixels);
+ vst2q_u8(outptr1 + 2 * colctr, output_pixels);
+ }
+ }
+}
diff --git a/simd/arm/common/jfdctfst-neon.c b/simd/arm/jfdctfst-neon.c
similarity index 88%
rename from simd/arm/common/jfdctfst-neon.c
rename to simd/arm/jfdctfst-neon.c
index e7b2e96..bb371be 100644
--- a/simd/arm/common/jfdctfst-neon.c
+++ b/simd/arm/jfdctfst-neon.c
@@ -1,7 +1,7 @@
/*
- * jfdctfst-neon.c - fast DCT (Arm NEON)
+ * jfdctfst-neon.c - fast integer FDCT (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,21 +21,21 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
-/*
- * 'jsimd_fdct_ifast_neon' performs a fast, not so accurate forward DCT
- * (Discrete Cosine Transform) on one block of samples. It uses the same
+
+/* jsimd_fdct_ifast_neon() performs a fast, not so accurate forward DCT
+ * (Discrete Cosine Transform) on one block of samples. It uses the same
* calculations and produces exactly the same output as IJG's original
- * 'jpeg_fdct_ifast' function, which can be found in jfdctfst.c.
+ * jpeg_fdct_ifast() function, which can be found in jfdctfst.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
* 0.382683433 = 12544 * 2^-15
@@ -43,9 +43,9 @@
* 0.707106781 = 23168 * 2^-15
* 0.306562965 = 9984 * 2^-15
*
- * See jfdctfst.c for further details of the IDCT algorithm. Where possible,
- * the variable names and comments here in 'jsimd_fdct_ifast_neon' match up
- * with those in 'jpeg_fdct_ifast'.
+ * See jfdctfst.c for further details of the DCT algorithm. Where possible,
+ * the variable names and comments here in jsimd_fdct_ifast_neon() match up
+ * with those in jpeg_fdct_ifast().
*/
#define F_0_382 12544
@@ -53,16 +53,17 @@
#define F_0_707 23168
#define F_0_306 9984
+
ALIGN(16) static const int16_t jsimd_fdct_ifast_neon_consts[] = {
F_0_382, F_0_541, F_0_707, F_0_306
};
void jsimd_fdct_ifast_neon(DCTELEM *data)
{
- /* Load an 8x8 block of samples into Neon registers. De-interleaving loads */
- /* are used followed by vuzp to transpose the block such that we have a */
- /* column of samples per vector - allowing all rows to be processed at */
- /* once. */
+ /* Load an 8x8 block of samples into Neon registers. De-interleaving loads
+ * are used, followed by vuzp to transpose the block such that we have a
+ * column of samples per vector - allowing all rows to be processed at once.
+ */
int16x8x4_t data1 = vld4q_s16(data);
int16x8x4_t data2 = vld4q_s16(data + 4 * DCTSIZE);
@@ -80,10 +81,11 @@
int16x8_t col6 = cols_26.val[1];
int16x8_t col7 = cols_37.val[1];
+ /* Pass 1: process rows. */
+
/* Load DCT conversion constants. */
const int16x4_t consts = vld1_s16(jsimd_fdct_ifast_neon_consts);
- /* Pass 1: process rows. */
int16x8_t tmp0 = vaddq_s16(col0, col7);
int16x8_t tmp7 = vsubq_s16(col0, col7);
int16x8_t tmp1 = vaddq_s16(col1, col6);
@@ -157,6 +159,7 @@
int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
/* Pass 2: process columns. */
+
tmp0 = vaddq_s16(row0, row7);
tmp7 = vsubq_s16(row0, row7);
tmp1 = vaddq_s16(row1, row6);
diff --git a/simd/arm/common/jfdctint-neon.c b/simd/arm/jfdctint-neon.c
similarity index 83%
rename from simd/arm/common/jfdctint-neon.c
rename to simd/arm/jfdctint-neon.c
index 55abb1b..ccfc07b 100644
--- a/simd/arm/common/jfdctint-neon.c
+++ b/simd/arm/jfdctint-neon.c
@@ -1,7 +1,8 @@
/*
- * jfdctint-neon.c - accurate DCT (Arm NEON)
+ * jfdctint-neon.c - accurate integer FDCT (Arm Neon)
*
- * Copyright 2020 The Chromium Aruthors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,21 +22,22 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
#include <arm_neon.h>
-/*
- * 'jsimd_fdct_islow_neon' performs a slow-but-accurate forward DCT (Discrete
- * Cosine Transform) on one block of samples. It uses the same calculations
- * and produces exactly the same output as IJG's original 'jpeg_fdct_islow'
- * function, which can be found in jfdctint.c.
+
+/* jsimd_fdct_islow_neon() performs a slower but more accurate forward DCT
+ * (Discrete Cosine Transform) on one block of samples. It uses the same
+ * calculations and produces exactly the same output as IJG's original
+ * jpeg_fdct_islow() function, which can be found in jfdctint.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
* 0.298631336 = 2446 * 2^-13
@@ -51,9 +53,9 @@
* 2.562915447 = 20995 * 2^-13
* 3.072711026 = 25172 * 2^-13
*
- * See jfdctint.c for further details of the DCT algorithm. Where possible,
- * the variable names and comments here in 'jsimd_fdct_islow_neon' match up
- * with those in 'jpeg_fdct_islow'.
+ * See jfdctint.c for further details of the DCT algorithm. Where possible,
+ * the variable names and comments here in jsimd_fdct_islow_neon() match up
+ * with those in jpeg_fdct_islow().
*/
#define CONST_BITS 13
@@ -75,6 +77,7 @@
#define F_2_562 20995
#define F_3_072 25172
+
ALIGN(16) static const int16_t jsimd_fdct_islow_neon_consts[] = {
F_0_298, -F_0_390, F_0_541, F_0_765,
-F_0_899, F_1_175, F_1_501, -F_1_847,
@@ -84,20 +87,20 @@
void jsimd_fdct_islow_neon(DCTELEM *data)
{
/* Load DCT constants. */
-#if defined(__clang__) || defined(_MSC_VER)
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_fdct_islow_neon_consts);
#else
/* GCC does not currently support the intrinsic vld1_<type>_x3(). */
const int16x4_t consts1 = vld1_s16(jsimd_fdct_islow_neon_consts);
const int16x4_t consts2 = vld1_s16(jsimd_fdct_islow_neon_consts + 4);
const int16x4_t consts3 = vld1_s16(jsimd_fdct_islow_neon_consts + 8);
- const int16x4x3_t consts = { consts1, consts2, consts3 };
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
#endif
- /* Load an 8x8 block of samples into Neon registers. De-interleaving loads */
- /* are used followed by vuzp to transpose the block such that we have a */
- /* column of samples per vector - allowing all rows to be processed at */
- /* once. */
+ /* Load an 8x8 block of samples into Neon registers. De-interleaving loads
+ * are used, followed by vuzp to transpose the block such that we have a
+ * column of samples per vector - allowing all rows to be processed at once.
+ */
int16x8x4_t s_rows_0123 = vld4q_s16(data);
int16x8x4_t s_rows_4567 = vld4q_s16(data + 4 * DCTSIZE);
@@ -116,6 +119,7 @@
int16x8_t col7 = cols_37.val[1];
/* Pass 1: process rows. */
+
int16x8_t tmp0 = vaddq_s16(col0, col7);
int16x8_t tmp7 = vsubq_s16(col0, col7);
int16x8_t tmp1 = vaddq_s16(col1, col6);
@@ -125,7 +129,7 @@
int16x8_t tmp3 = vaddq_s16(col3, col4);
int16x8_t tmp4 = vsubq_s16(col3, col4);
- /* Even part. */
+ /* Even part */
int16x8_t tmp10 = vaddq_s16(tmp0, tmp3);
int16x8_t tmp13 = vsubq_s16(tmp0, tmp3);
int16x8_t tmp11 = vaddq_s16(tmp1, tmp2);
@@ -135,26 +139,26 @@
col4 = vshlq_n_s16(vsubq_s16(tmp10, tmp11), PASS1_BITS);
int16x8_t tmp12_add_tmp13 = vaddq_s16(tmp12, tmp13);
- int32x4_t z1_l = vmull_lane_s16(vget_low_s16(tmp12_add_tmp13),
- consts.val[0], 2);
- int32x4_t z1_h = vmull_lane_s16(vget_high_s16(tmp12_add_tmp13),
- consts.val[0], 2);
+ int32x4_t z1_l =
+ vmull_lane_s16(vget_low_s16(tmp12_add_tmp13), consts.val[0], 2);
+ int32x4_t z1_h =
+ vmull_lane_s16(vget_high_s16(tmp12_add_tmp13), consts.val[0], 2);
- int32x4_t col2_scaled_l = vmlal_lane_s16(z1_l, vget_low_s16(tmp13),
- consts.val[0], 3);
- int32x4_t col2_scaled_h = vmlal_lane_s16(z1_h, vget_high_s16(tmp13),
- consts.val[0], 3);
+ int32x4_t col2_scaled_l =
+ vmlal_lane_s16(z1_l, vget_low_s16(tmp13), consts.val[0], 3);
+ int32x4_t col2_scaled_h =
+ vmlal_lane_s16(z1_h, vget_high_s16(tmp13), consts.val[0], 3);
col2 = vcombine_s16(vrshrn_n_s32(col2_scaled_l, DESCALE_P1),
vrshrn_n_s32(col2_scaled_h, DESCALE_P1));
- int32x4_t col6_scaled_l = vmlal_lane_s16(z1_l, vget_low_s16(tmp12),
- consts.val[1], 3);
- int32x4_t col6_scaled_h = vmlal_lane_s16(z1_h, vget_high_s16(tmp12),
- consts.val[1], 3);
+ int32x4_t col6_scaled_l =
+ vmlal_lane_s16(z1_l, vget_low_s16(tmp12), consts.val[1], 3);
+ int32x4_t col6_scaled_h =
+ vmlal_lane_s16(z1_h, vget_high_s16(tmp12), consts.val[1], 3);
col6 = vcombine_s16(vrshrn_n_s32(col6_scaled_l, DESCALE_P1),
vrshrn_n_s32(col6_scaled_h, DESCALE_P1));
- /* Odd part. */
+ /* Odd part */
int16x8_t z1 = vaddq_s16(tmp4, tmp7);
int16x8_t z2 = vaddq_s16(tmp5, tmp6);
int16x8_t z3 = vaddq_s16(tmp4, tmp6);
@@ -253,7 +257,8 @@
int16x8_t row6 = vreinterpretq_s16_s32(rows_26.val[1]);
int16x8_t row7 = vreinterpretq_s16_s32(rows_37.val[1]);
- /* Pass 2. */
+ /* Pass 2: process columns. */
+
tmp0 = vaddq_s16(row0, row7);
tmp7 = vsubq_s16(row0, row7);
tmp1 = vaddq_s16(row1, row6);
@@ -263,7 +268,7 @@
tmp3 = vaddq_s16(row3, row4);
tmp4 = vsubq_s16(row3, row4);
- /* Even part. */
+ /* Even part */
tmp10 = vaddq_s16(tmp0, tmp3);
tmp13 = vsubq_s16(tmp0, tmp3);
tmp11 = vaddq_s16(tmp1, tmp2);
@@ -276,21 +281,21 @@
z1_l = vmull_lane_s16(vget_low_s16(tmp12_add_tmp13), consts.val[0], 2);
z1_h = vmull_lane_s16(vget_high_s16(tmp12_add_tmp13), consts.val[0], 2);
- int32x4_t row2_scaled_l = vmlal_lane_s16(z1_l, vget_low_s16(tmp13),
- consts.val[0], 3);
- int32x4_t row2_scaled_h = vmlal_lane_s16(z1_h, vget_high_s16(tmp13),
- consts.val[0], 3);
+ int32x4_t row2_scaled_l =
+ vmlal_lane_s16(z1_l, vget_low_s16(tmp13), consts.val[0], 3);
+ int32x4_t row2_scaled_h =
+ vmlal_lane_s16(z1_h, vget_high_s16(tmp13), consts.val[0], 3);
row2 = vcombine_s16(vrshrn_n_s32(row2_scaled_l, DESCALE_P2),
vrshrn_n_s32(row2_scaled_h, DESCALE_P2));
- int32x4_t row6_scaled_l = vmlal_lane_s16(z1_l, vget_low_s16(tmp12),
- consts.val[1], 3);
- int32x4_t row6_scaled_h = vmlal_lane_s16(z1_h, vget_high_s16(tmp12),
- consts.val[1], 3);
+ int32x4_t row6_scaled_l =
+ vmlal_lane_s16(z1_l, vget_low_s16(tmp12), consts.val[1], 3);
+ int32x4_t row6_scaled_h =
+ vmlal_lane_s16(z1_h, vget_high_s16(tmp12), consts.val[1], 3);
row6 = vcombine_s16(vrshrn_n_s32(row6_scaled_l, DESCALE_P2),
vrshrn_n_s32(row6_scaled_h, DESCALE_P2));
- /* Odd part. */
+ /* Odd part */
z1 = vaddq_s16(tmp4, tmp7);
z2 = vaddq_s16(tmp5, tmp6);
z3 = vaddq_s16(tmp4, tmp6);
diff --git a/simd/arm/common/jidctfst-neon.c b/simd/arm/jidctfst-neon.c
similarity index 83%
rename from simd/arm/common/jidctfst-neon.c
rename to simd/arm/jidctfst-neon.c
index 87806fd..a91be53 100644
--- a/simd/arm/common/jidctfst-neon.c
+++ b/simd/arm/jidctfst-neon.c
@@ -1,7 +1,7 @@
/*
- * jidctfst-neon.c - fast IDCT (Arm NEON)
+ * jidctfst-neon.c - fast integer IDCT (Arm Neon)
*
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,20 +21,21 @@
*/
#define JPEG_INTERNALS
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
#include <arm_neon.h>
-/*
- * 'jsimd_idct_ifast_neon' performs dequantization and a fast, not so accurate
- * inverse DCT (Discrete Cosine Transform) on one block of coefficients. It
+
+/* jsimd_idct_ifast_neon() performs dequantization and a fast, not so accurate
+ * inverse DCT (Discrete Cosine Transform) on one block of coefficients. It
* uses the same calculations and produces exactly the same output as IJG's
- * original 'jpeg_idct_ifast' function, which can be found in jidctfst.c.
+ * original jpeg_idct_ifast() function, which can be found in jidctfst.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
* 0.082392200 = 2688 * 2^-15
@@ -42,9 +43,9 @@
* 0.847759065 = 27776 * 2^-15
* 0.613125930 = 20096 * 2^-15
*
- * See jidctfst.c for further details of the IDCT algorithm. Where possible,
- * the variable names and comments here in 'jsimd_idct_ifast_neon' match up
- * with those in 'jpeg_idct_ifast'.
+ * See jidctfst.c for further details of the IDCT algorithm. Where possible,
+ * the variable names and comments here in jsimd_idct_ifast_neon() match up
+ * with those in jpeg_idct_ifast().
*/
#define PASS1_BITS 2
@@ -54,10 +55,13 @@
#define F_0_847 27776
#define F_0_613 20096
-void jsimd_idct_ifast_neon(void *dct_table,
- JCOEFPTR coef_block,
- JSAMPARRAY output_buf,
- JDIMENSION output_col)
+
+ALIGN(16) static const int16_t jsimd_idct_ifast_neon_consts[] = {
+ F_0_082, F_0_414, F_0_847, F_0_613
+};
+
+void jsimd_idct_ifast_neon(void *dct_table, JCOEFPTR coef_block,
+ JSAMPARRAY output_buf, JDIMENSION output_col)
{
IFAST_MULT_TYPE *quantptr = dct_table;
@@ -87,9 +91,13 @@
int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
+ /* Load IDCT conversion constants. */
+ const int16x4_t consts = vld1_s16(jsimd_idct_ifast_neon_consts);
+
if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
- /* All AC coefficients are zero. */
- /* Compute DC values and duplicate into vectors. */
+ /* All AC coefficients are zero.
+ * Compute DC values and duplicate into vectors.
+ */
int16x8_t dcval = row0;
row1 = dcval;
row2 = dcval;
@@ -99,12 +107,14 @@
row6 = dcval;
row7 = dcval;
} else if (left_ac_bitmap == 0) {
- /* AC coefficients are zero for columns 0, 1, 2 and 3. */
- /* Use DC values for these columns. */
+ /* AC coefficients are zero for columns 0, 1, 2, and 3.
+ * Use DC values for these columns.
+ */
int16x4_t dcval = vget_low_s16(row0);
- /* Commence regular fast IDCT computation for columns 4, 5, 6 and 7. */
- /* Load quantization table.*/
+ /* Commence regular fast IDCT computation for columns 4, 5, 6, and 7. */
+
+ /* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
@@ -124,7 +134,7 @@
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
- int16x4_t tmp12 = vqdmulh_n_s16(tmp1_sub_tmp3, F_0_414);
+ int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
@@ -146,16 +156,16 @@
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
- tmp11 = vqdmulh_n_s16(z11_sub_z13, F_0_414);
+ tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
- int16x4_t z5 = vqdmulh_n_s16(z10_add_z12, F_0_847);
+ int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
- tmp10 = vqdmulh_n_s16(z12, F_0_082);
+ tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
- tmp12 = vqdmulh_n_s16(neg_z10, F_0_613);
+ tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
@@ -172,12 +182,14 @@
row4 = vcombine_s16(dcval, vadd_s16(tmp3, tmp4));
row3 = vcombine_s16(dcval, vsub_s16(tmp3, tmp4));
} else if (right_ac_bitmap == 0) {
- /* AC coefficients are zero for columns 4, 5, 6 and 7. */
- /* Use DC values for these columns. */
+ /* AC coefficients are zero for columns 4, 5, 6, and 7.
+ * Use DC values for these columns.
+ */
int16x4_t dcval = vget_high_s16(row0);
- /* Commence regular fast IDCT computation for columns 0, 1, 2 and 3. */
- /* Load quantization table.*/
+ /* Commence regular fast IDCT computation for columns 0, 1, 2, and 3. */
+
+ /* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
@@ -197,7 +209,7 @@
int16x4_t tmp13 = vadd_s16(tmp1, tmp3); /* phases 5-3 */
int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
- int16x4_t tmp12 = vqdmulh_n_s16(tmp1_sub_tmp3, F_0_414);
+ int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsub_s16(tmp12, tmp13);
@@ -219,16 +231,16 @@
tmp7 = vadd_s16(z11, z13); /* phase 5 */
int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
- tmp11 = vqdmulh_n_s16(z11_sub_z13, F_0_414);
+ tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vadd_s16(tmp11, z11_sub_z13);
int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
- int16x4_t z5 = vqdmulh_n_s16(z10_add_z12, F_0_847);
+ int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
z5 = vadd_s16(z5, z10_add_z12);
- tmp10 = vqdmulh_n_s16(z12, F_0_082);
+ tmp10 = vqdmulh_lane_s16(z12, consts, 0);
tmp10 = vadd_s16(tmp10, z12);
tmp10 = vsub_s16(tmp10, z5);
- tmp12 = vqdmulh_n_s16(neg_z10, F_0_613);
+ tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
tmp12 = vadd_s16(tmp12, z5);
@@ -246,7 +258,8 @@
row3 = vcombine_s16(vsub_s16(tmp3, tmp4), dcval);
} else {
/* Some AC coefficients are non-zero; full IDCT calculation required. */
- /* Load quantization table.*/
+
+ /* Load quantization table. */
int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
@@ -266,7 +279,7 @@
int16x8_t tmp13 = vaddq_s16(tmp1, tmp3); /* phases 5-3 */
int16x8_t tmp1_sub_tmp3 = vsubq_s16(tmp1, tmp3);
- int16x8_t tmp12 = vqdmulhq_n_s16(tmp1_sub_tmp3, F_0_414);
+ int16x8_t tmp12 = vqdmulhq_lane_s16(tmp1_sub_tmp3, consts, 1);
tmp12 = vaddq_s16(tmp12, tmp1_sub_tmp3);
tmp12 = vsubq_s16(tmp12, tmp13);
@@ -288,16 +301,16 @@
tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
- tmp11 = vqdmulhq_n_s16(z11_sub_z13, F_0_414);
+ tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
- int16x8_t z5 = vqdmulhq_n_s16(z10_add_z12, F_0_847);
+ int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
- tmp10 = vqdmulhq_n_s16(z12, F_0_082);
+ tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
- tmp12 = vqdmulhq_n_s16(neg_z10, F_0_613);
+ tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
@@ -315,7 +328,7 @@
row3 = vsubq_s16(tmp3, tmp4);
}
- /* Tranpose rows to work on columns in pass 2. */
+ /* Transpose rows to work on columns in pass 2. */
int16x8x2_t rows_01 = vtrnq_s16(row0, row1);
int16x8x2_t rows_23 = vtrnq_s16(row2, row3);
int16x8x2_t rows_45 = vtrnq_s16(row4, row5);
@@ -344,14 +357,15 @@
int16x8_t col6 = vreinterpretq_s16_s32(cols_26.val[1]);
int16x8_t col7 = vreinterpretq_s16_s32(cols_37.val[1]);
- /* 1-D IDCT, pass 2. */
- /* Even part. */
+ /* 1-D IDCT, pass 2 */
+
+ /* Even part */
int16x8_t tmp10 = vaddq_s16(col0, col4);
int16x8_t tmp11 = vsubq_s16(col0, col4);
int16x8_t tmp13 = vaddq_s16(col2, col6);
int16x8_t col2_sub_col6 = vsubq_s16(col2, col6);
- int16x8_t tmp12 = vqdmulhq_n_s16(col2_sub_col6, F_0_414);
+ int16x8_t tmp12 = vqdmulhq_lane_s16(col2_sub_col6, consts, 1);
tmp12 = vaddq_s16(tmp12, col2_sub_col6);
tmp12 = vsubq_s16(tmp12, tmp13);
@@ -360,7 +374,7 @@
int16x8_t tmp1 = vaddq_s16(tmp11, tmp12);
int16x8_t tmp2 = vsubq_s16(tmp11, tmp12);
- /* Odd part. */
+ /* Odd part */
int16x8_t z13 = vaddq_s16(col5, col3);
int16x8_t neg_z10 = vsubq_s16(col3, col5);
int16x8_t z11 = vaddq_s16(col1, col7);
@@ -368,16 +382,16 @@
int16x8_t tmp7 = vaddq_s16(z11, z13); /* phase 5 */
int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
- tmp11 = vqdmulhq_n_s16(z11_sub_z13, F_0_414);
+ tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
tmp11 = vaddq_s16(tmp11, z11_sub_z13);
int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
- int16x8_t z5 = vqdmulhq_n_s16(z10_add_z12, F_0_847);
+ int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
z5 = vaddq_s16(z5, z10_add_z12);
- tmp10 = vqdmulhq_n_s16(z12, F_0_082);
+ tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
tmp10 = vaddq_s16(tmp10, z12);
tmp10 = vsubq_s16(tmp10, z5);
- tmp12 = vqdmulhq_n_s16(neg_z10, F_0_613);
+ tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
tmp12 = vaddq_s16(tmp12, z5);
@@ -394,7 +408,7 @@
col4 = vaddq_s16(tmp3, tmp4);
col3 = vsubq_s16(tmp3, tmp4);
- /* Scale down by factor of 8, narrowing to 8-bit. */
+ /* Scale down by a factor of 8, narrowing to 8-bit. */
int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3),
vqshrn_n_s16(col1, PASS1_BITS + 3));
int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3),
@@ -404,16 +418,20 @@
int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3),
vqshrn_n_s16(col7, PASS1_BITS + 3));
/* Clamp to range [0-255]. */
- uint8x16_t cols_01 = vreinterpretq_u8_s8(
- vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
- uint8x16_t cols_45 = vreinterpretq_u8_s8(
- vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
- uint8x16_t cols_23 = vreinterpretq_u8_s8(
- vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
- uint8x16_t cols_67 = vreinterpretq_u8_s8(
- vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+ uint8x16_t cols_01 =
+ vreinterpretq_u8_s8
+ (vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+ uint8x16_t cols_45 =
+ vreinterpretq_u8_s8
+ (vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+ uint8x16_t cols_23 =
+ vreinterpretq_u8_s8
+ (vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
+ uint8x16_t cols_67 =
+ vreinterpretq_u8_s8
+ (vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
- /* Transpose block ready for store. */
+ /* Transpose block to prepare for store. */
uint32x4x2_t cols_0415 = vzipq_u32(vreinterpretq_u32_u8(cols_01),
vreinterpretq_u32_u8(cols_45));
uint32x4x2_t cols_2637 = vzipq_u32(vreinterpretq_u32_u8(cols_23),
diff --git a/simd/arm/common/jidctint-neon.c b/simd/arm/jidctint-neon.c
similarity index 74%
rename from simd/arm/common/jidctint-neon.c
rename to simd/arm/jidctint-neon.c
index 0fd4a36..043b652 100644
--- a/simd/arm/common/jidctint-neon.c
+++ b/simd/arm/jidctint-neon.c
@@ -1,7 +1,8 @@
/*
- * jidctint-neon.c - slow IDCT (Arm NEON)
+ * jidctint-neon.c - accurate integer IDCT (Arm Neon)
*
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,16 +22,19 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "jconfigint.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
#include <arm_neon.h>
+
#define CONST_BITS 13
#define PASS1_BITS 2
@@ -38,7 +42,7 @@
#define DESCALE_P2 (CONST_BITS + PASS1_BITS + 3)
/* The computation of the inverse DCT requires the use of constants known at
- * compile-time. Scaled integer constants are used to avoid floating-point
+ * compile time. Scaled integer constants are used to avoid floating-point
* arithmetic:
* 0.298631336 = 2446 * 2^-13
* 0.390180644 = 3196 * 2^-13
@@ -76,19 +80,21 @@
#define F_2_053_MINUS_2_562 (F_2_053 - F_2_562)
#define F_0_541_PLUS_0_765 (F_0_541 + F_0_765)
+
ALIGN(16) static const int16_t jsimd_idct_islow_neon_consts[] = {
- F_0_899, F_0_541,
- F_2_562, F_0_298_MINUS_0_899,
- F_1_501_MINUS_0_899, F_2_053_MINUS_2_562,
- F_0_541_PLUS_0_765, F_1_175,
- F_1_175_MINUS_0_390, F_0_541_MINUS_1_847,
- F_3_072_MINUS_2_562, F_1_175_MINUS_1_961,
- 0, 0, 0, 0
- };
+ F_0_899, F_0_541,
+ F_2_562, F_0_298_MINUS_0_899,
+ F_1_501_MINUS_0_899, F_2_053_MINUS_2_562,
+ F_0_541_PLUS_0_765, F_1_175,
+ F_1_175_MINUS_0_390, F_0_541_MINUS_1_847,
+ F_3_072_MINUS_2_562, F_1_175_MINUS_1_961,
+ 0, 0, 0, 0
+};
-/* Forward declaration of regular and sparse IDCT helper functions. */
-static inline void jsimd_idct_islow_pass1_regular(int16x4_t row0,
+/* Forward declaration of regular and sparse IDCT helper functions */
+
+static INLINE void jsimd_idct_islow_pass1_regular(int16x4_t row0,
int16x4_t row1,
int16x4_t row2,
int16x4_t row3,
@@ -107,7 +113,7 @@
int16_t *workspace_1,
int16_t *workspace_2);
-static inline void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
+static INLINE void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
int16x4_t row1,
int16x4_t row2,
int16x4_t row3,
@@ -118,32 +124,33 @@
int16_t *workspace_1,
int16_t *workspace_2);
-static inline void jsimd_idct_islow_pass2_regular(int16_t *workspace,
+static INLINE void jsimd_idct_islow_pass2_regular(int16_t *workspace,
JSAMPARRAY output_buf,
JDIMENSION output_col,
unsigned buf_offset);
-static inline void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
+static INLINE void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
JSAMPARRAY output_buf,
JDIMENSION output_col,
unsigned buf_offset);
-/* Performs dequantization and inverse DCT on one block of coefficients. For
- * reference, the C implementation 'jpeg_idct_slow' can be found jidctint.c.
+/* Perform dequantization and inverse DCT on one block of coefficients. For
+ * reference, the C implementation (jpeg_idct_slow()) can be found in
+ * jidctint.c.
*
- * Optimization techniques used for data access:
+ * Optimization techniques used for fast data access:
*
- * In each pass, the inverse DCT is computed on the left and right 4x8 halves
- * of the DCT block. This avoids spilling due to register pressure and the
- * increased granularity allows an optimized calculation depending on the
- * values of the DCT coefficients. Between passes, intermediate data is stored
+ * In each pass, the inverse DCT is computed for the left and right 4x8 halves
+ * of the DCT block. This avoids spilling due to register pressure, and the
+ * increased granularity allows for an optimized calculation depending on the
+ * values of the DCT coefficients. Between passes, intermediate data is stored
* in 4x8 workspace buffers.
*
* Transposing the 8x8 DCT block after each pass can be achieved by transposing
- * each of the four 4x4 quadrants, and swapping quadrants 1 and 2 (in the
- * diagram below.) Swapping quadrants is cheap as the second pass can just load
- * from the other workspace buffer.
+ * each of the four 4x4 quadrants and swapping quadrants 1 and 2 (refer to the
+ * diagram below.) Swapping quadrants is cheap, since the second pass can just
+ * swap the workspace buffer pointers.
*
* +-------+-------+ +-------+-------+
* | | | | | |
@@ -158,32 +165,30 @@
* Optimization techniques used to accelerate the inverse DCT calculation:
*
* In a DCT coefficient block, the coefficients are increasingly likely to be 0
- * moving diagonally from top left to bottom right. If whole rows of
- * coefficients are 0, the inverse DCT calculation can be simplified. In this
- * NEON implementation, on the first pass of the inverse DCT, we test for three
- * special cases before defaulting to a full 'regular' inverse DCT:
+ * as you move diagonally from top left to bottom right. If whole rows of
+ * coefficients are 0, then the inverse DCT calculation can be simplified. On
+ * the first pass of the inverse DCT, we test for three special cases before
+ * defaulting to a full "regular" inverse DCT:
*
- * i) AC and DC coefficients are all zero. (Only tested for the right 4x8
- * half of the DCT coefficient block.) In this case the inverse DCT result
- * is all zero. We do no work here, signalling that the 'sparse' case is
- * required in the second pass.
- * ii) AC coefficients (all but the top row) are zero. In this case, the value
- * of the inverse DCT of the AC coefficients is just the DC coefficients.
- * iii) Coefficients of rows 4, 5, 6 and 7 are all zero. In this case we opt to
- * execute a 'sparse' simplified inverse DCT.
+ * 1) Coefficients in rows 4-7 are all zero. In this case, we perform a
+ * "sparse" simplified inverse DCT on rows 0-3.
+ * 2) AC coefficients (rows 1-7) are all zero. In this case, the inverse DCT
+ * result is equal to the dequantized DC coefficients.
+ * 3) AC and DC coefficients are all zero. In this case, the inverse DCT
+ * result is all zero. For the left 4x8 half, this is handled identically
+ * to Case 2 above. For the right 4x8 half, we do no work and signal that
+ * the "sparse" algorithm is required for the second pass.
*
- * In the second pass, only a single special case is tested: whether the the AC
- * and DC coefficients were all zero in the right 4x8 block in the first pass
- * (case 'i'). If this is the case, a 'sparse' variant of the second pass
- * inverse DCT is executed for both the left and right halves of the DCT block.
- * (The transposition after the first pass would have made the bottom half of
- * the block all zero.)
+ * In the second pass, only a single special case is tested: whether the AC and
+ * DC coefficients were all zero in the right 4x8 block during the first pass
+ * (refer to Case 3 above.) If this is the case, then a "sparse" variant of
+ * the second pass is performed for both the left and right halves of the DCT
+ * block. (The transposition after the first pass means that the right 4x8
+ * block during the first pass becomes rows 4-7 during the second pass.)
*/
-void jsimd_idct_islow_neon(void *dct_table,
- JCOEFPTR coef_block,
- JSAMPARRAY output_buf,
- JDIMENSION output_col)
+void jsimd_idct_islow_neon(void *dct_table, JCOEFPTR coef_block,
+ JSAMPARRAY output_buf, JDIMENSION output_col)
{
ISLOW_MULT_TYPE *quantptr = dct_table;
@@ -191,6 +196,7 @@
int16_t workspace_r[8 * DCTSIZE / 2];
/* Compute IDCT first pass on left 4x8 coefficient block. */
+
/* Load DCT coefficients in left 4x8 block. */
int16x4_t row0 = vld1_s16(coef_block + 0 * DCTSIZE);
int16x4_t row1 = vld1_s16(coef_block + 1 * DCTSIZE);
@@ -225,7 +231,7 @@
if (left_ac_bitmap == 0) {
int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS);
- int16x4x4_t quadrant = { dcval, dcval, dcval, dcval };
+ int16x4x4_t quadrant = { { dcval, dcval, dcval, dcval } };
/* Store 4x4 blocks to workspace, transposing in the process. */
vst4_s16(workspace_l, quadrant);
vst4_s16(workspace_r, quadrant);
@@ -242,8 +248,9 @@
workspace_l, workspace_r);
}
- /* Compute IDCT first pass on right 4x8 coefficient block.*/
- /* Load DCT coefficients for right 4x8 block. */
+ /* Compute IDCT first pass on right 4x8 coefficient block. */
+
+ /* Load DCT coefficients in right 4x8 block. */
row0 = vld1_s16(coef_block + 0 * DCTSIZE + 4);
row1 = vld1_s16(coef_block + 1 * DCTSIZE + 4);
row2 = vld1_s16(coef_block + 2 * DCTSIZE + 4);
@@ -273,7 +280,7 @@
bitmap = vorr_s16(bitmap, row1);
int64_t right_ac_bitmap = vget_lane_s64(vreinterpret_s64_s16(bitmap), 0);
- /* Initialise to non-zero value: defaults to regular second pass. */
+ /* If this remains non-zero, a "regular" second pass will be performed. */
int64_t right_ac_dc_bitmap = 1;
if (right_ac_bitmap == 0) {
@@ -282,7 +289,7 @@
if (right_ac_dc_bitmap != 0) {
int16x4_t dcval = vshl_n_s16(vmul_s16(row0, quant_row0), PASS1_BITS);
- int16x4x4_t quadrant = { dcval, dcval, dcval, dcval };
+ int16x4x4_t quadrant = { { dcval, dcval, dcval, dcval } };
/* Store 4x4 blocks to workspace, transposing in the process. */
vst4_s16(workspace_l + 4 * DCTSIZE / 2, quadrant);
vst4_s16(workspace_r + 4 * DCTSIZE / 2, quadrant);
@@ -304,7 +311,8 @@
}
/* Second pass: compute IDCT on rows in workspace. */
- /* If all coefficients in right 4x8 block are 0, use 'sparse' second pass. */
+
+ /* If all coefficients in right 4x8 block are 0, use "sparse" second pass. */
if (right_ac_dc_bitmap == 0) {
jsimd_idct_islow_pass2_sparse(workspace_l, output_buf, output_col, 0);
jsimd_idct_islow_pass2_sparse(workspace_r, output_buf, output_col, 4);
@@ -315,19 +323,19 @@
}
-/* Performs dequantization and the first pass of the slow-but-accurate inverse
- * DCT on a 4x8 block of coefficients. (To process the full 8x8 DCT block this
- * function - or some other optimized variant - needs to be called on both the
- * right and left 4x8 blocks.)
+/* Perform dequantization and the first pass of the accurate inverse DCT on a
+ * 4x8 block of coefficients. (To process the full 8x8 DCT block, this
+ * function-- or some other optimized variant-- needs to be called for both the
+ * left and right 4x8 blocks.)
*
- * This 'regular' version assumes that no optimization can be made to the IDCT
- * calculation since no useful set of AC coefficients are all 0.
+ * This "regular" version assumes that no optimization can be made to the IDCT
+ * calculation, since no useful set of AC coefficients is all 0.
*
- * The original C implementation of the slow IDCT 'jpeg_idct_slow' can be found
- * in jidctint.c. Algorithmic changes made here are documented inline.
+ * The original C implementation of the accurate IDCT (jpeg_idct_slow()) can be
+ * found in jidctint.c. Algorithmic changes made here are documented inline.
*/
-static inline void jsimd_idct_islow_pass1_regular(int16x4_t row0,
+static INLINE void jsimd_idct_islow_pass1_regular(int16x4_t row0,
int16x4_t row1,
int16x4_t row2,
int16x4_t row3,
@@ -346,10 +354,17 @@
int16_t *workspace_1,
int16_t *workspace_2)
{
- /* Load constants for IDCT calculation. */
+ /* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+ const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+ const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+ const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
- /* Even part. */
+ /* Even part */
int16x4_t z2_s16 = vmul_s16(row2, quant_row2);
int16x4_t z3_s16 = vmul_s16(row6, quant_row6);
@@ -369,7 +384,7 @@
int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
- /* Odd part. */
+ /* Odd part */
int16x4_t tmp0_s16 = vmul_s16(row7, quant_row7);
int16x4_t tmp1_s16 = vmul_s16(row5, quant_row5);
int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3);
@@ -378,7 +393,7 @@
z3_s16 = vadd_s16(tmp0_s16, tmp2_s16);
int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16);
- /* Implementation as per 'jpeg_idct_islow' in jidctint.c:
+ /* Implementation as per jpeg_idct_islow() in jidctint.c:
* z5 = (z3 + z4) * 1.175875602;
* z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
* z3 += z5; z4 += z5;
@@ -393,7 +408,7 @@
z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
- /* Implementation as per 'jpeg_idct_islow' in jidctint.c:
+ /* Implementation as per jpeg_idct_islow() in jidctint.c:
* z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
* tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
* tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
@@ -426,33 +441,36 @@
tmp3 = vaddq_s32(tmp3, z4);
/* Final output stage: descale and narrow to 16-bit. */
- int16x4x4_t rows_0123 = { vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
- };
- int16x4x4_t rows_4567 = { vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
- };
+ int16x4x4_t rows_0123 = { {
+ vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
+ } };
+ int16x4x4_t rows_4567 = { {
+ vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
+ } };
- /* Store 4x4 blocks to the intermediate workspace ready for second pass. */
- /* (VST4 transposes the blocks - we need to operate on rows in next pass.) */
+ /* Store 4x4 blocks to the intermediate workspace, ready for the second pass.
+ * (VST4 transposes the blocks. We need to operate on rows in the next
+ * pass.)
+ */
vst4_s16(workspace_1, rows_0123);
vst4_s16(workspace_2, rows_4567);
}
-/* Performs dequantization and the first pass of the slow-but-accurate inverse
- * DCT on a 4x8 block of coefficients.
+/* Perform dequantization and the first pass of the accurate inverse DCT on a
+ * 4x8 block of coefficients.
*
- * This 'sparse' version assumes that the AC coefficients in rows 4, 5, 6 and 7
- * are all 0. This simplifies the IDCT calculation, accelerating overall
- * performance.
+ * This "sparse" version assumes that the AC coefficients in rows 4-7 are all
+ * 0. This simplifies the IDCT calculation, accelerating overall performance.
*/
-static inline void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
+static INLINE void jsimd_idct_islow_pass1_sparse(int16x4_t row0,
int16x4_t row1,
int16x4_t row2,
int16x4_t row3,
@@ -464,11 +482,17 @@
int16_t *workspace_2)
{
/* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+ const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+ const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+ const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
- /* Even part. */
+ /* Even part (z3 is all 0) */
int16x4_t z2_s16 = vmul_s16(row2, quant_row2);
- /* z3 is all 0. */
int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
@@ -482,8 +506,7 @@
int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
- /* Odd part. */
- /* tmp0 and tmp1 are both all 0. */
+ /* Odd part (tmp0 and tmp1 are both all 0) */
int16x4_t tmp2_s16 = vmul_s16(row3, quant_row3);
int16x4_t tmp3_s16 = vmul_s16(row1, quant_row1);
@@ -501,46 +524,58 @@
tmp3 = vmlal_lane_s16(z4, tmp3_s16, consts.val[1], 0);
/* Final output stage: descale and narrow to 16-bit. */
- int16x4x4_t rows_0123 = { vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
- vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
- };
- int16x4x4_t rows_4567 = { vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
- vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
- };
+ int16x4x4_t rows_0123 = { {
+ vrshrn_n_s32(vaddq_s32(tmp10, tmp3), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp11, tmp2), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp12, tmp1), DESCALE_P1),
+ vrshrn_n_s32(vaddq_s32(tmp13, tmp0), DESCALE_P1)
+ } };
+ int16x4x4_t rows_4567 = { {
+ vrshrn_n_s32(vsubq_s32(tmp13, tmp0), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp12, tmp1), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp11, tmp2), DESCALE_P1),
+ vrshrn_n_s32(vsubq_s32(tmp10, tmp3), DESCALE_P1)
+ } };
- /* Store 4x4 blocks to the intermediate workspace ready for second pass. */
- /* (VST4 transposes the blocks - we need to operate on rows in next pass.) */
+ /* Store 4x4 blocks to the intermediate workspace, ready for the second pass.
+ * (VST4 transposes the blocks. We need to operate on rows in the next
+ * pass.)
+ */
vst4_s16(workspace_1, rows_0123);
vst4_s16(workspace_2, rows_4567);
}
-/* Performs the second pass of the slow-but-accurate inverse DCT on a 4x8 block
- * of coefficients. (To process the full 8x8 DCT block this function - or some
- * other optimized variant - needs to be called on both the right and left 4x8
+/* Perform the second pass of the accurate inverse DCT on a 4x8 block of
+ * coefficients. (To process the full 8x8 DCT block, this function-- or some
+ * other optimized variant-- needs to be called for both the right and left 4x8
* blocks.)
*
- * This 'regular' version assumes that no optimization can be made to the IDCT
- * calculation since no useful set of coefficient values are all 0 after the
+ * This "regular" version assumes that no optimization can be made to the IDCT
+ * calculation, since no useful set of coefficient values are all 0 after the
* first pass.
*
- * Again, the original C implementation of the slow IDCT 'jpeg_idct_slow' can
- * be found in jidctint.c. Algorithmic changes made here are documented inline.
+ * Again, the original C implementation of the accurate IDCT (jpeg_idct_slow())
+ * can be found in jidctint.c. Algorithmic changes made here are documented
+ * inline.
*/
-static inline void jsimd_idct_islow_pass2_regular(int16_t *workspace,
+static INLINE void jsimd_idct_islow_pass2_regular(int16_t *workspace,
JSAMPARRAY output_buf,
JDIMENSION output_col,
unsigned buf_offset)
{
/* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+ const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+ const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+ const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
- /* Even part. */
+ /* Even part */
int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2);
int16x4_t z3_s16 = vld1_s16(workspace + 6 * DCTSIZE / 2);
@@ -560,7 +595,7 @@
int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
- /* Odd part. */
+ /* Odd part */
int16x4_t tmp0_s16 = vld1_s16(workspace + 7 * DCTSIZE / 2);
int16x4_t tmp1_s16 = vld1_s16(workspace + 5 * DCTSIZE / 2);
int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2);
@@ -569,7 +604,7 @@
z3_s16 = vadd_s16(tmp0_s16, tmp2_s16);
int16x4_t z4_s16 = vadd_s16(tmp1_s16, tmp3_s16);
- /* Implementation as per 'jpeg_idct_islow' in jidctint.c:
+ /* Implementation as per jpeg_idct_islow() in jidctint.c:
* z5 = (z3 + z4) * 1.175875602;
* z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
* z3 += z5; z4 += z5;
@@ -584,7 +619,7 @@
z3 = vmlal_lane_s16(z3, z4_s16, consts.val[1], 3);
z4 = vmlal_lane_s16(z4, z4_s16, consts.val[2], 0);
- /* Implementation as per 'jpeg_idct_islow' in jidctint.c:
+ /* Implementation as per jpeg_idct_islow() in jidctint.c:
* z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
* tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
* tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
@@ -640,15 +675,17 @@
uint8x8_t cols_57_u8 = vadd_u8(vreinterpret_u8_s8(cols_57_s8),
vdup_n_u8(CENTERJSAMPLE));
- /* Transpose 4x8 block and store to memory. */
- /* Zipping adjacent columns together allows us to store 16-bit elements. */
+ /* Transpose 4x8 block and store to memory. (Zipping adjacent columns
+ * together allows us to store 16-bit elements.)
+ */
uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8);
uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8);
- uint16x4x4_t cols_01_23_45_67 = { vreinterpret_u16_u8(cols_01_23.val[0]),
- vreinterpret_u16_u8(cols_01_23.val[1]),
- vreinterpret_u16_u8(cols_45_67.val[0]),
- vreinterpret_u16_u8(cols_45_67.val[1])
- };
+ uint16x4x4_t cols_01_23_45_67 = { {
+ vreinterpret_u16_u8(cols_01_23.val[0]),
+ vreinterpret_u16_u8(cols_01_23.val[1]),
+ vreinterpret_u16_u8(cols_45_67.val[0]),
+ vreinterpret_u16_u8(cols_45_67.val[1])
+ } };
JSAMPROW outptr0 = output_buf[buf_offset + 0] + output_col;
JSAMPROW outptr1 = output_buf[buf_offset + 1] + output_col;
@@ -662,25 +699,31 @@
}
-/* Performs the second pass of the slow-but-accurate inverse DCT on a 4x8 block
+/* Performs the second pass of the accurate inverse DCT on a 4x8 block
* of coefficients.
*
- * This 'sparse' version assumes that the coefficient values (after the first
- * pass) in rows 4, 5, 6 and 7 are all 0. This simplifies the IDCT calculation,
+ * This "sparse" version assumes that the coefficient values (after the first
+ * pass) in rows 4-7 are all 0. This simplifies the IDCT calculation,
* accelerating overall performance.
*/
-static inline void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
+static INLINE void jsimd_idct_islow_pass2_sparse(int16_t *workspace,
JSAMPARRAY output_buf,
JDIMENSION output_col,
unsigned buf_offset)
{
/* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_islow_neon_consts);
+#else
+ const int16x4_t consts1 = vld1_s16(jsimd_idct_islow_neon_consts);
+ const int16x4_t consts2 = vld1_s16(jsimd_idct_islow_neon_consts + 4);
+ const int16x4_t consts3 = vld1_s16(jsimd_idct_islow_neon_consts + 8);
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
- /* Even part. */
+ /* Even part (z3 is all 0) */
int16x4_t z2_s16 = vld1_s16(workspace + 2 * DCTSIZE / 2);
- /* z3 is all 0. */
int32x4_t tmp2 = vmull_lane_s16(z2_s16, consts.val[0], 1);
int32x4_t tmp3 = vmull_lane_s16(z2_s16, consts.val[1], 2);
@@ -694,8 +737,7 @@
int32x4_t tmp11 = vaddq_s32(tmp1, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp1, tmp2);
- /* Odd part. */
- /* tmp0 and tmp1 are both all 0. */
+ /* Odd part (tmp0 and tmp1 are both all 0) */
int16x4_t tmp2_s16 = vld1_s16(workspace + 3 * DCTSIZE / 2);
int16x4_t tmp3_s16 = vld1_s16(workspace + 1 * DCTSIZE / 2);
@@ -736,15 +778,17 @@
uint8x8_t cols_57_u8 = vadd_u8(vreinterpret_u8_s8(cols_57_s8),
vdup_n_u8(CENTERJSAMPLE));
- /* Transpose 4x8 block and store to memory. */
- /* Zipping adjacent columns together allow us to store 16-bit elements. */
+ /* Transpose 4x8 block and store to memory. (Zipping adjacent columns
+ * together allows us to store 16-bit elements.)
+ */
uint8x8x2_t cols_01_23 = vzip_u8(cols_02_u8, cols_13_u8);
uint8x8x2_t cols_45_67 = vzip_u8(cols_46_u8, cols_57_u8);
- uint16x4x4_t cols_01_23_45_67 = { vreinterpret_u16_u8(cols_01_23.val[0]),
- vreinterpret_u16_u8(cols_01_23.val[1]),
- vreinterpret_u16_u8(cols_45_67.val[0]),
- vreinterpret_u16_u8(cols_45_67.val[1])
- };
+ uint16x4x4_t cols_01_23_45_67 = { {
+ vreinterpret_u16_u8(cols_01_23.val[0]),
+ vreinterpret_u16_u8(cols_01_23.val[1]),
+ vreinterpret_u16_u8(cols_45_67.val[0]),
+ vreinterpret_u16_u8(cols_45_67.val[1])
+ } };
JSAMPROW outptr0 = output_buf[buf_offset + 0] + output_col;
JSAMPROW outptr1 = output_buf[buf_offset + 1] + output_col;
diff --git a/simd/arm/common/jidctred-neon.c b/simd/arm/jidctred-neon.c
similarity index 79%
rename from simd/arm/common/jidctred-neon.c
rename to simd/arm/jidctred-neon.c
index ed4232c..be9627e 100644
--- a/simd/arm/common/jidctred-neon.c
+++ b/simd/arm/jidctred-neon.c
@@ -1,7 +1,8 @@
/*
- * jidctred-neon.c - reduced-size IDCT (Arm NEON)
+ * jidctred-neon.c - reduced-size IDCT (Arm Neon)
*
- * Copyright 2019 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,16 +22,18 @@
*/
#define JPEG_INTERNALS
-#include "../../../jconfigint.h"
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
+#include "align.h"
+#include "neon-compat.h"
#include <arm_neon.h>
+
#define CONST_BITS 13
#define PASS1_BITS 2
@@ -49,10 +52,10 @@
#define F_2_562 20995
#define F_3_624 29692
-/*
- * 'jsimd_idct_2x2_neon' is an inverse-DCT function for getting reduced-size
- * 2x2 pixels output from an 8x8 DCT block. It uses the same calculations and
- * produces exactly the same output as IJG's original 'jpeg_idct_2x2' function
+
+/* jsimd_idct_2x2_neon() is an inverse DCT function that produces reduced-size
+ * 2x2 output from an 8x8 DCT block. It uses the same calculations and
+ * produces exactly the same output as IJG's original jpeg_idct_2x2() function
* from jpeg-6b, which can be found in jidctred.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
@@ -61,20 +64,17 @@
* 1.272758580 = 10426 * 2^-13
* 3.624509785 = 29692 * 2^-13
*
- * See jidctred.c for further details of the 2x2 reduced IDCT algorithm. Where
- * possible, the variable names and comments here in 'jsimd_idct_2x2_neon'
- * match up with those in 'jpeg_idct_2x2'.
- *
- * NOTE: jpeg-8 has an improved implementation of the 2x2 inverse-DCT which
- * requires fewer arithmetic operations and hence should be faster. The
- * primary purpose of this particular NEON optimized function is bit
- * exact compatibility with jpeg-6b.
+ * See jidctred.c for further details of the 2x2 IDCT algorithm. Where
+ * possible, the variable names and comments here in jsimd_idct_2x2_neon()
+ * match up with those in jpeg_idct_2x2().
*/
-void jsimd_idct_2x2_neon(void *dct_table,
- JCOEFPTR coef_block,
- JSAMPARRAY restrict output_buf,
- JDIMENSION output_col)
+ALIGN(16) static const int16_t jsimd_idct_2x2_neon_consts[] = {
+ -F_0_720, F_0_850, -F_1_272, F_3_624
+};
+
+void jsimd_idct_2x2_neon(void *dct_table, JCOEFPTR coef_block,
+ JSAMPARRAY output_buf, JDIMENSION output_col)
{
ISLOW_MULT_TYPE *quantptr = dct_table;
@@ -85,7 +85,7 @@
int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
- /* Load DCT quantization table. */
+ /* Load quantization table values. */
int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
@@ -99,20 +99,26 @@
row5 = vmulq_s16(row5, quant_row5);
row7 = vmulq_s16(row7, quant_row7);
- /* Pass 1: process input columns; put results in vectors row0 and row1. */
- /* Even part. */
+ /* Load IDCT conversion constants. */
+ const int16x4_t consts = vld1_s16(jsimd_idct_2x2_neon_consts);
+
+ /* Pass 1: process columns from input, put results in vectors row0 and
+ * row1.
+ */
+
+ /* Even part */
int32x4_t tmp10_l = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 2);
int32x4_t tmp10_h = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 2);
- /* Odd part. */
- int32x4_t tmp0_l = vmull_n_s16(vget_low_s16(row1), F_3_624);
- tmp0_l = vmlal_n_s16(tmp0_l, vget_low_s16(row3), -F_1_272);
- tmp0_l = vmlal_n_s16(tmp0_l, vget_low_s16(row5), F_0_850);
- tmp0_l = vmlal_n_s16(tmp0_l, vget_low_s16(row7), -F_0_720);
- int32x4_t tmp0_h = vmull_n_s16(vget_high_s16(row1), F_3_624);
- tmp0_h = vmlal_n_s16(tmp0_h, vget_high_s16(row3), -F_1_272);
- tmp0_h = vmlal_n_s16(tmp0_h, vget_high_s16(row5), F_0_850);
- tmp0_h = vmlal_n_s16(tmp0_h, vget_high_s16(row7), -F_0_720);
+ /* Odd part */
+ int32x4_t tmp0_l = vmull_lane_s16(vget_low_s16(row1), consts, 3);
+ tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row3), consts, 2);
+ tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row5), consts, 1);
+ tmp0_l = vmlal_lane_s16(tmp0_l, vget_low_s16(row7), consts, 0);
+ int32x4_t tmp0_h = vmull_lane_s16(vget_high_s16(row1), consts, 3);
+ tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row3), consts, 2);
+ tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row5), consts, 1);
+ tmp0_h = vmlal_lane_s16(tmp0_h, vget_high_s16(row7), consts, 0);
/* Final output stage: descale and narrow to 16-bit. */
row0 = vcombine_s16(vrshrn_n_s32(vaddq_s32(tmp10_l, tmp0_l), CONST_BITS),
@@ -120,7 +126,7 @@
row1 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp10_l, tmp0_l), CONST_BITS),
vrshrn_n_s32(vsubq_s32(tmp10_h, tmp0_h), CONST_BITS));
- /* Transpose two rows ready for second pass. */
+ /* Transpose two rows, ready for second pass. */
int16x8x2_t cols_0246_1357 = vtrnq_s16(row0, row1);
int16x8_t cols_0246 = cols_0246_1357.val[0];
int16x8_t cols_1357 = cols_0246_1357.val[1];
@@ -130,15 +136,18 @@
int16x8_t cols_1155 = vreinterpretq_s16_s32(cols_1155_3377.val[0]);
int16x8_t cols_3377 = vreinterpretq_s16_s32(cols_1155_3377.val[1]);
- /* Pass 2: process 2 rows, store to output array. */
- /* Even part: only interested in col0; top half of tmp10 is "don't care". */
+ /* Pass 2: process two rows, store to output array. */
+
+ /* Even part: we're only interested in col0; the top half of tmp10 is "don't
+ * care."
+ */
int32x4_t tmp10 = vshll_n_s16(vget_low_s16(cols_0246), CONST_BITS + 2);
- /* Odd part. Only interested in bottom half of tmp0. */
- int32x4_t tmp0 = vmull_n_s16(vget_low_s16(cols_1155), F_3_624);
- tmp0 = vmlal_n_s16(tmp0, vget_low_s16(cols_3377), -F_1_272);
- tmp0 = vmlal_n_s16(tmp0, vget_high_s16(cols_1155), F_0_850);
- tmp0 = vmlal_n_s16(tmp0, vget_high_s16(cols_3377), -F_0_720);
+ /* Odd part: we're only interested in the bottom half of tmp0. */
+ int32x4_t tmp0 = vmull_lane_s16(vget_low_s16(cols_1155), consts, 3);
+ tmp0 = vmlal_lane_s16(tmp0, vget_low_s16(cols_3377), consts, 2);
+ tmp0 = vmlal_lane_s16(tmp0, vget_high_s16(cols_1155), consts, 1);
+ tmp0 = vmlal_lane_s16(tmp0, vget_high_s16(cols_3377), consts, 0);
/* Final output stage: descale and clamp to range [0-255]. */
int16x8_t output_s16 = vcombine_s16(vaddhn_s32(tmp10, tmp0),
@@ -156,10 +165,9 @@
}
-/*
- * 'jsimd_idct_4x4_neon' is an inverse-DCT function for getting reduced-size
- * 4x4 pixels output from an 8x8 DCT block. It uses the same calculations and
- * produces exactly the same output as IJG's original 'jpeg_idct_4x4' function
+/* jsimd_idct_4x4_neon() is an inverse DCT function that produces reduced-size
+ * 4x4 output from an 8x8 DCT block. It uses the same calculations and
+ * produces exactly the same output as IJG's original jpeg_idct_4x4() function
* from jpeg-6b, which can be found in jidctred.c.
*
* Scaled integer constants are used to avoid floating-point arithmetic:
@@ -174,26 +182,19 @@
* 2.172734803 = 17799 * 2^-13
* 2.562915447 = 20995 * 2^-13
*
- * See jidctred.c for further details of the 4x4 reduced IDCT algorithm. Where
- * possible, the variable names and comments here in 'jsimd_idct_4x4_neon'
- * match up with those in 'jpeg_idct_4x4'.
- *
- * NOTE: jpeg-8 has an improved implementation of the 4x4 inverse-DCT which
- * requires fewer arithmetic operations and hence should be faster. The
- * primary purpose of this particular NEON optimized function is bit
- * exact compatibility with jpeg-6b.
+ * See jidctred.c for further details of the 4x4 IDCT algorithm. Where
+ * possible, the variable names and comments here in jsimd_idct_4x4_neon()
+ * match up with those in jpeg_idct_4x4().
*/
ALIGN(16) static const int16_t jsimd_idct_4x4_neon_consts[] = {
- F_1_847, -F_0_765, -F_0_211, F_1_451,
- -F_2_172, F_1_061, -F_0_509, -F_0_601,
- F_0_899, F_2_562, 0, 0
- };
+ F_1_847, -F_0_765, -F_0_211, F_1_451,
+ -F_2_172, F_1_061, -F_0_509, -F_0_601,
+ F_0_899, F_2_562, 0, 0
+};
-void jsimd_idct_4x4_neon(void *dct_table,
- JCOEFPTR coef_block,
- JSAMPARRAY restrict output_buf,
- JDIMENSION output_col)
+void jsimd_idct_4x4_neon(void *dct_table, JCOEFPTR coef_block,
+ JSAMPARRAY output_buf, JDIMENSION output_col)
{
ISLOW_MULT_TYPE *quantptr = dct_table;
@@ -207,7 +208,7 @@
int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
/* Load quantization table values for DC coefficients. */
- int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
+ int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
/* Dequantize DC coefficients. */
row0 = vmulq_s16(row0, quant_row0);
@@ -222,22 +223,33 @@
int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
/* Load constants for IDCT computation. */
+#ifdef HAVE_VLD1_S16_X3
const int16x4x3_t consts = vld1_s16_x3(jsimd_idct_4x4_neon_consts);
+#else
+ /* GCC does not currently support the intrinsic vld1_<type>_x3(). */
+ const int16x4_t consts1 = vld1_s16(jsimd_idct_4x4_neon_consts);
+ const int16x4_t consts2 = vld1_s16(jsimd_idct_4x4_neon_consts + 4);
+ const int16x4_t consts3 = vld1_s16(jsimd_idct_4x4_neon_consts + 8);
+ const int16x4x3_t consts = { { consts1, consts2, consts3 } };
+#endif
if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
- /* All AC coefficients are zero. */
- /* Compute DC values and duplicate into row vectors 0, 1, 2 and 3. */
+ /* All AC coefficients are zero.
+ * Compute DC values and duplicate into row vectors 0, 1, 2, and 3.
+ */
int16x8_t dcval = vshlq_n_s16(row0, PASS1_BITS);
row0 = dcval;
row1 = dcval;
row2 = dcval;
row3 = dcval;
} else if (left_ac_bitmap == 0) {
- /* AC coefficients are zero for columns 0, 1, 2 and 3. */
- /* Compute DC values for these columns. */
+ /* AC coefficients are zero for columns 0, 1, 2, and 3.
+ * Compute DC values for these columns.
+ */
int16x4_t dcval = vshl_n_s16(vget_low_s16(row0), PASS1_BITS);
- /* Commence regular IDCT computation for columns 4, 5, 6 and 7. */
+ /* Commence regular IDCT computation for columns 4, 5, 6, and 7. */
+
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
@@ -246,7 +258,7 @@
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
- /* Even part. */
+ /* Even part */
int32x4_t tmp0 = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 1);
int16x4_t z2 = vmul_s16(vget_high_s16(row2), quant_row2);
@@ -258,7 +270,7 @@
int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
- /* Odd part. */
+ /* Odd part */
int16x4_t z1 = vmul_s16(vget_high_s16(row7), quant_row7);
z2 = vmul_s16(vget_high_s16(row5), quant_row5);
z3 = vmul_s16(vget_high_s16(row3), quant_row3);
@@ -284,11 +296,13 @@
row2 = vcombine_s16(dcval, vrshrn_n_s32(vsubq_s32(tmp12, tmp0),
CONST_BITS - PASS1_BITS + 1));
} else if (right_ac_bitmap == 0) {
- /* AC coefficients are zero for columns 4, 5, 6 and 7. */
- /* Compute DC values for these columns. */
+ /* AC coefficients are zero for columns 4, 5, 6, and 7.
+ * Compute DC values for these columns.
+ */
int16x4_t dcval = vshl_n_s16(vget_high_s16(row0), PASS1_BITS);
- /* Commence regular IDCT computation for columns 0, 1, 2 and 3. */
+ /* Commence regular IDCT computation for columns 0, 1, 2, and 3. */
+
/* Load quantization table. */
int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
@@ -297,7 +311,7 @@
int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
- /* Even part. */
+ /* Even part */
int32x4_t tmp0 = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 1);
int16x4_t z2 = vmul_s16(vget_low_s16(row2), quant_row2);
@@ -309,7 +323,7 @@
int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
- /* Odd part. */
+ /* Odd part */
int16x4_t z1 = vmul_s16(vget_low_s16(row7), quant_row7);
z2 = vmul_s16(vget_low_s16(row5), quant_row5);
z3 = vmul_s16(vget_low_s16(row3), quant_row3);
@@ -343,7 +357,7 @@
int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
- /* Even part. */
+ /* Even part */
int32x4_t tmp0_l = vshll_n_s16(vget_low_s16(row0), CONST_BITS + 1);
int32x4_t tmp0_h = vshll_n_s16(vget_high_s16(row0), CONST_BITS + 1);
@@ -360,7 +374,7 @@
int32x4_t tmp12_l = vsubq_s32(tmp0_l, tmp2_l);
int32x4_t tmp12_h = vsubq_s32(tmp0_h, tmp2_h);
- /* Odd part. */
+ /* Odd part */
int16x8_t z1 = vmulq_s16(row7, quant_row7);
z2 = vmulq_s16(row5, quant_row5);
z3 = vmulq_s16(row3, quant_row3);
@@ -421,7 +435,8 @@
int16x4_t col7 = vreinterpret_s16_s32(vget_high_s32(cols_1537.val[1]));
/* Commence second pass of IDCT. */
- /* Even part. */
+
+ /* Even part */
int32x4_t tmp0 = vshll_n_s16(col0, CONST_BITS + 1);
int32x4_t tmp2 = vmull_lane_s16(col2, consts.val[0], 0);
tmp2 = vmlal_lane_s16(tmp2, col6, consts.val[0], 1);
@@ -429,7 +444,7 @@
int32x4_t tmp10 = vaddq_s32(tmp0, tmp2);
int32x4_t tmp12 = vsubq_s32(tmp0, tmp2);
- /* Odd part. */
+ /* Odd part */
tmp0 = vmull_lane_s16(col7, consts.val[0], 2);
tmp0 = vmlal_lane_s16(tmp0, col5, consts.val[0], 3);
tmp0 = vmlal_lane_s16(tmp0, col3, consts.val[1], 0);
@@ -449,13 +464,15 @@
CONST_BITS + PASS1_BITS + 3 + 1 - 16);
output_cols_13 = vrsraq_n_s16(vdupq_n_s16(CENTERJSAMPLE), output_cols_13,
CONST_BITS + PASS1_BITS + 3 + 1 - 16);
- /* Narrow to 8-bit and convert to unsigned while zipping 8-bit elements. */
- /* Interleaving store completes the transpose. */
+ /* Narrow to 8-bit and convert to unsigned while zipping 8-bit elements.
+ * An interleaving store completes the transpose.
+ */
uint8x8x2_t output_0123 = vzip_u8(vqmovun_s16(output_cols_02),
vqmovun_s16(output_cols_13));
- uint16x4x2_t output_01_23 = { vreinterpret_u16_u8(output_0123.val[0]),
- vreinterpret_u16_u8(output_0123.val[1])
- };
+ uint16x4x2_t output_01_23 = { {
+ vreinterpret_u16_u8(output_0123.val[0]),
+ vreinterpret_u16_u8(output_0123.val[1])
+ } };
/* Store 4x4 block to memory. */
JSAMPROW outptr0 = output_buf[0] + output_col;
diff --git a/simd/arm/common/jquanti-neon.c b/simd/arm/jquanti-neon.c
similarity index 75%
rename from simd/arm/common/jquanti-neon.c
rename to simd/arm/jquanti-neon.c
index 6f8a3ab..a7eb6f1 100644
--- a/simd/arm/common/jquanti-neon.c
+++ b/simd/arm/jquanti-neon.c
@@ -1,7 +1,7 @@
/*
- * jquanti-neon.c - sample conversion and integer quantization (Arm NEON)
+ * jquanti-neon.c - sample data conversion and quantization (Arm Neon)
*
- * Copyright 2020 The Chromium Authors. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
*
* This software is provided 'as-is', without any express or implied
* warranty. In no event will the authors be held liable for any damages
@@ -21,28 +21,28 @@
*/
#define JPEG_INTERNALS
-#include "../../../jinclude.h"
-#include "../../../jpeglib.h"
-#include "../../../jsimd.h"
-#include "../../../jdct.h"
-#include "../../../jsimddct.h"
+#include "../../jinclude.h"
+#include "../../jpeglib.h"
#include "../../jsimd.h"
+#include "../../jdct.h"
+#include "../../jsimddct.h"
+#include "../jsimd.h"
#include <arm_neon.h>
-/*
- * Pixel channel sample values have range [0,255]. The Discrete Cosine
- * Transform (DCT) operates on values centered around 0.
+
+/* After downsampling, the resulting sample values are in the range [0, 255],
+ * but the Discrete Cosine Transform (DCT) operates on values centered around
+ * 0.
*
* To prepare sample values for the DCT, load samples into a DCT workspace,
- * subtracting CENTREJSAMPLE (128). The samples, now in range [-128, 127],
+ * subtracting CENTERJSAMPLE (128). The samples, now in the range [-128, 127],
* are also widened from 8- to 16-bit.
*
- * The equivalent scalar C function 'convsamp' can be found in jcdctmgr.c.
+ * The equivalent scalar C function convsamp() can be found in jcdctmgr.c.
*/
-void jsimd_convsamp_neon(JSAMPARRAY sample_data,
- JDIMENSION start_col,
+void jsimd_convsamp_neon(JSAMPARRAY sample_data, JDIMENSION start_col,
DCTELEM *workspace)
{
uint8x8_t samp_row0 = vld1_u8(sample_data[0] + start_col);
@@ -54,22 +54,22 @@
uint8x8_t samp_row6 = vld1_u8(sample_data[6] + start_col);
uint8x8_t samp_row7 = vld1_u8(sample_data[7] + start_col);
- int16x8_t row0 = vreinterpretq_s16_u16(vsubl_u8(samp_row0,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row1 = vreinterpretq_s16_u16(vsubl_u8(samp_row1,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row2 = vreinterpretq_s16_u16(vsubl_u8(samp_row2,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row3 = vreinterpretq_s16_u16(vsubl_u8(samp_row3,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row4 = vreinterpretq_s16_u16(vsubl_u8(samp_row4,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row5 = vreinterpretq_s16_u16(vsubl_u8(samp_row5,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row6 = vreinterpretq_s16_u16(vsubl_u8(samp_row6,
- vdup_n_u8(CENTERJSAMPLE)));
- int16x8_t row7 = vreinterpretq_s16_u16(vsubl_u8(samp_row7,
- vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row0 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row0, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row1 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row1, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row2 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row2, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row3 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row3, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row4 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row4, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row5 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row5, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row6 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row6, vdup_n_u8(CENTERJSAMPLE)));
+ int16x8_t row7 =
+ vreinterpretq_s16_u16(vsubl_u8(samp_row7, vdup_n_u8(CENTERJSAMPLE)));
vst1q_s16(workspace + 0 * DCTSIZE, row0);
vst1q_s16(workspace + 1 * DCTSIZE, row1);
@@ -82,26 +82,25 @@
}
-/*
- * After the DCT, the resulting coefficient values need to be divided by a
- * quantization value.
+/* After the DCT, the resulting array of coefficient values needs to be divided
+ * by an array of quantization values.
*
* To avoid a slow division operation, the DCT coefficients are multiplied by
- * the (scaled) reciprocal of the quantization values and then right-shifted.
+ * the (scaled) reciprocals of the quantization values and then right-shifted.
*
- * The equivalent scalar C function 'quantize' can be found in jcdctmgr.c.
+ * The equivalent scalar C function quantize() can be found in jcdctmgr.c.
*/
-void jsimd_quantize_neon(JCOEFPTR coef_block,
- DCTELEM *divisors,
+void jsimd_quantize_neon(JCOEFPTR coef_block, DCTELEM *divisors,
DCTELEM *workspace)
{
JCOEFPTR out_ptr = coef_block;
UDCTELEM *recip_ptr = (UDCTELEM *)divisors;
UDCTELEM *corr_ptr = (UDCTELEM *)divisors + DCTSIZE2;
DCTELEM *shift_ptr = divisors + 3 * DCTSIZE2;
+ int i;
- for (int i = 0; i < DCTSIZE; i += DCTSIZE / 2) {
+ for (i = 0; i < DCTSIZE; i += DCTSIZE / 2) {
/* Load DCT coefficients. */
int16x8_t row0 = vld1q_s16(workspace + (i + 0) * DCTSIZE);
int16x8_t row1 = vld1q_s16(workspace + (i + 1) * DCTSIZE);
@@ -137,7 +136,7 @@
abs_row2 = vaddq_u16(abs_row2, corr2);
abs_row3 = vaddq_u16(abs_row3, corr3);
- /* Multiply DCT coefficients by quantization reciprocal. */
+ /* Multiply DCT coefficients by quantization reciprocals. */
int32x4_t row0_l = vreinterpretq_s32_u32(vmull_u16(vget_low_u16(abs_row0),
vget_low_u16(recip0)));
int32x4_t row0_h = vreinterpretq_s32_u32(vmull_u16(vget_high_u16(abs_row0),
@@ -160,8 +159,9 @@
row2 = vcombine_s16(vshrn_n_s32(row2_l, 16), vshrn_n_s32(row2_h, 16));
row3 = vcombine_s16(vshrn_n_s32(row3_l, 16), vshrn_n_s32(row3_h, 16));
- /* Since VSHR only supports an immediate as its second argument, negate */
- /* the shift value and shift left. */
+ /* Since VSHR only supports an immediate as its second argument, negate the
+ * shift value and shift left.
+ */
row0 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row0),
vnegq_s16(shift0)));
row1 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row1),
diff --git a/simd/arm/neon-compat.h b/simd/arm/neon-compat.h
new file mode 100644
index 0000000..3ce3bcb
--- /dev/null
+++ b/simd/arm/neon-compat.h
@@ -0,0 +1,37 @@
+/*
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#if defined(__clang__) || defined(_MSC_VER)
+#define HAVE_VLD1_S16_X3
+#define HAVE_VLD1_U16_X2
+#define HAVE_VLD1Q_U8_X4
+#endif
+
+/* Define compiler-independent count-leading-zeros macros */
+#if defined(_MSC_VER) && !defined(__clang__)
+#define BUILTIN_CLZ(x) _CountLeadingZeros(x)
+#define BUILTIN_CLZL(x) _CountLeadingZeros64(x)
+#elif defined(__clang__) || defined(__GNUC__)
+#define BUILTIN_CLZ(x) __builtin_clz(x)
+#define BUILTIN_CLZL(x) __builtin_clzl(x)
+#else
+#error "Unknown compiler"
+#endif
diff --git a/simd/arm/neon-compat.h.in b/simd/arm/neon-compat.h.in
new file mode 100644
index 0000000..e2347b9
--- /dev/null
+++ b/simd/arm/neon-compat.h.in
@@ -0,0 +1,35 @@
+/*
+ * Copyright (C) 2020, D. R. Commander. All Rights Reserved.
+ * Copyright (C) 2020, Arm Limited. All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#cmakedefine HAVE_VLD1_S16_X3
+#cmakedefine HAVE_VLD1_U16_X2
+#cmakedefine HAVE_VLD1Q_U8_X4
+
+/* Define compiler-independent count-leading-zeros macros */
+#if defined(_MSC_VER) && !defined(__clang__)
+#define BUILTIN_CLZ(x) _CountLeadingZeros(x)
+#define BUILTIN_CLZL(x) _CountLeadingZeros64(x)
+#elif defined(__clang__) || defined(__GNUC__)
+#define BUILTIN_CLZ(x) __builtin_clz(x)
+#define BUILTIN_CLZL(x) __builtin_clzl(x)
+#else
+#error "Unknown compiler"
+#endif
diff --git a/simd/gas-preprocessor.in b/simd/gas-preprocessor.in
deleted file mode 100755
index 560f788..0000000
--- a/simd/gas-preprocessor.in
+++ /dev/null
@@ -1 +0,0 @@
-gas-preprocessor.pl @CMAKE_ASM_COMPILER@ ${1+"$@"}
diff --git a/simd/i386/jchuff-sse2.asm b/simd/i386/jchuff-sse2.asm
index d0112e6..278cf5e 100644
--- a/simd/i386/jchuff-sse2.asm
+++ b/simd/i386/jchuff-sse2.asm
@@ -1,8 +1,9 @@
;
; jchuff-sse2.asm - Huffman entropy encoding (SSE2)
;
-; Copyright (C) 2009-2011, 2014-2017, D. R. Commander.
+; Copyright (C) 2009-2011, 2014-2017, 2019, D. R. Commander.
; Copyright (C) 2015, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -15,133 +16,255 @@
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
; This file contains an SSE2 implementation for Huffman coding of one block.
-; The following code is based directly on jchuff.c; see jchuff.c for more
-; details.
+; The following code is based on jchuff.c; see jchuff.c for more details.
%include "jsimdext.inc"
+struc working_state
+.next_output_byte: resp 1 ; => next byte to write in buffer
+.free_in_buffer: resp 1 ; # of byte spaces remaining in buffer
+.cur.put_buffer.simd resq 1 ; current bit accumulation buffer
+.cur.free_bits resd 1 ; # of bits available in it
+.cur.last_dc_val resd 4 ; last DC coef for each component
+.cinfo: resp 1 ; dump_buffer needs access to this
+endstruc
+
+struc c_derived_tbl
+.ehufco: resd 256 ; code for each symbol
+.ehufsi: resb 256 ; length of code for each symbol
+; If no code has been allocated for a symbol S, ehufsi[S] contains 0
+endstruc
+
; --------------------------------------------------------------------------
SECTION SEG_CONST
- alignz 32
GLOBAL_DATA(jconst_huff_encode_one_block)
- EXTERN EXTN(jpeg_nbits_table)
EXTN(jconst_huff_encode_one_block):
alignz 32
+jpeg_mask_bits dq 0x0000, 0x0001, 0x0003, 0x0007
+ dq 0x000f, 0x001f, 0x003f, 0x007f
+ dq 0x00ff, 0x01ff, 0x03ff, 0x07ff
+ dq 0x0fff, 0x1fff, 0x3fff, 0x7fff
+
+times 1 << 14 db 15
+times 1 << 13 db 14
+times 1 << 12 db 13
+times 1 << 11 db 12
+times 1 << 10 db 11
+times 1 << 9 db 10
+times 1 << 8 db 9
+times 1 << 7 db 8
+times 1 << 6 db 7
+times 1 << 5 db 6
+times 1 << 4 db 5
+times 1 << 3 db 4
+times 1 << 2 db 3
+times 1 << 1 db 2
+times 1 << 0 db 1
+times 1 db 0
+jpeg_nbits_table:
+times 1 db 0
+times 1 << 0 db 1
+times 1 << 1 db 2
+times 1 << 2 db 3
+times 1 << 3 db 4
+times 1 << 4 db 5
+times 1 << 5 db 6
+times 1 << 6 db 7
+times 1 << 7 db 8
+times 1 << 8 db 9
+times 1 << 9 db 10
+times 1 << 10 db 11
+times 1 << 11 db 12
+times 1 << 12 db 13
+times 1 << 13 db 14
+times 1 << 14 db 15
+
+ alignz 32
+
+%ifdef PIC
+%define NBITS(x) nbits_base + x
+%else
+%define NBITS(x) jpeg_nbits_table + x
+%endif
+%define MASK_BITS(x) NBITS((x) * 8) + (jpeg_mask_bits - jpeg_nbits_table)
+
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 32
-; These macros perform the same task as the emit_bits() function in the
-; original libjpeg code. In addition to reducing overhead by explicitly
-; inlining the code, additional performance is achieved by taking into
-; account the size of the bit buffer and waiting until it is almost full
-; before emptying it. This mostly benefits 64-bit platforms, since 6
-; bytes can be stored in a 64-bit bit buffer before it has to be emptied.
+%define mm_put_buffer mm0
+%define mm_all_0xff mm1
+%define mm_temp mm2
+%define mm_nbits mm3
+%define mm_code_bits mm3
+%define mm_code mm4
+%define mm_overflow_bits mm5
+%define mm_save_nbits mm6
-%macro EMIT_BYTE 0
- sub put_bits, 8 ; put_bits -= 8;
- mov edx, put_buffer
- mov ecx, put_bits
- shr edx, cl ; c = (JOCTET)GETJOCTET(put_buffer >> put_bits);
- mov byte [eax], dl ; *buffer++ = c;
- add eax, 1
- cmp dl, 0xFF ; need to stuff a zero byte?
- jne %%.EMIT_BYTE_END
- mov byte [eax], 0 ; *buffer++ = 0;
- add eax, 1
-%%.EMIT_BYTE_END:
-%endmacro
+; Shorthand used to describe SIMD operations:
+; wN: xmmN treated as eight signed 16-bit values
+; wN[i]: perform the same operation on all eight signed 16-bit values, i=0..7
+; bN: xmmN treated as 16 unsigned 8-bit values, or
+; mmN treated as eight unsigned 8-bit values
+; bN[i]: perform the same operation on all unsigned 8-bit values,
+; i=0..15 (SSE register) or i=0..7 (MMX register)
+; Contents of SIMD registers are shown in memory order.
-%macro PUT_BITS 1
- add put_bits, ecx ; put_bits += size;
- shl put_buffer, cl ; put_buffer = (put_buffer << size);
- or put_buffer, %1
-%endmacro
+; Fill the bit buffer to capacity with the leading bits from code, then output
+; the bit buffer and put the remaining bits from code into the bit buffer.
+;
+; Usage:
+; code - contains the bits to shift into the bit buffer (LSB-aligned)
+; %1 - temp register
+; %2 - low byte of temp register
+; %3 - second byte of temp register
+; %4-%8 (optional) - extra instructions to execute before the macro completes
+; %9 - the label to which to jump when the macro completes
+;
+; Upon completion, free_bits will be set to the number of remaining bits from
+; code, and put_buffer will contain those remaining bits. temp and code will
+; be clobbered.
+;
+; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE()
+; macro in jchuff.c.
-%macro CHECKBUF15 0
- cmp put_bits, 16 ; if (put_bits > 31) {
- jl %%.CHECKBUF15_END
- mov eax, POINTER [esp+buffer]
- EMIT_BYTE
- EMIT_BYTE
- mov POINTER [esp+buffer], eax
-%%.CHECKBUF15_END:
-%endmacro
-
-%macro EMIT_BITS 1
- PUT_BITS %1
- CHECKBUF15
-%endmacro
-
-%macro kloop_prepare 37 ;(ko, jno0, ..., jno31, xmm0, xmm1, xmm2, xmm3)
- pxor xmm4, xmm4 ; __m128i neg = _mm_setzero_si128();
- pxor xmm5, xmm5 ; __m128i neg = _mm_setzero_si128();
- pxor xmm6, xmm6 ; __m128i neg = _mm_setzero_si128();
- pxor xmm7, xmm7 ; __m128i neg = _mm_setzero_si128();
- pinsrw %34, word [esi + %2 * SIZEOF_WORD], 0 ; xmm_shadow[0] = block[jno0];
- pinsrw %35, word [esi + %10 * SIZEOF_WORD], 0 ; xmm_shadow[8] = block[jno8];
- pinsrw %36, word [esi + %18 * SIZEOF_WORD], 0 ; xmm_shadow[16] = block[jno16];
- pinsrw %37, word [esi + %26 * SIZEOF_WORD], 0 ; xmm_shadow[24] = block[jno24];
- pinsrw %34, word [esi + %3 * SIZEOF_WORD], 1 ; xmm_shadow[1] = block[jno1];
- pinsrw %35, word [esi + %11 * SIZEOF_WORD], 1 ; xmm_shadow[9] = block[jno9];
- pinsrw %36, word [esi + %19 * SIZEOF_WORD], 1 ; xmm_shadow[17] = block[jno17];
- pinsrw %37, word [esi + %27 * SIZEOF_WORD], 1 ; xmm_shadow[25] = block[jno25];
- pinsrw %34, word [esi + %4 * SIZEOF_WORD], 2 ; xmm_shadow[2] = block[jno2];
- pinsrw %35, word [esi + %12 * SIZEOF_WORD], 2 ; xmm_shadow[10] = block[jno10];
- pinsrw %36, word [esi + %20 * SIZEOF_WORD], 2 ; xmm_shadow[18] = block[jno18];
- pinsrw %37, word [esi + %28 * SIZEOF_WORD], 2 ; xmm_shadow[26] = block[jno26];
- pinsrw %34, word [esi + %5 * SIZEOF_WORD], 3 ; xmm_shadow[3] = block[jno3];
- pinsrw %35, word [esi + %13 * SIZEOF_WORD], 3 ; xmm_shadow[11] = block[jno11];
- pinsrw %36, word [esi + %21 * SIZEOF_WORD], 3 ; xmm_shadow[19] = block[jno19];
- pinsrw %37, word [esi + %29 * SIZEOF_WORD], 3 ; xmm_shadow[27] = block[jno27];
- pinsrw %34, word [esi + %6 * SIZEOF_WORD], 4 ; xmm_shadow[4] = block[jno4];
- pinsrw %35, word [esi + %14 * SIZEOF_WORD], 4 ; xmm_shadow[12] = block[jno12];
- pinsrw %36, word [esi + %22 * SIZEOF_WORD], 4 ; xmm_shadow[20] = block[jno20];
- pinsrw %37, word [esi + %30 * SIZEOF_WORD], 4 ; xmm_shadow[28] = block[jno28];
- pinsrw %34, word [esi + %7 * SIZEOF_WORD], 5 ; xmm_shadow[5] = block[jno5];
- pinsrw %35, word [esi + %15 * SIZEOF_WORD], 5 ; xmm_shadow[13] = block[jno13];
- pinsrw %36, word [esi + %23 * SIZEOF_WORD], 5 ; xmm_shadow[21] = block[jno21];
- pinsrw %37, word [esi + %31 * SIZEOF_WORD], 5 ; xmm_shadow[29] = block[jno29];
- pinsrw %34, word [esi + %8 * SIZEOF_WORD], 6 ; xmm_shadow[6] = block[jno6];
- pinsrw %35, word [esi + %16 * SIZEOF_WORD], 6 ; xmm_shadow[14] = block[jno14];
- pinsrw %36, word [esi + %24 * SIZEOF_WORD], 6 ; xmm_shadow[22] = block[jno22];
- pinsrw %37, word [esi + %32 * SIZEOF_WORD], 6 ; xmm_shadow[30] = block[jno30];
- pinsrw %34, word [esi + %9 * SIZEOF_WORD], 7 ; xmm_shadow[7] = block[jno7];
- pinsrw %35, word [esi + %17 * SIZEOF_WORD], 7 ; xmm_shadow[15] = block[jno15];
- pinsrw %36, word [esi + %25 * SIZEOF_WORD], 7 ; xmm_shadow[23] = block[jno23];
-%if %1 != 32
- pinsrw %37, word [esi + %33 * SIZEOF_WORD], 7 ; xmm_shadow[31] = block[jno31];
-%else
- pinsrw %37, ecx, 7 ; xmm_shadow[31] = block[jno31];
+%macro EMIT_QWORD 9
+%define %%temp %1
+%define %%tempb %2
+%define %%temph %3
+ add nbits, free_bits ; nbits += free_bits;
+ neg free_bits ; free_bits = -free_bits;
+ movq mm_temp, mm_code ; temp = code;
+ movd mm_nbits, nbits ; nbits --> MMX register
+ movd mm_overflow_bits, free_bits ; overflow_bits (temp register) = free_bits;
+ neg free_bits ; free_bits = -free_bits;
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ psrlq mm_temp, mm_overflow_bits ; temp >>= overflow_bits;
+ add free_bits, 64 ; free_bits += 64;
+ por mm_temp, mm_put_buffer ; temp |= put_buffer;
+%ifidn %%temp, nbits_base
+ movd mm_save_nbits, nbits_base ; save nbits_base
%endif
- pcmpgtw xmm4, %34 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm5, %35 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm6, %36 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm7, %37 ; neg = _mm_cmpgt_epi16(neg, x1);
- paddw %34, xmm4 ; x1 = _mm_add_epi16(x1, neg);
- paddw %35, xmm5 ; x1 = _mm_add_epi16(x1, neg);
- paddw %36, xmm6 ; x1 = _mm_add_epi16(x1, neg);
- paddw %37, xmm7 ; x1 = _mm_add_epi16(x1, neg);
- pxor %34, xmm4 ; x1 = _mm_xor_si128(x1, neg);
- pxor %35, xmm5 ; x1 = _mm_xor_si128(x1, neg);
- pxor %36, xmm6 ; x1 = _mm_xor_si128(x1, neg);
- pxor %37, xmm7 ; x1 = _mm_xor_si128(x1, neg);
- pxor xmm4, %34 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm5, %35 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm6, %36 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm7, %37 ; neg = _mm_xor_si128(neg, x1);
- movdqa XMMWORD [esp + t1 + %1 * SIZEOF_WORD], %34 ; _mm_storeu_si128((__m128i *)(t1 + ko), x1);
- movdqa XMMWORD [esp + t1 + (%1 + 8) * SIZEOF_WORD], %35 ; _mm_storeu_si128((__m128i *)(t1 + ko + 8), x1);
- movdqa XMMWORD [esp + t1 + (%1 + 16) * SIZEOF_WORD], %36 ; _mm_storeu_si128((__m128i *)(t1 + ko + 16), x1);
- movdqa XMMWORD [esp + t1 + (%1 + 24) * SIZEOF_WORD], %37 ; _mm_storeu_si128((__m128i *)(t1 + ko + 24), x1);
- movdqa XMMWORD [esp + t2 + %1 * SIZEOF_WORD], xmm4 ; _mm_storeu_si128((__m128i *)(t2 + ko), neg);
- movdqa XMMWORD [esp + t2 + (%1 + 8) * SIZEOF_WORD], xmm5 ; _mm_storeu_si128((__m128i *)(t2 + ko + 8), neg);
- movdqa XMMWORD [esp + t2 + (%1 + 16) * SIZEOF_WORD], xmm6 ; _mm_storeu_si128((__m128i *)(t2 + ko + 16), neg);
- movdqa XMMWORD [esp + t2 + (%1 + 24) * SIZEOF_WORD], xmm7 ; _mm_storeu_si128((__m128i *)(t2 + ko + 24), neg);
+ movq mm_code_bits, mm_temp ; code_bits (temp register) = temp;
+ movq mm_put_buffer, mm_code ; put_buffer = code;
+ pcmpeqb mm_temp, mm_all_0xff ; b_temp[i] = (b_temp[i] == 0xFF ? 0xFF : 0);
+ movq mm_code, mm_code_bits ; code = code_bits;
+ psrlq mm_code_bits, 32 ; code_bits >>= 32;
+ pmovmskb nbits, mm_temp ; nbits = 0; nbits |= ((b_temp[i] >> 7) << i);
+ movd %%temp, mm_code_bits ; temp = code_bits;
+ bswap %%temp ; temp = htonl(temp);
+ test nbits, nbits ; if (nbits != 0) /* Some 0xFF bytes */
+ jnz %%.SLOW ; goto %%.SLOW
+ mov dword [buffer], %%temp ; *(uint32_t)buffer = temp;
+%ifidn %%temp, nbits_base
+ movd nbits_base, mm_save_nbits ; restore nbits_base
+%endif
+ %4
+ movd nbits, mm_code ; nbits = (uint32_t)(code);
+ %5
+ bswap nbits ; nbits = htonl(nbits);
+ mov dword [buffer + 4], nbits ; *(uint32_t)(buffer + 4) = nbits;
+ lea buffer, [buffer + 8] ; buffer += 8;
+ %6
+ %7
+ %8
+ jmp %9 ; return
+%%.SLOW:
+ ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8
+ ; bytes in the qword.
+ mov byte [buffer], %%tempb ; buffer[0] = temp[0];
+ cmp %%tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], %%temph ; buffer[0] = temp[1];
+ cmp %%temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ shr %%temp, 16 ; temp >>= 16;
+ mov byte [buffer], %%tempb ; buffer[0] = temp[0];
+ cmp %%tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], %%temph ; buffer[0] = temp[1];
+ cmp %%temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ movd nbits, mm_code ; nbits (temp register) = (uint32_t)(code)
+%ifidn %%temp, nbits_base
+ movd nbits_base, mm_save_nbits ; restore nbits_base
+%endif
+ bswap nbits ; nbits = htonl(nbits)
+ mov byte [buffer], nbitsb ; buffer[0] = nbits[0];
+ cmp nbitsb, 0xFF ; Set CF if nbits[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], nbitsh ; buffer[0] = nbits[1];
+ cmp nbitsh, 0xFF ; Set CF if nbits[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
+ shr nbits, 16 ; nbits >>= 16;
+ mov byte [buffer], nbitsb ; buffer[0] = nbits[0];
+ cmp nbitsb, 0xFF ; Set CF if nbits[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (nbits[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], nbitsh ; buffer[0] = nbits[1];
+ %4
+ cmp nbitsh, 0xFF ; Set CF if nbits[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb buffer, -2 ; buffer -= (-2 + (nbits[1] < 0xFF ? 1 : 0));
+ %5
+ %6
+ %7
+ %8
+ jmp %9 ; return;
+%endmacro
+
+%macro PUSH 1
+ push %1
+%assign stack_offset stack_offset + 4
+%endmacro
+
+%macro POP 1
+ pop %1
+%assign stack_offset stack_offset - 4
+%endmacro
+
+; If PIC is defined, load the address of a symbol defined in this file into a
+; register. Equivalent to
+; get_GOT %1
+; lea %1, [GOTOFF(%1, %2)]
+; without using the GOT.
+;
+; Usage:
+; %1 - register into which to load the address of the symbol
+; %2 - symbol whose address should be loaded
+; %3 - optional multi-line macro to execute before the symbol address is loaded
+; %4 - optional multi-line macro to execute after the symbol address is loaded
+;
+; If PIC is not defined, then %3 and %4 are executed in order.
+
+%macro GET_SYM 2-4
+%ifdef PIC
+ call %%.geteip
+%%.ref:
+ %4
+ add %1, %2 - %%.ref
+ jmp short %%.done
+ align 32
+%%.geteip:
+ %3 4 ; must adjust stack pointer because of call
+ mov %1, POINTER [esp]
+ ret
+ align 32
+%%.done:
+%else
+ %3 0
+ %4
+%endif
%endmacro
;
@@ -152,272 +275,487 @@
; JCOEFPTR block, int last_dc_val,
; c_derived_tbl *dctbl, c_derived_tbl *actbl)
;
+; Stack layout:
+; Function args
+; Return address
+; Saved ebx
+; Saved ebp
+; Saved esi
+; Saved edi <-- esp_save
+; ...
+; esp_save
+; t_ 64*2 bytes (aligned to 128 bytes)
+;
+; esp is used (as t) to point into t_ (data in lower indices is not used once
+; esp passes over them, so this is signal-safe.) Aligning to 128 bytes allows
+; us to find the rest of the data again.
+;
+; NOTES:
+; When shuffling data, we try to avoid pinsrw as much as possible, since it is
+; slow on many CPUs. Its reciprocal throughput (issue latency) is 1 even on
+; modern CPUs, so chains of pinsrw instructions (even with different outputs)
+; can limit performance. pinsrw is a VectorPath instruction on AMD K8 and
+; requires 2 µops (with memory operand) on Intel. In either case, only one
+; pinsrw instruction can be decoded per cycle (and nothing else if they are
+; back-to-back), so out-of-order execution cannot be used to work around long
+; pinsrw chains (though for Sandy Bridge and later, this may be less of a
+; problem if the code runs from the µop cache.)
+;
+; We use tzcnt instead of bsf without checking for support. The instruction is
+; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to
+; rep bsf.) The destination (first) operand of bsf (and tzcnt on some CPUs) is
+; an input dependency (although the behavior is not formally defined, Intel
+; CPUs usually leave the destination unmodified if the source is zero.) This
+; can prevent out-of-order execution, so we clear the destination before
+; invoking tzcnt.
+;
+; Initial register allocation
+; eax - frame --> buffer
+; ebx - nbits_base (PIC) / emit_temp
+; ecx - dctbl --> size --> state
+; edx - block --> nbits
+; esi - code_temp --> state --> actbl
+; edi - index_temp --> free_bits
+; esp - t
+; ebp - index
-; eax + 8 = working_state *state
-; eax + 12 = JOCTET *buffer
-; eax + 16 = JCOEFPTR block
-; eax + 20 = int last_dc_val
-; eax + 24 = c_derived_tbl *dctbl
-; eax + 28 = c_derived_tbl *actbl
+%define frame eax
+%ifdef PIC
+%define nbits_base ebx
+%endif
+%define emit_temp ebx
+%define emit_tempb bl
+%define emit_temph bh
+%define dctbl ecx
+%define block edx
+%define code_temp esi
+%define index_temp edi
+%define t esp
+%define index ebp
-%define pad 6 * SIZEOF_DWORD ; Align to 16 bytes
-%define t1 pad
-%define t2 t1 + (DCTSIZE2 * SIZEOF_WORD)
-%define block t2 + (DCTSIZE2 * SIZEOF_WORD)
-%define actbl block + SIZEOF_DWORD
-%define buffer actbl + SIZEOF_DWORD
-%define temp buffer + SIZEOF_DWORD
-%define temp2 temp + SIZEOF_DWORD
-%define temp3 temp2 + SIZEOF_DWORD
-%define temp4 temp3 + SIZEOF_DWORD
-%define temp5 temp4 + SIZEOF_DWORD
-%define gotptr temp5 + SIZEOF_DWORD ; void *gotptr
-%define put_buffer ebx
-%define put_bits edi
+%assign save_frame DCTSIZE2 * SIZEOF_WORD
+
+; Step 1: Re-arrange input data according to jpeg_natural_order
+; xx 01 02 03 04 05 06 07 xx 01 08 16 09 02 03 10
+; 08 09 10 11 12 13 14 15 17 24 32 25 18 11 04 05
+; 16 17 18 19 20 21 22 23 12 19 26 33 40 48 41 34
+; 24 25 26 27 28 29 30 31 ==> 27 20 13 06 07 14 21 28
+; 32 33 34 35 36 37 38 39 35 42 49 56 57 50 43 36
+; 40 41 42 43 44 45 46 47 29 22 15 23 30 37 44 51
+; 48 49 50 51 52 53 54 55 58 59 52 45 38 31 39 46
+; 56 57 58 59 60 61 62 63 53 60 61 54 47 55 62 63
align 32
GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
EXTN(jsimd_huff_encode_one_block_sse2):
- push ebp
- mov eax, esp ; eax = original ebp
- sub esp, byte 4
- and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
- mov [esp], eax
- mov ebp, esp ; ebp = aligned ebp
- sub esp, temp5+9*SIZEOF_DWORD-pad
- push ebx
- push ecx
-; push edx ; need not be preserved
- push esi
- push edi
- push ebp
- mov esi, POINTER [eax+8] ; (working_state *state)
- mov put_buffer, dword [esi+8] ; put_buffer = state->cur.put_buffer;
- mov put_bits, dword [esi+12] ; put_bits = state->cur.put_bits;
- push esi ; esi is now scratch
+%assign stack_offset 0
+%define arg_state 4 + stack_offset
+%define arg_buffer 8 + stack_offset
+%define arg_block 12 + stack_offset
+%define arg_last_dc_val 16 + stack_offset
+%define arg_dctbl 20 + stack_offset
+%define arg_actbl 24 + stack_offset
- get_GOT edx ; get GOT address
- movpic POINTER [esp+gotptr], edx ; save GOT address
+ ;X: X = code stream
+ mov block, [esp + arg_block]
+ PUSH ebx
+ PUSH ebp
+ movups xmm3, XMMWORD [block + 0 * SIZEOF_WORD] ;D: w3 = xx 01 02 03 04 05 06 07
+ PUSH esi
+ PUSH edi
+ movdqa xmm0, xmm3 ;A: w0 = xx 01 02 03 04 05 06 07
+ mov frame, esp
+ lea t, [frame - (save_frame + 4)]
+ movups xmm1, XMMWORD [block + 8 * SIZEOF_WORD] ;B: w1 = 08 09 10 11 12 13 14 15
+ and t, -DCTSIZE2 * SIZEOF_WORD ; t = &t_[0]
+ mov [t + save_frame], frame
+ pxor xmm4, xmm4 ;A: w4[i] = 0;
+ punpckldq xmm0, xmm1 ;A: w0 = xx 01 08 09 02 03 10 11
+ pshuflw xmm0, xmm0, 11001001b ;A: w0 = 01 08 xx 09 02 03 10 11
+ pinsrw xmm0, word [block + 16 * SIZEOF_WORD], 2 ;A: w0 = 01 08 16 09 02 03 10 11
+ punpckhdq xmm3, xmm1 ;D: w3 = 04 05 12 13 06 07 14 15
+ punpcklqdq xmm1, xmm3 ;B: w1 = 08 09 10 11 04 05 12 13
+ pinsrw xmm0, word [block + 17 * SIZEOF_WORD], 7 ;A: w0 = 01 08 16 09 02 03 10 17
+ ;A: (Row 0, offset 1)
+ pcmpgtw xmm4, xmm0 ;A: w4[i] = (w0[i] < 0 ? -1 : 0);
+ paddw xmm0, xmm4 ;A: w0[i] += w4[i];
+ movaps XMMWORD [t + 0 * SIZEOF_WORD], xmm0 ;A: t[i] = w0[i];
- mov ecx, POINTER [eax+28]
- mov edx, POINTER [eax+16]
- mov esi, POINTER [eax+12]
- mov POINTER [esp+actbl], ecx
- mov POINTER [esp+block], edx
- mov POINTER [esp+buffer], esi
+ movq xmm2, qword [block + 24 * SIZEOF_WORD] ;B: w2 = 24 25 26 27 -- -- -- --
+ pshuflw xmm2, xmm2, 11011000b ;B: w2 = 24 26 25 27 -- -- -- --
+ pslldq xmm1, 1 * SIZEOF_WORD ;B: w1 = -- 08 09 10 11 04 05 12
+ movups xmm5, XMMWORD [block + 48 * SIZEOF_WORD] ;H: w5 = 48 49 50 51 52 53 54 55
+ movsd xmm1, xmm2 ;B: w1 = 24 26 25 27 11 04 05 12
+ punpcklqdq xmm2, xmm5 ;C: w2 = 24 26 25 27 48 49 50 51
+ pinsrw xmm1, word [block + 32 * SIZEOF_WORD], 1 ;B: w1 = 24 32 25 27 11 04 05 12
+ pxor xmm4, xmm4 ;A: w4[i] = 0;
+ psrldq xmm3, 2 * SIZEOF_WORD ;D: w3 = 12 13 06 07 14 15 -- --
+ pcmpeqw xmm0, xmm4 ;A: w0[i] = (w0[i] == 0 ? -1 : 0);
+ pinsrw xmm1, word [block + 18 * SIZEOF_WORD], 3 ;B: w1 = 24 32 25 18 11 04 05 12
+ ; (Row 1, offset 1)
+ pcmpgtw xmm4, xmm1 ;B: w4[i] = (w1[i] < 0 ? -1 : 0);
+ paddw xmm1, xmm4 ;B: w1[i] += w4[i];
+ movaps XMMWORD [t + 8 * SIZEOF_WORD], xmm1 ;B: t[i+8] = w1[i];
+ pxor xmm4, xmm4 ;B: w4[i] = 0;
+ pcmpeqw xmm1, xmm4 ;B: w1[i] = (w1[i] == 0 ? -1 : 0);
- ; Encode the DC coefficient difference per section F.1.2.1
- mov esi, POINTER [esp+block] ; block
- movsx ecx, word [esi] ; temp = temp2 = block[0] - last_dc_val;
- sub ecx, dword [eax+20]
- mov esi, ecx
+ packsswb xmm0, xmm1 ;AB: b0[i] = w0[i], b0[i+8] = w1[i]
+ ; w/ signed saturation
- ; This is a well-known technique for obtaining the absolute value
- ; with out a branch. It is derived from an assembly language technique
- ; presented in "How to Optimize for the Pentium Processors",
- ; Copyright (c) 1996, 1997 by Agner Fog.
- mov edx, ecx
- sar edx, 31 ; temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
- xor ecx, edx ; temp ^= temp3;
- sub ecx, edx ; temp -= temp3;
+ pinsrw xmm3, word [block + 20 * SIZEOF_WORD], 0 ;D: w3 = 20 13 06 07 14 15 -- --
+ pinsrw xmm3, word [block + 21 * SIZEOF_WORD], 5 ;D: w3 = 20 13 06 07 14 21 -- --
+ pinsrw xmm3, word [block + 28 * SIZEOF_WORD], 6 ;D: w3 = 20 13 06 07 14 21 28 --
+ pinsrw xmm3, word [block + 35 * SIZEOF_WORD], 7 ;D: w3 = 20 13 06 07 14 21 28 35
+ ; (Row 3, offset 1)
+ pcmpgtw xmm4, xmm3 ;D: w4[i] = (w3[i] < 0 ? -1 : 0);
+ paddw xmm3, xmm4 ;D: w3[i] += w4[i];
+ movaps XMMWORD [t + 24 * SIZEOF_WORD], xmm3 ;D: t[i+24] = w3[i];
+ pxor xmm4, xmm4 ;D: w4[i] = 0;
+ pcmpeqw xmm3, xmm4 ;D: w3[i] = (w3[i] == 0 ? -1 : 0);
- ; For a negative input, want temp2 = bitwise complement of abs(input)
- ; This code assumes we are on a two's complement machine
- add esi, edx ; temp2 += temp3;
- mov dword [esp+temp], esi ; backup temp2 in temp
+ pinsrw xmm2, word [block + 19 * SIZEOF_WORD], 0 ;C: w2 = 19 26 25 27 48 49 50 51
+ pinsrw xmm2, word [block + 33 * SIZEOF_WORD], 2 ;C: w2 = 19 26 33 27 48 49 50 51
+ pinsrw xmm2, word [block + 40 * SIZEOF_WORD], 3 ;C: w2 = 19 26 33 40 48 49 50 51
+ pinsrw xmm2, word [block + 41 * SIZEOF_WORD], 5 ;C: w2 = 19 26 33 40 48 41 50 51
+ pinsrw xmm2, word [block + 34 * SIZEOF_WORD], 6 ;C: w2 = 19 26 33 40 48 41 34 51
+ pinsrw xmm2, word [block + 27 * SIZEOF_WORD], 7 ;C: w2 = 19 26 33 40 48 41 34 27
+ ; (Row 2, offset 1)
+ pcmpgtw xmm4, xmm2 ;C: w4[i] = (w2[i] < 0 ? -1 : 0);
+ paddw xmm2, xmm4 ;C: w2[i] += w4[i];
+ movsx code_temp, word [block] ;Z: code_temp = block[0];
- ; Find the number of bits needed for the magnitude of the coefficient
- movpic ebp, POINTER [esp+gotptr] ; load GOT address (ebp)
- movzx edx, byte [GOTOFF(ebp, EXTN(jpeg_nbits_table) + ecx)] ; nbits = JPEG_NBITS(temp);
- mov dword [esp+temp2], edx ; backup nbits in temp2
+; %1 - stack pointer adjustment
+%macro GET_SYM_BEFORE 1
+ movaps XMMWORD [t + 16 * SIZEOF_WORD + %1], xmm2
+ ;C: t[i+16] = w2[i];
+ pxor xmm4, xmm4 ;C: w4[i] = 0;
+ pcmpeqw xmm2, xmm4 ;C: w2[i] = (w2[i] == 0 ? -1 : 0);
+ sub code_temp, [frame + arg_last_dc_val] ;Z: code_temp -= last_dc_val;
- ; Emit the Huffman-coded symbol for the number of bits
- mov ebp, POINTER [eax+24] ; After this point, arguments are not accessible anymore
- mov eax, INT [ebp + edx * 4] ; code = dctbl->ehufco[nbits];
- movzx ecx, byte [ebp + edx + 1024] ; size = dctbl->ehufsi[nbits];
- EMIT_BITS eax ; EMIT_BITS(code, size)
+ packsswb xmm2, xmm3 ;CD: b2[i] = w2[i], b2[i+8] = w3[i]
+ ; w/ signed saturation
- mov ecx, dword [esp+temp2] ; restore nbits
+ movdqa xmm3, xmm5 ;H: w3 = 48 49 50 51 52 53 54 55
+ pmovmskb index_temp, xmm2 ;Z: index_temp = 0; index_temp |= ((b2[i] >> 7) << i);
+ pmovmskb index, xmm0 ;Z: index = 0; index |= ((b0[i] >> 7) << i);
+ movups xmm0, XMMWORD [block + 56 * SIZEOF_WORD] ;H: w0 = 56 57 58 59 60 61 62 63
+ punpckhdq xmm3, xmm0 ;H: w3 = 52 53 60 61 54 55 62 63
+ shl index_temp, 16 ;Z: index_temp <<= 16;
+ psrldq xmm3, 1 * SIZEOF_WORD ;H: w3 = 53 60 61 54 55 62 63 --
+ pxor xmm2, xmm2 ;H: w2[i] = 0;
+ pshuflw xmm3, xmm3, 00111001b ;H: w3 = 60 61 54 53 55 62 63 --
+ or index, index_temp ;Z: index |= index_temp;
+%undef index_temp
+%define free_bits edi
+%endmacro
- ; Mask off any extra bits in code
- mov eax, 1
- shl eax, cl
- dec eax
- and eax, dword [esp+temp] ; temp2 &= (((JLONG)1)<<nbits) - 1;
+%macro GET_SYM_AFTER 0
+ movq xmm1, qword [block + 44 * SIZEOF_WORD] ;G: w1 = 44 45 46 47 -- -- -- --
+ unpcklps xmm5, xmm0 ;E: w5 = 48 49 56 57 50 51 58 59
+ pxor xmm0, xmm0 ;H: w0[i] = 0;
+ not index ;Z: index = ~index;
+ pinsrw xmm3, word [block + 47 * SIZEOF_WORD], 3 ;H: w3 = 60 61 54 47 55 62 63 --
+ ; (Row 7, offset 1)
+ pcmpgtw xmm2, xmm3 ;H: w2[i] = (w3[i] < 0 ? -1 : 0);
+ mov dctbl, [frame + arg_dctbl]
+ paddw xmm3, xmm2 ;H: w3[i] += w2[i];
+ movaps XMMWORD [t + 56 * SIZEOF_WORD], xmm3 ;H: t[i+56] = w3[i];
+ movq xmm4, qword [block + 36 * SIZEOF_WORD] ;G: w4 = 36 37 38 39 -- -- -- --
+ pcmpeqw xmm3, xmm0 ;H: w3[i] = (w3[i] == 0 ? -1 : 0);
+ punpckldq xmm4, xmm1 ;G: w4 = 36 37 44 45 38 39 46 47
+ movdqa xmm1, xmm4 ;F: w1 = 36 37 44 45 38 39 46 47
+ pcmpeqw mm_all_0xff, mm_all_0xff ;Z: all_0xff[i] = 0xFF;
+%endmacro
- ; Emit that number of bits of the value, if positive,
- ; or the complement of its magnitude, if negative.
- EMIT_BITS eax ; EMIT_BITS(temp2, nbits)
+ GET_SYM nbits_base, jpeg_nbits_table, GET_SYM_BEFORE, GET_SYM_AFTER
- ; Prepare data
- xor ecx, ecx
- mov esi, POINTER [esp+block]
- kloop_prepare 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, \
- 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, \
- 27, 20, 13, 6, 7, 14, 21, 28, 35, \
- xmm0, xmm1, xmm2, xmm3
- kloop_prepare 32, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, \
- 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, \
- 53, 60, 61, 54, 47, 55, 62, 63, 63, \
- xmm0, xmm1, xmm2, xmm3
+ psrldq xmm4, 1 * SIZEOF_WORD ;G: w4 = 37 44 45 38 39 46 47 --
+ shufpd xmm1, xmm5, 10b ;F: w1 = 36 37 44 45 50 51 58 59
+ pshufhw xmm4, xmm4, 11010011b ;G: w4 = 37 44 45 38 -- 39 46 --
+ pslldq xmm1, 1 * SIZEOF_WORD ;F: w1 = -- 36 37 44 45 50 51 58
+ pinsrw xmm4, word [block + 59 * SIZEOF_WORD], 0 ;G: w4 = 59 44 45 38 -- 39 46 --
+ pshufd xmm1, xmm1, 11011000b ;F: w1 = -- 36 45 50 37 44 51 58
+ cmp code_temp, 1 << 31 ;Z: Set CF if code_temp < 0x80000000,
+ ;Z: i.e. if code_temp is positive
+ pinsrw xmm4, word [block + 52 * SIZEOF_WORD], 1 ;G: w4 = 59 52 45 38 -- 39 46 --
+ movlps xmm1, qword [block + 20 * SIZEOF_WORD] ;F: w1 = 20 21 22 23 37 44 51 58
+ pinsrw xmm4, word [block + 31 * SIZEOF_WORD], 4 ;G: w4 = 59 52 45 38 31 39 46 --
+ pshuflw xmm1, xmm1, 01110010b ;F: w1 = 22 20 23 21 37 44 51 58
+ pinsrw xmm4, word [block + 53 * SIZEOF_WORD], 7 ;G: w4 = 59 52 45 38 31 39 46 53
+ ; (Row 6, offset 1)
+ adc code_temp, -1 ;Z: code_temp += -1 + (code_temp >= 0 ? 1 : 0);
+ pxor xmm2, xmm2 ;G: w2[i] = 0;
+ pcmpgtw xmm0, xmm4 ;G: w0[i] = (w4[i] < 0 ? -1 : 0);
+ pinsrw xmm1, word [block + 15 * SIZEOF_WORD], 1 ;F: w1 = 22 15 23 21 37 44 51 58
+ paddw xmm4, xmm0 ;G: w4[i] += w0[i];
+ movaps XMMWORD [t + 48 * SIZEOF_WORD], xmm4 ;G: t[48+i] = w4[i];
+ movd mm_temp, code_temp ;Z: temp = code_temp
+ pinsrw xmm1, word [block + 30 * SIZEOF_WORD], 3 ;F: w1 = 22 15 23 30 37 44 51 58
+ ; (Row 5, offset 1)
+ pcmpeqw xmm4, xmm2 ;G: w4[i] = (w4[i] == 0 ? -1 : 0);
- pxor xmm7, xmm7
- movdqa xmm0, XMMWORD [esp + t1 + 0 * SIZEOF_WORD] ; __m128i tmp0 = _mm_loadu_si128((__m128i *)(t1 + 0));
- movdqa xmm1, XMMWORD [esp + t1 + 8 * SIZEOF_WORD] ; __m128i tmp1 = _mm_loadu_si128((__m128i *)(t1 + 8));
- movdqa xmm2, XMMWORD [esp + t1 + 16 * SIZEOF_WORD] ; __m128i tmp2 = _mm_loadu_si128((__m128i *)(t1 + 16));
- movdqa xmm3, XMMWORD [esp + t1 + 24 * SIZEOF_WORD] ; __m128i tmp3 = _mm_loadu_si128((__m128i *)(t1 + 24));
- pcmpeqw xmm0, xmm7 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero);
- pcmpeqw xmm1, xmm7 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero);
- pcmpeqw xmm2, xmm7 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero);
- pcmpeqw xmm3, xmm7 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero);
- packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1);
- packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3);
- pmovmskb edx, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0;
- pmovmskb ecx, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16;
- shl ecx, 16
- or edx, ecx
- not edx ; index = ~index;
+ packsswb xmm4, xmm3 ;GH: b4[i] = w4[i], b4[i+8] = w3[i]
+ ; w/ signed saturation
- lea esi, [esp+t1]
- mov ebp, POINTER [esp+actbl] ; ebp = actbl
+ lea t, [t - SIZEOF_WORD] ;Z: t = &t[-1]
+ pxor xmm0, xmm0 ;F: w0[i] = 0;
+ pcmpgtw xmm2, xmm1 ;F: w2[i] = (w1[i] < 0 ? -1 : 0);
+ paddw xmm1, xmm2 ;F: w1[i] += w2[i];
+ movaps XMMWORD [t + (40+1) * SIZEOF_WORD], xmm1 ;F: t[40+i] = w1[i];
+ pcmpeqw xmm1, xmm0 ;F: w1[i] = (w1[i] == 0 ? -1 : 0);
+ pinsrw xmm5, word [block + 42 * SIZEOF_WORD], 0 ;E: w5 = 42 49 56 57 50 51 58 59
+ pinsrw xmm5, word [block + 43 * SIZEOF_WORD], 5 ;E: w5 = 42 49 56 57 50 43 58 59
+ pinsrw xmm5, word [block + 36 * SIZEOF_WORD], 6 ;E: w5 = 42 49 56 57 50 43 36 59
+ pinsrw xmm5, word [block + 29 * SIZEOF_WORD], 7 ;E: w5 = 42 49 56 57 50 43 36 29
+ ; (Row 4, offset 1)
+%undef block
+%define nbits edx
+%define nbitsb dl
+%define nbitsh dh
+ movzx nbits, byte [NBITS(code_temp)] ;Z: nbits = JPEG_NBITS(code_temp);
+%undef code_temp
+%define state esi
+ pxor xmm2, xmm2 ;E: w2[i] = 0;
+ mov state, [frame + arg_state]
+ movd mm_nbits, nbits ;Z: nbits --> MMX register
+ pcmpgtw xmm0, xmm5 ;E: w0[i] = (w5[i] < 0 ? -1 : 0);
+ movd mm_code, dword [dctbl + c_derived_tbl.ehufco + nbits * 4]
+ ;Z: code = dctbl->ehufco[nbits];
+%define size ecx
+%define sizeb cl
+%define sizeh ch
+ paddw xmm5, xmm0 ;E: w5[i] += w0[i];
+ movaps XMMWORD [t + (32+1) * SIZEOF_WORD], xmm5 ;E: t[32+i] = w5[i];
+ movzx size, byte [dctbl + c_derived_tbl.ehufsi + nbits]
+ ;Z: size = dctbl->ehufsi[nbits];
+%undef dctbl
+ pcmpeqw xmm5, xmm2 ;E: w5[i] = (w5[i] == 0 ? -1 : 0);
-.BLOOP:
- bsf ecx, edx ; r = __builtin_ctzl(index);
- jz near .ELOOP
- lea esi, [esi+ecx*2] ; k += r;
- shr edx, cl ; index >>= r;
- mov dword [esp+temp3], edx
-.BRLOOP:
- cmp ecx, 16 ; while (r > 15) {
- jl near .ERLOOP
- sub ecx, 16 ; r -= 16;
- mov dword [esp+temp], ecx
- mov eax, INT [ebp + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0];
- movzx ecx, byte [ebp + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0];
- EMIT_BITS eax ; EMIT_BITS(code_0xf0, size_0xf0)
- mov ecx, dword [esp+temp]
- jmp .BRLOOP
-.ERLOOP:
- movsx eax, word [esi] ; temp = t1[k];
- movpic edx, POINTER [esp+gotptr] ; load GOT address (edx)
- movzx eax, byte [GOTOFF(edx, EXTN(jpeg_nbits_table) + eax)] ; nbits = JPEG_NBITS(temp);
- mov dword [esp+temp2], eax
- ; Emit Huffman symbol for run length / number of bits
- shl ecx, 4 ; temp3 = (r << 4) + nbits;
- add ecx, eax
- mov eax, INT [ebp + ecx * 4] ; code = actbl->ehufco[temp3];
- movzx ecx, byte [ebp + ecx + 1024] ; size = actbl->ehufsi[temp3];
- EMIT_BITS eax
+ packsswb xmm5, xmm1 ;EF: b5[i] = w5[i], b5[i+8] = w1[i]
+ ; w/ signed saturation
- movsx edx, word [esi+DCTSIZE2*2] ; temp2 = t2[k];
- ; Mask off any extra bits in code
- mov ecx, dword [esp+temp2]
- mov eax, 1
- shl eax, cl
- dec eax
- and eax, edx ; temp2 &= (((JLONG)1)<<nbits) - 1;
- EMIT_BITS eax ; PUT_BITS(temp2, nbits)
- mov edx, dword [esp+temp3]
- add esi, 2 ; ++k;
- shr edx, 1 ; index >>= 1;
+ movq mm_put_buffer, [state + working_state.cur.put_buffer.simd]
+ ;Z: put_buffer = state->cur.put_buffer.simd;
+ mov free_bits, [state + working_state.cur.free_bits]
+ ;Z: free_bits = state->cur.free_bits;
+%undef state
+%define actbl esi
+ mov actbl, [frame + arg_actbl]
+%define buffer eax
+ mov buffer, [frame + arg_buffer]
+%undef frame
+ jmp .BEGIN
- jmp .BLOOP
-.ELOOP:
- movdqa xmm0, XMMWORD [esp + t1 + 32 * SIZEOF_WORD] ; __m128i tmp0 = _mm_loadu_si128((__m128i *)(t1 + 0));
- movdqa xmm1, XMMWORD [esp + t1 + 40 * SIZEOF_WORD] ; __m128i tmp1 = _mm_loadu_si128((__m128i *)(t1 + 8));
- movdqa xmm2, XMMWORD [esp + t1 + 48 * SIZEOF_WORD] ; __m128i tmp2 = _mm_loadu_si128((__m128i *)(t1 + 16));
- movdqa xmm3, XMMWORD [esp + t1 + 56 * SIZEOF_WORD] ; __m128i tmp3 = _mm_loadu_si128((__m128i *)(t1 + 24));
- pcmpeqw xmm0, xmm7 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero);
- pcmpeqw xmm1, xmm7 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero);
- pcmpeqw xmm2, xmm7 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero);
- pcmpeqw xmm3, xmm7 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero);
- packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1);
- packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3);
- pmovmskb edx, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0;
- pmovmskb ecx, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16;
- shl ecx, 16
- or edx, ecx
- not edx ; index = ~index;
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- lea eax, [esp + t1 + (DCTSIZE2/2) * 2]
- sub eax, esi
- shr eax, 1
- bsf ecx, edx ; r = __builtin_ctzl(index);
- jz near .ELOOP2
- shr edx, cl ; index >>= r;
- add ecx, eax
- lea esi, [esi+ecx*2] ; k += r;
- mov dword [esp+temp3], edx
- jmp .BRLOOP2
-.BLOOP2:
- bsf ecx, edx ; r = __builtin_ctzl(index);
- jz near .ELOOP2
- lea esi, [esi+ecx*2] ; k += r;
- shr edx, cl ; index >>= r;
- mov dword [esp+temp3], edx
-.BRLOOP2:
- cmp ecx, 16 ; while (r > 15) {
- jl near .ERLOOP2
- sub ecx, 16 ; r -= 16;
- mov dword [esp+temp], ecx
- mov eax, INT [ebp + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0];
- movzx ecx, byte [ebp + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0];
- EMIT_BITS eax ; EMIT_BITS(code_0xf0, size_0xf0)
- mov ecx, dword [esp+temp]
- jmp .BRLOOP2
-.ERLOOP2:
- movsx eax, word [esi] ; temp = t1[k];
- bsr eax, eax ; nbits = 32 - __builtin_clz(temp);
- inc eax
- mov dword [esp+temp2], eax
- ; Emit Huffman symbol for run length / number of bits
- shl ecx, 4 ; temp3 = (r << 4) + nbits;
- add ecx, eax
- mov eax, INT [ebp + ecx * 4] ; code = actbl->ehufco[temp3];
- movzx ecx, byte [ebp + ecx + 1024] ; size = actbl->ehufsi[temp3];
- EMIT_BITS eax
+ align 16
+; size <= 32, so this is not really a loop
+.BRLOOP1: ; .BRLOOP1:
+ movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+ ; nbits = actbl->ehufsi[0xf0];
+ movd mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+ ; code = actbl->ehufco[0xf0];
+ and index, 0x7ffffff ; clear index if size == 32
+ sub size, 16 ; size -= 16;
+ sub free_bits, nbits ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_BRLOOP1 ; goto .EMIT_BRLOOP1;
+ movd mm_nbits, nbits ; nbits --> MMX register
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ por mm_put_buffer, mm_code ; put_buffer |= code;
+ jmp .ERLOOP1 ; goto .ERLOOP1;
- movsx edx, word [esi+DCTSIZE2*2] ; temp2 = t2[k];
- ; Mask off any extra bits in code
- mov ecx, dword [esp+temp2]
- mov eax, 1
- shl eax, cl
- dec eax
- and eax, edx ; temp2 &= (((JLONG)1)<<nbits) - 1;
- EMIT_BITS eax ; PUT_BITS(temp2, nbits)
- mov edx, dword [esp+temp3]
- add esi, 2 ; ++k;
- shr edx, 1 ; index >>= 1;
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- jmp .BLOOP2
-.ELOOP2:
- ; If the last coef(s) were zero, emit an end-of-block code
- lea edx, [esp + t1 + (DCTSIZE2-1) * 2] ; r = DCTSIZE2-1-k;
- cmp edx, esi ; if (r > 0) {
- je .EFN
- mov eax, INT [ebp] ; code = actbl->ehufco[0];
- movzx ecx, byte [ebp + 1024] ; size = actbl->ehufsi[0];
- EMIT_BITS eax
-.EFN:
- mov eax, [esp+buffer]
- pop esi
- ; Save put_buffer & put_bits
- mov dword [esi+8], put_buffer ; state->cur.put_buffer = put_buffer;
- mov dword [esi+12], put_bits ; state->cur.put_bits = put_bits;
+ align 16
+%ifdef PIC
+ times 6 nop
+%else
+ times 2 nop
+%endif
+.BLOOP1: ; do { /* size = # of zero bits/elements to skip */
+; if size == 32, index remains unchanged. Correct in .BRLOOP.
+ shr index, sizeb ; index >>= size;
+ lea t, [t + size * SIZEOF_WORD] ; t += size;
+ cmp size, 16 ; if (size > 16)
+ jg .BRLOOP1 ; goto .BRLOOP1;
+.ERLOOP1: ; .ERLOOP1:
+ movsx nbits, word [t] ; nbits = *t;
+%ifdef PIC
+ add size, size ; size += size;
+%else
+ lea size, [size * 2] ; size += size;
+%endif
+ movd mm_temp, nbits ; temp = nbits;
+ movzx nbits, byte [NBITS(nbits)] ; nbits = JPEG_NBITS(nbits);
+ lea size, [size * 8 + nbits] ; size = size * 8 + nbits;
+ movd mm_nbits, nbits ; nbits --> MMX register
+ movd mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
+ ; code = actbl->ehufco[size-16];
+ movzx size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
+ ; size = actbl->ehufsi[size-16];
+.BEGIN: ; .BEGIN:
+ pand mm_temp, [MASK_BITS(nbits)] ; temp &= (1 << nbits) - 1;
+ psllq mm_code, mm_nbits ; code <<= nbits;
+ add nbits, size ; nbits += size;
+ por mm_code, mm_temp ; code |= temp;
+ sub free_bits, nbits ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_ERLOOP1 ; insert code, flush buffer, init size, goto .BLOOP1
+ xor size, size ; size = 0; /* kill tzcnt input dependency */
+ tzcnt size, index ; size = # of trailing 0 bits in index
+ movd mm_nbits, nbits ; nbits --> MMX register
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ inc size ; ++size;
+ por mm_put_buffer, mm_code ; put_buffer |= code;
+ test index, index
+ jnz .BLOOP1 ; } while (index != 0);
+; Round 2
+; t points to the last used word, possibly below t_ if the previous index had 32 zero bits.
+.ELOOP1: ; .ELOOP1:
+ pmovmskb size, xmm4 ; size = 0; size |= ((b4[i] >> 7) << i);
+ pmovmskb index, xmm5 ; index = 0; index |= ((b5[i] >> 7) << i);
+ shl size, 16 ; size <<= 16;
+ or index, size ; index |= size;
+ not index ; index = ~index;
+ lea nbits, [t + (1 + DCTSIZE2) * SIZEOF_WORD]
+ ; nbits = t + 1 + 64;
+ and nbits, -DCTSIZE2 * SIZEOF_WORD ; nbits &= -128; /* now points to &t_[64] */
+ sub nbits, t ; nbits -= t;
+ shr nbits, 1 ; nbits >>= 1; /* # of leading 0 bits in old index + 33 */
+ tzcnt size, index ; size = # of trailing 0 bits in index
+ inc size ; ++size;
+ test index, index ; if (index == 0)
+ jz .ELOOP2 ; goto .ELOOP2;
+; NOTE: size == 32 cannot happen, since the last element is always 0.
+ shr index, sizeb ; index >>= size;
+ lea size, [size + nbits - 33] ; size = size + nbits - 33;
+ lea t, [t + size * SIZEOF_WORD] ; t += size;
+ cmp size, 16 ; if (size <= 16)
+ jle .ERLOOP2 ; goto .ERLOOP2;
+.BRLOOP2: ; do {
+ movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+ ; nbits = actbl->ehufsi[0xf0];
+ sub size, 16 ; size -= 16;
+ movd mm_code, dword [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+ ; code = actbl->ehufco[0xf0];
+ sub free_bits, nbits ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_BRLOOP2 ; insert code and flush put_buffer
+ movd mm_nbits, nbits ; else { nbits --> MMX register
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ por mm_put_buffer, mm_code ; put_buffer |= code;
+ cmp size, 16 ; if (size <= 16)
+ jle .ERLOOP2 ; goto .ERLOOP2;
+ jmp .BRLOOP2 ; } while (1);
- pop ebp
- pop edi
- pop esi
-; pop edx ; need not be preserved
- pop ecx
- pop ebx
- mov esp, ebp ; esp <- aligned ebp
- pop esp ; esp <- original ebp
- pop ebp
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.BLOOP2: ; do { /* size = # of zero bits/elements to skip */
+ shr index, sizeb ; index >>= size;
+ lea t, [t + size * SIZEOF_WORD] ; t += size;
+ cmp size, 16 ; if (size > 16)
+ jg .BRLOOP2 ; goto .BRLOOP2;
+.ERLOOP2: ; .ERLOOP2:
+ movsx nbits, word [t] ; nbits = *t;
+ add size, size ; size += size;
+ movd mm_temp, nbits ; temp = nbits;
+ movzx nbits, byte [NBITS(nbits)] ; nbits = JPEG_NBITS(nbits);
+ movd mm_nbits, nbits ; nbits --> MMX register
+ lea size, [size * 8 + nbits] ; size = size * 8 + nbits;
+ movd mm_code, dword [actbl + c_derived_tbl.ehufco + (size - 16) * 4]
+ ; code = actbl->ehufco[size-16];
+ movzx size, byte [actbl + c_derived_tbl.ehufsi + (size - 16)]
+ ; size = actbl->ehufsi[size-16];
+ psllq mm_code, mm_nbits ; code <<= nbits;
+ pand mm_temp, [MASK_BITS(nbits)] ; temp &= (1 << nbits) - 1;
+ lea nbits, [nbits + size] ; nbits += size;
+ por mm_code, mm_temp ; code |= temp;
+ xor size, size ; size = 0; /* kill tzcnt input dependency */
+ sub free_bits, nbits ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_ERLOOP2 ; insert code, flush buffer, init size, goto .BLOOP2
+ tzcnt size, index ; size = # of trailing 0 bits in index
+ movd mm_nbits, nbits ; nbits --> MMX register
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ inc size ; ++size;
+ por mm_put_buffer, mm_code ; put_buffer |= code;
+ test index, index
+ jnz .BLOOP2 ; } while (index != 0);
+.ELOOP2: ; .ELOOP2:
+ mov nbits, t ; nbits = t;
+ lea t, [t + SIZEOF_WORD] ; t = &t[1];
+ and nbits, DCTSIZE2 * SIZEOF_WORD - 1 ; nbits &= 127;
+ and t, -DCTSIZE2 * SIZEOF_WORD ; t &= -128; /* t = &t_[0]; */
+ cmp nbits, (DCTSIZE2 - 2) * SIZEOF_WORD ; if (nbits != 62 * 2)
+ je .EFN ; {
+ movd mm_code, dword [actbl + c_derived_tbl.ehufco + 0]
+ ; code = actbl->ehufco[0];
+ movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0]
+ ; nbits = actbl->ehufsi[0];
+ sub free_bits, nbits ; if ((free_bits -= nbits) <= 0)
+ jg .EFN_SKIP_EMIT_CODE ; {
+ EMIT_QWORD size, sizeb, sizeh, , , , , , .EFN ; insert code, flush put_buffer
+ align 16
+.EFN_SKIP_EMIT_CODE: ; } else {
+ movd mm_nbits, nbits ; nbits --> MMX register
+ psllq mm_put_buffer, mm_nbits ; put_buffer <<= nbits;
+ por mm_put_buffer, mm_code ; put_buffer |= code;
+.EFN: ; } }
+%define frame esp
+ mov frame, [t + save_frame]
+%define state ecx
+ mov state, [frame + arg_state]
+ movq [state + working_state.cur.put_buffer.simd], mm_put_buffer
+ ; state->cur.put_buffer.simd = put_buffer;
+ emms
+ mov [state + working_state.cur.free_bits], free_bits
+ ; state->cur.free_bits = free_bits;
+ POP edi
+ POP esi
+ POP ebp
+ POP ebx
ret
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.EMIT_BRLOOP1:
+ EMIT_QWORD emit_temp, emit_tempb, emit_temph, , , , , , \
+ .ERLOOP1
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.EMIT_ERLOOP1:
+ EMIT_QWORD size, sizeb, sizeh, \
+ { xor size, size }, \
+ { tzcnt size, index }, \
+ { inc size }, \
+ { test index, index }, \
+ { jnz .BLOOP1 }, \
+ .ELOOP1
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.EMIT_BRLOOP2:
+ EMIT_QWORD emit_temp, emit_tempb, emit_temph, , , , \
+ { cmp size, 16 }, \
+ { jle .ERLOOP2 }, \
+ .BRLOOP2
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.EMIT_ERLOOP2:
+ EMIT_QWORD size, sizeb, sizeh, \
+ { xor size, size }, \
+ { tzcnt size, index }, \
+ { inc size }, \
+ { test index, index }, \
+ { jnz .BLOOP2 }, \
+ .ELOOP2
+
; For some reason, the OS X linker does not honor the request to align the
; segment unless we do this.
align 32
diff --git a/simd/i386/jfdctint-avx2.asm b/simd/i386/jfdctint-avx2.asm
index 97de230..23cf733 100644
--- a/simd/i386/jfdctint-avx2.asm
+++ b/simd/i386/jfdctint-avx2.asm
@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (AVX2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, 2018, D. R. Commander.
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; forward DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jfdctint.c; see the jfdctint.c for
; more details.
@@ -103,7 +103,7 @@
%endmacro
; --------------------------------------------------------------------------
-; In-place 8x8x16-bit slow integer forward DCT using AVX2 instructions
+; In-place 8x8x16-bit accurate integer forward DCT using AVX2 instructions
; %1-%4: Input/output registers
; %5-%8: Temp registers
; %9: Pass (1 or 2)
diff --git a/simd/i386/jfdctint-mmx.asm b/simd/i386/jfdctint-mmx.asm
index 3ade9d4..34a43b9 100644
--- a/simd/i386/jfdctint-mmx.asm
+++ b/simd/i386/jfdctint-mmx.asm
@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (MMX)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2016, D. R. Commander.
+; Copyright (C) 2016, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; forward DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jfdctint.c; see the jfdctint.c for
; more details.
diff --git a/simd/i386/jfdctint-sse2.asm b/simd/i386/jfdctint-sse2.asm
index 71b684c..6f8e18c 100644
--- a/simd/i386/jfdctint-sse2.asm
+++ b/simd/i386/jfdctint-sse2.asm
@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2016, D. R. Commander.
+; Copyright (C) 2016, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; forward DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jfdctint.c; see the jfdctint.c for
; more details.
diff --git a/simd/i386/jidctint-avx2.asm b/simd/i386/jidctint-avx2.asm
index c371985..199c7df 100644
--- a/simd/i386/jidctint-avx2.asm
+++ b/simd/i386/jidctint-avx2.asm
@@ -2,7 +2,7 @@
; jidctint.asm - accurate integer IDCT (AVX2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, 2018, D. R. Commander.
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; inverse DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jidctint.c; see the jidctint.c for
; more details.
@@ -113,7 +113,7 @@
%endmacro
; --------------------------------------------------------------------------
-; In-place 8x8x16-bit slow integer inverse DCT using AVX2 instructions
+; In-place 8x8x16-bit accurate integer inverse DCT using AVX2 instructions
; %1-%4: Input/output registers
; %5-%12: Temp registers
; %9: Pass (1 or 2)
diff --git a/simd/i386/jidctint-mmx.asm b/simd/i386/jidctint-mmx.asm
index 4f07f56..f15c8d3 100644
--- a/simd/i386/jidctint-mmx.asm
+++ b/simd/i386/jidctint-mmx.asm
@@ -2,7 +2,7 @@
; jidctint.asm - accurate integer IDCT (MMX)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2016, D. R. Commander.
+; Copyright (C) 2016, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; inverse DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jidctint.c; see the jidctint.c for
; more details.
diff --git a/simd/i386/jidctint-sse2.asm b/simd/i386/jidctint-sse2.asm
index e442fdd..43e3201 100644
--- a/simd/i386/jidctint-sse2.asm
+++ b/simd/i386/jidctint-sse2.asm
@@ -2,7 +2,7 @@
; jidctint.asm - accurate integer IDCT (SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2016, D. R. Commander.
+; Copyright (C) 2016, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; inverse DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jidctint.c; see the jidctint.c for
; more details.
diff --git a/simd/i386/jsimd.c b/simd/i386/jsimd.c
index 2f92f8c..563949a 100644
--- a/simd/i386/jsimd.c
+++ b/simd/i386/jsimd.c
@@ -543,12 +543,6 @@
return 0;
}
-GLOBAL(int)
-jsimd_can_h1v2_fancy_upsample(void)
-{
- return 0;
-}
-
GLOBAL(void)
jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
@@ -585,12 +579,6 @@
output_data_ptr);
}
-GLOBAL(void)
-jsimd_h1v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
- JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
-{
-}
-
GLOBAL(int)
jsimd_can_h2v2_merged_upsample(void)
{
diff --git a/simd/jsimd.h b/simd/jsimd.h
index 99c8801..64747c6 100644
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -2,11 +2,12 @@
* simd/jsimd.h
*
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
- * Copyright (C) 2011, 2014-2016, 2018, D. R. Commander.
+ * Copyright (C) 2011, 2014-2016, 2018, 2020, D. R. Commander.
* Copyright (C) 2013-2014, MIPS Technologies, Inc., California.
* Copyright (C) 2014, Linaro Limited.
* Copyright (C) 2015-2016, 2018, Matthieu Darbois.
- * Copyright (C) 2016-2017, Loongson Technology Corporation Limited, BeiJing.
+ * Copyright (C) 2016-2018, Loongson Technology Corporation Limited, BeiJing.
+ * Copyright (C) 2020, Arm Limited.
*
* Based on the x86 SIMD extension for IJG JPEG library,
* Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -121,6 +122,17 @@
(JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
JDIMENSION output_row, int num_rows);
+#ifndef NEON_INTRINSICS
+
+EXTERN(void) jsimd_extrgb_ycc_convert_neon_slowld3
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extbgr_ycc_convert_neon_slowld3
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+
+#endif
+
EXTERN(void) jsimd_rgb_ycc_convert_dspr2
(JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
JDIMENSION output_row, int num_rows);
@@ -300,6 +312,28 @@
(JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_rgb_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extrgb_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extrgbx_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extbgr_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extbgrx_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extxbgr_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+EXTERN(void) jsimd_extxrgb_gray_convert_mmi
+ (JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+ JDIMENSION output_row, int num_rows);
+
EXTERN(void) jsimd_rgb_gray_convert_altivec
(JDIMENSION img_width, JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
JDIMENSION output_row, int num_rows);
@@ -416,6 +450,17 @@
(JDIMENSION out_width, JSAMPIMAGE input_buf, JDIMENSION input_row,
JSAMPARRAY output_buf, int num_rows);
+#ifndef NEON_INTRINSICS
+
+EXTERN(void) jsimd_ycc_extrgb_convert_neon_slowst3
+ (JDIMENSION out_width, JSAMPIMAGE input_buf, JDIMENSION input_row,
+ JSAMPARRAY output_buf, int num_rows);
+EXTERN(void) jsimd_ycc_extbgr_convert_neon_slowst3
+ (JDIMENSION out_width, JSAMPIMAGE input_buf, JDIMENSION input_row,
+ JSAMPARRAY output_buf, int num_rows);
+
+#endif
+
EXTERN(void) jsimd_ycc_rgb_convert_dspr2
(JDIMENSION out_width, JSAMPIMAGE input_buf, JDIMENSION input_row,
JSAMPARRAY output_buf, int num_rows);
@@ -637,6 +682,9 @@
(int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,
JSAMPARRAY *output_data_ptr);
+EXTERN(void) jsimd_h2v1_fancy_upsample_mmi
+ (int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,
+ JSAMPARRAY *output_data_ptr);
EXTERN(void) jsimd_h2v2_fancy_upsample_mmi
(int max_v_samp_factor, JDIMENSION downsampled_width, JSAMPARRAY input_data,
JSAMPARRAY *output_data_ptr);
@@ -871,6 +919,50 @@
(JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf, JSAMPLE *range);
+EXTERN(void) jsimd_h2v1_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+
+EXTERN(void) jsimd_h2v2_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_mmi
+ (JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+ JSAMPARRAY output_buf);
+
EXTERN(void) jsimd_h2v1_merged_upsample_altivec
(JDIMENSION output_width, JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
JSAMPARRAY output_buf);
@@ -947,7 +1039,7 @@
EXTERN(void) jsimd_convsamp_float_dspr2
(JSAMPARRAY sample_data, JDIMENSION start_col, FAST_FLOAT *workspace);
-/* Slow Integer Forward DCT */
+/* Accurate Integer Forward DCT */
EXTERN(void) jsimd_fdct_islow_mmx(DCTELEM *data);
extern const int jconst_fdct_islow_sse2[];
@@ -974,6 +1066,8 @@
EXTERN(void) jsimd_fdct_ifast_dspr2(DCTELEM *data);
+EXTERN(void) jsimd_fdct_ifast_mmi(DCTELEM *data);
+
EXTERN(void) jsimd_fdct_ifast_altivec(DCTELEM *data);
/* Floating Point Forward DCT */
@@ -1054,7 +1148,7 @@
EXTERN(void) jsimd_idct_12x12_pass2_dspr2
(int *workspace, int *output);
-/* Slow Integer Inverse DCT */
+/* Accurate Integer Inverse DCT */
EXTERN(void) jsimd_idct_islow_mmx
(void *dct_table, JCOEFPTR coef_block, JSAMPARRAY output_buf,
JDIMENSION output_col);
@@ -1105,6 +1199,10 @@
(DCTELEM *wsptr, JSAMPARRAY output_buf, JDIMENSION output_col,
const int *idct_coefs);
+EXTERN(void) jsimd_idct_ifast_mmi
+ (void *dct_table, JCOEFPTR coef_block, JSAMPARRAY output_buf,
+ JDIMENSION output_col);
+
EXTERN(void) jsimd_idct_ifast_altivec
(void *dct_table, JCOEFPTR coef_block, JSAMPARRAY output_buf,
JDIMENSION output_col);
@@ -1134,15 +1232,27 @@
(void *state, JOCTET *buffer, JCOEFPTR block, int last_dc_val,
c_derived_tbl *dctbl, c_derived_tbl *actbl);
+#ifndef NEON_INTRINSICS
+
EXTERN(JOCTET *) jsimd_huff_encode_one_block_neon_slowtbl
(void *state, JOCTET *buffer, JCOEFPTR block, int last_dc_val,
c_derived_tbl *dctbl, c_derived_tbl *actbl);
+#endif
+
/* Progressive Huffman encoding */
EXTERN(void) jsimd_encode_mcu_AC_first_prepare_sse2
(const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
JCOEF *values, size_t *zerobits);
+EXTERN(void) jsimd_encode_mcu_AC_first_prepare_neon
+ (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+ JCOEF *values, size_t *zerobits);
+
EXTERN(int) jsimd_encode_mcu_AC_refine_prepare_sse2
(const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
JCOEF *absvalues, size_t *bits);
+
+EXTERN(int) jsimd_encode_mcu_AC_refine_prepare_neon
+ (const JCOEF *block, const int *jpeg_natural_order_start, int Sl, int Al,
+ JCOEF *absvalues, size_t *bits);
diff --git a/simd/nasm/jsimdcfg.inc.h b/simd/nasm/jsimdcfg.inc.h
index 7ff7e29..bf2a45a 100644
--- a/simd/nasm/jsimdcfg.inc.h
+++ b/simd/nasm/jsimdcfg.inc.h
@@ -1,8 +1,10 @@
-// This file generates the include file for the assembly
-// implementations by abusing the C preprocessor.
-//
-// Note: Some things are manually defined as they need to
-// be mapped to NASM types.
+/*
+ * This file generates the include file for the assembly
+ * implementations by abusing the C preprocessor.
+ *
+ * Note: Some things are manually defined as they need to
+ * be mapped to NASM types.
+ */
;
; Automatically generated include file from jsimdcfg.inc.h
diff --git a/simd/nasm/jsimdext.inc b/simd/nasm/jsimdext.inc
index 9930d80..e8d50b0 100644
--- a/simd/nasm/jsimdext.inc
+++ b/simd/nasm/jsimdext.inc
@@ -2,8 +2,9 @@
; jsimdext.inc - common declarations
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2010, 2016, 2019, D. R. Commander.
+; Copyright (C) 2010, 2016, 2018-2019, D. R. Commander.
; Copyright (C) 2018, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library - version 1.02
;
@@ -130,13 +131,53 @@
; Common types
;
%ifdef __x86_64__
+%ifnidn __OUTPUT_FORMAT__, elfx32
%define POINTER qword ; general pointer type
%define SIZEOF_POINTER SIZEOF_QWORD ; sizeof(POINTER)
%define POINTER_BIT QWORD_BIT ; sizeof(POINTER)*BYTE_BIT
-%else
+%define resp resq
+%define dp dq
+%define raxp rax
+%define rbxp rbx
+%define rcxp rcx
+%define rdxp rdx
+%define rsip rsi
+%define rdip rdi
+%define rbpp rbp
+%define rspp rsp
+%define r8p r8
+%define r9p r9
+%define r10p r10
+%define r11p r11
+%define r12p r12
+%define r13p r13
+%define r14p r14
+%define r15p r15
+%endif
+%endif
+%ifndef raxp
%define POINTER dword ; general pointer type
%define SIZEOF_POINTER SIZEOF_DWORD ; sizeof(POINTER)
%define POINTER_BIT DWORD_BIT ; sizeof(POINTER)*BYTE_BIT
+%define resp resd
+%define dp dd
+; x86_64 ILP32 ABI (x32)
+%define raxp eax
+%define rbxp ebx
+%define rcxp ecx
+%define rdxp edx
+%define rsip esi
+%define rdip edi
+%define rbpp ebp
+%define rspp esp
+%define r8p r8d
+%define r9p r9d
+%define r10p r10d
+%define r11p r11d
+%define r12p r12d
+%define r13p r13d
+%define r14p r14d
+%define r15p r15d
%endif
%define INT dword ; signed integer type
diff --git a/simd/x86_64/jccolext-avx2.asm b/simd/x86_64/jccolext-avx2.asm
index 10d2834..ffb527d 100644
--- a/simd/x86_64/jccolext-avx2.asm
+++ b/simd/x86_64/jccolext-avx2.asm
@@ -3,6 +3,7 @@
;
; Copyright (C) 2009, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -57,9 +58,9 @@
mov rsi, r12
mov ecx, r13d
- mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
+ mov rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
lea rdi, [rdi+rcx*SIZEOF_JSAMPROW]
lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
@@ -77,10 +78,10 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr0
- mov rbx, JSAMPROW [rbx] ; outptr1
- mov rdx, JSAMPROW [rdx] ; outptr2
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr0
+ mov rbxp, JSAMPROW [rbx] ; outptr1
+ mov rdxp, JSAMPROW [rdx] ; outptr2
cmp rcx, byte SIZEOF_YMMWORD
jae near .columnloop
diff --git a/simd/x86_64/jccolext-sse2.asm b/simd/x86_64/jccolext-sse2.asm
index 2c914d3..af70ed6 100644
--- a/simd/x86_64/jccolext-sse2.asm
+++ b/simd/x86_64/jccolext-sse2.asm
@@ -2,6 +2,7 @@
; jccolext.asm - colorspace conversion (64-bit SSE2)
;
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -56,9 +57,9 @@
mov rsi, r12
mov ecx, r13d
- mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
+ mov rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
lea rdi, [rdi+rcx*SIZEOF_JSAMPROW]
lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
@@ -76,10 +77,10 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr0
- mov rbx, JSAMPROW [rbx] ; outptr1
- mov rdx, JSAMPROW [rdx] ; outptr2
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr0
+ mov rbxp, JSAMPROW [rbx] ; outptr1
+ mov rdxp, JSAMPROW [rdx] ; outptr2
cmp rcx, byte SIZEOF_XMMWORD
jae near .columnloop
diff --git a/simd/x86_64/jcgryext-avx2.asm b/simd/x86_64/jcgryext-avx2.asm
index 175b60d..ddcc2c0 100644
--- a/simd/x86_64/jcgryext-avx2.asm
+++ b/simd/x86_64/jcgryext-avx2.asm
@@ -3,6 +3,7 @@
;
; Copyright (C) 2011, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -57,7 +58,7 @@
mov rsi, r12
mov ecx, r13d
- mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+ mov rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
lea rdi, [rdi+rcx*SIZEOF_JSAMPROW]
pop rcx
@@ -71,8 +72,8 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr0
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr0
cmp rcx, byte SIZEOF_YMMWORD
jae near .columnloop
diff --git a/simd/x86_64/jcgryext-sse2.asm b/simd/x86_64/jcgryext-sse2.asm
index 873be80..f1d399a 100644
--- a/simd/x86_64/jcgryext-sse2.asm
+++ b/simd/x86_64/jcgryext-sse2.asm
@@ -2,6 +2,7 @@
; jcgryext.asm - grayscale colorspace conversion (64-bit SSE2)
;
; Copyright (C) 2011, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -56,7 +57,7 @@
mov rsi, r12
mov ecx, r13d
- mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
+ mov rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
lea rdi, [rdi+rcx*SIZEOF_JSAMPROW]
pop rcx
@@ -70,8 +71,8 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr0
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr0
cmp rcx, byte SIZEOF_XMMWORD
jae near .columnloop
diff --git a/simd/x86_64/jchuff-sse2.asm b/simd/x86_64/jchuff-sse2.asm
index 7deab58..0072028 100644
--- a/simd/x86_64/jchuff-sse2.asm
+++ b/simd/x86_64/jchuff-sse2.asm
@@ -1,8 +1,9 @@
;
; jchuff-sse2.asm - Huffman entropy encoding (64-bit SSE2)
;
-; Copyright (C) 2009-2011, 2014-2016, D. R. Commander.
+; Copyright (C) 2009-2011, 2014-2016, 2019, D. R. Commander.
; Copyright (C) 2015, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -15,146 +16,164 @@
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
; This file contains an SSE2 implementation for Huffman coding of one block.
-; The following code is based directly on jchuff.c; see jchuff.c for more
-; details.
+; The following code is based on jchuff.c; see jchuff.c for more details.
%include "jsimdext.inc"
+struc working_state
+.next_output_byte: resp 1 ; => next byte to write in buffer
+.free_in_buffer: resp 1 ; # of byte spaces remaining in buffer
+.cur.put_buffer.simd resq 1 ; current bit accumulation buffer
+.cur.free_bits resd 1 ; # of bits available in it
+.cur.last_dc_val resd 4 ; last DC coef for each component
+.cinfo: resp 1 ; dump_buffer needs access to this
+endstruc
+
+struc c_derived_tbl
+.ehufco: resd 256 ; code for each symbol
+.ehufsi: resb 256 ; length of code for each symbol
+; If no code has been allocated for a symbol S, ehufsi[S] contains 0
+endstruc
+
; --------------------------------------------------------------------------
SECTION SEG_CONST
alignz 32
GLOBAL_DATA(jconst_huff_encode_one_block)
- EXTERN EXTN(jpeg_nbits_table)
EXTN(jconst_huff_encode_one_block):
+jpeg_mask_bits dd 0x0000, 0x0001, 0x0003, 0x0007
+ dd 0x000f, 0x001f, 0x003f, 0x007f
+ dd 0x00ff, 0x01ff, 0x03ff, 0x07ff
+ dd 0x0fff, 0x1fff, 0x3fff, 0x7fff
+
alignz 32
+times 1 << 14 db 15
+times 1 << 13 db 14
+times 1 << 12 db 13
+times 1 << 11 db 12
+times 1 << 10 db 11
+times 1 << 9 db 10
+times 1 << 8 db 9
+times 1 << 7 db 8
+times 1 << 6 db 7
+times 1 << 5 db 6
+times 1 << 4 db 5
+times 1 << 3 db 4
+times 1 << 2 db 3
+times 1 << 1 db 2
+times 1 << 0 db 1
+times 1 db 0
+jpeg_nbits_table:
+times 1 db 0
+times 1 << 0 db 1
+times 1 << 1 db 2
+times 1 << 2 db 3
+times 1 << 3 db 4
+times 1 << 4 db 5
+times 1 << 5 db 6
+times 1 << 6 db 7
+times 1 << 7 db 8
+times 1 << 8 db 9
+times 1 << 9 db 10
+times 1 << 10 db 11
+times 1 << 11 db 12
+times 1 << 12 db 13
+times 1 << 13 db 14
+times 1 << 14 db 15
+
+ alignz 32
+
+%define NBITS(x) nbits_base + x
+%define MASK_BITS(x) NBITS((x) * 4) + (jpeg_mask_bits - jpeg_nbits_table)
+
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 64
-; These macros perform the same task as the emit_bits() function in the
-; original libjpeg code. In addition to reducing overhead by explicitly
-; inlining the code, additional performance is achieved by taking into
-; account the size of the bit buffer and waiting until it is almost full
-; before emptying it. This mostly benefits 64-bit platforms, since 6
-; bytes can be stored in a 64-bit bit buffer before it has to be emptied.
+; Shorthand used to describe SIMD operations:
+; wN: xmmN treated as eight signed 16-bit values
+; wN[i]: perform the same operation on all eight signed 16-bit values, i=0..7
+; bN: xmmN treated as 16 unsigned 8-bit values
+; bN[i]: perform the same operation on all 16 unsigned 8-bit values, i=0..15
+; Contents of SIMD registers are shown in memory order.
-%macro EMIT_BYTE 0
- sub put_bits, 8 ; put_bits -= 8;
- mov rdx, put_buffer
- mov ecx, put_bits
- shr rdx, cl ; c = (JOCTET)GETJOCTET(put_buffer >> put_bits);
- mov byte [buffer], dl ; *buffer++ = c;
- add buffer, 1
- cmp dl, 0xFF ; need to stuff a zero byte?
- jne %%.EMIT_BYTE_END
- mov byte [buffer], 0 ; *buffer++ = 0;
- add buffer, 1
-%%.EMIT_BYTE_END:
-%endmacro
+; Fill the bit buffer to capacity with the leading bits from code, then output
+; the bit buffer and put the remaining bits from code into the bit buffer.
+;
+; Usage:
+; code - contains the bits to shift into the bit buffer (LSB-aligned)
+; %1 - the label to which to jump when the macro completes
+; %2 (optional) - extra instructions to execute after nbits has been set
+;
+; Upon completion, free_bits will be set to the number of remaining bits from
+; code, and put_buffer will contain those remaining bits. temp and code will
+; be clobbered.
+;
+; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE()
+; macro in jchuff.c.
-%macro PUT_BITS 1
- add put_bits, ecx ; put_bits += size;
- shl put_buffer, cl ; put_buffer = (put_buffer << size);
- or put_buffer, %1
-%endmacro
-
-%macro CHECKBUF31 0
- cmp put_bits, 32 ; if (put_bits > 31) {
- jl %%.CHECKBUF31_END
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
-%%.CHECKBUF31_END:
-%endmacro
-
-%macro CHECKBUF47 0
- cmp put_bits, 48 ; if (put_bits > 47) {
- jl %%.CHECKBUF47_END
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
- EMIT_BYTE
-%%.CHECKBUF47_END:
-%endmacro
-
-%macro EMIT_BITS 2
- CHECKBUF47
- mov ecx, %2
- PUT_BITS %1
-%endmacro
-
-%macro kloop_prepare 37 ;(ko, jno0, ..., jno31, xmm0, xmm1, xmm2, xmm3)
- pxor xmm8, xmm8 ; __m128i neg = _mm_setzero_si128();
- pxor xmm9, xmm9 ; __m128i neg = _mm_setzero_si128();
- pxor xmm10, xmm10 ; __m128i neg = _mm_setzero_si128();
- pxor xmm11, xmm11 ; __m128i neg = _mm_setzero_si128();
- pinsrw %34, word [r12 + %2 * SIZEOF_WORD], 0 ; xmm_shadow[0] = block[jno0];
- pinsrw %35, word [r12 + %10 * SIZEOF_WORD], 0 ; xmm_shadow[8] = block[jno8];
- pinsrw %36, word [r12 + %18 * SIZEOF_WORD], 0 ; xmm_shadow[16] = block[jno16];
- pinsrw %37, word [r12 + %26 * SIZEOF_WORD], 0 ; xmm_shadow[24] = block[jno24];
- pinsrw %34, word [r12 + %3 * SIZEOF_WORD], 1 ; xmm_shadow[1] = block[jno1];
- pinsrw %35, word [r12 + %11 * SIZEOF_WORD], 1 ; xmm_shadow[9] = block[jno9];
- pinsrw %36, word [r12 + %19 * SIZEOF_WORD], 1 ; xmm_shadow[17] = block[jno17];
- pinsrw %37, word [r12 + %27 * SIZEOF_WORD], 1 ; xmm_shadow[25] = block[jno25];
- pinsrw %34, word [r12 + %4 * SIZEOF_WORD], 2 ; xmm_shadow[2] = block[jno2];
- pinsrw %35, word [r12 + %12 * SIZEOF_WORD], 2 ; xmm_shadow[10] = block[jno10];
- pinsrw %36, word [r12 + %20 * SIZEOF_WORD], 2 ; xmm_shadow[18] = block[jno18];
- pinsrw %37, word [r12 + %28 * SIZEOF_WORD], 2 ; xmm_shadow[26] = block[jno26];
- pinsrw %34, word [r12 + %5 * SIZEOF_WORD], 3 ; xmm_shadow[3] = block[jno3];
- pinsrw %35, word [r12 + %13 * SIZEOF_WORD], 3 ; xmm_shadow[11] = block[jno11];
- pinsrw %36, word [r12 + %21 * SIZEOF_WORD], 3 ; xmm_shadow[19] = block[jno19];
- pinsrw %37, word [r12 + %29 * SIZEOF_WORD], 3 ; xmm_shadow[27] = block[jno27];
- pinsrw %34, word [r12 + %6 * SIZEOF_WORD], 4 ; xmm_shadow[4] = block[jno4];
- pinsrw %35, word [r12 + %14 * SIZEOF_WORD], 4 ; xmm_shadow[12] = block[jno12];
- pinsrw %36, word [r12 + %22 * SIZEOF_WORD], 4 ; xmm_shadow[20] = block[jno20];
- pinsrw %37, word [r12 + %30 * SIZEOF_WORD], 4 ; xmm_shadow[28] = block[jno28];
- pinsrw %34, word [r12 + %7 * SIZEOF_WORD], 5 ; xmm_shadow[5] = block[jno5];
- pinsrw %35, word [r12 + %15 * SIZEOF_WORD], 5 ; xmm_shadow[13] = block[jno13];
- pinsrw %36, word [r12 + %23 * SIZEOF_WORD], 5 ; xmm_shadow[21] = block[jno21];
- pinsrw %37, word [r12 + %31 * SIZEOF_WORD], 5 ; xmm_shadow[29] = block[jno29];
- pinsrw %34, word [r12 + %8 * SIZEOF_WORD], 6 ; xmm_shadow[6] = block[jno6];
- pinsrw %35, word [r12 + %16 * SIZEOF_WORD], 6 ; xmm_shadow[14] = block[jno14];
- pinsrw %36, word [r12 + %24 * SIZEOF_WORD], 6 ; xmm_shadow[22] = block[jno22];
- pinsrw %37, word [r12 + %32 * SIZEOF_WORD], 6 ; xmm_shadow[30] = block[jno30];
- pinsrw %34, word [r12 + %9 * SIZEOF_WORD], 7 ; xmm_shadow[7] = block[jno7];
- pinsrw %35, word [r12 + %17 * SIZEOF_WORD], 7 ; xmm_shadow[15] = block[jno15];
- pinsrw %36, word [r12 + %25 * SIZEOF_WORD], 7 ; xmm_shadow[23] = block[jno23];
-%if %1 != 32
- pinsrw %37, word [r12 + %33 * SIZEOF_WORD], 7 ; xmm_shadow[31] = block[jno31];
-%else
- pinsrw %37, ebx, 7 ; xmm_shadow[31] = block[jno31];
-%endif
- pcmpgtw xmm8, %34 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm9, %35 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm10, %36 ; neg = _mm_cmpgt_epi16(neg, x1);
- pcmpgtw xmm11, %37 ; neg = _mm_cmpgt_epi16(neg, x1);
- paddw %34, xmm8 ; x1 = _mm_add_epi16(x1, neg);
- paddw %35, xmm9 ; x1 = _mm_add_epi16(x1, neg);
- paddw %36, xmm10 ; x1 = _mm_add_epi16(x1, neg);
- paddw %37, xmm11 ; x1 = _mm_add_epi16(x1, neg);
- pxor %34, xmm8 ; x1 = _mm_xor_si128(x1, neg);
- pxor %35, xmm9 ; x1 = _mm_xor_si128(x1, neg);
- pxor %36, xmm10 ; x1 = _mm_xor_si128(x1, neg);
- pxor %37, xmm11 ; x1 = _mm_xor_si128(x1, neg);
- pxor xmm8, %34 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm9, %35 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm10, %36 ; neg = _mm_xor_si128(neg, x1);
- pxor xmm11, %37 ; neg = _mm_xor_si128(neg, x1);
- movdqa XMMWORD [t1 + %1 * SIZEOF_WORD], %34 ; _mm_storeu_si128((__m128i *)(t1 + ko), x1);
- movdqa XMMWORD [t1 + (%1 + 8) * SIZEOF_WORD], %35 ; _mm_storeu_si128((__m128i *)(t1 + ko + 8), x1);
- movdqa XMMWORD [t1 + (%1 + 16) * SIZEOF_WORD], %36 ; _mm_storeu_si128((__m128i *)(t1 + ko + 16), x1);
- movdqa XMMWORD [t1 + (%1 + 24) * SIZEOF_WORD], %37 ; _mm_storeu_si128((__m128i *)(t1 + ko + 24), x1);
- movdqa XMMWORD [t2 + %1 * SIZEOF_WORD], xmm8 ; _mm_storeu_si128((__m128i *)(t2 + ko), neg);
- movdqa XMMWORD [t2 + (%1 + 8) * SIZEOF_WORD], xmm9 ; _mm_storeu_si128((__m128i *)(t2 + ko + 8), neg);
- movdqa XMMWORD [t2 + (%1 + 16) * SIZEOF_WORD], xmm10 ; _mm_storeu_si128((__m128i *)(t2 + ko + 16), neg);
- movdqa XMMWORD [t2 + (%1 + 24) * SIZEOF_WORD], xmm11 ; _mm_storeu_si128((__m128i *)(t2 + ko + 24), neg);
+%macro EMIT_QWORD 1-2
+ add nbitsb, free_bitsb ; nbits += free_bits;
+ neg free_bitsb ; free_bits = -free_bits;
+ mov tempd, code ; temp = code;
+ shl put_buffer, nbitsb ; put_buffer <<= nbits;
+ mov nbitsb, free_bitsb ; nbits = free_bits;
+ neg free_bitsb ; free_bits = -free_bits;
+ shr tempd, nbitsb ; temp >>= nbits;
+ or tempq, put_buffer ; temp |= put_buffer;
+ movq xmm0, tempq ; xmm0.u64 = { temp, 0 };
+ bswap tempq ; temp = htonl(temp);
+ mov put_buffer, codeq ; put_buffer = code;
+ pcmpeqb xmm0, xmm1 ; b0[i] = (b0[i] == 0xFF ? 0xFF : 0);
+ %2
+ pmovmskb code, xmm0 ; code = 0; code |= ((b0[i] >> 7) << i);
+ mov qword [buffer], tempq ; memcpy(buffer, &temp, 8);
+ ; (speculative; will be overwritten if
+ ; code contains any 0xFF bytes)
+ add free_bitsb, 64 ; free_bits += 64;
+ add bufferp, 8 ; buffer += 8;
+ test code, code ; if (code == 0) /* No 0xFF bytes */
+ jz %1 ; return;
+ ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8
+ ; bytes in the qword.
+ cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer-7], 0 ; buffer[-7] = 0;
+ sbb bufferp, 6 ; buffer -= (6 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], temph ; buffer[0] = temp[1];
+ cmp temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ shr tempq, 16 ; temp >>= 16;
+ mov byte [buffer], tempb ; buffer[0] = temp[0];
+ cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], temph ; buffer[0] = temp[1];
+ cmp temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ shr tempq, 16 ; temp >>= 16;
+ mov byte [buffer], tempb ; buffer[0] = temp[0];
+ cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], temph ; buffer[0] = temp[1];
+ cmp temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ shr tempd, 16 ; temp >>= 16;
+ mov byte [buffer], tempb ; buffer[0] = temp[0];
+ cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0));
+ mov byte [buffer], temph ; buffer[0] = temp[1];
+ cmp temph, 0xFF ; Set CF if temp[1] < 0xFF
+ mov byte [buffer+1], 0 ; buffer[1] = 0;
+ sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0));
+ jmp %1 ; return;
%endmacro
;
@@ -165,181 +184,399 @@
; JCOEFPTR block, int last_dc_val,
; c_derived_tbl *dctbl, c_derived_tbl *actbl)
;
+; NOTES:
+; When shuffling data, we try to avoid pinsrw as much as possible, since it is
+; slow on many CPUs. Its reciprocal throughput (issue latency) is 1 even on
+; modern CPUs, so chains of pinsrw instructions (even with different outputs)
+; can limit performance. pinsrw is a VectorPath instruction on AMD K8 and
+; requires 2 µops (with memory operand) on Intel. In either case, only one
+; pinsrw instruction can be decoded per cycle (and nothing else if they are
+; back-to-back), so out-of-order execution cannot be used to work around long
+; pinsrw chains (though for Sandy Bridge and later, this may be less of a
+; problem if the code runs from the µop cache.)
+;
+; We use tzcnt instead of bsf without checking for support. The instruction is
+; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to
+; rep bsf.) The destination (first) operand of bsf (and tzcnt on some CPUs) is
+; an input dependency (although the behavior is not formally defined, Intel
+; CPUs usually leave the destination unmodified if the source is zero.) This
+; can prevent out-of-order execution, so we clear the destination before
+; invoking tzcnt.
+;
+; Initial register allocation
+; rax - buffer
+; rbx - temp
+; rcx - nbits
+; rdx - block --> free_bits
+; rsi - nbits_base
+; rdi - t
+; rbp - code
+; r8 - dctbl --> code_temp
+; r9 - actbl
+; r10 - state
+; r11 - index
+; r12 - put_buffer
-; r10 = working_state *state
-; r11 = JOCTET *buffer
-; r12 = JCOEFPTR block
-; r13d = int last_dc_val
-; r14 = c_derived_tbl *dctbl
-; r15 = c_derived_tbl *actbl
+%define buffer rax
+%ifdef WIN64
+%define bufferp rax
+%else
+%define bufferp raxp
+%endif
+%define tempq rbx
+%define tempd ebx
+%define tempb bl
+%define temph bh
+%define nbitsq rcx
+%define nbits ecx
+%define nbitsb cl
+%define block rdx
+%define nbits_base rsi
+%define t rdi
+%define td edi
+%define codeq rbp
+%define code ebp
+%define dctbl r8
+%define actbl r9
+%define state r10
+%define index r11
+%define indexd r11d
+%define put_buffer r12
+%define put_bufferd r12d
-%define t1 rbp - (DCTSIZE2 * SIZEOF_WORD)
-%define t2 t1 - (DCTSIZE2 * SIZEOF_WORD)
-%define put_buffer r8
-%define put_bits r9d
-%define buffer rax
+; Step 1: Re-arrange input data according to jpeg_natural_order
+; xx 01 02 03 04 05 06 07 xx 01 08 16 09 02 03 10
+; 08 09 10 11 12 13 14 15 17 24 32 25 18 11 04 05
+; 16 17 18 19 20 21 22 23 12 19 26 33 40 48 41 34
+; 24 25 26 27 28 29 30 31 ==> 27 20 13 06 07 14 21 28
+; 32 33 34 35 36 37 38 39 35 42 49 56 57 50 43 36
+; 40 41 42 43 44 45 46 47 29 22 15 23 30 37 44 51
+; 48 49 50 51 52 53 54 55 58 59 52 45 38 31 39 46
+; 56 57 58 59 60 61 62 63 53 60 61 54 47 55 62 63
align 32
GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
EXTN(jsimd_huff_encode_one_block_sse2):
- push rbp
- mov rax, rsp ; rax = original rbp
- sub rsp, byte 4
- and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
- mov [rsp], rax
- mov rbp, rsp ; rbp = aligned rbp
- lea rsp, [t2]
- push_xmm 4
- collect_args 6
+
+%ifdef WIN64
+
+; rcx = working_state *state
+; rdx = JOCTET *buffer
+; r8 = JCOEFPTR block
+; r9 = int last_dc_val
+; [rax+48] = c_derived_tbl *dctbl
+; [rax+56] = c_derived_tbl *actbl
+
+ ;X: X = code stream
+ mov buffer, rdx
+ mov block, r8
+ movups xmm3, XMMWORD [block + 0 * SIZEOF_WORD] ;D: w3 = xx 01 02 03 04 05 06 07
push rbx
+ push rbp
+ movdqa xmm0, xmm3 ;A: w0 = xx 01 02 03 04 05 06 07
+ push rsi
+ push rdi
+ push r12
+ movups xmm1, XMMWORD [block + 8 * SIZEOF_WORD] ;B: w1 = 08 09 10 11 12 13 14 15
+ mov state, rcx
+ movsx code, word [block] ;Z: code = block[0];
+ pxor xmm4, xmm4 ;A: w4[i] = 0;
+ sub code, r9d ;Z: code -= last_dc_val;
+ mov dctbl, POINTER [rsp+6*8+4*8]
+ mov actbl, POINTER [rsp+6*8+5*8]
+ punpckldq xmm0, xmm1 ;A: w0 = xx 01 08 09 02 03 10 11
+ lea nbits_base, [rel jpeg_nbits_table]
+ add rsp, -DCTSIZE2 * SIZEOF_WORD
+ mov t, rsp
- mov buffer, r11 ; r11 is now sratch
+%else
- mov put_buffer, MMWORD [r10+16] ; put_buffer = state->cur.put_buffer;
- mov put_bits, dword [r10+24] ; put_bits = state->cur.put_bits;
- push r10 ; r10 is now scratch
+; rdi = working_state *state
+; rsi = JOCTET *buffer
+; rdx = JCOEFPTR block
+; rcx = int last_dc_val
+; r8 = c_derived_tbl *dctbl
+; r9 = c_derived_tbl *actbl
- ; Encode the DC coefficient difference per section F.1.2.1
- movsx edi, word [r12] ; temp = temp2 = block[0] - last_dc_val;
- sub edi, r13d ; r13 is not used anymore
- mov ebx, edi
+ ;X: X = code stream
+ movups xmm3, XMMWORD [block + 0 * SIZEOF_WORD] ;D: w3 = xx 01 02 03 04 05 06 07
+ push rbx
+ push rbp
+ movdqa xmm0, xmm3 ;A: w0 = xx 01 02 03 04 05 06 07
+ push r12
+ mov state, rdi
+ mov buffer, rsi
+ movups xmm1, XMMWORD [block + 8 * SIZEOF_WORD] ;B: w1 = 08 09 10 11 12 13 14 15
+ movsx codeq, word [block] ;Z: code = block[0];
+ lea nbits_base, [rel jpeg_nbits_table]
+ pxor xmm4, xmm4 ;A: w4[i] = 0;
+ sub codeq, rcx ;Z: code -= last_dc_val;
+ punpckldq xmm0, xmm1 ;A: w0 = xx 01 08 09 02 03 10 11
+ lea t, [rsp - DCTSIZE2 * SIZEOF_WORD] ; use red zone for t_
- ; This is a well-known technique for obtaining the absolute value
- ; without a branch. It is derived from an assembly language technique
- ; presented in "How to Optimize for the Pentium Processors",
- ; Copyright (c) 1996, 1997 by Agner Fog.
- mov esi, edi
- sar esi, 31 ; temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
- xor edi, esi ; temp ^= temp3;
- sub edi, esi ; temp -= temp3;
+%endif
- ; For a negative input, want temp2 = bitwise complement of abs(input)
- ; This code assumes we are on a two's complement machine
- add ebx, esi ; temp2 += temp3;
+ pshuflw xmm0, xmm0, 11001001b ;A: w0 = 01 08 xx 09 02 03 10 11
+ pinsrw xmm0, word [block + 16 * SIZEOF_WORD], 2 ;A: w0 = 01 08 16 09 02 03 10 11
+ punpckhdq xmm3, xmm1 ;D: w3 = 04 05 12 13 06 07 14 15
+ punpcklqdq xmm1, xmm3 ;B: w1 = 08 09 10 11 04 05 12 13
+ pinsrw xmm0, word [block + 17 * SIZEOF_WORD], 7 ;A: w0 = 01 08 16 09 02 03 10 17
+ ;A: (Row 0, offset 1)
+ pcmpgtw xmm4, xmm0 ;A: w4[i] = (w0[i] < 0 ? -1 : 0);
+ paddw xmm0, xmm4 ;A: w0[i] += w4[i];
+ movaps XMMWORD [t + 0 * SIZEOF_WORD], xmm0 ;A: t[i] = w0[i];
- ; Find the number of bits needed for the magnitude of the coefficient
- lea r11, [rel EXTN(jpeg_nbits_table)]
- movzx rdi, byte [r11 + rdi] ; nbits = JPEG_NBITS(temp);
- ; Emit the Huffman-coded symbol for the number of bits
- mov r11d, INT [r14 + rdi * 4] ; code = dctbl->ehufco[nbits];
- movzx esi, byte [r14 + rdi + 1024] ; size = dctbl->ehufsi[nbits];
- EMIT_BITS r11, esi ; EMIT_BITS(code, size)
+ movq xmm2, qword [block + 24 * SIZEOF_WORD] ;B: w2 = 24 25 26 27 -- -- -- --
+ pshuflw xmm2, xmm2, 11011000b ;B: w2 = 24 26 25 27 -- -- -- --
+ pslldq xmm1, 1 * SIZEOF_WORD ;B: w1 = -- 08 09 10 11 04 05 12
+ movups xmm5, XMMWORD [block + 48 * SIZEOF_WORD] ;H: w5 = 48 49 50 51 52 53 54 55
+ movsd xmm1, xmm2 ;B: w1 = 24 26 25 27 11 04 05 12
+ punpcklqdq xmm2, xmm5 ;C: w2 = 24 26 25 27 48 49 50 51
+ pinsrw xmm1, word [block + 32 * SIZEOF_WORD], 1 ;B: w1 = 24 32 25 27 11 04 05 12
+ pxor xmm4, xmm4 ;A: w4[i] = 0;
+ psrldq xmm3, 2 * SIZEOF_WORD ;D: w3 = 12 13 06 07 14 15 -- --
+ pcmpeqw xmm0, xmm4 ;A: w0[i] = (w0[i] == 0 ? -1 : 0);
+ pinsrw xmm1, word [block + 18 * SIZEOF_WORD], 3 ;B: w1 = 24 32 25 18 11 04 05 12
+ ; (Row 1, offset 1)
+ pcmpgtw xmm4, xmm1 ;B: w4[i] = (w1[i] < 0 ? -1 : 0);
+ paddw xmm1, xmm4 ;B: w1[i] += w4[i];
+ movaps XMMWORD [t + 8 * SIZEOF_WORD], xmm1 ;B: t[i+8] = w1[i];
+ pxor xmm4, xmm4 ;B: w4[i] = 0;
+ pcmpeqw xmm1, xmm4 ;B: w1[i] = (w1[i] == 0 ? -1 : 0);
- ; Mask off any extra bits in code
- mov esi, 1
- mov ecx, edi
- shl esi, cl
- dec esi
- and ebx, esi ; temp2 &= (((JLONG)1)<<nbits) - 1;
+ packsswb xmm0, xmm1 ;AB: b0[i] = w0[i], b0[i+8] = w1[i]
+ ; w/ signed saturation
- ; Emit that number of bits of the value, if positive,
- ; or the complement of its magnitude, if negative.
- EMIT_BITS rbx, edi ; EMIT_BITS(temp2, nbits)
+ pinsrw xmm3, word [block + 20 * SIZEOF_WORD], 0 ;D: w3 = 20 13 06 07 14 15 -- --
+ pinsrw xmm3, word [block + 21 * SIZEOF_WORD], 5 ;D: w3 = 20 13 06 07 14 21 -- --
+ pinsrw xmm3, word [block + 28 * SIZEOF_WORD], 6 ;D: w3 = 20 13 06 07 14 21 28 --
+ pinsrw xmm3, word [block + 35 * SIZEOF_WORD], 7 ;D: w3 = 20 13 06 07 14 21 28 35
+ ; (Row 3, offset 1)
+ pcmpgtw xmm4, xmm3 ;D: w4[i] = (w3[i] < 0 ? -1 : 0);
+ paddw xmm3, xmm4 ;D: w3[i] += w4[i];
+ movaps XMMWORD [t + 24 * SIZEOF_WORD], xmm3 ;D: t[i+24] = w3[i];
+ pxor xmm4, xmm4 ;D: w4[i] = 0;
+ pcmpeqw xmm3, xmm4 ;D: w3[i] = (w3[i] == 0 ? -1 : 0);
- ; Prepare data
- xor ebx, ebx
- kloop_prepare 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, \
- 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, \
- 27, 20, 13, 6, 7, 14, 21, 28, 35, \
- xmm0, xmm1, xmm2, xmm3
- kloop_prepare 32, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, \
- 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, \
- 53, 60, 61, 54, 47, 55, 62, 63, 63, \
- xmm4, xmm5, xmm6, xmm7
+ pinsrw xmm2, word [block + 19 * SIZEOF_WORD], 0 ;C: w2 = 19 26 25 27 48 49 50 51
+ cmp code, 1 << 31 ;Z: Set CF if code < 0x80000000,
+ ;Z: i.e. if code is positive
+ pinsrw xmm2, word [block + 33 * SIZEOF_WORD], 2 ;C: w2 = 19 26 33 27 48 49 50 51
+ pinsrw xmm2, word [block + 40 * SIZEOF_WORD], 3 ;C: w2 = 19 26 33 40 48 49 50 51
+ adc code, -1 ;Z: code += -1 + (code >= 0 ? 1 : 0);
+ pinsrw xmm2, word [block + 41 * SIZEOF_WORD], 5 ;C: w2 = 19 26 33 40 48 41 50 51
+ pinsrw xmm2, word [block + 34 * SIZEOF_WORD], 6 ;C: w2 = 19 26 33 40 48 41 34 51
+ movsxd codeq, code ;Z: sign extend code
+ pinsrw xmm2, word [block + 27 * SIZEOF_WORD], 7 ;C: w2 = 19 26 33 40 48 41 34 27
+ ; (Row 2, offset 1)
+ pcmpgtw xmm4, xmm2 ;C: w4[i] = (w2[i] < 0 ? -1 : 0);
+ paddw xmm2, xmm4 ;C: w2[i] += w4[i];
+ movaps XMMWORD [t + 16 * SIZEOF_WORD], xmm2 ;C: t[i+16] = w2[i];
+ pxor xmm4, xmm4 ;C: w4[i] = 0;
+ pcmpeqw xmm2, xmm4 ;C: w2[i] = (w2[i] == 0 ? -1 : 0);
- pxor xmm8, xmm8
- pcmpeqw xmm0, xmm8 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero);
- pcmpeqw xmm1, xmm8 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero);
- pcmpeqw xmm2, xmm8 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero);
- pcmpeqw xmm3, xmm8 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero);
- pcmpeqw xmm4, xmm8 ; tmp4 = _mm_cmpeq_epi16(tmp4, zero);
- pcmpeqw xmm5, xmm8 ; tmp5 = _mm_cmpeq_epi16(tmp5, zero);
- pcmpeqw xmm6, xmm8 ; tmp6 = _mm_cmpeq_epi16(tmp6, zero);
- pcmpeqw xmm7, xmm8 ; tmp7 = _mm_cmpeq_epi16(tmp7, zero);
- packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1);
- packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3);
- packsswb xmm4, xmm5 ; tmp4 = _mm_packs_epi16(tmp4, tmp5);
- packsswb xmm6, xmm7 ; tmp6 = _mm_packs_epi16(tmp6, tmp7);
- pmovmskb r11d, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0;
- pmovmskb r12d, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16;
- pmovmskb r13d, xmm4 ; index = ((uint64_t)_mm_movemask_epi8(tmp4)) << 32;
- pmovmskb r14d, xmm6 ; index = ((uint64_t)_mm_movemask_epi8(tmp6)) << 48;
- shl r12, 16
- shl r14, 16
- or r11, r12
- or r13, r14
- shl r13, 32
- or r11, r13
- not r11 ; index = ~index;
+ packsswb xmm2, xmm3 ;CD: b2[i] = w2[i], b2[i+8] = w3[i]
+ ; w/ signed saturation
- ;mov MMWORD [ t1 + DCTSIZE2 * SIZEOF_WORD ], r11
- ;jmp .EFN
+ movzx nbitsq, byte [NBITS(codeq)] ;Z: nbits = JPEG_NBITS(code);
+ movdqa xmm3, xmm5 ;H: w3 = 48 49 50 51 52 53 54 55
+ pmovmskb tempd, xmm2 ;Z: temp = 0; temp |= ((b2[i] >> 7) << i);
+ pmovmskb put_bufferd, xmm0 ;Z: put_buffer = 0; put_buffer |= ((b0[i] >> 7) << i);
+ movups xmm0, XMMWORD [block + 56 * SIZEOF_WORD] ;H: w0 = 56 57 58 59 60 61 62 63
+ punpckhdq xmm3, xmm0 ;H: w3 = 52 53 60 61 54 55 62 63
+ shl tempd, 16 ;Z: temp <<= 16;
+ psrldq xmm3, 1 * SIZEOF_WORD ;H: w3 = 53 60 61 54 55 62 63 --
+ pxor xmm2, xmm2 ;H: w2[i] = 0;
+ or put_bufferd, tempd ;Z: put_buffer |= temp;
+ pshuflw xmm3, xmm3, 00111001b ;H: w3 = 60 61 54 53 55 62 63 --
+ movq xmm1, qword [block + 44 * SIZEOF_WORD] ;G: w1 = 44 45 46 47 -- -- -- --
+ unpcklps xmm5, xmm0 ;E: w5 = 48 49 56 57 50 51 58 59
+ pxor xmm0, xmm0 ;H: w0[i] = 0;
+ pinsrw xmm3, word [block + 47 * SIZEOF_WORD], 3 ;H: w3 = 60 61 54 47 55 62 63 --
+ ; (Row 7, offset 1)
+ pcmpgtw xmm2, xmm3 ;H: w2[i] = (w3[i] < 0 ? -1 : 0);
+ paddw xmm3, xmm2 ;H: w3[i] += w2[i];
+ movaps XMMWORD [t + 56 * SIZEOF_WORD], xmm3 ;H: t[i+56] = w3[i];
+ movq xmm4, qword [block + 36 * SIZEOF_WORD] ;G: w4 = 36 37 38 39 -- -- -- --
+ pcmpeqw xmm3, xmm0 ;H: w3[i] = (w3[i] == 0 ? -1 : 0);
+ punpckldq xmm4, xmm1 ;G: w4 = 36 37 44 45 38 39 46 47
+ mov tempd, [dctbl + c_derived_tbl.ehufco + nbitsq * 4]
+ ;Z: temp = dctbl->ehufco[nbits];
+ movdqa xmm1, xmm4 ;F: w1 = 36 37 44 45 38 39 46 47
+ psrldq xmm4, 1 * SIZEOF_WORD ;G: w4 = 37 44 45 38 39 46 47 --
+ shufpd xmm1, xmm5, 10b ;F: w1 = 36 37 44 45 50 51 58 59
+ and code, dword [MASK_BITS(nbitsq)] ;Z: code &= (1 << nbits) - 1;
+ pshufhw xmm4, xmm4, 11010011b ;G: w4 = 37 44 45 38 -- 39 46 --
+ pslldq xmm1, 1 * SIZEOF_WORD ;F: w1 = -- 36 37 44 45 50 51 58
+ shl tempq, nbitsb ;Z: temp <<= nbits;
+ pinsrw xmm4, word [block + 59 * SIZEOF_WORD], 0 ;G: w4 = 59 44 45 38 -- 39 46 --
+ pshufd xmm1, xmm1, 11011000b ;F: w1 = -- 36 45 50 37 44 51 58
+ pinsrw xmm4, word [block + 52 * SIZEOF_WORD], 1 ;G: w4 = 59 52 45 38 -- 39 46 --
+ or code, tempd ;Z: code |= temp;
+ movlps xmm1, qword [block + 20 * SIZEOF_WORD] ;F: w1 = 20 21 22 23 37 44 51 58
+ pinsrw xmm4, word [block + 31 * SIZEOF_WORD], 4 ;G: w4 = 59 52 45 38 31 39 46 --
+ pshuflw xmm1, xmm1, 01110010b ;F: w1 = 22 20 23 21 37 44 51 58
+ pinsrw xmm4, word [block + 53 * SIZEOF_WORD], 7 ;G: w4 = 59 52 45 38 31 39 46 53
+ ; (Row 6, offset 1)
+ pxor xmm2, xmm2 ;G: w2[i] = 0;
+ pcmpgtw xmm0, xmm4 ;G: w0[i] = (w4[i] < 0 ? -1 : 0);
+ pinsrw xmm1, word [block + 15 * SIZEOF_WORD], 1 ;F: w1 = 22 15 23 21 37 44 51 58
+ paddw xmm4, xmm0 ;G: w4[i] += w0[i];
+ movaps XMMWORD [t + 48 * SIZEOF_WORD], xmm4 ;G: t[48+i] = w4[i];
+ pinsrw xmm1, word [block + 30 * SIZEOF_WORD], 3 ;F: w1 = 22 15 23 30 37 44 51 58
+ ; (Row 5, offset 1)
+ pcmpeqw xmm4, xmm2 ;G: w4[i] = (w4[i] == 0 ? -1 : 0);
+ pinsrw xmm5, word [block + 42 * SIZEOF_WORD], 0 ;E: w5 = 42 49 56 57 50 51 58 59
- mov r13d, INT [r15 + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0];
- movzx r14d, byte [r15 + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0];
- lea rsi, [t1]
-.BLOOP:
- bsf r12, r11 ; r = __builtin_ctzl(index);
- jz .ELOOP
- mov rcx, r12
- lea rsi, [rsi+r12*2] ; k += r;
- shr r11, cl ; index >>= r;
- movzx rdi, word [rsi] ; temp = t1[k];
- lea rbx, [rel EXTN(jpeg_nbits_table)]
- movzx rdi, byte [rbx + rdi] ; nbits = JPEG_NBITS(temp);
-.BRLOOP:
- cmp r12, 16 ; while (r > 15) {
- jl .ERLOOP
- EMIT_BITS r13, r14d ; EMIT_BITS(code_0xf0, size_0xf0)
- sub r12, 16 ; r -= 16;
- jmp .BRLOOP
-.ERLOOP:
- ; Emit Huffman symbol for run length / number of bits
- CHECKBUF31 ; uses rcx, rdx
+ packsswb xmm4, xmm3 ;GH: b4[i] = w4[i], b4[i+8] = w3[i]
+ ; w/ signed saturation
- shl r12, 4 ; temp3 = (r << 4) + nbits;
- add r12, rdi
- mov ebx, INT [r15 + r12 * 4] ; code = actbl->ehufco[temp3];
- movzx ecx, byte [r15 + r12 + 1024] ; size = actbl->ehufsi[temp3];
- PUT_BITS rbx
+ pxor xmm0, xmm0 ;F: w0[i] = 0;
+ pinsrw xmm5, word [block + 43 * SIZEOF_WORD], 5 ;E: w5 = 42 49 56 57 50 43 58 59
+ pcmpgtw xmm2, xmm1 ;F: w2[i] = (w1[i] < 0 ? -1 : 0);
+ pmovmskb tempd, xmm4 ;Z: temp = 0; temp |= ((b4[i] >> 7) << i);
+ pinsrw xmm5, word [block + 36 * SIZEOF_WORD], 6 ;E: w5 = 42 49 56 57 50 43 36 59
+ paddw xmm1, xmm2 ;F: w1[i] += w2[i];
+ movaps XMMWORD [t + 40 * SIZEOF_WORD], xmm1 ;F: t[40+i] = w1[i];
+ pinsrw xmm5, word [block + 29 * SIZEOF_WORD], 7 ;E: w5 = 42 49 56 57 50 43 36 29
+ ; (Row 4, offset 1)
+%undef block
+%define free_bitsq rdx
+%define free_bitsd edx
+%define free_bitsb dl
+ pcmpeqw xmm1, xmm0 ;F: w1[i] = (w1[i] == 0 ? -1 : 0);
+ shl tempq, 48 ;Z: temp <<= 48;
+ pxor xmm2, xmm2 ;E: w2[i] = 0;
+ pcmpgtw xmm0, xmm5 ;E: w0[i] = (w5[i] < 0 ? -1 : 0);
+ paddw xmm5, xmm0 ;E: w5[i] += w0[i];
+ or tempq, put_buffer ;Z: temp |= put_buffer;
+ movaps XMMWORD [t + 32 * SIZEOF_WORD], xmm5 ;E: t[32+i] = w5[i];
+ lea t, [dword t - 2] ;Z: t = &t[-1];
+ pcmpeqw xmm5, xmm2 ;E: w5[i] = (w5[i] == 0 ? -1 : 0);
- ;EMIT_CODE(code, size)
+ packsswb xmm5, xmm1 ;EF: b5[i] = w5[i], b5[i+8] = w1[i]
+ ; w/ signed saturation
- movsx ebx, word [rsi-DCTSIZE2*2] ; temp2 = t2[k];
- ; Mask off any extra bits in code
- mov rcx, rdi
- mov rdx, 1
- shl rdx, cl
- dec rdx
- and rbx, rdx ; temp2 &= (((JLONG)1)<<nbits) - 1;
- PUT_BITS rbx ; PUT_BITS(temp2, nbits)
+ add nbitsb, byte [dctbl + c_derived_tbl.ehufsi + nbitsq]
+ ;Z: nbits += dctbl->ehufsi[nbits];
+%undef dctbl
+%define code_temp r8d
+ pmovmskb indexd, xmm5 ;Z: index = 0; index |= ((b5[i] >> 7) << i);
+ mov free_bitsd, [state+working_state.cur.free_bits]
+ ;Z: free_bits = state->cur.free_bits;
+ pcmpeqw xmm1, xmm1 ;Z: b1[i] = 0xFF;
+ shl index, 32 ;Z: index <<= 32;
+ mov put_buffer, [state+working_state.cur.put_buffer.simd]
+ ;Z: put_buffer = state->cur.put_buffer.simd;
+ or index, tempq ;Z: index |= temp;
+ not index ;Z: index = ~index;
+ sub free_bitsb, nbitsb ;Z: if ((free_bits -= nbits) >= 0)
+ jnl .ENTRY_SKIP_EMIT_CODE ;Z: goto .ENTRY_SKIP_EMIT_CODE;
+ align 16
+.EMIT_CODE: ;Z: .EMIT_CODE:
+ EMIT_QWORD .BLOOP_COND ;Z: insert code, flush buffer, goto .BLOOP_COND
- shr r11, 1 ; index >>= 1;
- add rsi, 2 ; ++k;
- jmp .BLOOP
-.ELOOP:
- ; If the last coef(s) were zero, emit an end-of-block code
- lea rdi, [t1 + (DCTSIZE2-1) * 2] ; r = DCTSIZE2-1-k;
- cmp rdi, rsi ; if (r > 0) {
- je .EFN
- mov ebx, INT [r15] ; code = actbl->ehufco[0];
- movzx r12d, byte [r15 + 1024] ; size = actbl->ehufsi[0];
- EMIT_BITS rbx, r12d
-.EFN:
- pop r10
- ; Save put_buffer & put_bits
- mov MMWORD [r10+16], put_buffer ; state->cur.put_buffer = put_buffer;
- mov dword [r10+24], put_bits ; state->cur.put_bits = put_bits;
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- pop rbx
- uncollect_args 6
- pop_xmm 4
- mov rsp, rbp ; rsp <- aligned rbp
- pop rsp ; rsp <- original rbp
+ align 16
+.BRLOOP: ; do {
+ lea code_temp, [nbitsq - 16] ; code_temp = nbits - 16;
+ movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0]
+ ; nbits = actbl->ehufsi[0xf0];
+ mov code, [actbl + c_derived_tbl.ehufco + 0xf0 * 4]
+ ; code = actbl->ehufco[0xf0];
+ sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_BRLOOP_CODE ; goto .EMIT_BRLOOP_CODE;
+ shl put_buffer, nbitsb ; put_buffer <<= nbits;
+ mov nbits, code_temp ; nbits = code_temp;
+ or put_buffer, codeq ; put_buffer |= code;
+ cmp nbits, 16 ; if (nbits <= 16)
+ jle .ERLOOP ; break;
+ jmp .BRLOOP ; } while (1);
+
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+ times 5 nop
+.ENTRY_SKIP_EMIT_CODE: ; .ENTRY_SKIP_EMIT_CODE:
+ shl put_buffer, nbitsb ; put_buffer <<= nbits;
+ or put_buffer, codeq ; put_buffer |= code;
+.BLOOP_COND: ; .BLOOP_COND:
+ test index, index ; if (index != 0)
+ jz .ELOOP ; {
+.BLOOP: ; do {
+ xor nbits, nbits ; nbits = 0; /* kill tzcnt input dependency */
+ tzcnt nbitsq, index ; nbits = # of trailing 0 bits in index
+ inc nbits ; ++nbits;
+ lea t, [t + nbitsq * 2] ; t = &t[nbits];
+ shr index, nbitsb ; index >>= nbits;
+.EMIT_BRLOOP_CODE_END: ; .EMIT_BRLOOP_CODE_END:
+ cmp nbits, 16 ; if (nbits > 16)
+ jg .BRLOOP ; goto .BRLOOP;
+.ERLOOP: ; .ERLOOP:
+ movsx codeq, word [t] ; code = *t;
+ lea tempd, [nbitsq * 2] ; temp = nbits * 2;
+ movzx nbits, byte [NBITS(codeq)] ; nbits = JPEG_NBITS(code);
+ lea tempd, [nbitsq + tempq * 8] ; temp = temp * 8 + nbits;
+ mov code_temp, [actbl + c_derived_tbl.ehufco + (tempq - 16) * 4]
+ ; code_temp = actbl->ehufco[temp-16];
+ shl code_temp, nbitsb ; code_temp <<= nbits;
+ and code, dword [MASK_BITS(nbitsq)] ; code &= (1 << nbits) - 1;
+ add nbitsb, [actbl + c_derived_tbl.ehufsi + (tempq - 16)]
+ ; free_bits -= actbl->ehufsi[temp-16];
+ or code, code_temp ; code |= code_temp;
+ sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0)
+ jle .EMIT_CODE ; goto .EMIT_CODE;
+ shl put_buffer, nbitsb ; put_buffer <<= nbits;
+ or put_buffer, codeq ; put_buffer |= code;
+ test index, index
+ jnz .BLOOP ; } while (index != 0);
+.ELOOP: ; } /* index != 0 */
+ sub td, esp ; t -= (WIN64: &t_[0], UNIX: &t_[64]);
+%ifdef WIN64
+ cmp td, (DCTSIZE2 - 2) * SIZEOF_WORD ; if (t != 62)
+%else
+ cmp td, -2 * SIZEOF_WORD ; if (t != -2)
+%endif
+ je .EFN ; {
+ movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0]
+ ; nbits = actbl->ehufsi[0];
+ mov code, [actbl + c_derived_tbl.ehufco + 0] ; code = actbl->ehufco[0];
+ sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0)
+ jg .EFN_SKIP_EMIT_CODE ; {
+ EMIT_QWORD .EFN ; insert code, flush buffer
+ align 16
+.EFN_SKIP_EMIT_CODE: ; } else {
+ shl put_buffer, nbitsb ; put_buffer <<= nbits;
+ or put_buffer, codeq ; put_buffer |= code;
+.EFN: ; } }
+ mov [state + working_state.cur.put_buffer.simd], put_buffer
+ ; state->cur.put_buffer.simd = put_buffer;
+ mov byte [state + working_state.cur.free_bits], free_bitsb
+ ; state->cur.free_bits = free_bits;
+%ifdef WIN64
+ sub rsp, -DCTSIZE2 * SIZEOF_WORD
+ pop r12
+ pop rdi
+ pop rsi
pop rbp
+ pop rbx
+%else
+ pop r12
+ pop rbp
+ pop rbx
+%endif
ret
+; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ align 16
+.EMIT_BRLOOP_CODE:
+ EMIT_QWORD .EMIT_BRLOOP_CODE_END, { mov nbits, code_temp }
+ ; insert code, flush buffer,
+ ; nbits = code_temp, goto .EMIT_BRLOOP_CODE_END
+
; For some reason, the OS X linker does not honor the request to align the
; segment unless we do this.
align 32
diff --git a/simd/x86_64/jcsample-avx2.asm b/simd/x86_64/jcsample-avx2.asm
index d9922bb..b32527a 100644
--- a/simd/x86_64/jcsample-avx2.asm
+++ b/simd/x86_64/jcsample-avx2.asm
@@ -4,6 +4,7 @@
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -71,7 +72,7 @@
push rax
push rcx
- mov rdi, JSAMPROW [rsi]
+ mov rdip, JSAMPROW [rsi]
add rdi, rdx
mov al, JSAMPLE [rdi-1]
@@ -107,8 +108,8 @@
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
cmp rcx, byte SIZEOF_YMMWORD
jae short .columnloop
@@ -233,7 +234,7 @@
push rax
push rcx
- mov rdi, JSAMPROW [rsi]
+ mov rdip, JSAMPROW [rsi]
add rdi, rdx
mov al, JSAMPLE [rdi-1]
@@ -269,9 +270,9 @@
push rdi
push rsi
- mov rdx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
- mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rdxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
+ mov rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1
+ mov rdip, JSAMPROW [rdi] ; outptr
cmp rcx, byte SIZEOF_YMMWORD
jae short .columnloop
diff --git a/simd/x86_64/jcsample-sse2.asm b/simd/x86_64/jcsample-sse2.asm
index 0f107e9..2fcfe45 100644
--- a/simd/x86_64/jcsample-sse2.asm
+++ b/simd/x86_64/jcsample-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -70,7 +71,7 @@
push rax
push rcx
- mov rdi, JSAMPROW [rsi]
+ mov rdip, JSAMPROW [rsi]
add rdi, rdx
mov al, JSAMPLE [rdi-1]
@@ -105,8 +106,8 @@
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
cmp rcx, byte SIZEOF_XMMWORD
jae short .columnloop
@@ -215,7 +216,7 @@
push rax
push rcx
- mov rdi, JSAMPROW [rsi]
+ mov rdip, JSAMPROW [rsi]
add rdi, rdx
mov al, JSAMPLE [rdi-1]
@@ -250,9 +251,9 @@
push rdi
push rsi
- mov rdx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
- mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rdxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
+ mov rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1
+ mov rdip, JSAMPROW [rdi] ; outptr
cmp rcx, byte SIZEOF_XMMWORD
jae short .columnloop
diff --git a/simd/x86_64/jdcolext-avx2.asm b/simd/x86_64/jdcolext-avx2.asm
index 677b8ed..2370fda 100644
--- a/simd/x86_64/jdcolext-avx2.asm
+++ b/simd/x86_64/jdcolext-avx2.asm
@@ -4,6 +4,7 @@
; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2012, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -58,9 +59,9 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
@@ -79,10 +80,10 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr0
- mov rbx, JSAMPROW [rbx] ; inptr1
- mov rdx, JSAMPROW [rdx] ; inptr2
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr0
+ mov rbxp, JSAMPROW [rbx] ; inptr1
+ mov rdxp, JSAMPROW [rdx] ; inptr2
+ mov rdip, JSAMPROW [rdi] ; outptr
.columnloop:
vmovdqu ymm5, YMMWORD [rbx] ; ymm5=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
diff --git a/simd/x86_64/jdcolext-sse2.asm b/simd/x86_64/jdcolext-sse2.asm
index 071aa62..e07c8d7 100644
--- a/simd/x86_64/jdcolext-sse2.asm
+++ b/simd/x86_64/jdcolext-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -57,9 +58,9 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
@@ -78,10 +79,10 @@
push rsi
push rcx ; col
- mov rsi, JSAMPROW [rsi] ; inptr0
- mov rbx, JSAMPROW [rbx] ; inptr1
- mov rdx, JSAMPROW [rdx] ; inptr2
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr0
+ mov rbxp, JSAMPROW [rbx] ; inptr1
+ mov rdxp, JSAMPROW [rdx] ; inptr2
+ mov rdip, JSAMPROW [rdi] ; outptr
.columnloop:
movdqa xmm5, XMMWORD [rbx] ; xmm5=Cb(0123456789ABCDEF)
diff --git a/simd/x86_64/jdmrgext-avx2.asm b/simd/x86_64/jdmrgext-avx2.asm
index bb733c5..8b264b4 100644
--- a/simd/x86_64/jdmrgext-avx2.asm
+++ b/simd/x86_64/jdmrgext-avx2.asm
@@ -4,6 +4,7 @@
; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2012, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -58,14 +59,14 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
mov rdi, r13
- mov rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
- mov rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
- mov rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
+ mov rbxp, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
+ mov rdxp, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
+ mov rdip, JSAMPROW [rdi] ; outptr
pop rcx ; col
@@ -514,15 +515,16 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
mov rdi, r13
lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
+ sub rsp, SIZEOF_JSAMPARRAY*4
+ mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
+ mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
+ mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
mov rbx, rsp
push rdi
@@ -546,16 +548,16 @@
pop rax
pop rcx
pop rdi
- pop rsi
- pop rbx
- pop rdx
+ mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
add rdi, byte SIZEOF_JSAMPROW ; outptr1
add rsi, byte SIZEOF_JSAMPROW ; inptr01
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
+ mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
+ mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
+ mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
mov rbx, rsp
push rdi
@@ -579,9 +581,10 @@
pop rax
pop rcx
pop rdi
- pop rsi
- pop rbx
- pop rdx
+ mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+ add rsp, SIZEOF_JSAMPARRAY*4
pop rbx
uncollect_args 4
diff --git a/simd/x86_64/jdmrgext-sse2.asm b/simd/x86_64/jdmrgext-sse2.asm
index b176a4c..eb3ab9d 100644
--- a/simd/x86_64/jdmrgext-sse2.asm
+++ b/simd/x86_64/jdmrgext-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2012, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -57,14 +58,14 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
mov rdi, r13
- mov rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
- mov rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
- mov rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
+ mov rbxp, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
+ mov rdxp, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
+ mov rdip, JSAMPROW [rdi] ; outptr
pop rcx ; col
@@ -456,15 +457,16 @@
mov rdi, r11
mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
+ mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
mov rdi, r13
lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
+ sub rsp, SIZEOF_JSAMPARRAY*4
+ mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
+ mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
+ mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
mov rbx, rsp
push rdi
@@ -488,16 +490,16 @@
pop rax
pop rcx
pop rdi
- pop rsi
- pop rbx
- pop rdx
+ mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
add rdi, byte SIZEOF_JSAMPROW ; outptr1
add rsi, byte SIZEOF_JSAMPROW ; inptr01
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
+ mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
+ mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
+ mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
mov rbx, rsp
push rdi
@@ -521,9 +523,10 @@
pop rax
pop rcx
pop rdi
- pop rsi
- pop rbx
- pop rdx
+ mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
+ mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
+ mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
+ add rsp, SIZEOF_JSAMPARRAY*4
pop rbx
uncollect_args 4
diff --git a/simd/x86_64/jdsample-avx2.asm b/simd/x86_64/jdsample-avx2.asm
index fc274a9..1e4979f 100644
--- a/simd/x86_64/jdsample-avx2.asm
+++ b/simd/x86_64/jdsample-avx2.asm
@@ -4,6 +4,7 @@
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
; Copyright (C) 2015, Intel Corporation.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -76,7 +77,7 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
vpxor ymm0, ymm0, ymm0 ; ymm0=(all 0's)
vpcmpeqb xmm9, xmm9, xmm9
@@ -90,8 +91,8 @@
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
test rax, SIZEOF_YMMWORD-1
jz short .skip
@@ -235,18 +236,18 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rax ; colctr
push rcx
push rdi
push rsi
- mov rcx, JSAMPROW [rsi-1*SIZEOF_JSAMPROW] ; inptr1(above)
- mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
- mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1(below)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
- mov rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+ mov rcxp, JSAMPROW [rsi-1*SIZEOF_JSAMPROW] ; inptr1(above)
+ mov rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
+ mov rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1(below)
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+ mov rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
vpxor ymm8, ymm8, ymm8 ; ymm8=(all 0's)
vpcmpeqb xmm9, xmm9, xmm9
@@ -539,13 +540,13 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
mov rax, rdx ; colctr
.columnloop:
@@ -629,14 +630,14 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rbx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
- mov rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rbxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+ mov rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
mov rax, rdx ; colctr
.columnloop:
diff --git a/simd/x86_64/jdsample-sse2.asm b/simd/x86_64/jdsample-sse2.asm
index 20e0767..38dbcee 100644
--- a/simd/x86_64/jdsample-sse2.asm
+++ b/simd/x86_64/jdsample-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -74,14 +75,14 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rax ; colctr
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
test rax, SIZEOF_XMMWORD-1
jz short .skip
@@ -221,18 +222,18 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rax ; colctr
push rcx
push rdi
push rsi
- mov rcx, JSAMPROW [rsi-1*SIZEOF_JSAMPROW] ; inptr1(above)
- mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
- mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1(below)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
- mov rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+ mov rcxp, JSAMPROW [rsi-1*SIZEOF_JSAMPROW] ; inptr1(above)
+ mov rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0
+ mov rsip, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1(below)
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+ mov rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
test rax, SIZEOF_XMMWORD-1
jz short .skip
@@ -512,13 +513,13 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rdi, JSAMPROW [rdi] ; outptr
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rdip, JSAMPROW [rdi] ; outptr
mov rax, rdx ; colctr
.columnloop:
@@ -600,14 +601,14 @@
mov rsi, r12 ; input_data
mov rdi, r13
- mov rdi, JSAMPARRAY [rdi] ; output_data
+ mov rdip, JSAMPARRAY [rdi] ; output_data
.rowloop:
push rdi
push rsi
- mov rsi, JSAMPROW [rsi] ; inptr
- mov rbx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
- mov rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
+ mov rsip, JSAMPROW [rsi] ; inptr
+ mov rbxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] ; outptr0
+ mov rdip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] ; outptr1
mov rax, rdx ; colctr
.columnloop:
diff --git a/simd/x86_64/jfdctint-avx2.asm b/simd/x86_64/jfdctint-avx2.asm
index 6ad4cf0..e56258b 100644
--- a/simd/x86_64/jfdctint-avx2.asm
+++ b/simd/x86_64/jfdctint-avx2.asm
@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (64-bit AVX2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, 2018, D. R. Commander.
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; forward DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jfdctint.c; see the jfdctint.c for
; more details.
@@ -103,7 +103,7 @@
%endmacro
; --------------------------------------------------------------------------
-; In-place 8x8x16-bit slow integer forward DCT using AVX2 instructions
+; In-place 8x8x16-bit accurate integer forward DCT using AVX2 instructions
; %1-%4: Input/output registers
; %5-%8: Temp registers
; %9: Pass (1 or 2)
diff --git a/simd/x86_64/jfdctint-sse2.asm b/simd/x86_64/jfdctint-sse2.asm
index 5d0de3c..ec1f383 100644
--- a/simd/x86_64/jfdctint-sse2.asm
+++ b/simd/x86_64/jfdctint-sse2.asm
@@ -2,7 +2,7 @@
; jfdctint.asm - accurate integer FDCT (64-bit SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2009, 2016, 2020, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +14,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; forward DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jfdctint.c; see the jfdctint.c for
; more details.
diff --git a/simd/x86_64/jidctflt-sse2.asm b/simd/x86_64/jidctflt-sse2.asm
index ab95e1a..60bf961 100644
--- a/simd/x86_64/jidctflt-sse2.asm
+++ b/simd/x86_64/jidctflt-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -455,12 +456,12 @@
pshufd xmm5, xmm6, 0x4E ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
pshufd xmm3, xmm7, 0x4E ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
- mov rbx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+ mov rbxp, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
movq XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm7
- mov rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
- mov rbx, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+ mov rbxp, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
movq XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm3
diff --git a/simd/x86_64/jidctfst-sse2.asm b/simd/x86_64/jidctfst-sse2.asm
index a66a681..cb97fdf 100644
--- a/simd/x86_64/jidctfst-sse2.asm
+++ b/simd/x86_64/jidctfst-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -460,21 +461,21 @@
pshufd xmm6, xmm4, 0x4E ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
pshufd xmm2, xmm7, 0x4E ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
- mov rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
- mov rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
- mov rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
diff --git a/simd/x86_64/jidctint-avx2.asm b/simd/x86_64/jidctint-avx2.asm
index 50270f4..ca7e317 100644
--- a/simd/x86_64/jidctint-avx2.asm
+++ b/simd/x86_64/jidctint-avx2.asm
@@ -2,7 +2,8 @@
; jidctint.asm - accurate integer IDCT (64-bit AVX2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, 2018, D. R. Commander.
+; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +15,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; inverse DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jidctint.c; see the jidctint.c for
; more details.
@@ -113,7 +114,7 @@
%endmacro
; --------------------------------------------------------------------------
-; In-place 8x8x16-bit slow integer inverse DCT using AVX2 instructions
+; In-place 8x8x16-bit accurate integer inverse DCT using AVX2 instructions
; %1-%4: Input/output registers
; %5-%12: Temp registers
; %9: Pass (1 or 2)
@@ -387,23 +388,23 @@
mov eax, r13d
- mov rdx, JSAMPROW [r12+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rsi, JSAMPROW [r12+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [r12+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r12+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm0
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
- mov rdx, JSAMPROW [r12+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rsi, JSAMPROW [r12+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [r12+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r12+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
- mov rdx, JSAMPROW [r12+4*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rsi, JSAMPROW [r12+5*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [r12+4*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r12+5*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
- mov rdx, JSAMPROW [r12+6*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rsi, JSAMPROW [r12+7*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [r12+6*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r12+7*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
diff --git a/simd/x86_64/jidctint-sse2.asm b/simd/x86_64/jidctint-sse2.asm
index 034530c..7aa869b 100644
--- a/simd/x86_64/jidctint-sse2.asm
+++ b/simd/x86_64/jidctint-sse2.asm
@@ -2,7 +2,8 @@
; jidctint.asm - accurate integer IDCT (64-bit SSE2)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
-; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2009, 2016, 2020, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -14,7 +15,7 @@
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
-; This file contains a slow-but-accurate integer implementation of the
+; This file contains a slower but more accurate integer implementation of the
; inverse DCT (Discrete Cosine Transform). The following code is based
; directly on the IJG's original jidctint.c; see the jidctint.c for
; more details.
@@ -817,21 +818,21 @@
pshufd xmm2, xmm4, 0x4E ; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
pshufd xmm5, xmm3, 0x4E ; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
- mov rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
- mov rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
- mov rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
diff --git a/simd/x86_64/jidctred-sse2.asm b/simd/x86_64/jidctred-sse2.asm
index 7fbfcc5..4ece9d8 100644
--- a/simd/x86_64/jidctred-sse2.asm
+++ b/simd/x86_64/jidctred-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -379,12 +380,12 @@
pshufd xmm1, xmm4, 0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
pshufd xmm3, xmm4, 0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
- mov rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
@@ -558,8 +559,8 @@
pextrw ebx, xmm6, 0x00 ; ebx=(C0 D0 -- --)
pextrw ecx, xmm6, 0x01 ; ecx=(C1 D1 -- --)
- mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
- mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
+ mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
+ mov rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
mov word [rdx+rax*SIZEOF_JSAMPLE], bx
mov word [rsi+rax*SIZEOF_JSAMPLE], cx
diff --git a/simd/x86_64/jquantf-sse2.asm b/simd/x86_64/jquantf-sse2.asm
index 83596a9..ab2e395 100644
--- a/simd/x86_64/jquantf-sse2.asm
+++ b/simd/x86_64/jquantf-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -51,8 +52,8 @@
mov rdi, r12
mov rcx, DCTSIZE/2
.convloop:
- mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]
movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]
diff --git a/simd/x86_64/jquanti-avx2.asm b/simd/x86_64/jquanti-avx2.asm
index 5f04d22..70fe811 100644
--- a/simd/x86_64/jquanti-avx2.asm
+++ b/simd/x86_64/jquanti-avx2.asm
@@ -4,6 +4,7 @@
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, 2018, D. R. Commander.
; Copyright (C) 2016, Matthieu Darbois.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -44,23 +45,23 @@
mov eax, r11d
- mov rsi, JSAMPROW [r10+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdi, JSAMPROW [r10+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r10+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdip, JSAMPROW [r10+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm0, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
pinsrq xmm0, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
- mov rsi, JSAMPROW [r10+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdi, JSAMPROW [r10+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r10+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdip, JSAMPROW [r10+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm1, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
pinsrq xmm1, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
- mov rsi, JSAMPROW [r10+4*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdi, JSAMPROW [r10+5*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r10+4*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdip, JSAMPROW [r10+5*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm2, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
pinsrq xmm2, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
- mov rsi, JSAMPROW [r10+6*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdi, JSAMPROW [r10+7*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rsip, JSAMPROW [r10+6*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdip, JSAMPROW [r10+7*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm3, XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE]
pinsrq xmm3, XMM_MMWORD [rdi+rax*SIZEOF_JSAMPLE], 1
diff --git a/simd/x86_64/jquanti-sse2.asm b/simd/x86_64/jquanti-sse2.asm
index bb6fa69..3ee4420 100644
--- a/simd/x86_64/jquanti-sse2.asm
+++ b/simd/x86_64/jquanti-sse2.asm
@@ -3,6 +3,7 @@
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2009, 2016, D. R. Commander.
+; Copyright (C) 2018, Matthias Räncker.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
@@ -51,14 +52,14 @@
mov rdi, r12
mov rcx, DCTSIZE/4
.convloop:
- mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm0=(01234567)
movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF)
- mov rbx, JSAMPROW [rsi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
- mov rdx, JSAMPROW [rsi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rbxp, JSAMPROW [rsi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
+ mov rdxp, JSAMPROW [rsi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN)
movq xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV)
diff --git a/simd/x86_64/jsimd.c b/simd/x86_64/jsimd.c
index dc639fc..eb76679 100644
--- a/simd/x86_64/jsimd.c
+++ b/simd/x86_64/jsimd.c
@@ -472,12 +472,6 @@
return 0;
}
-GLOBAL(int)
-jsimd_can_h1v2_fancy_upsample(void)
-{
- return 0;
-}
-
GLOBAL(void)
jsimd_h2v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
@@ -506,12 +500,6 @@
output_data_ptr);
}
-GLOBAL(void)
-jsimd_h1v2_fancy_upsample(j_decompress_ptr cinfo, jpeg_component_info *compptr,
- JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr)
-{
-}
-
GLOBAL(int)
jsimd_can_h2v2_merged_upsample(void)
{
@@ -1043,8 +1031,6 @@
return 0;
if (sizeof(JCOEF) != 2)
return 0;
- if (SIZEOF_SIZE_T != 8)
- return 0;
if (simd_support & JSIMD_SSE2)
return 1;
@@ -1069,8 +1055,6 @@
return 0;
if (sizeof(JCOEF) != 2)
return 0;
- if (SIZEOF_SIZE_T != 8)
- return 0;
if (simd_support & JSIMD_SSE2)
return 1;
diff --git a/structure.txt b/structure.txt
index c0792a3..15b8d37 100644
--- a/structure.txt
+++ b/structure.txt
@@ -548,13 +548,9 @@
typedef JSAMPROW *JSAMPARRAY; ptr to a list of rows
typedef JSAMPARRAY *JSAMPIMAGE; ptr to a list of color-component arrays
-The basic element type JSAMPLE will typically be one of unsigned char,
-(signed) char, or short. Short will be used if samples wider than 8 bits are
-to be supported (this is a compile-time option). Otherwise, unsigned char is
-used if possible. If the compiler only supports signed chars, then it is
-necessary to mask off the value when reading. Thus, all reads of JSAMPLE
-values must be coded as "GETJSAMPLE(value)", where the macro will be defined
-as "((value) & 0xFF)" on signed-char machines and "((int) (value))" elsewhere.
+The basic element type JSAMPLE will be one of unsigned char or short. Short
+will be used if samples wider than 8 bits are to be supported (this is a
+compile-time option). Otherwise, unsigned char is used.
With these conventions, JSAMPLE values can be assumed to be >= 0. This helps
simplify correct rounding during downsampling, etc. The JPEG standard's
@@ -587,7 +583,7 @@
is helpful when dealing with noninterleaved JPEG files.
In general, a specific sample value is accessed by code such as
- GETJSAMPLE(image[colorcomponent][row][col])
+ image[colorcomponent][row][col]
where col is measured from the image left edge, but row is measured from the
first sample row currently in memory. Either of the first two indexings can
be precomputed by copying the relevant pointer.
diff --git a/tjexampletest.in b/tjexampletest.in
deleted file mode 100755
index 0d3047e..0000000
--- a/tjexampletest.in
+++ /dev/null
@@ -1,149 +0,0 @@
-#!/bin/bash
-
-set -u
-set -e
-trap onexit INT
-trap onexit TERM
-trap onexit EXIT
-
-onexit()
-{
- if [ -d $OUTDIR ]; then
- rm -rf $OUTDIR
- fi
-}
-
-runme()
-{
- echo \*\*\* $*
- $*
-}
-
-IMAGES="vgl_5674_0098.bmp vgl_6434_0018a.bmp vgl_6548_0026a.bmp nightshot_iso_100.bmp"
-IMGDIR=@CMAKE_CURRENT_SOURCE_DIR@/testimages
-OUTDIR=`mktemp -d /tmp/__tjexampletest_output.XXXXXX`
-EXEDIR=@CMAKE_CURRENT_BINARY_DIR@
-
-if [ -d $OUTDIR ]; then
- rm -rf $OUTDIR
-fi
-mkdir -p $OUTDIR
-
-exec >$EXEDIR/tjexampletest.log
-
-for image in $IMAGES; do
-
- cp $IMGDIR/$image $OUTDIR
- basename=`basename $image .bmp`
- runme $EXEDIR/cjpeg -quality 95 -dct fast -grayscale -outfile $OUTDIR/${basename}_GRAY_fast_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct fast -sample 2x2 -outfile $OUTDIR/${basename}_420_fast_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct fast -sample 2x1 -outfile $OUTDIR/${basename}_422_fast_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct fast -sample 1x1 -outfile $OUTDIR/${basename}_444_fast_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct int -grayscale -outfile $OUTDIR/${basename}_GRAY_accurate_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct int -sample 2x2 -outfile $OUTDIR/${basename}_420_accurate_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct int -sample 2x1 -outfile $OUTDIR/${basename}_422_accurate_cjpeg.jpg $IMGDIR/${basename}.bmp
- runme $EXEDIR/cjpeg -quality 95 -dct int -sample 1x1 -outfile $OUTDIR/${basename}_444_accurate_cjpeg.jpg $IMGDIR/${basename}.bmp
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/djpeg -rgb -bmp -outfile $OUTDIR/${basename}_${samp}_default_djpeg.bmp $OUTDIR/${basename}_${samp}_fast_cjpeg.jpg
- runme $EXEDIR/djpeg -dct fast -rgb -bmp -outfile $OUTDIR/${basename}_${samp}_fast_djpeg.bmp $OUTDIR/${basename}_${samp}_fast_cjpeg.jpg
- runme $EXEDIR/djpeg -dct int -rgb -bmp -outfile $OUTDIR/${basename}_${samp}_accurate_djpeg.bmp $OUTDIR/${basename}_${samp}_accurate_cjpeg.jpg
- done
- for samp in 420 422; do
- runme $EXEDIR/djpeg -nosmooth -bmp -outfile $OUTDIR/${basename}_${samp}_default_nosmooth_djpeg.bmp $OUTDIR/${basename}_${samp}_fast_cjpeg.jpg
- runme $EXEDIR/djpeg -dct fast -nosmooth -bmp -outfile $OUTDIR/${basename}_${samp}_fast_nosmooth_djpeg.bmp $OUTDIR/${basename}_${samp}_fast_cjpeg.jpg
- runme $EXEDIR/djpeg -dct int -nosmooth -bmp -outfile $OUTDIR/${basename}_${samp}_accurate_nosmooth_djpeg.bmp $OUTDIR/${basename}_${samp}_accurate_cjpeg.jpg
- done
-
- # Compression
- for dct in fast accurate; do
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/tjexample $OUTDIR/$image $OUTDIR/${basename}_${samp}_${dct}.jpg -q 95 -subsamp ${samp} -${dct}dct
- runme cmp $OUTDIR/${basename}_${samp}_${dct}.jpg $OUTDIR/${basename}_${samp}_${dct}_cjpeg.jpg
- done
- done
-
- # Decompression
- for dct in fast accurate default; do
- srcdct=${dct}
- dctarg=-${dct}dct
- if [ "${dct}" = "default" ]; then
- srcdct=fast
- dctarg=
- fi
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_${srcdct}.jpg $OUTDIR/${basename}_${samp}_${dct}.bmp ${dctarg}
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${dct}.bmp $OUTDIR/${basename}_${samp}_${dct}_djpeg.bmp
- rm $OUTDIR/${basename}_${samp}_${dct}.bmp
- done
- for samp in 420 422; do
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_${srcdct}.jpg $OUTDIR/${basename}_${samp}_${dct}_nosmooth.bmp -fastupsample ${dctarg}
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${dct}_nosmooth.bmp $OUTDIR/${basename}_${samp}_${dct}_nosmooth_djpeg.bmp
- rm $OUTDIR/${basename}_${samp}_${dct}_nosmooth.bmp
- done
- done
-
- # Scaled decompression
- for scale in 2_1 15_8 7_4 13_8 3_2 11_8 5_4 9_8 7_8 3_4 5_8 1_2 3_8 1_4 1_8; do
- scalearg=`echo $scale | sed 's/\_/\//g'`
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/djpeg -rgb -bmp -scale ${scalearg} -outfile $OUTDIR/${basename}_${samp}_${scale}_djpeg.bmp $OUTDIR/${basename}_${samp}_fast_cjpeg.jpg
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${scale}.bmp -scale ${scalearg}
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${scale}.bmp $OUTDIR/${basename}_${samp}_${scale}_djpeg.bmp
- rm $OUTDIR/${basename}_${samp}_${scale}.bmp
- done
- done
-
- # Transforms
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -flip horizontal -trim -outfile $OUTDIR/${basename}_${samp}_hflip_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -flip vertical -trim -outfile $OUTDIR/${basename}_${samp}_vflip_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -transpose -trim -outfile $OUTDIR/${basename}_${samp}_transpose_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -transverse -trim -outfile $OUTDIR/${basename}_${samp}_transverse_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -rotate 90 -trim -outfile $OUTDIR/${basename}_${samp}_rot90_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -rotate 180 -trim -outfile $OUTDIR/${basename}_${samp}_rot180_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- runme $EXEDIR/jpegtran -crop 70x60+16+16 -rotate 270 -trim -outfile $OUTDIR/${basename}_${samp}_rot270_jpegtran.jpg $OUTDIR/${basename}_${samp}_fast.jpg
- done
- for xform in hflip vflip transpose transverse rot90 rot180 rot270; do
- for samp in GRAY 420 422 444; do
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}.jpg -$xform -crop 70x60+16+16
- runme cmp $OUTDIR/${basename}_${samp}_${xform}.jpg $OUTDIR/${basename}_${samp}_${xform}_jpegtran.jpg
- runme $EXEDIR/djpeg -rgb -bmp -outfile $OUTDIR/${basename}_${samp}_${xform}_jpegtran.bmp $OUTDIR/${basename}_${samp}_${xform}_jpegtran.jpg
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}.bmp -$xform -crop 70x60+16+16
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${xform}.bmp $OUTDIR/${basename}_${samp}_${xform}_jpegtran.bmp
- rm $OUTDIR/${basename}_${samp}_${xform}.bmp
- done
- for samp in 420 422; do
- runme $EXEDIR/djpeg -nosmooth -rgb -bmp -outfile $OUTDIR/${basename}_${samp}_${xform}_jpegtran.bmp $OUTDIR/${basename}_${samp}_${xform}_jpegtran.jpg
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}.bmp -$xform -crop 70x60+16+16 -fastupsample
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${xform}.bmp $OUTDIR/${basename}_${samp}_${xform}_jpegtran.bmp
- rm $OUTDIR/${basename}_${samp}_${xform}.bmp
- done
- done
-
- # Grayscale transform
- for xform in hflip vflip transpose transverse rot90 rot180 rot270; do
- for samp in GRAY 444 422 420; do
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}.jpg -$xform -grayscale -crop 70x60+16+16
- runme cmp $OUTDIR/${basename}_${samp}_${xform}.jpg $OUTDIR/${basename}_GRAY_${xform}_jpegtran.jpg
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}.bmp -$xform -grayscale -crop 70x60+16+16
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${xform}.bmp $OUTDIR/${basename}_GRAY_${xform}_jpegtran.bmp
- rm $OUTDIR/${basename}_${samp}_${xform}.bmp
- done
- done
-
- # Transforms with scaling
- for xform in hflip vflip transpose transverse rot90 rot180 rot270; do
- for samp in GRAY 444 422 420; do
- for scale in 2_1 15_8 7_4 13_8 3_2 11_8 5_4 9_8 7_8 3_4 5_8 1_2 3_8 1_4 1_8; do
- scalearg=`echo $scale | sed 's/\_/\//g'`
- runme $EXEDIR/djpeg -rgb -bmp -scale ${scalearg} -outfile $OUTDIR/${basename}_${samp}_${xform}_${scale}_jpegtran.bmp $OUTDIR/${basename}_${samp}_${xform}_jpegtran.jpg
- runme $EXEDIR/tjexample $OUTDIR/${basename}_${samp}_fast.jpg $OUTDIR/${basename}_${samp}_${xform}_${scale}.bmp -$xform -scale ${scalearg} -crop 70x60+16+16
- runme cmp -i 54:54 $OUTDIR/${basename}_${samp}_${xform}_${scale}.bmp $OUTDIR/${basename}_${samp}_${xform}_${scale}_jpegtran.bmp
- rm $OUTDIR/${basename}_${samp}_${xform}_${scale}.bmp
- done
- done
- done
-
-done
-
-echo SUCCESS!
diff --git a/transupp.c b/transupp.c
index f3370ac..6e86077 100644
--- a/transupp.c
+++ b/transupp.c
@@ -2,7 +2,7 @@
* transupp.c
*
* This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1997-2011, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 1997-2019, Thomas G. Lane, Guido Vollbeding.
* libjpeg-turbo Modifications:
* Copyright (C) 2010, 2017, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
@@ -89,6 +89,189 @@
LOCAL(void)
+dequant_comp(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+ jvirt_barray_ptr coef_array, JQUANT_TBL *qtblptr1)
+{
+ JDIMENSION blk_x, blk_y;
+ int offset_y, k;
+ JQUANT_TBL *qtblptr;
+ JBLOCKARRAY buffer;
+ JBLOCKROW block;
+ JCOEFPTR ptr;
+
+ qtblptr = compptr->quant_table;
+ for (blk_y = 0; blk_y < compptr->height_in_blocks;
+ blk_y += compptr->v_samp_factor) {
+ buffer = (*cinfo->mem->access_virt_barray)
+ ((j_common_ptr)cinfo, coef_array, blk_y,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ block = buffer[offset_y];
+ for (blk_x = 0; blk_x < compptr->width_in_blocks; blk_x++) {
+ ptr = block[blk_x];
+ for (k = 0; k < DCTSIZE2; k++)
+ if (qtblptr->quantval[k] != qtblptr1->quantval[k])
+ ptr[k] *= qtblptr->quantval[k] / qtblptr1->quantval[k];
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+requant_comp(j_decompress_ptr cinfo, jpeg_component_info *compptr,
+ jvirt_barray_ptr coef_array, JQUANT_TBL *qtblptr1)
+{
+ JDIMENSION blk_x, blk_y;
+ int offset_y, k;
+ JQUANT_TBL *qtblptr;
+ JBLOCKARRAY buffer;
+ JBLOCKROW block;
+ JCOEFPTR ptr;
+ JCOEF temp, qval;
+
+ qtblptr = compptr->quant_table;
+ for (blk_y = 0; blk_y < compptr->height_in_blocks;
+ blk_y += compptr->v_samp_factor) {
+ buffer = (*cinfo->mem->access_virt_barray)
+ ((j_common_ptr)cinfo, coef_array, blk_y,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ block = buffer[offset_y];
+ for (blk_x = 0; blk_x < compptr->width_in_blocks; blk_x++) {
+ ptr = block[blk_x];
+ for (k = 0; k < DCTSIZE2; k++) {
+ temp = qtblptr->quantval[k];
+ qval = qtblptr1->quantval[k];
+ if (temp != qval) {
+ temp *= ptr[k];
+ /* The following quantization code is copied from jcdctmgr.c */
+#ifdef FAST_DIVIDE
+#define DIVIDE_BY(a, b) a /= b
+#else
+#define DIVIDE_BY(a, b) if (a >= b) a /= b; else a = 0
+#endif
+ if (temp < 0) {
+ temp = -temp;
+ temp += qval >> 1; /* for rounding */
+ DIVIDE_BY(temp, qval);
+ temp = -temp;
+ } else {
+ temp += qval >> 1; /* for rounding */
+ DIVIDE_BY(temp, qval);
+ }
+ ptr[k] = temp;
+ }
+ }
+ }
+ }
+ }
+}
+
+
+/*
+ * Calculate largest common denominator using Euclid's algorithm.
+ */
+LOCAL(JCOEF)
+largest_common_denominator(JCOEF a, JCOEF b)
+{
+ JCOEF c;
+
+ do {
+ c = a % b;
+ a = b;
+ b = c;
+ } while (c);
+
+ return a;
+}
+
+
+LOCAL(void)
+adjust_quant(j_decompress_ptr srcinfo, jvirt_barray_ptr *src_coef_arrays,
+ j_decompress_ptr dropinfo, jvirt_barray_ptr *drop_coef_arrays,
+ boolean trim, j_compress_ptr dstinfo)
+{
+ jpeg_component_info *compptr1, *compptr2;
+ JQUANT_TBL *qtblptr1, *qtblptr2, *qtblptr3;
+ int ci, k;
+
+ for (ci = 0; ci < dstinfo->num_components && ci < dropinfo->num_components;
+ ci++) {
+ compptr1 = srcinfo->comp_info + ci;
+ compptr2 = dropinfo->comp_info + ci;
+ qtblptr1 = compptr1->quant_table;
+ qtblptr2 = compptr2->quant_table;
+ for (k = 0; k < DCTSIZE2; k++) {
+ if (qtblptr1->quantval[k] != qtblptr2->quantval[k]) {
+ if (trim)
+ requant_comp(dropinfo, compptr2, drop_coef_arrays[ci], qtblptr1);
+ else {
+ qtblptr3 = dstinfo->quant_tbl_ptrs[compptr1->quant_tbl_no];
+ for (k = 0; k < DCTSIZE2; k++)
+ if (qtblptr1->quantval[k] != qtblptr2->quantval[k])
+ qtblptr3->quantval[k] =
+ largest_common_denominator(qtblptr1->quantval[k],
+ qtblptr2->quantval[k]);
+ dequant_comp(srcinfo, compptr1, src_coef_arrays[ci], qtblptr3);
+ dequant_comp(dropinfo, compptr2, drop_coef_arrays[ci], qtblptr3);
+ }
+ break;
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_drop(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ j_decompress_ptr dropinfo, jvirt_barray_ptr *drop_coef_arrays,
+ JDIMENSION drop_width, JDIMENSION drop_height)
+/* Drop (insert) the contents of another image into the source image. If the
+ * number of components in the drop image is smaller than the number of
+ * components in the destination image, then we fill in the remaining
+ * components with zero. This allows for dropping the contents of grayscale
+ * images into (arbitrarily sampled) color images.
+ */
+{
+ JDIMENSION comp_width, comp_height;
+ JDIMENSION blk_y, x_drop_blocks, y_drop_blocks;
+ int ci, offset_y;
+ JBLOCKARRAY src_buffer, dst_buffer;
+ jpeg_component_info *compptr;
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ comp_width = drop_width * compptr->h_samp_factor;
+ comp_height = drop_height * compptr->v_samp_factor;
+ x_drop_blocks = x_crop_offset * compptr->h_samp_factor;
+ y_drop_blocks = y_crop_offset * compptr->v_samp_factor;
+ for (blk_y = 0; blk_y < comp_height; blk_y += compptr->v_samp_factor) {
+ dst_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], blk_y + y_drop_blocks,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ if (ci < dropinfo->num_components) {
+ src_buffer = (*dropinfo->mem->access_virt_barray)
+ ((j_common_ptr)dropinfo, drop_coef_arrays[ci], blk_y,
+ (JDIMENSION)compptr->v_samp_factor, FALSE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ jcopy_block_row(src_buffer[offset_y],
+ dst_buffer[offset_y] + x_drop_blocks, comp_width);
+ }
+ } else {
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(dst_buffer[offset_y] + x_drop_blocks,
+ comp_width * sizeof(JBLOCK));
+ }
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
do_crop(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
jvirt_barray_ptr *src_coef_arrays,
@@ -113,13 +296,422 @@
((j_common_ptr)srcinfo, dst_coef_arrays[ci], dst_blk_y,
(JDIMENSION)compptr->v_samp_factor, TRUE);
src_buffer = (*srcinfo->mem->access_virt_barray)
- ((j_common_ptr)srcinfo, src_coef_arrays[ci],
- dst_blk_y + y_crop_blocks,
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], dst_blk_y + y_crop_blocks,
(JDIMENSION)compptr->v_samp_factor, FALSE);
for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
jcopy_block_row(src_buffer[offset_y] + x_crop_blocks,
- dst_buffer[offset_y],
- compptr->width_in_blocks);
+ dst_buffer[offset_y], compptr->width_in_blocks);
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_crop_ext_zero(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ jvirt_barray_ptr *dst_coef_arrays)
+/* Crop. This is only used when no rotate/flip is requested with the crop.
+ * Extension: If the destination size is larger than the source, we fill in the
+ * expanded region with zero (neutral gray). Note that we also have to zero
+ * partial iMCUs at the right and bottom edge of the source image area in this
+ * case.
+ */
+{
+ JDIMENSION MCU_cols, MCU_rows, comp_width, comp_height;
+ JDIMENSION dst_blk_y, x_crop_blocks, y_crop_blocks;
+ int ci, offset_y;
+ JBLOCKARRAY src_buffer, dst_buffer;
+ jpeg_component_info *compptr;
+
+ MCU_cols = srcinfo->output_width /
+ (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+ MCU_rows = srcinfo->output_height /
+ (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ comp_width = MCU_cols * compptr->h_samp_factor;
+ comp_height = MCU_rows * compptr->v_samp_factor;
+ x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+ y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+ for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+ dst_blk_y += compptr->v_samp_factor) {
+ dst_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, dst_coef_arrays[ci], dst_blk_y,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ if (dstinfo->_jpeg_height > srcinfo->output_height) {
+ if (dst_blk_y < y_crop_blocks ||
+ dst_blk_y >= y_crop_blocks + comp_height) {
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(dst_buffer[offset_y],
+ compptr->width_in_blocks * sizeof(JBLOCK));
+ }
+ continue;
+ }
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y - y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ } else {
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y + y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ }
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ if (dstinfo->_jpeg_width > srcinfo->output_width) {
+ if (x_crop_blocks > 0) {
+ MEMZERO(dst_buffer[offset_y], x_crop_blocks * sizeof(JBLOCK));
+ }
+ jcopy_block_row(src_buffer[offset_y],
+ dst_buffer[offset_y] + x_crop_blocks, comp_width);
+ if (compptr->width_in_blocks > x_crop_blocks + comp_width) {
+ MEMZERO(dst_buffer[offset_y] + x_crop_blocks + comp_width,
+ (compptr->width_in_blocks - x_crop_blocks - comp_width) *
+ sizeof(JBLOCK));
+ }
+ } else {
+ jcopy_block_row(src_buffer[offset_y] + x_crop_blocks,
+ dst_buffer[offset_y], compptr->width_in_blocks);
+ }
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_crop_ext_flat(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ jvirt_barray_ptr *dst_coef_arrays)
+/* Crop. This is only used when no rotate/flip is requested with the crop.
+ * Extension: The destination width is larger than the source, and we fill in
+ * the expanded region with the DC coefficient of the adjacent block. Note
+ * that we also have to fill partial iMCUs at the right and bottom edge of the
+ * source image area in this case.
+ */
+{
+ JDIMENSION MCU_cols, MCU_rows, comp_width, comp_height;
+ JDIMENSION dst_blk_x, dst_blk_y, x_crop_blocks, y_crop_blocks;
+ int ci, offset_y;
+ JCOEF dc;
+ JBLOCKARRAY src_buffer, dst_buffer;
+ jpeg_component_info *compptr;
+
+ MCU_cols = srcinfo->output_width /
+ (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+ MCU_rows = srcinfo->output_height /
+ (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ comp_width = MCU_cols * compptr->h_samp_factor;
+ comp_height = MCU_rows * compptr->v_samp_factor;
+ x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+ y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+ for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+ dst_blk_y += compptr->v_samp_factor) {
+ dst_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, dst_coef_arrays[ci], dst_blk_y,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ if (dstinfo->_jpeg_height > srcinfo->output_height) {
+ if (dst_blk_y < y_crop_blocks ||
+ dst_blk_y >= y_crop_blocks + comp_height) {
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(dst_buffer[offset_y],
+ compptr->width_in_blocks * sizeof(JBLOCK));
+ }
+ continue;
+ }
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y - y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ } else {
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y + y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ }
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ if (x_crop_blocks > 0) {
+ MEMZERO(dst_buffer[offset_y], x_crop_blocks * sizeof(JBLOCK));
+ dc = src_buffer[offset_y][0][0];
+ for (dst_blk_x = 0; dst_blk_x < x_crop_blocks; dst_blk_x++) {
+ dst_buffer[offset_y][dst_blk_x][0] = dc;
+ }
+ }
+ jcopy_block_row(src_buffer[offset_y],
+ dst_buffer[offset_y] + x_crop_blocks, comp_width);
+ if (compptr->width_in_blocks > x_crop_blocks + comp_width) {
+ MEMZERO(dst_buffer[offset_y] + x_crop_blocks + comp_width,
+ (compptr->width_in_blocks - x_crop_blocks - comp_width) *
+ sizeof(JBLOCK));
+ dc = src_buffer[offset_y][comp_width - 1][0];
+ for (dst_blk_x = x_crop_blocks + comp_width;
+ dst_blk_x < compptr->width_in_blocks; dst_blk_x++) {
+ dst_buffer[offset_y][dst_blk_x][0] = dc;
+ }
+ }
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_crop_ext_reflect(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ jvirt_barray_ptr *dst_coef_arrays)
+/* Crop. This is only used when no rotate/flip is requested with the crop.
+ * Extension: The destination width is larger than the source, and we fill in
+ * the expanded region with repeated reflections of the source image. Note
+ * that we also have to fill partial iMCUs at the right and bottom edge of the
+ * source image area in this case.
+ */
+{
+ JDIMENSION MCU_cols, MCU_rows, comp_width, comp_height, src_blk_x;
+ JDIMENSION dst_blk_x, dst_blk_y, x_crop_blocks, y_crop_blocks;
+ int ci, k, offset_y;
+ JBLOCKARRAY src_buffer, dst_buffer;
+ JBLOCKROW src_row_ptr, dst_row_ptr;
+ JCOEFPTR src_ptr, dst_ptr;
+ jpeg_component_info *compptr;
+
+ MCU_cols = srcinfo->output_width /
+ (dstinfo->max_h_samp_factor * dstinfo_min_DCT_h_scaled_size);
+ MCU_rows = srcinfo->output_height /
+ (dstinfo->max_v_samp_factor * dstinfo_min_DCT_v_scaled_size);
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ comp_width = MCU_cols * compptr->h_samp_factor;
+ comp_height = MCU_rows * compptr->v_samp_factor;
+ x_crop_blocks = x_crop_offset * compptr->h_samp_factor;
+ y_crop_blocks = y_crop_offset * compptr->v_samp_factor;
+ for (dst_blk_y = 0; dst_blk_y < compptr->height_in_blocks;
+ dst_blk_y += compptr->v_samp_factor) {
+ dst_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, dst_coef_arrays[ci], dst_blk_y,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ if (dstinfo->_jpeg_height > srcinfo->output_height) {
+ if (dst_blk_y < y_crop_blocks ||
+ dst_blk_y >= y_crop_blocks + comp_height) {
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(dst_buffer[offset_y],
+ compptr->width_in_blocks * sizeof(JBLOCK));
+ }
+ continue;
+ }
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y - y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ } else {
+ src_buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci],
+ dst_blk_y + y_crop_blocks, (JDIMENSION)compptr->v_samp_factor,
+ FALSE);
+ }
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ /* Copy source region */
+ jcopy_block_row(src_buffer[offset_y],
+ dst_buffer[offset_y] + x_crop_blocks, comp_width);
+ if (x_crop_blocks > 0) {
+ /* Reflect to left */
+ dst_row_ptr = dst_buffer[offset_y] + x_crop_blocks;
+ for (dst_blk_x = x_crop_blocks; dst_blk_x > 0;) {
+ src_row_ptr = dst_row_ptr; /* (re)set axis of reflection */
+ for (src_blk_x = comp_width; src_blk_x > 0 && dst_blk_x > 0;
+ src_blk_x--, dst_blk_x--) {
+ dst_ptr = *(--dst_row_ptr); /* destination goes left */
+ src_ptr = *src_row_ptr++; /* source goes right */
+ /* This unrolled loop doesn't need to know which row it's on. */
+ for (k = 0; k < DCTSIZE2; k += 2) {
+ *dst_ptr++ = *src_ptr++; /* copy even column */
+ *dst_ptr++ = -(*src_ptr++); /* copy odd column with sign
+ change */
+ }
+ }
+ }
+ }
+ if (compptr->width_in_blocks > x_crop_blocks + comp_width) {
+ /* Reflect to right */
+ dst_row_ptr = dst_buffer[offset_y] + x_crop_blocks + comp_width;
+ for (dst_blk_x = compptr->width_in_blocks - x_crop_blocks - comp_width;
+ dst_blk_x > 0;) {
+ src_row_ptr = dst_row_ptr; /* (re)set axis of reflection */
+ for (src_blk_x = comp_width; src_blk_x > 0 && dst_blk_x > 0;
+ src_blk_x--, dst_blk_x--) {
+ dst_ptr = *dst_row_ptr++; /* destination goes right */
+ src_ptr = *(--src_row_ptr); /* source goes left */
+ /* This unrolled loop doesn't need to know which row it's on. */
+ for (k = 0; k < DCTSIZE2; k += 2) {
+ *dst_ptr++ = *src_ptr++; /* copy even column */
+ *dst_ptr++ = -(*src_ptr++); /* copy odd column with sign
+ change */
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_wipe(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ JDIMENSION drop_width, JDIMENSION drop_height)
+/* Wipe - discard image contents of specified region and fill with zero
+ * (neutral gray)
+ */
+{
+ JDIMENSION x_wipe_blocks, wipe_width;
+ JDIMENSION y_wipe_blocks, wipe_bottom;
+ int ci, offset_y;
+ JBLOCKARRAY buffer;
+ jpeg_component_info *compptr;
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ x_wipe_blocks = x_crop_offset * compptr->h_samp_factor;
+ wipe_width = drop_width * compptr->h_samp_factor;
+ y_wipe_blocks = y_crop_offset * compptr->v_samp_factor;
+ wipe_bottom = drop_height * compptr->v_samp_factor + y_wipe_blocks;
+ for (; y_wipe_blocks < wipe_bottom;
+ y_wipe_blocks += compptr->v_samp_factor) {
+ buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], y_wipe_blocks,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(buffer[offset_y] + x_wipe_blocks, wipe_width * sizeof(JBLOCK));
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_flatten(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, JDIMENSION y_crop_offset,
+ jvirt_barray_ptr *src_coef_arrays,
+ JDIMENSION drop_width, JDIMENSION drop_height)
+/* Flatten - discard image contents of specified region, similarly to wipe,
+ * but fill with the average of adjacent blocks instead of zero.
+ */
+{
+ JDIMENSION x_wipe_blocks, wipe_width, wipe_right;
+ JDIMENSION y_wipe_blocks, wipe_bottom, blk_x;
+ int ci, offset_y, dc_left_value, dc_right_value, average;
+ JBLOCKARRAY buffer;
+ jpeg_component_info *compptr;
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ x_wipe_blocks = x_crop_offset * compptr->h_samp_factor;
+ wipe_width = drop_width * compptr->h_samp_factor;
+ wipe_right = wipe_width + x_wipe_blocks;
+ y_wipe_blocks = y_crop_offset * compptr->v_samp_factor;
+ wipe_bottom = drop_height * compptr->v_samp_factor + y_wipe_blocks;
+ for (; y_wipe_blocks < wipe_bottom;
+ y_wipe_blocks += compptr->v_samp_factor) {
+ buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], y_wipe_blocks,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ MEMZERO(buffer[offset_y] + x_wipe_blocks, wipe_width * sizeof(JBLOCK));
+ if (x_wipe_blocks > 0) {
+ dc_left_value = buffer[offset_y][x_wipe_blocks - 1][0];
+ if (wipe_right < compptr->width_in_blocks) {
+ dc_right_value = buffer[offset_y][wipe_right][0];
+ average = (dc_left_value + dc_right_value) >> 1;
+ } else {
+ average = dc_left_value;
+ }
+ } else if (wipe_right < compptr->width_in_blocks) {
+ average = buffer[offset_y][wipe_right][0];
+ } else continue;
+ for (blk_x = x_wipe_blocks; blk_x < wipe_right; blk_x++) {
+ buffer[offset_y][blk_x][0] = (JCOEF)average;
+ }
+ }
+ }
+ }
+}
+
+
+LOCAL(void)
+do_reflect(j_decompress_ptr srcinfo, j_compress_ptr dstinfo,
+ JDIMENSION x_crop_offset, jvirt_barray_ptr *src_coef_arrays,
+ JDIMENSION drop_width, JDIMENSION drop_height)
+/* Reflect - discard image contents of specified region, similarly to wipe,
+ * but fill with repeated reflections of the outside region instead of zero.
+ * NB: y_crop_offset is assumed to be zero.
+ */
+{
+ JDIMENSION x_wipe_blocks, wipe_width;
+ JDIMENSION y_wipe_blocks, wipe_bottom;
+ JDIMENSION src_blk_x, dst_blk_x;
+ int ci, k, offset_y;
+ JBLOCKARRAY buffer;
+ JBLOCKROW src_row_ptr, dst_row_ptr;
+ JCOEFPTR src_ptr, dst_ptr;
+ jpeg_component_info *compptr;
+
+ for (ci = 0; ci < dstinfo->num_components; ci++) {
+ compptr = dstinfo->comp_info + ci;
+ x_wipe_blocks = x_crop_offset * compptr->h_samp_factor;
+ wipe_width = drop_width * compptr->h_samp_factor;
+ wipe_bottom = drop_height * compptr->v_samp_factor;
+ for (y_wipe_blocks = 0; y_wipe_blocks < wipe_bottom;
+ y_wipe_blocks += compptr->v_samp_factor) {
+ buffer = (*srcinfo->mem->access_virt_barray)
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], y_wipe_blocks,
+ (JDIMENSION)compptr->v_samp_factor, TRUE);
+ for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
+ if (x_wipe_blocks > 0) {
+ /* Reflect from left */
+ dst_row_ptr = buffer[offset_y] + x_wipe_blocks;
+ for (dst_blk_x = wipe_width; dst_blk_x > 0;) {
+ src_row_ptr = dst_row_ptr; /* (re)set axis of reflection */
+ for (src_blk_x = x_wipe_blocks;
+ src_blk_x > 0 && dst_blk_x > 0; src_blk_x--, dst_blk_x--) {
+ dst_ptr = *dst_row_ptr++; /* destination goes right */
+ src_ptr = *(--src_row_ptr); /* source goes left */
+ /* this unrolled loop doesn't need to know which row it's on... */
+ for (k = 0; k < DCTSIZE2; k += 2) {
+ *dst_ptr++ = *src_ptr++; /* copy even column */
+ *dst_ptr++ = -(*src_ptr++); /* copy odd column with sign change */
+ }
+ }
+ }
+ } else if (compptr->width_in_blocks > x_wipe_blocks + wipe_width) {
+ /* Reflect from right */
+ dst_row_ptr = buffer[offset_y] + x_wipe_blocks + wipe_width;
+ for (dst_blk_x = wipe_width; dst_blk_x > 0;) {
+ src_row_ptr = dst_row_ptr; /* (re)set axis of reflection */
+ src_blk_x = compptr->width_in_blocks - x_wipe_blocks - wipe_width;
+ for (; src_blk_x > 0 && dst_blk_x > 0; src_blk_x--, dst_blk_x--) {
+ dst_ptr = *(--dst_row_ptr); /* destination goes left */
+ src_ptr = *src_row_ptr++; /* source goes right */
+ /* this unrolled loop doesn't need to know which row it's on... */
+ for (k = 0; k < DCTSIZE2; k += 2) {
+ *dst_ptr++ = *src_ptr++; /* copy even column */
+ *dst_ptr++ = -(*src_ptr++); /* copy odd column with sign change */
+ }
+ }
+ }
+ } else {
+ MEMZERO(buffer[offset_y] + x_wipe_blocks,
+ wipe_width * sizeof(JBLOCK));
+ }
}
}
}
@@ -224,8 +816,7 @@
((j_common_ptr)srcinfo, dst_coef_arrays[ci], dst_blk_y,
(JDIMENSION)compptr->v_samp_factor, TRUE);
src_buffer = (*srcinfo->mem->access_virt_barray)
- ((j_common_ptr)srcinfo, src_coef_arrays[ci],
- dst_blk_y + y_crop_blocks,
+ ((j_common_ptr)srcinfo, src_coef_arrays[ci], dst_blk_y + y_crop_blocks,
(JDIMENSION)compptr->v_samp_factor, FALSE);
for (offset_y = 0; offset_y < compptr->v_samp_factor; offset_y++) {
dst_row_ptr = dst_buffer[offset_y];
@@ -238,8 +829,9 @@
src_ptr = src_row_ptr[comp_width - x_crop_blocks - dst_blk_x - 1];
/* this unrolled loop doesn't need to know which row it's on... */
for (k = 0; k < DCTSIZE2; k += 2) {
- *dst_ptr++ = *src_ptr++; /* copy even column */
- *dst_ptr++ = - *src_ptr++; /* copy odd column with sign change */
+ *dst_ptr++ = *src_ptr++; /* copy even column */
+ *dst_ptr++ = -(*src_ptr++); /* copy odd column with sign
+ change */
}
} else {
/* Copy last partial block(s) verbatim */
@@ -318,14 +910,13 @@
*dst_ptr++ = *src_ptr++;
/* copy odd row with sign change */
for (j = 0; j < DCTSIZE; j++)
- *dst_ptr++ = - *src_ptr++;
+ *dst_ptr++ = -(*src_ptr++);
}
}
} else {
/* Just copy row verbatim. */
jcopy_block_row(src_buffer[offset_y] + x_crop_blocks,
- dst_buffer[offset_y],
- compptr->width_in_blocks);
+ dst_buffer[offset_y], compptr->width_in_blocks);
}
}
}
@@ -599,11 +1190,11 @@
/* For even row, negate every odd column. */
for (j = 0; j < DCTSIZE; j += 2) {
*dst_ptr++ = *src_ptr++;
- *dst_ptr++ = - *src_ptr++;
+ *dst_ptr++ = -(*src_ptr++);
}
/* For odd row, negate every even column. */
for (j = 0; j < DCTSIZE; j += 2) {
- *dst_ptr++ = - *src_ptr++;
+ *dst_ptr++ = -(*src_ptr++);
*dst_ptr++ = *src_ptr++;
}
}
@@ -614,7 +1205,7 @@
for (j = 0; j < DCTSIZE; j++)
*dst_ptr++ = *src_ptr++;
for (j = 0; j < DCTSIZE; j++)
- *dst_ptr++ = - *src_ptr++;
+ *dst_ptr++ = -(*src_ptr++);
}
}
}
@@ -630,7 +1221,7 @@
src_row_ptr[comp_width - x_crop_blocks - dst_blk_x - 1];
for (i = 0; i < DCTSIZE2; i += 2) {
*dst_ptr++ = *src_ptr++;
- *dst_ptr++ = - *src_ptr++;
+ *dst_ptr++ = -(*src_ptr++);
}
} else {
/* Any remaining right-edge blocks are only copied. */
@@ -786,7 +1377,7 @@
* The routine returns TRUE if the spec string is valid, FALSE if not.
*
* The crop spec string should have the format
- * <width>[f]x<height>[f]{+-}<xoffset>{+-}<yoffset>
+ * <width>[{fr}]x<height>[{fr}]{+-}<xoffset>{+-}<yoffset>
* where width, height, xoffset, and yoffset are unsigned integers.
* Each of the elements can be omitted to indicate a default value.
* (A weakness of this style is that it is not possible to omit xoffset
@@ -811,6 +1402,9 @@
if (*spec == 'f' || *spec == 'F') {
spec++;
info->crop_width_set = JCROP_FORCE;
+ } else if (*spec == 'r' || *spec == 'R') {
+ spec++;
+ info->crop_width_set = JCROP_REFLECT;
} else
info->crop_width_set = JCROP_POS;
}
@@ -822,6 +1416,9 @@
if (*spec == 'f' || *spec == 'F') {
spec++;
info->crop_height_set = JCROP_FORCE;
+ } else if (*spec == 'r' || *spec == 'R') {
+ spec++;
+ info->crop_height_set = JCROP_REFLECT;
} else
info->crop_height_set = JCROP_POS;
}
@@ -896,10 +1493,10 @@
jvirt_barray_ptr *coef_arrays;
boolean need_workspace, transpose_it;
jpeg_component_info *compptr;
- JDIMENSION xoffset, yoffset;
+ JDIMENSION xoffset, yoffset, dtemp;
JDIMENSION width_in_iMCUs, height_in_iMCUs;
JDIMENSION width_in_blocks, height_in_blocks;
- int ci, h_samp_factor, v_samp_factor;
+ int itemp, ci, h_samp_factor, v_samp_factor;
/* Determine number of components in output image */
if (info->force_grayscale &&
@@ -985,39 +1582,129 @@
info->crop_xoffset = 0; /* default to +0 */
if (info->crop_yoffset_set == JCROP_UNSET)
info->crop_yoffset = 0; /* default to +0 */
- if (info->crop_xoffset >= info->output_width ||
- info->crop_yoffset >= info->output_height)
- ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
- if (info->crop_width_set == JCROP_UNSET)
+ if (info->crop_width_set == JCROP_UNSET) {
+ if (info->crop_xoffset >= info->output_width)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
info->crop_width = info->output_width - info->crop_xoffset;
- if (info->crop_height_set == JCROP_UNSET)
+ } else {
+ /* Check for crop extension */
+ if (info->crop_width > info->output_width) {
+ /* Crop extension does not work when transforming! */
+ if (info->transform != JXFORM_NONE ||
+ info->crop_xoffset >= info->crop_width ||
+ info->crop_xoffset > info->crop_width - info->output_width)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+ } else {
+ if (info->crop_xoffset >= info->output_width ||
+ info->crop_width <= 0 ||
+ info->crop_xoffset > info->output_width - info->crop_width)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+ }
+ }
+ if (info->crop_height_set == JCROP_UNSET) {
+ if (info->crop_yoffset >= info->output_height)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
info->crop_height = info->output_height - info->crop_yoffset;
- /* Ensure parameters are valid */
- if (info->crop_width <= 0 || info->crop_width > info->output_width ||
- info->crop_height <= 0 || info->crop_height > info->output_height ||
- info->crop_xoffset > info->output_width - info->crop_width ||
- info->crop_yoffset > info->output_height - info->crop_height)
- ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+ } else {
+ /* Check for crop extension */
+ if (info->crop_height > info->output_height) {
+ /* Crop extension does not work when transforming! */
+ if (info->transform != JXFORM_NONE ||
+ info->crop_yoffset >= info->crop_height ||
+ info->crop_yoffset > info->crop_height - info->output_height)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+ } else {
+ if (info->crop_yoffset >= info->output_height ||
+ info->crop_height <= 0 ||
+ info->crop_yoffset > info->output_height - info->crop_height)
+ ERREXIT(srcinfo, JERR_BAD_CROP_SPEC);
+ }
+ }
/* Convert negative crop offsets into regular offsets */
- if (info->crop_xoffset_set == JCROP_NEG)
- xoffset = info->output_width - info->crop_width - info->crop_xoffset;
- else
+ if (info->crop_xoffset_set != JCROP_NEG)
xoffset = info->crop_xoffset;
- if (info->crop_yoffset_set == JCROP_NEG)
- yoffset = info->output_height - info->crop_height - info->crop_yoffset;
+ else if (info->crop_width > info->output_width) /* crop extension */
+ xoffset = info->crop_width - info->output_width - info->crop_xoffset;
else
+ xoffset = info->output_width - info->crop_width - info->crop_xoffset;
+ if (info->crop_yoffset_set != JCROP_NEG)
yoffset = info->crop_yoffset;
+ else if (info->crop_height > info->output_height) /* crop extension */
+ yoffset = info->crop_height - info->output_height - info->crop_yoffset;
+ else
+ yoffset = info->output_height - info->crop_height - info->crop_yoffset;
/* Now adjust so that upper left corner falls at an iMCU boundary */
- if (info->crop_width_set == JCROP_FORCE)
- info->output_width = info->crop_width;
- else
- info->output_width =
- info->crop_width + (xoffset % info->iMCU_sample_width);
- if (info->crop_height_set == JCROP_FORCE)
- info->output_height = info->crop_height;
- else
- info->output_height =
- info->crop_height + (yoffset % info->iMCU_sample_height);
+ switch (info->transform) {
+ case JXFORM_DROP:
+ /* Ensure the effective drop region will not exceed the requested */
+ itemp = info->iMCU_sample_width;
+ dtemp = itemp - 1 - ((xoffset + itemp - 1) % itemp);
+ xoffset += dtemp;
+ if (info->crop_width <= dtemp)
+ info->drop_width = 0;
+ else if (xoffset + info->crop_width - dtemp == info->output_width)
+ /* Matching right edge: include partial iMCU */
+ info->drop_width = (info->crop_width - dtemp + itemp - 1) / itemp;
+ else
+ info->drop_width = (info->crop_width - dtemp) / itemp;
+ itemp = info->iMCU_sample_height;
+ dtemp = itemp - 1 - ((yoffset + itemp - 1) % itemp);
+ yoffset += dtemp;
+ if (info->crop_height <= dtemp)
+ info->drop_height = 0;
+ else if (yoffset + info->crop_height - dtemp == info->output_height)
+ /* Matching bottom edge: include partial iMCU */
+ info->drop_height = (info->crop_height - dtemp + itemp - 1) / itemp;
+ else
+ info->drop_height = (info->crop_height - dtemp) / itemp;
+ /* Check if sampling factors match for dropping */
+ if (info->drop_width != 0 && info->drop_height != 0)
+ for (ci = 0; ci < info->num_components &&
+ ci < info->drop_ptr->num_components; ci++) {
+ if (info->drop_ptr->comp_info[ci].h_samp_factor *
+ srcinfo->max_h_samp_factor !=
+ srcinfo->comp_info[ci].h_samp_factor *
+ info->drop_ptr->max_h_samp_factor)
+ ERREXIT6(srcinfo, JERR_BAD_DROP_SAMPLING, ci,
+ info->drop_ptr->comp_info[ci].h_samp_factor,
+ info->drop_ptr->max_h_samp_factor,
+ srcinfo->comp_info[ci].h_samp_factor,
+ srcinfo->max_h_samp_factor, 'h');
+ if (info->drop_ptr->comp_info[ci].v_samp_factor *
+ srcinfo->max_v_samp_factor !=
+ srcinfo->comp_info[ci].v_samp_factor *
+ info->drop_ptr->max_v_samp_factor)
+ ERREXIT6(srcinfo, JERR_BAD_DROP_SAMPLING, ci,
+ info->drop_ptr->comp_info[ci].v_samp_factor,
+ info->drop_ptr->max_v_samp_factor,
+ srcinfo->comp_info[ci].v_samp_factor,
+ srcinfo->max_v_samp_factor, 'v');
+ }
+ break;
+ case JXFORM_WIPE:
+ /* Ensure the effective wipe region will cover the requested */
+ info->drop_width = (JDIMENSION)jdiv_round_up
+ ((long)(info->crop_width + (xoffset % info->iMCU_sample_width)),
+ (long)info->iMCU_sample_width);
+ info->drop_height = (JDIMENSION)jdiv_round_up
+ ((long)(info->crop_height + (yoffset % info->iMCU_sample_height)),
+ (long)info->iMCU_sample_height);
+ break;
+ default:
+ /* Ensure the effective crop region will cover the requested */
+ if (info->crop_width_set == JCROP_FORCE ||
+ info->crop_width > info->output_width)
+ info->output_width = info->crop_width;
+ else
+ info->output_width =
+ info->crop_width + (xoffset % info->iMCU_sample_width);
+ if (info->crop_height_set == JCROP_FORCE ||
+ info->crop_height > info->output_height)
+ info->output_height = info->crop_height;
+ else
+ info->output_height =
+ info->crop_height + (yoffset % info->iMCU_sample_height);
+ }
/* Save x/y offsets measured in iMCUs */
info->x_crop_offset = xoffset / info->iMCU_sample_width;
info->y_crop_offset = yoffset / info->iMCU_sample_height;
@@ -1033,7 +1720,9 @@
transpose_it = FALSE;
switch (info->transform) {
case JXFORM_NONE:
- if (info->x_crop_offset != 0 || info->y_crop_offset != 0)
+ if (info->x_crop_offset != 0 || info->y_crop_offset != 0 ||
+ info->output_width > srcinfo->output_width ||
+ info->output_height > srcinfo->output_height)
need_workspace = TRUE;
/* No workspace needed if neither cropping nor transforming */
break;
@@ -1087,6 +1776,10 @@
need_workspace = TRUE;
transpose_it = TRUE;
break;
+ case JXFORM_WIPE:
+ break;
+ case JXFORM_DROP:
+ break;
}
/* Allocate workspace if needed.
@@ -1190,47 +1883,47 @@
if (length < 12) return; /* Length of an IFD entry */
/* Discover byte order */
- if (GETJOCTET(data[0]) == 0x49 && GETJOCTET(data[1]) == 0x49)
+ if (data[0] == 0x49 && data[1] == 0x49)
is_motorola = FALSE;
- else if (GETJOCTET(data[0]) == 0x4D && GETJOCTET(data[1]) == 0x4D)
+ else if (data[0] == 0x4D && data[1] == 0x4D)
is_motorola = TRUE;
else
return;
/* Check Tag Mark */
if (is_motorola) {
- if (GETJOCTET(data[2]) != 0) return;
- if (GETJOCTET(data[3]) != 0x2A) return;
+ if (data[2] != 0) return;
+ if (data[3] != 0x2A) return;
} else {
- if (GETJOCTET(data[3]) != 0) return;
- if (GETJOCTET(data[2]) != 0x2A) return;
+ if (data[3] != 0) return;
+ if (data[2] != 0x2A) return;
}
/* Get first IFD offset (offset to IFD0) */
if (is_motorola) {
- if (GETJOCTET(data[4]) != 0) return;
- if (GETJOCTET(data[5]) != 0) return;
- firstoffset = GETJOCTET(data[6]);
+ if (data[4] != 0) return;
+ if (data[5] != 0) return;
+ firstoffset = data[6];
firstoffset <<= 8;
- firstoffset += GETJOCTET(data[7]);
+ firstoffset += data[7];
} else {
- if (GETJOCTET(data[7]) != 0) return;
- if (GETJOCTET(data[6]) != 0) return;
- firstoffset = GETJOCTET(data[5]);
+ if (data[7] != 0) return;
+ if (data[6] != 0) return;
+ firstoffset = data[5];
firstoffset <<= 8;
- firstoffset += GETJOCTET(data[4]);
+ firstoffset += data[4];
}
if (firstoffset > length - 2) return; /* check end of data segment */
/* Get the number of directory entries contained in this IFD */
if (is_motorola) {
- number_of_tags = GETJOCTET(data[firstoffset]);
+ number_of_tags = data[firstoffset];
number_of_tags <<= 8;
- number_of_tags += GETJOCTET(data[firstoffset + 1]);
+ number_of_tags += data[firstoffset + 1];
} else {
- number_of_tags = GETJOCTET(data[firstoffset + 1]);
+ number_of_tags = data[firstoffset + 1];
number_of_tags <<= 8;
- number_of_tags += GETJOCTET(data[firstoffset]);
+ number_of_tags += data[firstoffset];
}
if (number_of_tags == 0) return;
firstoffset += 2;
@@ -1240,13 +1933,13 @@
if (firstoffset > length - 12) return; /* check end of data segment */
/* Get Tag number */
if (is_motorola) {
- tagnum = GETJOCTET(data[firstoffset]);
+ tagnum = data[firstoffset];
tagnum <<= 8;
- tagnum += GETJOCTET(data[firstoffset + 1]);
+ tagnum += data[firstoffset + 1];
} else {
- tagnum = GETJOCTET(data[firstoffset + 1]);
+ tagnum = data[firstoffset + 1];
tagnum <<= 8;
- tagnum += GETJOCTET(data[firstoffset]);
+ tagnum += data[firstoffset];
}
if (tagnum == 0x8769) break; /* found ExifSubIFD offset Tag */
if (--number_of_tags == 0) return;
@@ -1255,29 +1948,29 @@
/* Get the ExifSubIFD offset */
if (is_motorola) {
- if (GETJOCTET(data[firstoffset + 8]) != 0) return;
- if (GETJOCTET(data[firstoffset + 9]) != 0) return;
- offset = GETJOCTET(data[firstoffset + 10]);
+ if (data[firstoffset + 8] != 0) return;
+ if (data[firstoffset + 9] != 0) return;
+ offset = data[firstoffset + 10];
offset <<= 8;
- offset += GETJOCTET(data[firstoffset + 11]);
+ offset += data[firstoffset + 11];
} else {
- if (GETJOCTET(data[firstoffset + 11]) != 0) return;
- if (GETJOCTET(data[firstoffset + 10]) != 0) return;
- offset = GETJOCTET(data[firstoffset + 9]);
+ if (data[firstoffset + 11] != 0) return;
+ if (data[firstoffset + 10] != 0) return;
+ offset = data[firstoffset + 9];
offset <<= 8;
- offset += GETJOCTET(data[firstoffset + 8]);
+ offset += data[firstoffset + 8];
}
if (offset > length - 2) return; /* check end of data segment */
/* Get the number of directory entries contained in this SubIFD */
if (is_motorola) {
- number_of_tags = GETJOCTET(data[offset]);
+ number_of_tags = data[offset];
number_of_tags <<= 8;
- number_of_tags += GETJOCTET(data[offset + 1]);
+ number_of_tags += data[offset + 1];
} else {
- number_of_tags = GETJOCTET(data[offset + 1]);
+ number_of_tags = data[offset + 1];
number_of_tags <<= 8;
- number_of_tags += GETJOCTET(data[offset]);
+ number_of_tags += data[offset];
}
if (number_of_tags < 2) return;
offset += 2;
@@ -1287,13 +1980,13 @@
if (offset > length - 12) return; /* check end of data segment */
/* Get Tag number */
if (is_motorola) {
- tagnum = GETJOCTET(data[offset]);
+ tagnum = data[offset];
tagnum <<= 8;
- tagnum += GETJOCTET(data[offset + 1]);
+ tagnum += data[offset + 1];
} else {
- tagnum = GETJOCTET(data[offset + 1]);
+ tagnum = data[offset + 1];
tagnum <<= 8;
- tagnum += GETJOCTET(data[offset]);
+ tagnum += data[offset];
}
if (tagnum == 0xA002 || tagnum == 0xA003) {
if (tagnum == 0xA002)
@@ -1387,7 +2080,7 @@
dstinfo->jpeg_height = info->output_height;
#endif
- /* Transpose destination image parameters */
+ /* Transpose destination image parameters, adjust quantization */
switch (info->transform) {
case JXFORM_TRANSPOSE:
case JXFORM_TRANSVERSE:
@@ -1399,6 +2092,12 @@
#endif
transpose_critical_parameters(dstinfo);
break;
+ case JXFORM_DROP:
+ if (info->drop_width != 0 && info->drop_height != 0)
+ adjust_quant(srcinfo, src_coef_arrays,
+ info->drop_ptr, info->drop_coef_arrays,
+ info->trim, dstinfo);
+ break;
default:
#if JPEG_LIB_VERSION < 80
dstinfo->image_width = info->output_width;
@@ -1411,12 +2110,12 @@
if (srcinfo->marker_list != NULL &&
srcinfo->marker_list->marker == JPEG_APP0 + 1 &&
srcinfo->marker_list->data_length >= 6 &&
- GETJOCTET(srcinfo->marker_list->data[0]) == 0x45 &&
- GETJOCTET(srcinfo->marker_list->data[1]) == 0x78 &&
- GETJOCTET(srcinfo->marker_list->data[2]) == 0x69 &&
- GETJOCTET(srcinfo->marker_list->data[3]) == 0x66 &&
- GETJOCTET(srcinfo->marker_list->data[4]) == 0 &&
- GETJOCTET(srcinfo->marker_list->data[5]) == 0) {
+ srcinfo->marker_list->data[0] == 0x45 &&
+ srcinfo->marker_list->data[1] == 0x78 &&
+ srcinfo->marker_list->data[2] == 0x69 &&
+ srcinfo->marker_list->data[3] == 0x66 &&
+ srcinfo->marker_list->data[4] == 0 &&
+ srcinfo->marker_list->data[5] == 0) {
/* Suppress output of JFIF marker */
dstinfo->write_JFIF_header = FALSE;
/* Adjust Exif image parameters */
@@ -1465,7 +2164,23 @@
*/
switch (info->transform) {
case JXFORM_NONE:
- if (info->x_crop_offset != 0 || info->y_crop_offset != 0)
+ if (info->output_width > srcinfo->output_width ||
+ info->output_height > srcinfo->output_height) {
+ if (info->output_width > srcinfo->output_width &&
+ info->crop_width_set == JCROP_REFLECT)
+ do_crop_ext_reflect(srcinfo, dstinfo,
+ info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, dst_coef_arrays);
+ else if (info->output_width > srcinfo->output_width &&
+ info->crop_width_set == JCROP_FORCE)
+ do_crop_ext_flat(srcinfo, dstinfo,
+ info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, dst_coef_arrays);
+ else
+ do_crop_ext_zero(srcinfo, dstinfo,
+ info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, dst_coef_arrays);
+ } else if (info->x_crop_offset != 0 || info->y_crop_offset != 0)
do_crop(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
src_coef_arrays, dst_coef_arrays);
break;
@@ -1501,6 +2216,30 @@
do_rot_270(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
src_coef_arrays, dst_coef_arrays);
break;
+ case JXFORM_WIPE:
+ if (info->crop_width_set == JCROP_REFLECT &&
+ info->y_crop_offset == 0 && info->drop_height ==
+ (JDIMENSION)jdiv_round_up
+ ((long)info->output_height, (long)info->iMCU_sample_height) &&
+ (info->x_crop_offset == 0 ||
+ info->x_crop_offset + info->drop_width ==
+ (JDIMENSION)jdiv_round_up
+ ((long)info->output_width, (long)info->iMCU_sample_width)))
+ do_reflect(srcinfo, dstinfo, info->x_crop_offset,
+ src_coef_arrays, info->drop_width, info->drop_height);
+ else if (info->crop_width_set == JCROP_FORCE)
+ do_flatten(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, info->drop_width, info->drop_height);
+ else
+ do_wipe(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, info->drop_width, info->drop_height);
+ break;
+ case JXFORM_DROP:
+ if (info->drop_width != 0 && info->drop_height != 0)
+ do_drop(srcinfo, dstinfo, info->x_crop_offset, info->y_crop_offset,
+ src_coef_arrays, info->drop_ptr, info->drop_coef_arrays,
+ info->drop_width, info->drop_height);
+ break;
}
}
@@ -1607,20 +2346,20 @@
if (dstinfo->write_JFIF_header &&
marker->marker == JPEG_APP0 &&
marker->data_length >= 5 &&
- GETJOCTET(marker->data[0]) == 0x4A &&
- GETJOCTET(marker->data[1]) == 0x46 &&
- GETJOCTET(marker->data[2]) == 0x49 &&
- GETJOCTET(marker->data[3]) == 0x46 &&
- GETJOCTET(marker->data[4]) == 0)
+ marker->data[0] == 0x4A &&
+ marker->data[1] == 0x46 &&
+ marker->data[2] == 0x49 &&
+ marker->data[3] == 0x46 &&
+ marker->data[4] == 0)
continue; /* reject duplicate JFIF */
if (dstinfo->write_Adobe_marker &&
marker->marker == JPEG_APP0 + 14 &&
marker->data_length >= 5 &&
- GETJOCTET(marker->data[0]) == 0x41 &&
- GETJOCTET(marker->data[1]) == 0x64 &&
- GETJOCTET(marker->data[2]) == 0x6F &&
- GETJOCTET(marker->data[3]) == 0x62 &&
- GETJOCTET(marker->data[4]) == 0x65)
+ marker->data[0] == 0x41 &&
+ marker->data[1] == 0x64 &&
+ marker->data[2] == 0x6F &&
+ marker->data[3] == 0x62 &&
+ marker->data[4] == 0x65)
continue; /* reject duplicate Adobe */
jpeg_write_marker(dstinfo, marker->marker,
marker->data, marker->data_length);
diff --git a/transupp.h b/transupp.h
index 80264cc..ea6be1f 100644
--- a/transupp.h
+++ b/transupp.h
@@ -2,7 +2,7 @@
* transupp.h
*
* This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1997-2011, Thomas G. Lane, Guido Vollbeding.
+ * Copyright (C) 1997-2019, Thomas G. Lane, Guido Vollbeding.
* libjpeg-turbo Modifications:
* Copyright (C) 2017, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
@@ -62,6 +62,17 @@
* output image covers at least the requested region, but may cover more.)
* The adjustment of the region dimensions may be optionally disabled.
*
+ * A complementary lossless wipe option is provided to discard (gray out) data
+ * inside a given image region while losslessly preserving what is outside.
+ * A lossless drop option is also provided, which allows another JPEG image to
+ * be inserted ("dropped") into the source image data at a given position,
+ * replacing the existing image data at that position. Both the source image
+ * and the drop image must have the same subsampling level. It is best if they
+ * also have the same quantization (quality.) Otherwise, the quantization of
+ * the output image will be adapted to accommodate the higher of the source
+ * image quality and the drop image quality. The trim option can be used with
+ * the drop option to requantize the drop image to match the source image.
+ *
* We also provide a lossless-resize option, which is kind of a lossless-crop
* operation in the DCT coefficient block domain - it discards higher-order
* coefficients and losslessly preserves lower-order coefficients of a
@@ -92,20 +103,23 @@
JXFORM_TRANSVERSE, /* transpose across UR-to-LL axis */
JXFORM_ROT_90, /* 90-degree clockwise rotation */
JXFORM_ROT_180, /* 180-degree rotation */
- JXFORM_ROT_270 /* 270-degree clockwise (or 90 ccw) */
+ JXFORM_ROT_270, /* 270-degree clockwise (or 90 ccw) */
+ JXFORM_WIPE, /* wipe */
+ JXFORM_DROP /* drop */
} JXFORM_CODE;
/*
* Codes for crop parameters, which can individually be unspecified,
* positive or negative for xoffset or yoffset,
- * positive or forced for width or height.
+ * positive or force or reflect for width or height.
*/
typedef enum {
JCROP_UNSET,
JCROP_POS,
JCROP_NEG,
- JCROP_FORCE
+ JCROP_FORCE,
+ JCROP_REFLECT
} JCROP_CODE;
/*
@@ -120,7 +134,7 @@
boolean perfect; /* if TRUE, fail if partial MCUs are requested */
boolean trim; /* if TRUE, trim partial MCUs as needed */
boolean force_grayscale; /* if TRUE, convert color image to grayscale */
- boolean crop; /* if TRUE, crop source image */
+ boolean crop; /* if TRUE, crop or wipe source image, or drop */
boolean slow_hflip; /* For best performance, the JXFORM_FLIP_H transform
normally modifies the source coefficients in place.
Setting this to TRUE will instead use a slower,
@@ -133,14 +147,18 @@
* These can be filled in by jtransform_parse_crop_spec().
*/
JDIMENSION crop_width; /* Width of selected region */
- JCROP_CODE crop_width_set; /* (forced disables adjustment) */
+ JCROP_CODE crop_width_set; /* (force-disables adjustment) */
JDIMENSION crop_height; /* Height of selected region */
- JCROP_CODE crop_height_set; /* (forced disables adjustment) */
+ JCROP_CODE crop_height_set; /* (force-disables adjustment) */
JDIMENSION crop_xoffset; /* X offset of selected region */
JCROP_CODE crop_xoffset_set; /* (negative measures from right edge) */
JDIMENSION crop_yoffset; /* Y offset of selected region */
JCROP_CODE crop_yoffset_set; /* (negative measures from bottom edge) */
+ /* Drop parameters: set by caller for drop request */
+ j_decompress_ptr drop_ptr;
+ jvirt_barray_ptr *drop_coef_arrays;
+
/* Internal workspace: caller should not touch these */
int num_components; /* # of components in workspace */
jvirt_barray_ptr *workspace_coef_arrays; /* workspace for transformations */
@@ -148,6 +166,8 @@
JDIMENSION output_height;
JDIMENSION x_crop_offset; /* destination crop offsets measured in iMCUs */
JDIMENSION y_crop_offset;
+ JDIMENSION drop_width; /* drop/wipe dimensions measured in iMCUs */
+ JDIMENSION drop_height;
int iMCU_sample_width; /* destination iMCU size */
int iMCU_sample_height;
} jpeg_transform_info;
diff --git a/turbojpeg-jni.c b/turbojpeg-jni.c
index 9363450..1b728e3 100644
--- a/turbojpeg-jni.c
+++ b/turbojpeg-jni.c
@@ -1,5 +1,5 @@
/*
- * Copyright (C)2011-2019 D. R. Commander. All Rights Reserved.
+ * Copyright (C)2011-2020 D. R. Commander. All Rights Reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
@@ -326,9 +326,11 @@
tjhandle handle = 0;
unsigned long jpegSize = 0;
jbyteArray jSrcPlanes[3] = { NULL, NULL, NULL };
- const unsigned char *srcPlanes[3];
+ const unsigned char *srcPlanesTmp[3] = { NULL, NULL, NULL };
+ const unsigned char *srcPlanes[3] = { NULL, NULL, NULL };
+ int *srcOffsetsTmp = NULL, srcOffsets[3] = { 0, 0, 0 };
+ int *srcStridesTmp = NULL, srcStrides[3] = { 0, 0, 0 };
unsigned char *jpegBuf = NULL;
- int *srcOffsets = NULL, *srcStrides = NULL;
int nc = (subsamp == org_libjpegturbo_turbojpeg_TJ_SAMP_GRAY ? 1 : 3), i;
GET_HANDLE();
@@ -351,56 +353,49 @@
if (ProcessSystemProperties(env) < 0) goto bailout;
-#define RELEASE_ARRAYS_COMPRESSFROMYUV() { \
- SAFE_RELEASE(dst, jpegBuf); \
- for (i = 0; i < nc; i++) \
- SAFE_RELEASE(jSrcPlanes[i], srcPlanes[i]); \
- SAFE_RELEASE(jSrcStrides, srcStrides); \
- SAFE_RELEASE(jSrcOffsets, srcOffsets); \
-}
+ BAILIF0(srcOffsetsTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jSrcOffsets, 0));
+ for (i = 0; i < nc; i++) srcOffsets[i] = srcOffsetsTmp[i];
+ SAFE_RELEASE(jSrcOffsets, srcOffsetsTmp);
- BAILIF0(srcOffsets = (*env)->GetPrimitiveArrayCritical(env, jSrcOffsets, 0));
- BAILIF0(srcStrides = (*env)->GetPrimitiveArrayCritical(env, jSrcStrides, 0));
+ BAILIF0(srcStridesTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jSrcStrides, 0));
+ for (i = 0; i < nc; i++) srcStrides[i] = srcStridesTmp[i];
+ SAFE_RELEASE(jSrcStrides, srcStridesTmp);
+
for (i = 0; i < nc; i++) {
int planeSize = tjPlaneSizeYUV(i, width, srcStrides[i], height, subsamp);
int pw = tjPlaneWidth(i, width, subsamp);
- if (planeSize < 0 || pw < 0) {
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ if (planeSize < 0 || pw < 0)
THROW_ARG(tjGetErrorStr());
- }
- if (srcOffsets[i] < 0) {
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ if (srcOffsets[i] < 0)
THROW_ARG("Invalid argument in compressFromYUV()");
- }
- if (srcStrides[i] < 0 && srcOffsets[i] - planeSize + pw < 0) {
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ if (srcStrides[i] < 0 && srcOffsets[i] - planeSize + pw < 0)
THROW_ARG("Negative plane stride would cause memory to be accessed below plane boundary");
- }
BAILIF0(jSrcPlanes[i] = (*env)->GetObjectArrayElement(env, srcobjs, i));
if ((*env)->GetArrayLength(env, jSrcPlanes[i]) <
- srcOffsets[i] + planeSize) {
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ srcOffsets[i] + planeSize)
THROW_ARG("Source plane is not large enough");
- }
- BAILIF0(srcPlanes[i] =
+ BAILIF0(srcPlanesTmp[i] =
(*env)->GetPrimitiveArrayCritical(env, jSrcPlanes[i], 0));
- srcPlanes[i] = &srcPlanes[i][srcOffsets[i]];
+ srcPlanes[i] = &srcPlanesTmp[i][srcOffsets[i]];
+ SAFE_RELEASE(jSrcPlanes[i], srcPlanesTmp[i]);
}
BAILIF0(jpegBuf = (*env)->GetPrimitiveArrayCritical(env, dst, 0));
if (tjCompressFromYUVPlanes(handle, srcPlanes, width, srcStrides, height,
subsamp, &jpegBuf, &jpegSize, jpegQual,
flags | TJFLAG_NOREALLOC) == -1) {
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ SAFE_RELEASE(dst, jpegBuf);
THROW_TJ();
}
bailout:
- RELEASE_ARRAYS_COMPRESSFROMYUV();
+ SAFE_RELEASE(dst, jpegBuf);
return (jint)jpegSize;
}
@@ -411,9 +406,12 @@
{
tjhandle handle = 0;
jsize arraySize = 0, actualPitch;
+ unsigned char *srcBuf = NULL;
jbyteArray jDstPlanes[3] = { NULL, NULL, NULL };
- unsigned char *srcBuf = NULL, *dstPlanes[3];
- int *dstOffsets = NULL, *dstStrides = NULL;
+ unsigned char *dstPlanesTmp[3] = { NULL, NULL, NULL };
+ unsigned char *dstPlanes[3] = { NULL, NULL, NULL };
+ int *dstOffsetsTmp = NULL, dstOffsets[3] = { 0, 0, 0 };
+ int *dstStridesTmp = NULL, dstStrides[3] = { 0, 0, 0 };
int nc = (subsamp == org_libjpegturbo_turbojpeg_TJ_SAMP_GRAY ? 1 : 3), i;
GET_HANDLE();
@@ -438,56 +436,49 @@
if ((*env)->GetArrayLength(env, src) * srcElementSize < arraySize)
THROW_ARG("Source buffer is not large enough");
-#define RELEASE_ARRAYS_ENCODEYUV() { \
- SAFE_RELEASE(src, srcBuf); \
- for (i = 0; i < nc; i++) \
- SAFE_RELEASE(jDstPlanes[i], dstPlanes[i]); \
- SAFE_RELEASE(jDstStrides, dstStrides); \
- SAFE_RELEASE(jDstOffsets, dstOffsets); \
-}
+ BAILIF0(dstOffsetsTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jDstOffsets, 0));
+ for (i = 0; i < nc; i++) dstOffsets[i] = dstOffsetsTmp[i];
+ SAFE_RELEASE(jDstOffsets, dstOffsetsTmp);
- BAILIF0(dstOffsets = (*env)->GetPrimitiveArrayCritical(env, jDstOffsets, 0));
- BAILIF0(dstStrides = (*env)->GetPrimitiveArrayCritical(env, jDstStrides, 0));
+ BAILIF0(dstStridesTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jDstStrides, 0));
+ for (i = 0; i < nc; i++) dstStrides[i] = dstStridesTmp[i];
+ SAFE_RELEASE(jDstStrides, dstStridesTmp);
+
for (i = 0; i < nc; i++) {
int planeSize = tjPlaneSizeYUV(i, width, dstStrides[i], height, subsamp);
int pw = tjPlaneWidth(i, width, subsamp);
- if (planeSize < 0 || pw < 0) {
- RELEASE_ARRAYS_ENCODEYUV();
+ if (planeSize < 0 || pw < 0)
THROW_ARG(tjGetErrorStr());
- }
- if (dstOffsets[i] < 0) {
- RELEASE_ARRAYS_ENCODEYUV();
+ if (dstOffsets[i] < 0)
THROW_ARG("Invalid argument in encodeYUV()");
- }
- if (dstStrides[i] < 0 && dstOffsets[i] - planeSize + pw < 0) {
- RELEASE_ARRAYS_ENCODEYUV();
+ if (dstStrides[i] < 0 && dstOffsets[i] - planeSize + pw < 0)
THROW_ARG("Negative plane stride would cause memory to be accessed below plane boundary");
- }
BAILIF0(jDstPlanes[i] = (*env)->GetObjectArrayElement(env, dstobjs, i));
if ((*env)->GetArrayLength(env, jDstPlanes[i]) <
- dstOffsets[i] + planeSize) {
- RELEASE_ARRAYS_ENCODEYUV();
+ dstOffsets[i] + planeSize)
THROW_ARG("Destination plane is not large enough");
- }
- BAILIF0(dstPlanes[i] =
+ BAILIF0(dstPlanesTmp[i] =
(*env)->GetPrimitiveArrayCritical(env, jDstPlanes[i], 0));
- dstPlanes[i] = &dstPlanes[i][dstOffsets[i]];
+ dstPlanes[i] = &dstPlanesTmp[i][dstOffsets[i]];
+ SAFE_RELEASE(jDstPlanes[i], dstPlanesTmp[i]);
}
BAILIF0(srcBuf = (*env)->GetPrimitiveArrayCritical(env, src, 0));
if (tjEncodeYUVPlanes(handle, &srcBuf[y * actualPitch + x * tjPixelSize[pf]],
width, pitch, height, pf, dstPlanes, dstStrides,
subsamp, flags) == -1) {
- RELEASE_ARRAYS_ENCODEYUV();
+ SAFE_RELEASE(src, srcBuf);
THROW_TJ();
}
bailout:
- RELEASE_ARRAYS_ENCODEYUV();
+ SAFE_RELEASE(src, srcBuf);
}
/* TurboJPEG 1.4.x: TJCompressor::encodeYUV() byte source */
@@ -785,9 +776,12 @@
jintArray jDstStrides, jint desiredHeight, jint flags)
{
tjhandle handle = 0;
+ unsigned char *jpegBuf = NULL;
jbyteArray jDstPlanes[3] = { NULL, NULL, NULL };
- unsigned char *jpegBuf = NULL, *dstPlanes[3];
- int *dstOffsets = NULL, *dstStrides = NULL;
+ unsigned char *dstPlanesTmp[3] = { NULL, NULL, NULL };
+ unsigned char *dstPlanes[3] = { NULL, NULL, NULL };
+ int *dstOffsetsTmp = NULL, dstOffsets[3] = { 0, 0, 0 };
+ int *dstStridesTmp = NULL, dstStrides[3] = { 0, 0, 0 };
int jpegSubsamp = -1, jpegWidth = 0, jpegHeight = 0;
int nc = 0, i, width, height, scaledWidth, scaledHeight, nsf = 0;
tjscalingfactor *sf;
@@ -821,57 +815,50 @@
if (i >= nsf)
THROW_ARG("Could not scale down to desired image dimensions");
-#define RELEASE_ARRAYS_DECOMPRESSTOYUV() { \
- SAFE_RELEASE(src, jpegBuf); \
- for (i = 0; i < nc; i++) \
- SAFE_RELEASE(jDstPlanes[i], dstPlanes[i]); \
- SAFE_RELEASE(jDstStrides, dstStrides); \
- SAFE_RELEASE(jDstOffsets, dstOffsets); \
-}
+ BAILIF0(dstOffsetsTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jDstOffsets, 0));
+ for (i = 0; i < nc; i++) dstOffsets[i] = dstOffsetsTmp[i];
+ SAFE_RELEASE(jDstOffsets, dstOffsetsTmp);
- BAILIF0(dstOffsets = (*env)->GetPrimitiveArrayCritical(env, jDstOffsets, 0));
- BAILIF0(dstStrides = (*env)->GetPrimitiveArrayCritical(env, jDstStrides, 0));
+ BAILIF0(dstStridesTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jDstStrides, 0));
+ for (i = 0; i < nc; i++) dstStrides[i] = dstStridesTmp[i];
+ SAFE_RELEASE(jDstStrides, dstStridesTmp);
+
for (i = 0; i < nc; i++) {
int planeSize = tjPlaneSizeYUV(i, scaledWidth, dstStrides[i], scaledHeight,
jpegSubsamp);
int pw = tjPlaneWidth(i, scaledWidth, jpegSubsamp);
- if (planeSize < 0 || pw < 0) {
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ if (planeSize < 0 || pw < 0)
THROW_ARG(tjGetErrorStr());
- }
- if (dstOffsets[i] < 0) {
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ if (dstOffsets[i] < 0)
THROW_ARG("Invalid argument in decompressToYUV()");
- }
- if (dstStrides[i] < 0 && dstOffsets[i] - planeSize + pw < 0) {
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ if (dstStrides[i] < 0 && dstOffsets[i] - planeSize + pw < 0)
THROW_ARG("Negative plane stride would cause memory to be accessed below plane boundary");
- }
BAILIF0(jDstPlanes[i] = (*env)->GetObjectArrayElement(env, dstobjs, i));
if ((*env)->GetArrayLength(env, jDstPlanes[i]) <
- dstOffsets[i] + planeSize) {
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ dstOffsets[i] + planeSize)
THROW_ARG("Destination plane is not large enough");
- }
- BAILIF0(dstPlanes[i] =
+ BAILIF0(dstPlanesTmp[i] =
(*env)->GetPrimitiveArrayCritical(env, jDstPlanes[i], 0));
- dstPlanes[i] = &dstPlanes[i][dstOffsets[i]];
+ dstPlanes[i] = &dstPlanesTmp[i][dstOffsets[i]];
+ SAFE_RELEASE(jDstPlanes[i], dstPlanesTmp[i]);
}
BAILIF0(jpegBuf = (*env)->GetPrimitiveArrayCritical(env, src, 0));
if (tjDecompressToYUVPlanes(handle, jpegBuf, (unsigned long)jpegSize,
dstPlanes, desiredWidth, dstStrides,
desiredHeight, flags) == -1) {
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ SAFE_RELEASE(src, jpegBuf);
THROW_TJ();
}
bailout:
- RELEASE_ARRAYS_DECOMPRESSTOYUV();
+ SAFE_RELEASE(src, jpegBuf);
}
/* TurboJPEG 1.2.x: TJDecompressor::decompressToYUV() */
@@ -920,9 +907,11 @@
tjhandle handle = 0;
jsize arraySize = 0, actualPitch;
jbyteArray jSrcPlanes[3] = { NULL, NULL, NULL };
- const unsigned char *srcPlanes[3];
+ const unsigned char *srcPlanesTmp[3] = { NULL, NULL, NULL };
+ const unsigned char *srcPlanes[3] = { NULL, NULL, NULL };
+ int *srcOffsetsTmp = NULL, srcOffsets[3] = { 0, 0, 0 };
+ int *srcStridesTmp = NULL, srcStrides[3] = { 0, 0, 0 };
unsigned char *dstBuf = NULL;
- int *srcOffsets = NULL, *srcStrides = NULL;
int nc = (subsamp == org_libjpegturbo_turbojpeg_TJ_SAMP_GRAY ? 1 : 3), i;
GET_HANDLE();
@@ -946,56 +935,49 @@
if ((*env)->GetArrayLength(env, dst) * dstElementSize < arraySize)
THROW_ARG("Destination buffer is not large enough");
-#define RELEASE_ARRAYS_DECODEYUV() { \
- SAFE_RELEASE(dst, dstBuf); \
- for (i = 0; i < nc; i++) \
- SAFE_RELEASE(jSrcPlanes[i], srcPlanes[i]); \
- SAFE_RELEASE(jSrcStrides, srcStrides); \
- SAFE_RELEASE(jSrcOffsets, srcOffsets); \
-}
+ BAILIF0(srcOffsetsTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jSrcOffsets, 0));
+ for (i = 0; i < nc; i++) srcOffsets[i] = srcOffsetsTmp[i];
+ SAFE_RELEASE(jSrcOffsets, srcOffsetsTmp);
- BAILIF0(srcOffsets = (*env)->GetPrimitiveArrayCritical(env, jSrcOffsets, 0));
- BAILIF0(srcStrides = (*env)->GetPrimitiveArrayCritical(env, jSrcStrides, 0));
+ BAILIF0(srcStridesTmp =
+ (*env)->GetPrimitiveArrayCritical(env, jSrcStrides, 0));
+ for (i = 0; i < nc; i++) srcStrides[i] = srcStridesTmp[i];
+ SAFE_RELEASE(jSrcStrides, srcStridesTmp);
+
for (i = 0; i < nc; i++) {
int planeSize = tjPlaneSizeYUV(i, width, srcStrides[i], height, subsamp);
int pw = tjPlaneWidth(i, width, subsamp);
- if (planeSize < 0 || pw < 0) {
- RELEASE_ARRAYS_DECODEYUV();
+ if (planeSize < 0 || pw < 0)
THROW_ARG(tjGetErrorStr());
- }
- if (srcOffsets[i] < 0) {
- RELEASE_ARRAYS_DECODEYUV();
+ if (srcOffsets[i] < 0)
THROW_ARG("Invalid argument in decodeYUV()");
- }
- if (srcStrides[i] < 0 && srcOffsets[i] - planeSize + pw < 0) {
- RELEASE_ARRAYS_DECODEYUV();
+ if (srcStrides[i] < 0 && srcOffsets[i] - planeSize + pw < 0)
THROW_ARG("Negative plane stride would cause memory to be accessed below plane boundary");
- }
BAILIF0(jSrcPlanes[i] = (*env)->GetObjectArrayElement(env, srcobjs, i));
if ((*env)->GetArrayLength(env, jSrcPlanes[i]) <
- srcOffsets[i] + planeSize) {
- RELEASE_ARRAYS_DECODEYUV();
+ srcOffsets[i] + planeSize)
THROW_ARG("Source plane is not large enough");
- }
- BAILIF0(srcPlanes[i] =
+ BAILIF0(srcPlanesTmp[i] =
(*env)->GetPrimitiveArrayCritical(env, jSrcPlanes[i], 0));
- srcPlanes[i] = &srcPlanes[i][srcOffsets[i]];
+ srcPlanes[i] = &srcPlanesTmp[i][srcOffsets[i]];
+ SAFE_RELEASE(jSrcPlanes[i], srcPlanesTmp[i]);
}
BAILIF0(dstBuf = (*env)->GetPrimitiveArrayCritical(env, dst, 0));
if (tjDecodeYUVPlanes(handle, srcPlanes, srcStrides, subsamp,
&dstBuf[y * actualPitch + x * tjPixelSize[pf]], width,
pitch, height, pf, flags) == -1) {
- RELEASE_ARRAYS_DECODEYUV();
+ SAFE_RELEASE(dst, dstBuf);
THROW_TJ();
}
bailout:
- RELEASE_ARRAYS_DECODEYUV();
+ SAFE_RELEASE(dst, dstBuf);
}
/* TurboJPEG 1.4.x: TJDecompressor::decodeYUV() byte destination */
diff --git a/turbojpeg.c b/turbojpeg.c
index e63d113..8260555 100644
--- a/turbojpeg.c
+++ b/turbojpeg.c
@@ -234,10 +234,10 @@
return -1;
}
-static int setCompDefaults(struct jpeg_compress_struct *cinfo, int pixelFormat,
- int subsamp, int jpegQual, int flags)
+static void setCompDefaults(struct jpeg_compress_struct *cinfo,
+ int pixelFormat, int subsamp, int jpegQual,
+ int flags)
{
- int retval = 0;
#ifndef NO_GETENV
char *env = NULL;
#endif
@@ -300,8 +300,6 @@
cinfo->comp_info[2].v_samp_factor = 1;
if (cinfo->num_components > 3)
cinfo->comp_info[3].v_samp_factor = tjMCUHeight[subsamp] / 8;
-
- return retval;
}
@@ -676,8 +674,7 @@
alloc = 0; *jpegSize = tjBufSize(width, height, jpegSubsamp);
}
jpeg_mem_dest_tj(cinfo, jpegBuf, jpegSize, alloc);
- if (setCompDefaults(cinfo, pixelFormat, jpegSubsamp, jpegQual, flags) == -1)
- return -1;
+ setCompDefaults(cinfo, pixelFormat, jpegSubsamp, jpegQual, flags);
jpeg_start_compress(cinfo, TRUE);
for (i = 0; i < height; i++) {
@@ -772,7 +769,7 @@
else if (flags & TJFLAG_FORCESSE2) putenv("JSIMD_FORCESSE2=1");
#endif
- if (setCompDefaults(cinfo, pixelFormat, subsamp, -1, flags) == -1) return -1;
+ setCompDefaults(cinfo, pixelFormat, subsamp, -1, flags);
/* Execute only the parts of jpeg_start_compress() that we need. If we
were to call the whole jpeg_start_compress() function, then it would try
@@ -986,8 +983,7 @@
alloc = 0; *jpegSize = tjBufSize(width, height, subsamp);
}
jpeg_mem_dest_tj(cinfo, jpegBuf, jpegSize, alloc);
- if (setCompDefaults(cinfo, TJPF_RGB, subsamp, jpegQual, flags) == -1)
- return -1;
+ setCompDefaults(cinfo, TJPF_RGB, subsamp, jpegQual, flags);
cinfo->raw_data_in = TRUE;
jpeg_start_compress(cinfo, TRUE);
diff --git a/usage.txt b/usage.txt
index 3cbdb53..f7fa3c0 100644
--- a/usage.txt
+++ b/usage.txt
@@ -50,11 +50,9 @@
This syntax works on all systems, so it is useful for scripts.
The currently supported image file formats are: PPM (PBMPLUS color format),
-PGM (PBMPLUS grayscale format), BMP, Targa, and RLE (Utah Raster Toolkit
-format). (RLE is supported only if the URT library is available, which it
-isn't on most non-Unix systems.) cjpeg recognizes the input image format
-automatically, with the exception of some Targa files. You have to tell djpeg
-which format to generate.
+PGM (PBMPLUS grayscale format), BMP, GIF, and Targa. cjpeg recognizes the
+input image format automatically, with the exception of some Targa files. You
+have to tell djpeg which format to generate.
JPEG files are in the defacto standard JFIF file format. There are other,
less widely used JPEG-based file formats, but we don't support them.
@@ -76,10 +74,10 @@
-grayscale Create monochrome JPEG file from color input.
Be sure to use this switch when compressing a grayscale
- BMP file, because cjpeg isn't bright enough to notice
- whether a BMP file uses only shades of gray. By
- saying -grayscale, you'll get a smaller JPEG file that
- takes less time to process.
+ BMP or GIF file, because cjpeg isn't bright enough to
+ notice whether a BMP or GIF file uses only shades of
+ gray. By saying -grayscale, you'll get a smaller JPEG
+ file that takes less time to process.
-rgb Create RGB JPEG file.
Using this switch suppresses the conversion from RGB
@@ -170,35 +168,43 @@
be unable to view an arithmetic coded JPEG file at
all.
- -dct int Use integer DCT method (default).
- -dct fast Use fast integer DCT (less accurate).
- In libjpeg-turbo, the fast method is generally about
- 5-15% faster than the int method when using the
- x86/x86-64 SIMD extensions (results may vary with other
- SIMD implementations, or when using libjpeg-turbo
- without SIMD extensions.) For quality levels of 90 and
- below, there should be little or no perceptible
- difference between the two algorithms. For quality
- levels above 90, however, the difference between
- the fast and the int methods becomes more pronounced.
- With quality=97, for instance, the fast method incurs
- generally about a 1-3 dB loss (in PSNR) relative to
- the int method, but this can be larger for some images.
- Do not use the fast method with quality levels above
- 97. The algorithm often degenerates at quality=98 and
- above and can actually produce a more lossy image than
- if lower quality levels had been used. Also, in
- libjpeg-turbo, the fast method is not fully accerated
- for quality levels above 97, so it will be slower than
- the int method.
- -dct float Use floating-point DCT method.
- The float method is mainly a legacy feature. It does
- not produce significantly more accurate results than
- the int method, and it is much slower. The float
- method may also give different results on different
- machines due to varying roundoff behavior, whereas the
- integer methods should give the same results on all
- machines.
+ -dct int Use accurate integer DCT method (default).
+ -dct fast Use less accurate integer DCT method [legacy feature].
+ When the Independent JPEG Group's software was first
+ released in 1991, the compression time for a
+ 1-megapixel JPEG image on a mainstream PC was measured
+ in minutes. Thus, the fast integer DCT algorithm
+ provided noticeable performance benefits. On modern
+ CPUs running libjpeg-turbo, however, the compression
+ time for a 1-megapixel JPEG image is measured in
+ milliseconds, and thus the performance benefits of the
+ fast algorithm are much less noticeable. On modern
+ x86/x86-64 CPUs that support AVX2 instructions, the
+ fast and int methods have similar performance. On
+ other types of CPUs, the fast method is generally about
+ 5-15% faster than the int method.
+
+ For quality levels of 90 and below, there should be
+ little or no perceptible quality difference between the
+ two algorithms. For quality levels above 90, however,
+ the difference between the fast and int methods becomes
+ more pronounced. With quality=97, for instance, the
+ fast method incurs generally about a 1-3 dB loss in
+ PSNR relative to the int method, but this can be larger
+ for some images. Do not use the fast method with
+ quality levels above 97. The algorithm often
+ degenerates at quality=98 and above and can actually
+ produce a more lossy image than if lower quality levels
+ had been used. Also, in libjpeg-turbo, the fast method
+ is not fully accelerated for quality levels above 97,
+ so it will be slower than the int method.
+ -dct float Use floating-point DCT method [legacy feature].
+ The float method does not produce significantly more
+ accurate results than the int method, and it is much
+ slower. The float method may also give different
+ results on different machines due to varying roundoff
+ behavior, whereas the integer methods should give the
+ same results on all machines.
-restart N Emit a JPEG restart marker every N MCU rows, or every
N MCU blocks if "B" is attached to the number.
@@ -290,10 +296,17 @@
is specified, or if the JPEG file is grayscale;
otherwise, 24-bit full-color format is emitted.
- -gif Select GIF output format. Since GIF does not support
- more than 256 colors, -colors 256 is assumed (unless
- you specify a smaller number of colors). If you
- specify -fast, the default number of colors is 216.
+ -gif Select GIF output format (LZW-compressed). Since GIF
+ does not support more than 256 colors, -colors 256 is
+ assumed (unless you specify a smaller number of
+ colors). If you specify -fast, the default number of
+ colors is 216.
+
+ -gif0 Select GIF output format (uncompressed). Since GIF
+ does not support more than 256 colors, -colors 256 is
+ assumed (unless you specify a smaller number of
+ colors). If you specify -fast, the default number of
+ colors is 216.
-os2 Select BMP output format (OS/2 1.x flavor). 8-bit
colormapped format is emitted if -colors or -grayscale
@@ -305,8 +318,6 @@
grayscale or if -grayscale is specified; otherwise
PPM is emitted.
- -rle Select RLE output format. (Requires URT library.)
-
-targa Select Targa output format. Grayscale format is
emitted if the JPEG file is grayscale or if
-grayscale is specified; otherwise, colormapped format
@@ -315,36 +326,45 @@
Switches for advanced users:
- -dct int Use integer DCT method (default).
- -dct fast Use fast integer DCT (less accurate).
- In libjpeg-turbo, the fast method is generally about
- 5-15% faster than the int method when using the
- x86/x86-64 SIMD extensions (results may vary with other
- SIMD implementations, or when using libjpeg-turbo
- without SIMD extensions.) If the JPEG image was
- compressed using a quality level of 85 or below, then
- there should be little or no perceptible difference
- between the two algorithms. When decompressing images
- that were compressed using quality levels above 85,
- however, the difference between the fast and int
- methods becomes more pronounced. With images
- compressed using quality=97, for instance, the fast
- method incurs generally about a 4-6 dB loss (in PSNR)
- relative to the int method, but this can be larger for
- some images. If you can avoid it, do not use the fast
- method when decompressing images that were compressed
- using quality levels above 97. The algorithm often
- degenerates for such images and can actually produce
- a more lossy output image than if the JPEG image had
- been compressed using lower quality levels.
- -dct float Use floating-point DCT method.
- The float method is mainly a legacy feature. It does
- not produce significantly more accurate results than
- the int method, and it is much slower. The float
- method may also give different results on different
- machines due to varying roundoff behavior, whereas the
- integer methods should give the same results on all
- machines.
+ -dct int Use accurate integer DCT method (default).
+ -dct fast Use less accurate integer DCT method [legacy feature].
+ When the Independent JPEG Group's software was first
+ released in 1991, the decompression time for a
+ 1-megapixel JPEG image on a mainstream PC was measured
+ in minutes. Thus, the fast integer DCT algorithm
+ provided noticeable performance benefits. On modern
+ CPUs running libjpeg-turbo, however, the decompression
+ time for a 1-megapixel JPEG image is measured in
+ milliseconds, and thus the performance benefits of the
+ fast algorithm are much less noticeable. On modern
+ x86/x86-64 CPUs that support AVX2 instructions, the
+ fast and int methods have similar performance. On
+ other types of CPUs, the fast method is generally about
+ 5-15% faster than the int method.
+
+ If the JPEG image was compressed using a quality level
+ of 85 or below, then there should be little or no
+ perceptible quality difference between the two
+ algorithms. When decompressing images that were
+ compressed using quality levels above 85, however, the
+ difference between the fast and int methods becomes
+ more pronounced. With images compressed using
+ quality=97, for instance, the fast method incurs
+ generally about a 4-6 dB loss in PSNR relative to the
+ int method, but this can be larger for some images. If
+ you can avoid it, do not use the fast method when
+ decompressing images that were compressed using quality
+ levels above 97. The algorithm often degenerates for
+ such images and can actually produce a more lossy
+ output image than if the JPEG image had been compressed
+ using lower quality levels.
+ -dct float Use floating-point DCT method [legacy feature].
+ The float method does not produce significantly more
+ accurate results than the int method, and it is much
+ slower. The float method may also give different
+ results on different machines due to varying roundoff
+ behavior, whereas the integer methods should give the
+ same results on all machines.
-dither fs Use Floyd-Steinberg dithering in color quantization.
-dither ordered Use ordered dithering in color quantization.
@@ -404,11 +424,6 @@
is often a lot more than it is on larger files. (At present, -optimize
mode is always selected when generating progressive JPEG files.)
-Support for GIF input files was removed in cjpeg v6b due to concerns over
-the Unisys LZW patent. Although this patent expired in 2006, cjpeg still
-lacks GIF support, for these historical reasons. (Conversion of GIF files to
-JPEG is usually a bad idea anyway.)
-
HINTS FOR DJPEG
@@ -423,10 +438,6 @@
much lower quality than the default behavior. "-dither none" may give
acceptable results in two-pass mode, but is seldom tolerable in one-pass mode.
-To avoid the Unisys LZW patent (now expired), djpeg produces uncompressed GIF
-files. These are larger than they should be, but are readable by standard GIF
-decoders.
-
HINTS FOR BOTH PROGRAMS
@@ -533,6 +544,43 @@
-crop WxH+X+Y Crop to a rectangular region of width W and height H,
starting at point X,Y.
+If W or H is larger than the width/height of the input image, then the output
+image is expanded in size, and the expanded region is filled in with zeros
+(neutral gray). Attaching an 'f' character ("flatten") to the width number
+will cause each block in the expanded region to be filled in with the DC
+coefficient of the nearest block in the input image rather than grayed out.
+Attaching an 'r' character ("reflect") to the width number will cause the
+expanded region to be filled in with repeated reflections of the input image
+rather than grayed out.
+
+A complementary lossless wipe option is provided to discard (gray out) data
+inside a given image region while losslessly preserving what is outside:
+ -wipe WxH+X+Y Wipe (gray out) a rectangular region of width W and
+ height H from the input image, starting at point X,Y.
+
+Attaching an 'f' character ("flatten") to the width number will cause the
+region to be filled with the average of adjacent blocks rather than grayed out.
+If the wipe region and the region outside the wipe region, when adjusted to the
+nearest iMCU boundary, form two horizontally adjacent rectangles, then
+attaching an 'r' character ("reflect") to the width number will cause the wipe
+region to be filled with repeated reflections of the outside region rather than
+grayed out.
+
+A lossless drop option is also provided, which allows another JPEG image to be
+inserted ("dropped") into the input image data at a given position, replacing
+the existing image data at that position:
+ -drop +X+Y filename Drop (insert) another image at point X,Y
+
+Both the input image and the drop image must have the same subsampling level.
+It is best if they also have the same quantization (quality.) Otherwise, the
+quantization of the output image will be adapted to accommodate the higher of
+the input image quality and the drop image quality. The trim option can be
+used with the drop option to requantize the drop image to match the input
+image. Note that a grayscale image can be dropped into a full-color image or
+vice versa, as long as the full-color image has no vertical subsampling. If
+the input image is grayscale and the drop image is full-color, then the
+chrominance channels from the drop image will be discarded.
+
Other not-strictly-lossless transformation switches are:
-grayscale Force grayscale output.
diff --git a/wrbmp.c b/wrbmp.c
index 239f64e..408a722 100644
--- a/wrbmp.c
+++ b/wrbmp.c
@@ -141,7 +141,6 @@
}
} else if (cinfo->out_color_space == JCS_CMYK) {
for (col = cinfo->output_width; col > 0; col--) {
- /* can omit GETJSAMPLE() safely */
JSAMPLE c = *inptr++, m = *inptr++, y = *inptr++, k = *inptr++;
cmyk_to_rgb(c, m, y, k, outptr + 2, outptr + 1, outptr);
outptr += 3;
@@ -153,7 +152,6 @@
register int ps = rgb_pixelsize[cinfo->out_color_space];
for (col = cinfo->output_width; col > 0; col--) {
- /* can omit GETJSAMPLE() safely */
outptr[0] = inptr[bindex];
outptr[1] = inptr[gindex];
outptr[2] = inptr[rindex];
@@ -372,18 +370,18 @@
if (cinfo->out_color_components == 3) {
/* Normal case with RGB colormap */
for (i = 0; i < num_colors; i++) {
- putc(GETJSAMPLE(colormap[2][i]), outfile);
- putc(GETJSAMPLE(colormap[1][i]), outfile);
- putc(GETJSAMPLE(colormap[0][i]), outfile);
+ putc(colormap[2][i], outfile);
+ putc(colormap[1][i], outfile);
+ putc(colormap[0][i], outfile);
if (map_entry_size == 4)
putc(0, outfile);
}
} else {
/* Grayscale colormap (only happens with grayscale quantization) */
for (i = 0; i < num_colors; i++) {
- putc(GETJSAMPLE(colormap[0][i]), outfile);
- putc(GETJSAMPLE(colormap[0][i]), outfile);
- putc(GETJSAMPLE(colormap[0][i]), outfile);
+ putc(colormap[0][i], outfile);
+ putc(colormap[0][i], outfile);
+ putc(colormap[0][i], outfile);
if (map_entry_size == 4)
putc(0, outfile);
}
@@ -438,7 +436,6 @@
JSAMPARRAY image_ptr;
register JSAMPROW data_ptr;
JDIMENSION row;
- register JDIMENSION col;
cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
if (dest->use_inversion_array) {
@@ -459,10 +456,7 @@
((j_common_ptr)cinfo, dest->whole_image, row - 1, (JDIMENSION)1,
FALSE);
data_ptr = image_ptr[0];
- for (col = dest->row_width; col > 0; col--) {
- putc(GETJSAMPLE(*data_ptr), outfile);
- data_ptr++;
- }
+ (void)JFWRITE(outfile, data_ptr, dest->row_width);
}
if (progress != NULL)
progress->completed_extra_passes++;
diff --git a/wrgif.c b/wrgif.c
index 1804e0b..82a2429 100644
--- a/wrgif.c
+++ b/wrgif.c
@@ -3,6 +3,7 @@
*
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2015-2019 by Guido Vollbeding.
* libjpeg-turbo Modifications:
* Copyright (C) 2015, 2017, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
@@ -10,12 +11,6 @@
*
* This file contains routines to write output images in GIF format.
*
- **************************************************************************
- * NOTE: to avoid entanglements with Unisys' patent on LZW compression, *
- * this code has been modified to output "uncompressed GIF" files. *
- * There is no trace of the LZW algorithm in this file. *
- **************************************************************************
- *
* These routines may need modification for non-Unix environments or
* specialized applications. As they stand, they assume output to
* an ordinary stdio stream.
@@ -33,11 +28,6 @@
* copyright notice and this permission notice appear in supporting
* documentation. This software is provided "as is" without express or
* implied warranty.
- *
- * We are also required to state that
- * "The Graphics Interchange Format(c) is the Copyright property of
- * CompuServe Incorporated. GIF(sm) is a Service Mark property of
- * CompuServe Incorporated."
*/
#include "cdjpeg.h" /* Common decls for cjpeg/djpeg applications */
@@ -45,6 +35,37 @@
#ifdef GIF_SUPPORTED
+#define MAX_LZW_BITS 12 /* maximum LZW code size (4096 symbols) */
+
+typedef INT16 code_int; /* must hold -1 .. 2**MAX_LZW_BITS */
+
+#define LZW_TABLE_SIZE ((code_int)1 << MAX_LZW_BITS)
+
+#define HSIZE 5003 /* hash table size for 80% occupancy */
+
+typedef int hash_int; /* must hold -2*HSIZE..2*HSIZE */
+
+#define MAXCODE(n_bits) (((code_int)1 << (n_bits)) - 1)
+
+
+/*
+ * The LZW hash table consists of two parallel arrays:
+ * hash_code[i] code of symbol in slot i, or 0 if empty slot
+ * hash_value[i] symbol's value; undefined if empty slot
+ * where slot values (i) range from 0 to HSIZE-1. The symbol value is
+ * its prefix symbol's code concatenated with its suffix character.
+ *
+ * Algorithm: use open addressing double hashing (no chaining) on the
+ * prefix code / suffix character combination. We do a variant of Knuth's
+ * algorithm D (vol. 3, sec. 6.4) along with G. Knott's relatively-prime
+ * secondary probe.
+ */
+
+typedef int hash_entry; /* must hold (code_int << 8) | byte */
+
+#define HASH_ENTRY(prefix, suffix) ((((hash_entry)(prefix)) << 8) | (suffix))
+
+
/* Private version of data destination object */
typedef struct {
@@ -54,14 +75,24 @@
/* State for packing variable-width codes into a bitstream */
int n_bits; /* current number of bits/code */
- int maxcode; /* maximum code, given n_bits */
- long cur_accum; /* holds bits not yet output */
+ code_int maxcode; /* maximum code, given n_bits */
+ int init_bits; /* initial n_bits ... restored after clear */
+ int cur_accum; /* holds bits not yet output */
int cur_bits; /* # of bits in cur_accum */
+ /* LZW string construction */
+ code_int waiting_code; /* symbol not yet output; may be extendable */
+ boolean first_byte; /* if TRUE, waiting_code is not valid */
+
/* State for GIF code assignment */
- int ClearCode; /* clear code (doesn't change) */
- int EOFCode; /* EOF code (ditto) */
- int code_counter; /* counts output symbols */
+ code_int ClearCode; /* clear code (doesn't change) */
+ code_int EOFCode; /* EOF code (ditto) */
+ code_int free_code; /* LZW: first not-yet-used symbol code */
+ code_int code_counter; /* not LZW: counts output symbols */
+
+ /* LZW hash table */
+ code_int *hash_code; /* => hash table of symbol codes */
+ hash_entry *hash_value; /* => hash table of symbol values */
/* GIF data packet construction buffer */
int bytesinpkt; /* # of bytes in current packet */
@@ -71,9 +102,6 @@
typedef gif_dest_struct *gif_dest_ptr;
-/* Largest value that will fit in N bits */
-#define MAXCODE(n_bits) ((1 << (n_bits)) - 1)
-
/*
* Routines to package finished data bytes into GIF data blocks.
@@ -105,7 +133,7 @@
/* Routine to convert variable-width codes into a byte stream */
LOCAL(void)
-output(gif_dest_ptr dinfo, int code)
+output(gif_dest_ptr dinfo, code_int code)
/* Emit a code of n_bits bits */
/* Uses cur_accum and cur_bits to reblock into 8-bit bytes */
{
@@ -117,74 +145,76 @@
dinfo->cur_accum >>= 8;
dinfo->cur_bits -= 8;
}
+
+ /*
+ * If the next entry is going to be too big for the code size,
+ * then increase it, if possible. We do this here to ensure
+ * that it's done in sync with the decoder's codesize increases.
+ */
+ if (dinfo->free_code > dinfo->maxcode) {
+ dinfo->n_bits++;
+ if (dinfo->n_bits == MAX_LZW_BITS)
+ dinfo->maxcode = LZW_TABLE_SIZE; /* free_code will never exceed this */
+ else
+ dinfo->maxcode = MAXCODE(dinfo->n_bits);
+ }
}
-/* The pseudo-compression algorithm.
- *
- * In this module we simply output each pixel value as a separate symbol;
- * thus, no compression occurs. In fact, there is expansion of one bit per
- * pixel, because we use a symbol width one bit wider than the pixel width.
- *
- * GIF ordinarily uses variable-width symbols, and the decoder will expect
- * to ratchet up the symbol width after a fixed number of symbols.
- * To simplify the logic and keep the expansion penalty down, we emit a
- * GIF Clear code to reset the decoder just before the width would ratchet up.
- * Thus, all the symbols in the output file will have the same bit width.
- * Note that emitting the Clear codes at the right times is a mere matter of
- * counting output symbols and is in no way dependent on the LZW patent.
- *
- * With a small basic pixel width (low color count), Clear codes will be
- * needed very frequently, causing the file to expand even more. So this
- * simplistic approach wouldn't work too well on bilevel images, for example.
- * But for output of JPEG conversions the pixel width will usually be 8 bits
- * (129 to 256 colors), so the overhead added by Clear symbols is only about
- * one symbol in every 256.
- */
+/* Compression initialization & termination */
+
+
+LOCAL(void)
+clear_hash(gif_dest_ptr dinfo)
+/* Fill the hash table with empty entries */
+{
+ /* It's sufficient to zero hash_code[] */
+ MEMZERO(dinfo->hash_code, HSIZE * sizeof(code_int));
+}
+
+
+LOCAL(void)
+clear_block(gif_dest_ptr dinfo)
+/* Reset compressor and issue a Clear code */
+{
+ clear_hash(dinfo); /* delete all the symbols */
+ dinfo->free_code = dinfo->ClearCode + 2;
+ output(dinfo, dinfo->ClearCode); /* inform decoder */
+ dinfo->n_bits = dinfo->init_bits; /* reset code size */
+ dinfo->maxcode = MAXCODE(dinfo->n_bits);
+}
+
LOCAL(void)
compress_init(gif_dest_ptr dinfo, int i_bits)
-/* Initialize pseudo-compressor */
+/* Initialize compressor */
{
/* init all the state variables */
- dinfo->n_bits = i_bits;
+ dinfo->n_bits = dinfo->init_bits = i_bits;
dinfo->maxcode = MAXCODE(dinfo->n_bits);
- dinfo->ClearCode = (1 << (i_bits - 1));
+ dinfo->ClearCode = ((code_int) 1 << (i_bits - 1));
dinfo->EOFCode = dinfo->ClearCode + 1;
- dinfo->code_counter = dinfo->ClearCode + 2;
+ dinfo->code_counter = dinfo->free_code = dinfo->ClearCode + 2;
+ dinfo->first_byte = TRUE; /* no waiting symbol yet */
/* init output buffering vars */
dinfo->bytesinpkt = 0;
dinfo->cur_accum = 0;
dinfo->cur_bits = 0;
+ /* clear hash table */
+ if (dinfo->hash_code != NULL)
+ clear_hash(dinfo);
/* GIF specifies an initial Clear code */
output(dinfo, dinfo->ClearCode);
}
LOCAL(void)
-compress_pixel(gif_dest_ptr dinfo, int c)
-/* Accept and "compress" one pixel value.
- * The given value must be less than n_bits wide.
- */
-{
- /* Output the given pixel value as a symbol. */
- output(dinfo, c);
- /* Issue Clear codes often enough to keep the reader from ratcheting up
- * its symbol size.
- */
- if (dinfo->code_counter < dinfo->maxcode) {
- dinfo->code_counter++;
- } else {
- output(dinfo, dinfo->ClearCode);
- dinfo->code_counter = dinfo->ClearCode + 2; /* reset the counter */
- }
-}
-
-
-LOCAL(void)
compress_term(gif_dest_ptr dinfo)
/* Clean up at end */
{
+ /* Flush out the buffered LZW code */
+ if (!dinfo->first_byte)
+ output(dinfo, dinfo->waiting_code);
/* Send an EOF code */
output(dinfo, dinfo->EOFCode);
/* Flush the bit-packing buffer */
@@ -221,7 +251,7 @@
LOCAL(void)
emit_header(gif_dest_ptr dinfo, int num_colors, JSAMPARRAY colormap)
/* Output the GIF file header, including color map */
-/* If colormap==NULL, synthesize a grayscale colormap */
+/* If colormap == NULL, synthesize a grayscale colormap */
{
int BitsPerPixel, ColorMapSize, InitCodeSize, FlagByte;
int cshift = dinfo->cinfo->data_precision - 8;
@@ -265,12 +295,12 @@
if (colormap != NULL) {
if (dinfo->cinfo->out_color_space == JCS_RGB) {
/* Normal case: RGB color map */
- putc(GETJSAMPLE(colormap[0][i]) >> cshift, dinfo->pub.output_file);
- putc(GETJSAMPLE(colormap[1][i]) >> cshift, dinfo->pub.output_file);
- putc(GETJSAMPLE(colormap[2][i]) >> cshift, dinfo->pub.output_file);
+ putc(colormap[0][i] >> cshift, dinfo->pub.output_file);
+ putc(colormap[1][i] >> cshift, dinfo->pub.output_file);
+ putc(colormap[2][i] >> cshift, dinfo->pub.output_file);
} else {
/* Grayscale "color map": possible if quantizing grayscale image */
- put_3bytes(dinfo, GETJSAMPLE(colormap[0][i]) >> cshift);
+ put_3bytes(dinfo, colormap[0][i] >> cshift);
}
} else {
/* Create a grayscale map of num_colors values, range 0..255 */
@@ -278,7 +308,7 @@
}
} else {
/* fill out the map to a power of 2 */
- put_3bytes(dinfo, 0);
+ put_3bytes(dinfo, CENTERJSAMPLE >> cshift);
}
}
/* Write image separator and Image Descriptor */
@@ -292,7 +322,7 @@
/* Write Initial Code Size byte */
putc(InitCodeSize, dinfo->pub.output_file);
- /* Initialize for "compression" of image data */
+ /* Initialize for compression of image data */
compress_init(dinfo, InitCodeSize + 1);
}
@@ -318,17 +348,139 @@
* In this module rows_supplied will always be 1.
*/
+
+/*
+ * The LZW algorithm proper
+ */
+
METHODDEF(void)
-put_pixel_rows(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
- JDIMENSION rows_supplied)
+put_LZW_pixel_rows(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+ JDIMENSION rows_supplied)
{
gif_dest_ptr dest = (gif_dest_ptr)dinfo;
register JSAMPROW ptr;
register JDIMENSION col;
+ code_int c;
+ register hash_int i;
+ register hash_int disp;
+ register hash_entry probe_value;
ptr = dest->pub.buffer[0];
for (col = cinfo->output_width; col > 0; col--) {
- compress_pixel(dest, GETJSAMPLE(*ptr++));
+ /* Accept and compress one 8-bit byte */
+ c = (code_int)(*ptr++);
+
+ if (dest->first_byte) { /* need to initialize waiting_code */
+ dest->waiting_code = c;
+ dest->first_byte = FALSE;
+ continue;
+ }
+
+ /* Probe hash table to see if a symbol exists for
+ * waiting_code followed by c.
+ * If so, replace waiting_code by that symbol and continue.
+ */
+ i = ((hash_int)c << (MAX_LZW_BITS - 8)) + dest->waiting_code;
+ /* i is less than twice 2**MAX_LZW_BITS, therefore less than twice HSIZE */
+ if (i >= HSIZE)
+ i -= HSIZE;
+
+ probe_value = HASH_ENTRY(dest->waiting_code, c);
+
+ if (dest->hash_code[i] == 0) {
+ /* hit empty slot; desired symbol not in table */
+ output(dest, dest->waiting_code);
+ if (dest->free_code < LZW_TABLE_SIZE) {
+ dest->hash_code[i] = dest->free_code++; /* add symbol to hashtable */
+ dest->hash_value[i] = probe_value;
+ } else
+ clear_block(dest);
+ dest->waiting_code = c;
+ continue;
+ }
+ if (dest->hash_value[i] == probe_value) {
+ dest->waiting_code = dest->hash_code[i];
+ continue;
+ }
+
+ if (i == 0) /* secondary hash (after G. Knott) */
+ disp = 1;
+ else
+ disp = HSIZE - i;
+ for (;;) {
+ i -= disp;
+ if (i < 0)
+ i += HSIZE;
+ if (dest->hash_code[i] == 0) {
+ /* hit empty slot; desired symbol not in table */
+ output(dest, dest->waiting_code);
+ if (dest->free_code < LZW_TABLE_SIZE) {
+ dest->hash_code[i] = dest->free_code++; /* add symbol to hashtable */
+ dest->hash_value[i] = probe_value;
+ } else
+ clear_block(dest);
+ dest->waiting_code = c;
+ break;
+ }
+ if (dest->hash_value[i] == probe_value) {
+ dest->waiting_code = dest->hash_code[i];
+ break;
+ }
+ }
+ }
+}
+
+
+/*
+ * The pseudo-compression algorithm.
+ *
+ * In this version we simply output each pixel value as a separate symbol;
+ * thus, no compression occurs. In fact, there is expansion of one bit per
+ * pixel, because we use a symbol width one bit wider than the pixel width.
+ *
+ * GIF ordinarily uses variable-width symbols, and the decoder will expect
+ * to ratchet up the symbol width after a fixed number of symbols.
+ * To simplify the logic and keep the expansion penalty down, we emit a
+ * GIF Clear code to reset the decoder just before the width would ratchet up.
+ * Thus, all the symbols in the output file will have the same bit width.
+ * Note that emitting the Clear codes at the right times is a mere matter of
+ * counting output symbols and is in no way dependent on the LZW algorithm.
+ *
+ * With a small basic pixel width (low color count), Clear codes will be
+ * needed very frequently, causing the file to expand even more. So this
+ * simplistic approach wouldn't work too well on bilevel images, for example.
+ * But for output of JPEG conversions the pixel width will usually be 8 bits
+ * (129 to 256 colors), so the overhead added by Clear symbols is only about
+ * one symbol in every 256.
+ */
+
+METHODDEF(void)
+put_raw_pixel_rows(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
+ JDIMENSION rows_supplied)
+{
+ gif_dest_ptr dest = (gif_dest_ptr)dinfo;
+ register JSAMPROW ptr;
+ register JDIMENSION col;
+ code_int c;
+
+ ptr = dest->pub.buffer[0];
+ for (col = cinfo->output_width; col > 0; col--) {
+ c = (code_int)(*ptr++);
+ /* Accept and output one pixel value.
+ * The given value must be less than n_bits wide.
+ */
+
+ /* Output the given pixel value as a symbol. */
+ output(dest, c);
+ /* Issue Clear codes often enough to keep the reader from ratcheting up
+ * its symbol size.
+ */
+ if (dest->code_counter < dest->maxcode) {
+ dest->code_counter++;
+ } else {
+ output(dest, dest->ClearCode);
+ dest->code_counter = dest->ClearCode + 2; /* reset the counter */
+ }
}
}
@@ -342,7 +494,7 @@
{
gif_dest_ptr dest = (gif_dest_ptr)dinfo;
- /* Flush "compression" mechanism */
+ /* Flush compression mechanism */
compress_term(dest);
/* Write a zero-length data block to end the series */
putc(0, dest->pub.output_file);
@@ -370,7 +522,7 @@
*/
GLOBAL(djpeg_dest_ptr)
-jinit_write_gif(j_decompress_ptr cinfo)
+jinit_write_gif(j_decompress_ptr cinfo, boolean is_lzw)
{
gif_dest_ptr dest;
@@ -380,7 +532,6 @@
sizeof(gif_dest_struct));
dest->cinfo = cinfo; /* make back link for subroutines */
dest->pub.start_output = start_output_gif;
- dest->pub.put_pixel_rows = put_pixel_rows;
dest->pub.finish_output = finish_output_gif;
dest->pub.calc_buffer_dimensions = calc_buffer_dimensions_gif;
@@ -407,6 +558,22 @@
((j_common_ptr)cinfo, JPOOL_IMAGE, cinfo->output_width, (JDIMENSION)1);
dest->pub.buffer_height = 1;
+ if (is_lzw) {
+ dest->pub.put_pixel_rows = put_LZW_pixel_rows;
+ /* Allocate space for hash table */
+ dest->hash_code = (code_int *)
+ (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ HSIZE * sizeof(code_int));
+ dest->hash_value = (hash_entry *)
+ (*cinfo->mem->alloc_large) ((j_common_ptr)cinfo, JPOOL_IMAGE,
+ HSIZE * sizeof(hash_entry));
+ } else {
+ dest->pub.put_pixel_rows = put_raw_pixel_rows;
+ /* Mark tables unused */
+ dest->hash_code = NULL;
+ dest->hash_value = NULL;
+ }
+
return (djpeg_dest_ptr)dest;
}
diff --git a/wrppm.c b/wrppm.c
index 69f91e8..3081ec3 100644
--- a/wrppm.c
+++ b/wrppm.c
@@ -5,7 +5,7 @@
* Copyright (C) 1991-1996, Thomas G. Lane.
* Modified 2009 by Guido Vollbeding.
* libjpeg-turbo Modifications:
- * Copyright (C) 2017, 2019, D. R. Commander.
+ * Copyright (C) 2017, 2019-2020, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -108,17 +108,17 @@
ppm_dest_ptr dest = (ppm_dest_ptr)dinfo;
register char *bufferptr;
register JSAMPROW ptr;
-#if BITS_IN_JSAMPLE != 8 || (!defined(HAVE_UNSIGNED_CHAR) && !defined(__CHAR_UNSIGNED__))
+#if BITS_IN_JSAMPLE != 8
register JDIMENSION col;
#endif
ptr = dest->pub.buffer[0];
bufferptr = dest->iobuffer;
-#if BITS_IN_JSAMPLE == 8 && (defined(HAVE_UNSIGNED_CHAR) || defined(__CHAR_UNSIGNED__))
+#if BITS_IN_JSAMPLE == 8
MEMCOPY(bufferptr, ptr, dest->samples_per_row);
#else
for (col = dest->samples_per_row; col > 0; col--) {
- PUTPPMSAMPLE(bufferptr, GETJSAMPLE(*ptr++));
+ PUTPPMSAMPLE(bufferptr, *ptr++);
}
#endif
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
@@ -200,10 +200,10 @@
ptr = dest->pub.buffer[0];
bufferptr = dest->iobuffer;
for (col = cinfo->output_width; col > 0; col--) {
- pixval = GETJSAMPLE(*ptr++);
- PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map0[pixval]));
- PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map1[pixval]));
- PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map2[pixval]));
+ pixval = *ptr++;
+ PUTPPMSAMPLE(bufferptr, color_map0[pixval]);
+ PUTPPMSAMPLE(bufferptr, color_map1[pixval]);
+ PUTPPMSAMPLE(bufferptr, color_map2[pixval]);
}
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
}
@@ -222,7 +222,7 @@
ptr = dest->pub.buffer[0];
bufferptr = dest->iobuffer;
for (col = cinfo->output_width; col > 0; col--) {
- PUTPPMSAMPLE(bufferptr, GETJSAMPLE(color_map[GETJSAMPLE(*ptr++)]));
+ PUTPPMSAMPLE(bufferptr, color_map[*ptr++]);
}
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
}
@@ -326,11 +326,12 @@
if (cinfo->quantize_colors || BITS_IN_JSAMPLE != 8 ||
sizeof(JSAMPLE) != sizeof(char) ||
- (cinfo->out_color_space != JCS_EXT_RGB
#if RGB_RED == 0 && RGB_GREEN == 1 && RGB_BLUE == 2 && RGB_PIXELSIZE == 3
- && cinfo->out_color_space != JCS_RGB
+ (cinfo->out_color_space != JCS_EXT_RGB &&
+ cinfo->out_color_space != JCS_RGB)) {
+#else
+ cinfo->out_color_space != JCS_EXT_RGB) {
#endif
- )) {
/* When quantizing, we need an output buffer for colormap indexes
* that's separate from the physical I/O buffer. We also need a
* separate buffer if pixel format translation must take place.
diff --git a/wrrle.c b/wrrle.c
deleted file mode 100644
index 5c98ec0..0000000
--- a/wrrle.c
+++ /dev/null
@@ -1,309 +0,0 @@
-/*
- * wrrle.c
- *
- * This file was part of the Independent JPEG Group's software:
- * Copyright (C) 1991-1996, Thomas G. Lane.
- * libjpeg-turbo Modifications:
- * Copyright (C) 2017, D. R. Commander.
- * For conditions of distribution and use, see the accompanying README.ijg
- * file.
- *
- * This file contains routines to write output images in RLE format.
- * The Utah Raster Toolkit library is required (version 3.1 or later).
- *
- * These routines may need modification for non-Unix environments or
- * specialized applications. As they stand, they assume output to
- * an ordinary stdio stream.
- *
- * Based on code contributed by Mike Lijewski,
- * with updates from Robert Hutchinson.
- */
-
-#include "cdjpeg.h" /* Common decls for cjpeg/djpeg applications */
-
-#ifdef RLE_SUPPORTED
-
-/* rle.h is provided by the Utah Raster Toolkit. */
-
-#include <rle.h>
-
-/*
- * We assume that JSAMPLE has the same representation as rle_pixel,
- * to wit, "unsigned char". Hence we can't cope with 12- or 16-bit samples.
- */
-
-#if BITS_IN_JSAMPLE != 8
- Sorry, this code only copes with 8-bit JSAMPLEs. /* deliberate syntax err */
-#endif
-
-
-/*
- * Since RLE stores scanlines bottom-to-top, we have to invert the image
- * from JPEG's top-to-bottom order. To do this, we save the outgoing data
- * in a virtual array during put_pixel_row calls, then actually emit the
- * RLE file during finish_output.
- */
-
-
-/*
- * For now, if we emit an RLE color map then it is always 256 entries long,
- * though not all of the entries need be used.
- */
-
-#define CMAPBITS 8
-#define CMAPLENGTH (1 << (CMAPBITS))
-
-typedef struct {
- struct djpeg_dest_struct pub; /* public fields */
-
- jvirt_sarray_ptr image; /* virtual array to store the output image */
- rle_map *colormap; /* RLE-style color map, or NULL if none */
- rle_pixel **rle_row; /* To pass rows to rle_putrow() */
-
-} rle_dest_struct;
-
-typedef rle_dest_struct *rle_dest_ptr;
-
-/* Forward declarations */
-METHODDEF(void) rle_put_pixel_rows(j_decompress_ptr cinfo,
- djpeg_dest_ptr dinfo,
- JDIMENSION rows_supplied);
-
-
-/*
- * Write the file header.
- *
- * In this module it's easier to wait till finish_output to write anything.
- */
-
-METHODDEF(void)
-start_output_rle(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
-{
- rle_dest_ptr dest = (rle_dest_ptr)dinfo;
- size_t cmapsize;
- int i, ci;
-#ifdef PROGRESS_REPORT
- cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
-#endif
-
- /*
- * Make sure the image can be stored in RLE format.
- *
- * - RLE stores image dimensions as *signed* 16 bit integers. JPEG
- * uses unsigned, so we have to check the width.
- *
- * - Colorspace is expected to be grayscale or RGB.
- *
- * - The number of channels (components) is expected to be 1 (grayscale/
- * pseudocolor) or 3 (truecolor/directcolor).
- * (could be 2 or 4 if using an alpha channel, but we aren't)
- */
-
- if (cinfo->output_width > 32767 || cinfo->output_height > 32767)
- ERREXIT2(cinfo, JERR_RLE_DIMENSIONS, cinfo->output_width,
- cinfo->output_height);
-
- if (cinfo->out_color_space != JCS_GRAYSCALE &&
- cinfo->out_color_space != JCS_RGB)
- ERREXIT(cinfo, JERR_RLE_COLORSPACE);
-
- if (cinfo->output_components != 1 && cinfo->output_components != 3)
- ERREXIT1(cinfo, JERR_RLE_TOOMANYCHANNELS, cinfo->num_components);
-
- /* Convert colormap, if any, to RLE format. */
-
- dest->colormap = NULL;
-
- if (cinfo->quantize_colors) {
- /* Allocate storage for RLE-style cmap, zero any extra entries */
- cmapsize = cinfo->out_color_components * CMAPLENGTH * sizeof(rle_map);
- dest->colormap = (rle_map *)(*cinfo->mem->alloc_small)
- ((j_common_ptr)cinfo, JPOOL_IMAGE, cmapsize);
- MEMZERO(dest->colormap, cmapsize);
-
- /* Save away data in RLE format --- note 8-bit left shift! */
- /* Shifting would need adjustment for JSAMPLEs wider than 8 bits. */
- for (ci = 0; ci < cinfo->out_color_components; ci++) {
- for (i = 0; i < cinfo->actual_number_of_colors; i++) {
- dest->colormap[ci * CMAPLENGTH + i] =
- GETJSAMPLE(cinfo->colormap[ci][i]) << 8;
- }
- }
- }
-
- /* Set the output buffer to the first row */
- dest->pub.buffer = (*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, dest->image, (JDIMENSION)0, (JDIMENSION)1, TRUE);
- dest->pub.buffer_height = 1;
-
- dest->pub.put_pixel_rows = rle_put_pixel_rows;
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->total_extra_passes++; /* count file writing as separate pass */
- }
-#endif
-}
-
-
-/*
- * Write some pixel data.
- *
- * This routine just saves the data away in a virtual array.
- */
-
-METHODDEF(void)
-rle_put_pixel_rows(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo,
- JDIMENSION rows_supplied)
-{
- rle_dest_ptr dest = (rle_dest_ptr)dinfo;
-
- if (cinfo->output_scanline < cinfo->output_height) {
- dest->pub.buffer = (*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, dest->image,
- cinfo->output_scanline, (JDIMENSION)1, TRUE);
- }
-}
-
-/*
- * Finish up at the end of the file.
- *
- * Here is where we really output the RLE file.
- */
-
-METHODDEF(void)
-finish_output_rle(j_decompress_ptr cinfo, djpeg_dest_ptr dinfo)
-{
- rle_dest_ptr dest = (rle_dest_ptr)dinfo;
- rle_hdr header; /* Output file information */
- rle_pixel **rle_row, *red, *green, *blue;
- JSAMPROW output_row;
- char cmapcomment[80];
- int row, col;
- int ci;
-#ifdef PROGRESS_REPORT
- cd_progress_ptr progress = (cd_progress_ptr)cinfo->progress;
-#endif
-
- /* Initialize the header info */
- header = *rle_hdr_init(NULL);
- header.rle_file = dest->pub.output_file;
- header.xmin = 0;
- header.xmax = cinfo->output_width - 1;
- header.ymin = 0;
- header.ymax = cinfo->output_height - 1;
- header.alpha = 0;
- header.ncolors = cinfo->output_components;
- for (ci = 0; ci < cinfo->output_components; ci++) {
- RLE_SET_BIT(header, ci);
- }
- if (cinfo->quantize_colors) {
- header.ncmap = cinfo->out_color_components;
- header.cmaplen = CMAPBITS;
- header.cmap = dest->colormap;
- /* Add a comment to the output image with the true colormap length. */
- sprintf(cmapcomment, "color_map_length=%d",
- cinfo->actual_number_of_colors);
- rle_putcom(cmapcomment, &header);
- }
-
- /* Emit the RLE header and color map (if any) */
- rle_put_setup(&header);
-
- /* Now output the RLE data from our virtual array.
- * We assume here that rle_pixel is represented the same as JSAMPLE.
- */
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_limit = cinfo->output_height;
- progress->pub.pass_counter = 0;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
-
- if (cinfo->output_components == 1) {
- for (row = cinfo->output_height - 1; row >= 0; row--) {
- rle_row = (rle_pixel **)(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, dest->image,
- (JDIMENSION)row, (JDIMENSION)1, FALSE);
- rle_putrow(rle_row, (int)cinfo->output_width, &header);
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_counter++;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
- }
- } else {
- for (row = cinfo->output_height - 1; row >= 0; row--) {
- rle_row = (rle_pixel **)dest->rle_row;
- output_row = *(*cinfo->mem->access_virt_sarray)
- ((j_common_ptr)cinfo, dest->image,
- (JDIMENSION)row, (JDIMENSION)1, FALSE);
- red = rle_row[0];
- green = rle_row[1];
- blue = rle_row[2];
- for (col = cinfo->output_width; col > 0; col--) {
- *red++ = GETJSAMPLE(*output_row++);
- *green++ = GETJSAMPLE(*output_row++);
- *blue++ = GETJSAMPLE(*output_row++);
- }
- rle_putrow(rle_row, (int)cinfo->output_width, &header);
-#ifdef PROGRESS_REPORT
- if (progress != NULL) {
- progress->pub.pass_counter++;
- (*progress->pub.progress_monitor) ((j_common_ptr)cinfo);
- }
-#endif
- }
- }
-
-#ifdef PROGRESS_REPORT
- if (progress != NULL)
- progress->completed_extra_passes++;
-#endif
-
- /* Emit file trailer */
- rle_puteof(&header);
- fflush(dest->pub.output_file);
- if (ferror(dest->pub.output_file))
- ERREXIT(cinfo, JERR_FILE_WRITE);
-}
-
-
-/*
- * The module selection routine for RLE format output.
- */
-
-GLOBAL(djpeg_dest_ptr)
-jinit_write_rle(j_decompress_ptr cinfo)
-{
- rle_dest_ptr dest;
-
- /* Create module interface object, fill in method pointers */
- dest = (rle_dest_ptr)
- (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE,
- sizeof(rle_dest_struct));
- dest->pub.start_output = start_output_rle;
- dest->pub.finish_output = finish_output_rle;
- dest->pub.calc_buffer_dimensions = NULL;
-
- /* Calculate output image dimensions so we can allocate space */
- jpeg_calc_output_dimensions(cinfo);
-
- /* Allocate a work array for output to the RLE library. */
- dest->rle_row = (*cinfo->mem->alloc_sarray)
- ((j_common_ptr)cinfo, JPOOL_IMAGE,
- cinfo->output_width, (JDIMENSION)cinfo->output_components);
-
- /* Allocate a virtual array to hold the image. */
- dest->image = (*cinfo->mem->request_virt_sarray)
- ((j_common_ptr)cinfo, JPOOL_IMAGE, FALSE,
- (JDIMENSION)(cinfo->output_width * cinfo->output_components),
- cinfo->output_height, (JDIMENSION)1);
-
- return (djpeg_dest_ptr)dest;
-}
-
-#endif /* RLE_SUPPORTED */
diff --git a/wrtarga.c b/wrtarga.c
index 9dfa920..7a654ff 100644
--- a/wrtarga.c
+++ b/wrtarga.c
@@ -4,7 +4,7 @@
* This file was part of the Independent JPEG Group's software:
* Copyright (C) 1991-1996, Thomas G. Lane.
* libjpeg-turbo Modifications:
- * Copyright (C) 2017, D. R. Commander.
+ * Copyright (C) 2017, 2019, D. R. Commander.
* For conditions of distribution and use, see the accompanying README.ijg
* file.
*
@@ -102,9 +102,9 @@
inptr = dest->pub.buffer[0];
outptr = dest->iobuffer;
for (col = cinfo->output_width; col > 0; col--) {
- outptr[0] = (char)GETJSAMPLE(inptr[2]); /* RGB to BGR order */
- outptr[1] = (char)GETJSAMPLE(inptr[1]);
- outptr[2] = (char)GETJSAMPLE(inptr[0]);
+ outptr[0] = inptr[2]; /* RGB to BGR order */
+ outptr[1] = inptr[1];
+ outptr[2] = inptr[0];
inptr += 3, outptr += 3;
}
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
@@ -118,13 +118,10 @@
tga_dest_ptr dest = (tga_dest_ptr)dinfo;
register JSAMPROW inptr;
register char *outptr;
- register JDIMENSION col;
inptr = dest->pub.buffer[0];
outptr = dest->iobuffer;
- for (col = cinfo->output_width; col > 0; col--) {
- *outptr++ = (char)GETJSAMPLE(*inptr++);
- }
+ MEMCOPY(outptr, inptr, cinfo->output_width);
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
}
@@ -147,7 +144,7 @@
inptr = dest->pub.buffer[0];
outptr = dest->iobuffer;
for (col = cinfo->output_width; col > 0; col--) {
- *outptr++ = (char)GETJSAMPLE(color_map0[GETJSAMPLE(*inptr++)]);
+ *outptr++ = color_map0[*inptr++];
}
(void)JFWRITE(dest->pub.output_file, dest->iobuffer, dest->buffer_width);
}
@@ -182,9 +179,9 @@
/* Write the colormap. Note Targa uses BGR byte order */
outfile = dest->pub.output_file;
for (i = 0; i < num_colors; i++) {
- putc(GETJSAMPLE(cinfo->colormap[2][i]), outfile);
- putc(GETJSAMPLE(cinfo->colormap[1][i]), outfile);
- putc(GETJSAMPLE(cinfo->colormap[0][i]), outfile);
+ putc(cinfo->colormap[2][i], outfile);
+ putc(cinfo->colormap[1][i], outfile);
+ putc(cinfo->colormap[0][i], outfile);
}
dest->pub.put_pixel_rows = put_gray_rows;
} else {
