blob: a9d338a2a0892aeadc2bfa4a6e09e4617e4c419e [file] [log] [blame]
/* stb_image - v2.08 - public domain image loader - http://nothings.org/stb_image.h
no warranty implied; use at your own risk
Do this:
#define STB_IMAGE_IMPLEMENTATION
before you include this file in *one* C or C++ file to create the implementation.
// i.e. it should look like this:
#include ...
#include ...
#include ...
#define STB_IMAGE_IMPLEMENTATION
#include "stb_image.h"
You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
QUICK NOTES:
Primarily of interest to game developers and other people who can
avoid problematic images and only need the trivial interface
JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
PNG 1/2/4/8-bit-per-channel (16 bpc not supported)
TGA (not sure what subset, if a subset)
BMP non-1bpp, non-RLE
PSD (composited view only, no extra channels, 8/16 bit-per-channel)
GIF (*comp always reports as 4-channel)
HDR (radiance rgbE format)
PIC (Softimage PIC)
PNM (PPM and PGM binary only)
Animated GIF still needs a proper API, but here's one way to do it:
http://gist.github.com/urraka/685d9a6340b26b830d49
- decode from memory or through FILE (define STBI_NO_STDIO to remove code)
- decode from arbitrary I/O callbacks
- SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
Full documentation under "DOCUMENTATION" below.
Revision 2.00 release notes:
- Progressive JPEG is now supported.
- PPM and PGM binary formats are now supported, thanks to Ken Miller.
- x86 platforms now make use of SSE2 SIMD instructions for
JPEG decoding, and ARM platforms can use NEON SIMD if requested.
This work was done by Fabian "ryg" Giesen. SSE2 is used by
default, but NEON must be enabled explicitly; see docs.
With other JPEG optimizations included in this version, we see
2x speedup on a JPEG on an x86 machine, and a 1.5x speedup
on a JPEG on an ARM machine, relative to previous versions of this
library. The same results will not obtain for all JPGs and for all
x86/ARM machines. (Note that progressive JPEGs are significantly
slower to decode than regular JPEGs.) This doesn't mean that this
is the fastest JPEG decoder in the land; rather, it brings it
closer to parity with standard libraries. If you want the fastest
decode, look elsewhere. (See "Philosophy" section of docs below.)
See final bullet items below for more info on SIMD.
- Added STBI_MALLOC, STBI_REALLOC, and STBI_FREE macros for replacing
the memory allocator. Unlike other STBI libraries, these macros don't
support a context parameter, so if you need to pass a context in to
the allocator, you'll have to store it in a global or a thread-local
variable.
- Split existing STBI_NO_HDR flag into two flags, STBI_NO_HDR and
STBI_NO_LINEAR.
STBI_NO_HDR: suppress implementation of .hdr reader format
STBI_NO_LINEAR: suppress high-dynamic-range light-linear float API
- You can suppress implementation of any of the decoders to reduce
your code footprint by #defining one or more of the following
symbols before creating the implementation.
STBI_NO_JPEG
STBI_NO_PNG
STBI_NO_BMP
STBI_NO_PSD
STBI_NO_TGA
STBI_NO_GIF
STBI_NO_HDR
STBI_NO_PIC
STBI_NO_PNM (.ppm and .pgm)
- You can request *only* certain decoders and suppress all other ones
(this will be more forward-compatible, as addition of new decoders
doesn't require you to disable them explicitly):
STBI_ONLY_JPEG
STBI_ONLY_PNG
STBI_ONLY_BMP
STBI_ONLY_PSD
STBI_ONLY_TGA
STBI_ONLY_GIF
STBI_ONLY_HDR
STBI_ONLY_PIC
STBI_ONLY_PNM (.ppm and .pgm)
Note that you can define multiples of these, and you will get all
of them ("only x" and "only y" is interpreted to mean "only x&y").
- If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
- Compilation of all SIMD code can be suppressed with
#define STBI_NO_SIMD
It should not be necessary to disable SIMD unless you have issues
compiling (e.g. using an x86 compiler which doesn't support SSE
intrinsics or that doesn't support the method used to detect
SSE2 support at run-time), and even those can be reported as
bugs so I can refine the built-in compile-time checking to be
smarter.
- The old STBI_SIMD system which allowed installing a user-defined
IDCT etc. has been removed. If you need this, don't upgrade. My
assumption is that almost nobody was doing this, and those who
were will find the built-in SIMD more satisfactory anyway.
- RGB values computed for JPEG images are slightly different from
previous versions of stb_image. (This is due to using less
integer precision in SIMD.) The C code has been adjusted so
that the same RGB values will be computed regardless of whether
SIMD support is available, so your app should always produce
consistent results. But these results are slightly different from
previous versions. (Specifically, about 3% of available YCbCr values
will compute different RGB results from pre-1.49 versions by +-1;
most of the deviating values are one smaller in the G channel.)
- If you must produce consistent results with previous versions of
stb_image, #define STBI_JPEG_OLD and you will get the same results
you used to; however, you will not get the SIMD speedups for
the YCbCr-to-RGB conversion step (although you should still see
significant JPEG speedup from the other changes).
Please note that STBI_JPEG_OLD is a temporary feature; it will be
removed in future versions of the library. It is only intended for
near-term back-compatibility use.
Latest revision history:
2.08 (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
2.07 (2015-09-13) partial animated GIF support
limited 16-bit PSD support
minor bugs, code cleanup, and compiler warnings
2.06 (2015-04-19) fix bug where PSD returns wrong '*comp' value
2.05 (2015-04-19) fix bug in progressive JPEG handling, fix warning
2.04 (2015-04-15) try to re-enable SIMD on MinGW 64-bit
2.03 (2015-04-12) additional corruption checking
stbi_set_flip_vertically_on_load
fix NEON support; fix mingw support
2.02 (2015-01-19) fix incorrect assert, fix warning
2.01 (2015-01-17) fix various warnings
2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
2.00 (2014-12-25) optimize JPEG, including x86 SSE2 & ARM NEON SIMD
progressive JPEG
PGM/PPM support
STBI_MALLOC,STBI_REALLOC,STBI_FREE
STBI_NO_*, STBI_ONLY_*
GIF bugfix
1.48 (2014-12-14) fix incorrectly-named assert()
1.47 (2014-12-14) 1/2/4-bit PNG support (both grayscale and paletted)
optimize PNG
fix bug in interlaced PNG with user-specified channel count
See end of file for full revision history.
============================ Contributors =========================
Image formats Bug fixes & warning fixes
Sean Barrett (jpeg, png, bmp) Marc LeBlanc
Nicolas Schulz (hdr, psd) Christpher Lloyd
Jonathan Dummer (tga) Dave Moore
Jean-Marc Lienher (gif) Won Chun
Tom Seddon (pic) the Horde3D community
Thatcher Ulrich (psd) Janez Zemva
Ken Miller (pgm, ppm) Jonathan Blow
urraka@github (animated gif) Laurent Gomila
Aruelien Pocheville
Ryamond Barbiero
David Woo
Extensions, features Martin Golini
Jetro Lauha (stbi_info) Roy Eltham
Martin "SpartanJ" Golini (stbi_info) Luke Graham
James "moose2000" Brown (iPhone PNG) Thomas Ruf
Ben "Disch" Wenger (io callbacks) John Bartholomew
Omar Cornut (1/2/4-bit PNG) Ken Hamada
Nicolas Guillemot (vertical flip) Cort Stratton
Richard Mitton (16-bit PSD) Blazej Dariusz Roszkowski
Thibault Reuille
Paul Du Bois
Guillaume George
Jerry Jansson
Hayaki Saito
Johan Duparc
Ronny Chevalier
Optimizations & bugfixes Michal Cichon
Fabian "ryg" Giesen Tero Hanninen
Arseny Kapoulkine Sergio Gonzalez
Cass Everitt
Engin Manap
If your name should be here but Martins Mozeiko
isn't, let Sean know. Joseph Thomson
Phil Jordan
Nathan Reed
Michaelangel007@github
Nick Verigakis
LICENSE
This software is in the public domain. Where that dedication is not
recognized, you are granted a perpetual, irrevocable license to copy,
distribute, and modify this file as you see fit.
*/
#ifndef STBI_INCLUDE_STB_IMAGE_H
#define STBI_INCLUDE_STB_IMAGE_H
// DOCUMENTATION
//
// Limitations:
// - no 16-bit-per-channel PNG
// - no 12-bit-per-channel JPEG
// - no JPEGs with arithmetic coding
// - no 1-bit BMP
// - GIF always returns *comp=4
//
// Basic usage (see HDR discussion below for HDR usage):
// int x,y,n;
// unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
// // ... process data if not NULL ...
// // ... x = width, y = height, n = # 8-bit components per pixel ...
// // ... replace '0' with '1'..'4' to force that many components per pixel
// // ... but 'n' will always be the number that it would have been if you said 0
// stbi_image_free(data)
//
// Standard parameters:
// int *x -- outputs image width in pixels
// int *y -- outputs image height in pixels
// int *comp -- outputs # of image components in image file
// int req_comp -- if non-zero, # of image components requested in result
//
// The return value from an image loader is an 'unsigned char *' which points
// to the pixel data, or NULL on an allocation failure or if the image is
// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
// with each pixel consisting of N interleaved 8-bit components; the first
// pixel pointed to is top-left-most in the image. There is no padding between
// image scanlines or between pixels, regardless of format. The number of
// components N is 'req_comp' if req_comp is non-zero, or *comp otherwise.
// If req_comp is non-zero, *comp has the number of components that _would_
// have been output otherwise. E.g. if you set req_comp to 4, you will always
// get RGBA output, but you can check *comp to see if it's trivially opaque
// because e.g. there were only 3 channels in the source image.
//
// An output image with N components has the following components interleaved
// in this order in each pixel:
//
// N=#comp components
// 1 grey
// 2 grey, alpha
// 3 red, green, blue
// 4 red, green, blue, alpha
//
// If image loading fails for any reason, the return value will be NULL,
// and *x, *y, *comp will be unchanged. The function stbi_failure_reason()
// can be queried for an extremely brief, end-user unfriendly explanation
// of why the load failed. Define STBI_NO_FAILURE_STRINGS to avoid
// compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
// more user-friendly ones.
//
// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
//
// ===========================================================================
//
// Philosophy
//
// stb libraries are designed with the following priorities:
//
// 1. easy to use
// 2. easy to maintain
// 3. good performance
//
// Sometimes I let "good performance" creep up in priority over "easy to maintain",
// and for best performance I may provide less-easy-to-use APIs that give higher
// performance, in addition to the easy to use ones. Nevertheless, it's important
// to keep in mind that from the standpoint of you, a client of this library,
// all you care about is #1 and #3, and stb libraries do not emphasize #3 above all.
//
// Some secondary priorities arise directly from the first two, some of which
// make more explicit reasons why performance can't be emphasized.
//
// - Portable ("ease of use")
// - Small footprint ("easy to maintain")
// - No dependencies ("ease of use")
//
// ===========================================================================
//
// I/O callbacks
//
// I/O callbacks allow you to read from arbitrary sources, like packaged
// files or some other source. Data read from callbacks are processed
// through a small internal buffer (currently 128 bytes) to try to reduce
// overhead.
//
// The three functions you must define are "read" (reads some bytes of data),
// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
//
// ===========================================================================
//
// SIMD support
//
// The JPEG decoder will try to automatically use SIMD kernels on x86 when
// supported by the compiler. For ARM Neon support, you must explicitly
// request it.
//
// (The old do-it-yourself SIMD API is no longer supported in the current
// code.)
//
// On x86, SSE2 will automatically be used when available based on a run-time
// test; if not, the generic C versions are used as a fall-back. On ARM targets,
// the typical path is to have separate builds for NEON and non-NEON devices
// (at least this is true for iOS and Android). Therefore, the NEON support is
// toggled by a build flag: define STBI_NEON to get NEON loops.
//
// The output of the JPEG decoder is slightly different from versions where
// SIMD support was introduced (that is, for versions before 1.49). The
// difference is only +-1 in the 8-bit RGB channels, and only on a small
// fraction of pixels. You can force the pre-1.49 behavior by defining
// STBI_JPEG_OLD, but this will disable some of the SIMD decoding path
// and hence cost some performance.
//
// If for some reason you do not want to use any of SIMD code, or if
// you have issues compiling it, you can disable it entirely by
// defining STBI_NO_SIMD.
//
// ===========================================================================
//
// HDR image support (disable by defining STBI_NO_HDR)
//
// stb_image now supports loading HDR images in general, and currently
// the Radiance .HDR file format, although the support is provided
// generically. You can still load any file through the existing interface;
// if you attempt to load an HDR file, it will be automatically remapped to
// LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
// both of these constants can be reconfigured through this interface:
//
// stbi_hdr_to_ldr_gamma(2.2f);
// stbi_hdr_to_ldr_scale(1.0f);
//
// (note, do not use _inverse_ constants; stbi_image will invert them
// appropriately).
//
// Additionally, there is a new, parallel interface for loading files as
// (linear) floats to preserve the full dynamic range:
//
// float *data = stbi_loadf(filename, &x, &y, &n, 0);
//
// If you load LDR images through this interface, those images will
// be promoted to floating point values, run through the inverse of
// constants corresponding to the above:
//
// stbi_ldr_to_hdr_scale(1.0f);
// stbi_ldr_to_hdr_gamma(2.2f);
//
// Finally, given a filename (or an open file or memory block--see header
// file for details) containing image data, you can query for the "most
// appropriate" interface to use (that is, whether the image is HDR or
// not), using:
//
// stbi_is_hdr(char *filename);
//
// ===========================================================================
//
// iPhone PNG support:
//
// By default we convert iphone-formatted PNGs back to RGB, even though
// they are internally encoded differently. You can disable this conversion
// by by calling stbi_convert_iphone_png_to_rgb(0), in which case
// you will always just get the native iphone "format" through (which
// is BGR stored in RGB).
//
// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
// pixel to remove any premultiplied alpha *only* if the image file explicitly
// says there's premultiplied data (currently only happens in iPhone images,
// and only if iPhone convert-to-rgb processing is on).
//
#ifndef STBI_NO_STDIO
#include <stdio.h>
#endif // STBI_NO_STDIO
#define STBI_VERSION 1
enum
{
STBI_default = 0, // only used for req_comp
STBI_grey = 1,
STBI_grey_alpha = 2,
STBI_rgb = 3,
STBI_rgb_alpha = 4
};
typedef unsigned char stbi_uc;
#ifdef __cplusplus
extern "C" {
#endif
#ifdef STB_IMAGE_STATIC
#define STBIDEF static
#else
#define STBIDEF extern
#endif
//////////////////////////////////////////////////////////////////////////////
//
// PRIMARY API - works on images of any type
//
//
// load image by filename, open file, or memory buffer
//
typedef struct
{
int (*read) (void *user,char *data,int size); // fill 'data' with 'size' bytes. return number of bytes actually read
void (*skip) (void *user,int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
int (*eof) (void *user); // returns nonzero if we are at end of file/data
} stbi_io_callbacks;
STBIDEF stbi_uc *stbi_load (char const *filename, int *x, int *y, int *comp, int req_comp);
STBIDEF stbi_uc *stbi_load_from_memory (stbi_uc const *buffer, int len , int *x, int *y, int *comp, int req_comp);
STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk , void *user, int *x, int *y, int *comp, int req_comp);
#ifndef STBI_NO_STDIO
STBIDEF stbi_uc *stbi_load_from_file (FILE *f, int *x, int *y, int *comp, int req_comp);
// for stbi_load_from_file, file pointer is left pointing immediately after image
#endif
#ifndef STBI_NO_LINEAR
STBIDEF float *stbi_loadf (char const *filename, int *x, int *y, int *comp, int req_comp);
STBIDEF float *stbi_loadf_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp);
STBIDEF float *stbi_loadf_from_callbacks (stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp);
#ifndef STBI_NO_STDIO
STBIDEF float *stbi_loadf_from_file (FILE *f, int *x, int *y, int *comp, int req_comp);
#endif
#endif
#ifndef STBI_NO_HDR
STBIDEF void stbi_hdr_to_ldr_gamma(float gamma);
STBIDEF void stbi_hdr_to_ldr_scale(float scale);
#endif
#ifndef STBI_NO_LINEAR
STBIDEF void stbi_ldr_to_hdr_gamma(float gamma);
STBIDEF void stbi_ldr_to_hdr_scale(float scale);
#endif // STBI_NO_HDR
// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user);
STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len);
#ifndef STBI_NO_STDIO
STBIDEF int stbi_is_hdr (char const *filename);
STBIDEF int stbi_is_hdr_from_file(FILE *f);
#endif // STBI_NO_STDIO
// get a VERY brief reason for failure
// NOT THREADSAFE
STBIDEF const char *stbi_failure_reason (void);
// free the loaded image -- this is just free()
STBIDEF void stbi_image_free (void *retval_from_stbi_load);
// get image dimensions & components without fully decoding
STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp);
STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp);
#ifndef STBI_NO_STDIO
STBIDEF int stbi_info (char const *filename, int *x, int *y, int *comp);
STBIDEF int stbi_info_from_file (FILE *f, int *x, int *y, int *comp);
#endif
// for image formats that explicitly notate that they have premultiplied alpha,
// we just return the colors as stored in the file. set this flag to force
// unpremultiplication. results are undefined if the unpremultiply overflow.
STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
// indicate whether we should process iphone images back to canonical format,
// or just pass them through "as-is"
STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
// flip the image vertically, so the first pixel in the output array is the bottom left
STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
// ZLIB client - used by PNG, available for other purposes
STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen);
STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header);
STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen);
STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen);
STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
#ifdef __cplusplus
}
#endif
//
//
//// end header file /////////////////////////////////////////////////////
#endif // STBI_INCLUDE_STB_IMAGE_H
#ifdef STB_IMAGE_IMPLEMENTATION
#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \
|| defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \
|| defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \
|| defined(STBI_ONLY_ZLIB)
#ifndef STBI_ONLY_JPEG
#define STBI_NO_JPEG
#endif
#ifndef STBI_ONLY_PNG
#define STBI_NO_PNG
#endif
#ifndef STBI_ONLY_BMP
#define STBI_NO_BMP
#endif
#ifndef STBI_ONLY_PSD
#define STBI_NO_PSD
#endif
#ifndef STBI_ONLY_TGA
#define STBI_NO_TGA
#endif
#ifndef STBI_ONLY_GIF
#define STBI_NO_GIF
#endif
#ifndef STBI_ONLY_HDR
#define STBI_NO_HDR
#endif
#ifndef STBI_ONLY_PIC
#define STBI_NO_PIC
#endif
#ifndef STBI_ONLY_PNM
#define STBI_NO_PNM
#endif
#endif
#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
#define STBI_NO_ZLIB
#endif
#include <stdarg.h>
#include <stddef.h> // ptrdiff_t on osx
#include <stdlib.h>
#include <string.h>
#include <limits.h>
#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
#include <math.h> // ldexp
#endif
#ifndef STBI_NO_STDIO
#include <stdio.h>
#endif
#ifndef STBI_ASSERT
#include <assert.h>
#define STBI_ASSERT(x) assert(x)
#endif
#ifndef _MSC_VER
#ifdef __cplusplus
#define stbi_inline inline
#else
#define stbi_inline
#endif
#else
#define stbi_inline __forceinline
#endif
#ifdef _MSC_VER
typedef unsigned short stbi__uint16;
typedef signed short stbi__int16;
typedef unsigned int stbi__uint32;
typedef signed int stbi__int32;
#else
#include <stdint.h>
typedef uint16_t stbi__uint16;
typedef int16_t stbi__int16;
typedef uint32_t stbi__uint32;
typedef int32_t stbi__int32;
#endif
// should produce compiler error if size is wrong
typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
#ifdef _MSC_VER
#define STBI_NOTUSED(v) (void)(v)
#else
#define STBI_NOTUSED(v) (void)sizeof(v)
#endif
#ifdef _MSC_VER
#define STBI_HAS_LROTL
#endif
#ifdef STBI_HAS_LROTL
#define stbi_lrot(x,y) _lrotl(x,y)
#else
#define stbi_lrot(x,y) (((x) << (y)) | ((x) >> (32 - (y))))
#endif
#if defined(STBI_MALLOC) && defined(STBI_FREE) && defined(STBI_REALLOC)
// ok
#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC)
// ok
#else
#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC."
#endif
#ifndef STBI_MALLOC
#define STBI_MALLOC(sz) malloc(sz)
#define STBI_REALLOC(p,sz) realloc(p,sz)
#define STBI_FREE(p) free(p)
#endif
// x86/x64 detection
#if defined(__x86_64__) || defined(_M_X64)
#define STBI__X64_TARGET
#elif defined(__i386) || defined(_M_IX86)
#define STBI__X86_TARGET
#endif
#if defined(__GNUC__) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET)) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
// NOTE: not clear do we actually need this for the 64-bit path?
// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
// (but compiling with -msse2 allows the compiler to use SSE2 everywhere;
// this is just broken and gcc are jerks for not fixing it properly
// http://www.virtualdub.org/blog/pivot/entry.php?id=363 )
#define STBI_NO_SIMD
#endif
#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
//
// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
// simultaneously enabling "-mstackrealign".
//
// See https://github.com/nothings/stb/issues/81 for more information.
//
// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
#define STBI_NO_SIMD
#endif
#if !defined(STBI_NO_SIMD) && defined(STBI__X86_TARGET)
#define STBI_SSE2
#include <emmintrin.h>
#ifdef _MSC_VER
#if _MSC_VER >= 1400 // not VC6
#include <intrin.h> // __cpuid
static int stbi__cpuid3(void)
{
int info[4];
__cpuid(info,1);
return info[3];
}
#else
static int stbi__cpuid3(void)
{
int res;
__asm {
mov eax,1
cpuid
mov res,edx
}
return res;
}
#endif
#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
static int stbi__sse2_available()
{
int info3 = stbi__cpuid3();
return ((info3 >> 26) & 1) != 0;
}
#else // assume GCC-style if not VC++
#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
static int stbi__sse2_available()
{
#if defined(__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__) >= 408 // GCC 4.8 or later
// GCC 4.8+ has a nice way to do this
return __builtin_cpu_supports("sse2");
#else
// portable way to do this, preferably without using GCC inline ASM?
// just bail for now.
return 0;
#endif
}
#endif
#endif
// ARM NEON
#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
#undef STBI_NEON
#endif
#ifdef STBI_NEON
#include <arm_neon.h>
// assume GCC or Clang on ARM targets
#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
#endif
#ifndef STBI_SIMD_ALIGN
#define STBI_SIMD_ALIGN(type, name) type name
#endif
///////////////////////////////////////////////
//
// stbi__context struct and start_xxx functions
// stbi__context structure is our basic context used by all images, so it
// contains all the IO context, plus some basic image information
typedef struct
{
stbi__uint32 img_x, img_y;
int img_n, img_out_n;
stbi_io_callbacks io;
void *io_user_data;
int read_from_callbacks;
int buflen;
stbi_uc buffer_start[128];
stbi_uc *img_buffer, *img_buffer_end;
stbi_uc *img_buffer_original, *img_buffer_original_end;
} stbi__context;
static void stbi__refill_buffer(stbi__context *s);
// initialize a memory-decode context
static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len)
{
s->io.read = NULL;
s->read_from_callbacks = 0;
s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer;
s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len;
}
// initialize a callback-based context
static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user)
{
s->io = *c;
s->io_user_data = user;
s->buflen = sizeof(s->buffer_start);
s->read_from_callbacks = 1;
s->img_buffer_original = s->buffer_start;
stbi__refill_buffer(s);
s->img_buffer_original_end = s->img_buffer_end;
}
#ifndef STBI_NO_STDIO
static int stbi__stdio_read(void *user, char *data, int size)
{
return (int) fread(data,1,size,(FILE*) user);
}
static void stbi__stdio_skip(void *user, int n)
{
fseek((FILE*) user, n, SEEK_CUR);
}
static int stbi__stdio_eof(void *user)
{
return feof((FILE*) user);
}
static stbi_io_callbacks stbi__stdio_callbacks =
{
stbi__stdio_read,
stbi__stdio_skip,
stbi__stdio_eof,
};
static void stbi__start_file(stbi__context *s, FILE *f)
{
stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f);
}
//static void stop_file(stbi__context *s) { }
#endif // !STBI_NO_STDIO
static void stbi__rewind(stbi__context *s)
{
// conceptually rewind SHOULD rewind to the beginning of the stream,
// but we just rewind to the beginning of the initial buffer, because
// we only use it after doing 'test', which only ever looks at at most 92 bytes
s->img_buffer = s->img_buffer_original;
s->img_buffer_end = s->img_buffer_original_end;
}
#ifndef STBI_NO_JPEG
static int stbi__jpeg_test(stbi__context *s);
static stbi_uc *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_PNG
static int stbi__png_test(stbi__context *s);
static stbi_uc *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_BMP
static int stbi__bmp_test(stbi__context *s);
static stbi_uc *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_TGA
static int stbi__tga_test(stbi__context *s);
static stbi_uc *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_PSD
static int stbi__psd_test(stbi__context *s);
static stbi_uc *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_HDR
static int stbi__hdr_test(stbi__context *s);
static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_PIC
static int stbi__pic_test(stbi__context *s);
static stbi_uc *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_GIF
static int stbi__gif_test(stbi__context *s);
static stbi_uc *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
#endif
#ifndef STBI_NO_PNM
static int stbi__pnm_test(stbi__context *s);
static stbi_uc *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp);
static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
#endif
// this is not threadsafe
static const char *stbi__g_failure_reason;
STBIDEF const char *stbi_failure_reason(void)
{
return stbi__g_failure_reason;
}
static int stbi__err(const char *str)
{
stbi__g_failure_reason = str;
return 0;
}
static void *stbi__malloc(size_t size)
{
return STBI_MALLOC(size);
}
// stbi__err - error
// stbi__errpf - error returning pointer to float
// stbi__errpuc - error returning pointer to unsigned char
#ifdef STBI_NO_FAILURE_STRINGS
#define stbi__err(x,y) 0
#elif defined(STBI_FAILURE_USERMSG)
#define stbi__err(x,y) stbi__err(y)
#else
#define stbi__err(x,y) stbi__err(x)
#endif
#define stbi__errpf(x,y) ((float *)(size_t) (stbi__err(x,y)?NULL:NULL))
#define stbi__errpuc(x,y) ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL))
STBIDEF void stbi_image_free(void *retval_from_stbi_load)
{
STBI_FREE(retval_from_stbi_load);
}
#ifndef STBI_NO_LINEAR
static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp);
#endif
#ifndef STBI_NO_HDR
static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp);
#endif
static int stbi__vertically_flip_on_load = 0;
STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip)
{
stbi__vertically_flip_on_load = flag_true_if_should_flip;
}
static unsigned char *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
{
#ifndef STBI_NO_JPEG
if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_PNG
if (stbi__png_test(s)) return stbi__png_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_BMP
if (stbi__bmp_test(s)) return stbi__bmp_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_GIF
if (stbi__gif_test(s)) return stbi__gif_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_PSD
if (stbi__psd_test(s)) return stbi__psd_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_PIC
if (stbi__pic_test(s)) return stbi__pic_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_PNM
if (stbi__pnm_test(s)) return stbi__pnm_load(s,x,y,comp,req_comp);
#endif
#ifndef STBI_NO_HDR
if (stbi__hdr_test(s)) {
float *hdr = stbi__hdr_load(s, x,y,comp,req_comp);
return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
}
#endif
#ifndef STBI_NO_TGA
// test tga last because it's a crappy test!
if (stbi__tga_test(s))
return stbi__tga_load(s,x,y,comp,req_comp);
#endif
return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
}
static unsigned char *stbi__load_flip(stbi__context *s, int *x, int *y, int *comp, int req_comp)
{
unsigned char *result = stbi__load_main(s, x, y, comp, req_comp);
if (stbi__vertically_flip_on_load && result != NULL) {
int w = *x, h = *y;
int depth = req_comp ? req_comp : *comp;
int row,col,z;
stbi_uc temp;
// @OPTIMIZE: use a bigger temp buffer and memcpy multiple pixels at once
for (row = 0; row < (h>>1); row++) {
for (col = 0; col < w; col++) {
for (z = 0; z < depth; z++) {
temp = result[(row * w + col) * depth + z];
result[(row * w + col) * depth + z] = result[((h - row - 1) * w + col) * depth + z];
result[((h - row - 1) * w + col) * depth + z] = temp;
}
}
}
}
return result;
}
#ifndef STBI_NO_HDR
static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp)
{
if (stbi__vertically_flip_on_load && result != NULL) {
int w = *x, h = *y;
int depth = req_comp ? req_comp : *comp;
int row,col,z;
float temp;
// @OPTIMIZE: use a bigger temp buffer and memcpy multiple pixels at once
for (row = 0; row < (h>>1); row++) {
for (col = 0; col < w; col++) {
for (z = 0; z < depth; z++) {
temp = result[(row * w + col) * depth + z];
result[(row * w + col) * depth + z] = result[((h - row - 1) * w + col) * depth + z];
result[((h - row - 1) * w + col) * depth + z] = temp;
}
}
}
}
}
#endif
#ifndef STBI_NO_STDIO
static FILE *stbi__fopen(char const *filename, char const *mode)
{
FILE *f;
#if defined(_MSC_VER) && _MSC_VER >= 1400
if (0 != fopen_s(&f, filename, mode))
f=0;
#else
f = fopen(filename, mode);
#endif
return f;
}
STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp)
{
FILE *f = stbi__fopen(filename, "rb");
unsigned char *result;
if (!f) return stbi__errpuc("can't fopen", "Unable to open file");
result = stbi_load_from_file(f,x,y,comp,req_comp);
fclose(f);
return result;
}
STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
{
unsigned char *result;
stbi__context s;
stbi__start_file(&s,f);
result = stbi__load_flip(&s,x,y,comp,req_comp);
if (result) {
// need to 'unget' all the characters in the IO buffer
fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
}
return result;
}
#endif //!STBI_NO_STDIO
STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
{
stbi__context s;
stbi__start_mem(&s,buffer,len);
return stbi__load_flip(&s,x,y,comp,req_comp);
}
STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
{
stbi__context s;
stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
return stbi__load_flip(&s,x,y,comp,req_comp);
}
#ifndef STBI_NO_LINEAR
static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
{
unsigned char *data;
#ifndef STBI_NO_HDR
if (stbi__hdr_test(s)) {
float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp);
if (hdr_data)
stbi__float_postprocess(hdr_data,x,y,comp,req_comp);
return hdr_data;
}
#endif
data = stbi__load_flip(s, x, y, comp, req_comp);
if (data)
return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
}
STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
{
stbi__context s;
stbi__start_mem(&s,buffer,len);
return stbi__loadf_main(&s,x,y,comp,req_comp);
}
STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
{
stbi__context s;
stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
return stbi__loadf_main(&s,x,y,comp,req_comp);
}
#ifndef STBI_NO_STDIO
STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp)
{
float *result;
FILE *f = stbi__fopen(filename, "rb");
if (!f) return stbi__errpf("can't fopen", "Unable to open file");
result = stbi_loadf_from_file(f,x,y,comp,req_comp);
fclose(f);
return result;
}
STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
{
stbi__context s;
stbi__start_file(&s,f);
return stbi__loadf_main(&s,x,y,comp,req_comp);
}
#endif // !STBI_NO_STDIO
#endif // !STBI_NO_LINEAR
// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
// reports false!
STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len)
{
#ifndef STBI_NO_HDR
stbi__context s;
stbi__start_mem(&s,buffer,len);
return stbi__hdr_test(&s);
#else
STBI_NOTUSED(buffer);
STBI_NOTUSED(len);
return 0;
#endif
}
#ifndef STBI_NO_STDIO
STBIDEF int stbi_is_hdr (char const *filename)
{
FILE *f = stbi__fopen(filename, "rb");
int result=0;
if (f) {
result = stbi_is_hdr_from_file(f);
fclose(f);
}
return result;
}
STBIDEF int stbi_is_hdr_from_file(FILE *f)
{
#ifndef STBI_NO_HDR
stbi__context s;
stbi__start_file(&s,f);
return stbi__hdr_test(&s);
#else
STBI_NOTUSED(f);
return 0;
#endif
}
#endif // !STBI_NO_STDIO
STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user)
{
#ifndef STBI_NO_HDR
stbi__context s;
stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
return stbi__hdr_test(&s);
#else
STBI_NOTUSED(clbk);
STBI_NOTUSED(user);
return 0;
#endif
}
static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f;
static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f;
#ifndef STBI_NO_LINEAR
STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
#endif
STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; }
STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; }
//////////////////////////////////////////////////////////////////////////////
//
// Common code used by all image loaders
//
enum
{
STBI__SCAN_load=0,
STBI__SCAN_type,
STBI__SCAN_header
};
static void stbi__refill_buffer(stbi__context *s)
{
int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen);
if (n == 0) {
// at end of file, treat same as if from memory, but need to handle case
// where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
s->read_from_callbacks = 0;
s->img_buffer = s->buffer_start;
s->img_buffer_end = s->buffer_start+1;
*s->img_buffer = 0;
} else {
s->img_buffer = s->buffer_start;
s->img_buffer_end = s->buffer_start + n;
}
}
stbi_inline static stbi_uc stbi__get8(stbi__context *s)
{
if (s->img_buffer < s->img_buffer_end)
return *s->img_buffer++;
if (s->read_from_callbacks) {
stbi__refill_buffer(s);
return *s->img_buffer++;
}
return 0;
}
stbi_inline static int stbi__at_eof(stbi__context *s)
{
if (s->io.read) {
if (!(s->io.eof)(s->io_user_data)) return 0;
// if feof() is true, check if buffer = end
// special case: we've only got the special 0 character at the end
if (s->read_from_callbacks == 0) return 1;
}
return s->img_buffer >= s->img_buffer_end;
}
static void stbi__skip(stbi__context *s, int n)
{
if (n < 0) {
s->img_buffer = s->img_buffer_end;
return;
}
if (s->io.read) {
int blen = (int) (s->img_buffer_end - s->img_buffer);
if (blen < n) {
s->img_buffer = s->img_buffer_end;
(s->io.skip)(s->io_user_data, n - blen);
return;
}
}
s->img_buffer += n;
}
static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n)
{
if (s->io.read) {
int blen = (int) (s->img_buffer_end - s->img_buffer);
if (blen < n) {
int res, count;
memcpy(buffer, s->img_buffer, blen);
count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen);
res = (count == (n-blen));
s->img_buffer = s->img_buffer_end;
return res;
}
}
if (s->img_buffer+n <= s->img_buffer_end) {
memcpy(buffer, s->img_buffer, n);
s->img_buffer += n;
return 1;
} else
return 0;
}
static int stbi__get16be(stbi__context *s)
{
int z = stbi__get8(s);
return (z << 8) + stbi__get8(s);
}
static stbi__uint32 stbi__get32be(stbi__context *s)
{
stbi__uint32 z = stbi__get16be(s);
return (z << 16) + stbi__get16be(s);
}
#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
// nothing
#else
static int stbi__get16le(stbi__context *s)
{
int z = stbi__get8(s);
return z + (stbi__get8(s) << 8);
}
#endif
#ifndef STBI_NO_BMP
static stbi__uint32 stbi__get32le(stbi__context *s)
{
stbi__uint32 z = stbi__get16le(s);
return z + (stbi__get16le(s) << 16);
}
#endif
#define STBI__BYTECAST(x) ((stbi_uc) ((x) & 255)) // truncate int to byte without warnings
//////////////////////////////////////////////////////////////////////////////
//
// generic converter from built-in img_n to req_comp
// individual types do this automatically as much as possible (e.g. jpeg
// does all cases internally since it needs to colorspace convert anyway,
// and it never has alpha, so very few cases ). png can automatically
// interleave an alpha=255 channel, but falls back to this for other cases
//
// assume data buffer is malloced, so malloc a new one and free that one
// only failure mode is malloc failing
static stbi_uc stbi__compute_y(int r, int g, int b)
{
return (stbi_uc) (((r*77) + (g*150) + (29*b)) >> 8);
}
static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y)
{
int i,j;
unsigned char *good;
if (req_comp == img_n) return data;
STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
if (x == 0 || y == 0 || req_comp <= 0 || (req_comp > INT_MAX / x / y))
return stbi__errpuc("Integer OverFlow", "x or y is bad");
good = (unsigned char *) stbi__malloc(req_comp * x * y);
if (good == NULL) {
STBI_FREE(data);
return stbi__errpuc("outofmem", "Out of memory");
}
for (j=0; j < (int) y; ++j) {
unsigned char *src = data + j * x * img_n ;
unsigned char *dest = good + j * x * req_comp;
#define COMBO(a,b) ((a)*8+(b))
#define CASE(a,b) case COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
// convert source image with img_n components to one with req_comp components;
// avoid switch per pixel, so use switch per scanline and massive macros
switch (COMBO(img_n, req_comp)) {
CASE(1,2) dest[0]=src[0], dest[1]=255; break;
CASE(1,3) dest[0]=dest[1]=dest[2]=src[0]; break;
CASE(1,4) dest[0]=dest[1]=dest[2]=src[0], dest[3]=255; break;
CASE(2,1) dest[0]=src[0]; break;
CASE(2,3) dest[0]=dest[1]=dest[2]=src[0]; break;
CASE(2,4) dest[0]=dest[1]=dest[2]=src[0], dest[3]=src[1]; break;
CASE(3,4) dest[0]=src[0],dest[1]=src[1],dest[2]=src[2],dest[3]=255; break;
CASE(3,1) dest[0]=stbi__compute_y(src[0],src[1],src[2]); break;
CASE(3,2) dest[0]=stbi__compute_y(src[0],src[1],src[2]), dest[1] = 255; break;
CASE(4,1) dest[0]=stbi__compute_y(src[0],src[1],src[2]); break;
CASE(4,2) dest[0]=stbi__compute_y(src[0],src[1],src[2]), dest[1] = src[3]; break;
CASE(4,3) dest[0]=src[0],dest[1]=src[1],dest[2]=src[2]; break;
default: STBI_ASSERT(0);
}
#undef CASE
}
STBI_FREE(data);
return good;
}
#ifndef STBI_NO_LINEAR
static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp)
{
int i,k,n;
if (x <= 0 || y <= 0 || comp <= 0 ||
(sizeof(float) > INT_MAX / x / y / comp))
return stbi__errpf("Integer OverFlow", "x , y or comp is too large");
float *output = (float *) stbi__malloc(x * y * comp * sizeof(float));
if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); }
// compute number of non-alpha components
if (comp & 1) n = comp; else n = comp-1;
for (i=0; i < x*y; ++i) {
for (k=0; k < n; ++k) {
output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale);
}
if (k < comp) output[i*comp + k] = data[i*comp+k]/255.0f;
}
STBI_FREE(data);
return output;
}
#endif
#ifndef STBI_NO_HDR
#define stbi__float2int(x) ((int) (x))
static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp)
{
int i,k,n;
if (x <= 0 || y <= 0 || comp <= 0 ||
(comp > INT_MAX / x / y))
return stbi__errpuc("Integer OverFlow", "x or y is too large");
stbi_uc *output = (stbi_uc *) stbi__malloc(x * y * comp);
if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); }
// compute number of non-alpha components
if (comp & 1) n = comp; else n = comp-1;
for (i=0; i < x*y; ++i) {
for (k=0; k < n; ++k) {
float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
if (z < 0) z = 0;
if (z > 255) z = 255;
output[i*comp + k] = (stbi_uc) stbi__float2int(z);
}
if (k < comp) {
float z = data[i*comp+k] * 255 + 0.5f;
if (z < 0) z = 0;
if (z > 255) z = 255;
output[i*comp + k] = (stbi_uc) stbi__float2int(z);
}
}
STBI_FREE(data);
return output;
}
#endif
//////////////////////////////////////////////////////////////////////////////
//
// "baseline" JPEG/JFIF decoder
//
// simple implementation
// - doesn't support delayed output of y-dimension
// - simple interface (only one output format: 8-bit interleaved RGB)
// - doesn't try to recover corrupt jpegs
// - doesn't allow partial loading, loading multiple at once
// - still fast on x86 (copying globals into locals doesn't help x86)
// - allocates lots of intermediate memory (full size of all components)
// - non-interleaved case requires this anyway
// - allows good upsampling (see next)
// high-quality
// - upsampled channels are bilinearly interpolated, even across blocks
// - quality integer IDCT derived from IJG's 'slow'
// performance
// - fast huffman; reasonable integer IDCT
// - some SIMD kernels for common paths on targets with SSE2/NEON
// - uses a lot of intermediate memory, could cache poorly
#ifndef STBI_NO_JPEG
// huffman decoding acceleration
#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache
typedef struct
{
stbi_uc fast[1 << FAST_BITS];
// weirdly, repacking this into AoS is a 10% speed loss, instead of a win
stbi__uint16 code[256];
stbi_uc values[256];
stbi_uc size[257];
unsigned int maxcode[18];
int delta[17]; // old 'firstsymbol' - old 'firstcode'
} stbi__huffman;
typedef struct
{
stbi__context *s;
stbi__huffman huff_dc[4];
stbi__huffman huff_ac[4];
stbi_uc dequant[4][64];
stbi__int16 fast_ac[4][1 << FAST_BITS];
// sizes for components, interleaved MCUs
int img_h_max, img_v_max;
int img_mcu_x, img_mcu_y;
int img_mcu_w, img_mcu_h;
// definition of jpeg image component
struct
{
int id;
int h,v;
int tq;
int hd,ha;
int dc_pred;
int x,y,w2,h2;
stbi_uc *data;
void *raw_data, *raw_coeff;
stbi_uc *linebuf;
short *coeff; // progressive only
int coeff_w, coeff_h; // number of 8x8 coefficient blocks
} img_comp[4];
stbi__uint32 code_buffer; // jpeg entropy-coded buffer
int code_bits; // number of valid bits
unsigned char marker; // marker seen while filling entropy buffer
int nomore; // flag if we saw a marker so must stop
int progressive;
int spec_start;
int spec_end;
int succ_high;
int succ_low;
int eob_run;
int scan_n, order[4];
int restart_interval, todo;
// kernels
void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]);
void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step);
stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs);
} stbi__jpeg;
static int stbi__build_huffman(stbi__huffman *h, int *count)
{
int i,j,k=0,code;
// build size list for each symbol (from JPEG spec)
for (i=0; i < 16; ++i)
for (j=0; j < count[i]; ++j)
h->size[k++] = (stbi_uc) (i+1);
h->size[k] = 0;
// compute actual symbols (from jpeg spec)
code = 0;
k = 0;
for(j=1; j <= 16; ++j) {
// compute delta to add to code to compute symbol id
h->delta[j] = k - code;
if (h->size[k] == j) {
while (h->size[k] == j)
h->code[k++] = (stbi__uint16) (code++);
if (code-1 >= (1 << j)) return stbi__err("bad code lengths","Corrupt JPEG");
}
// compute largest code + 1 for this size, preshifted as needed later
h->maxcode[j] = code << (16-j);
code <<= 1;
}
h->maxcode[j] = 0xffffffff;
// build non-spec acceleration table; 255 is flag for not-accelerated
memset(h->fast, 255, 1 << FAST_BITS);
for (i=0; i < k; ++i) {
int s = h->size[i];
if (s <= FAST_BITS) {
int c = h->code[i] << (FAST_BITS-s);
int m = 1 << (FAST_BITS-s);
for (j=0; j < m; ++j) {
h->fast[c+j] = (stbi_uc) i;
}
}
}
return 1;
}
// build a table that decodes both magnitude and value of small ACs in
// one go.
static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h)
{
int i;
for (i=0; i < (1 << FAST_BITS); ++i) {
stbi_uc fast = h->fast[i];
fast_ac[i] = 0;
if (fast < 255) {
int rs = h->values[fast];
int run = (rs >> 4) & 15;
int magbits = rs & 15;
int len = h->size[fast];
if (magbits && len + magbits <= FAST_BITS) {
// magnitude code followed by receive_extend code
int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
int m = 1 << (magbits - 1);
if (k < m) k += (-1 << magbits) + 1;
// if the result is small enough, we can fit it in fast_ac table
if (k >= -128 && k <= 127)
fast_ac[i] = (stbi__int16) ((k << 8) + (run << 4) + (len + magbits));
}
}
}
}
static void stbi__grow_buffer_unsafe(stbi__jpeg *j)
{
do {
int b = j->nomore ? 0 : stbi__get8(j->s);
if (b == 0xff) {
int c = stbi__get8(j->s);
if (c != 0) {
j->marker = (unsigned char) c;
j->nomore = 1;
return;
}
}
j->code_buffer |= b << (24 - j->code_bits);
j->code_bits += 8;
} while (j->code_bits <= 24);
}
// (1 << n) - 1
static stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535};
// decode a jpeg huffman value from the bitstream
stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h)
{
unsigned int temp;
int c,k;
if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
// look at the top FAST_BITS and determine what symbol ID it is,
// if the code is <= FAST_BITS
c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
k = h->fast[c];
if (k < 255) {
int s = h->size[k];
if (s > j->code_bits)
return -1;
j->code_buffer <<= s;
j->code_bits -= s;
return h->values[k];
}
// naive test is to shift the code_buffer down so k bits are
// valid, then test against maxcode. To speed this up, we've
// preshifted maxcode left so that it has (16-k) 0s at the
// end; in other words, regardless of the number of bits, it
// wants to be compared against something shifted to have 16;
// that way we don't need to shift inside the loop.
temp = j->code_buffer >> 16;
for (k=FAST_BITS+1 ; ; ++k)
if (temp < h->maxcode[k])
break;
if (k == 17) {
// error! code not found
j->code_bits -= 16;
return -1;
}
if (k > j->code_bits)
return -1;
// convert the huffman code to the symbol id
c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
// convert the id to a symbol
j->code_bits -= k;
j->code_buffer <<= k;
return h->values[c];
}
// bias[n] = (-1<<n) + 1
static int const stbi__jbias[16] = {0,-1,-3,-7,-15,-31,-63,-127,-255,-511,-1023,-2047,-4095,-8191,-16383,-32767};
// combined JPEG 'receive' and JPEG 'extend', since baseline
// always extends everything it receives.
stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
{
unsigned int k;
int sgn;
if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
sgn = (stbi__int32)j->code_buffer >> 31; // sign bit is always in MSB
k = stbi_lrot(j->code_buffer, n);
STBI_ASSERT(n >= 0 && n < (int) (sizeof(stbi__bmask)/sizeof(*stbi__bmask)));
j->code_buffer = k & ~stbi__bmask[n];
k &= stbi__bmask[n];
j->code_bits -= n;
return k + (stbi__jbias[n] & ~sgn);
}
// get some unsigned bits
stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n)
{
unsigned int k;
if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
k = stbi_lrot(j->code_buffer, n);
j->code_buffer = k & ~stbi__bmask[n];
k &= stbi__bmask[n];
j->code_bits -= n;
return k;
}
stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j)
{
unsigned int k;
if (j->code_bits < 1) stbi__grow_buffer_unsafe(j);
k = j->code_buffer;
j->code_buffer <<= 1;
--j->code_bits;
return k & 0x80000000;
}
// given a value that's at position X in the zigzag stream,
// where does it appear in the 8x8 matrix coded as row-major?
static stbi_uc stbi__jpeg_dezigzag[64+15] =
{
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, 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,
// let corrupt input sample past end
63, 63, 63, 63, 63, 63, 63, 63,
63, 63, 63, 63, 63, 63, 63
};
// decode one 64-entry block--
static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi_uc *dequant)
{
int diff,dc,k;
int t;
if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
t = stbi__jpeg_huff_decode(j, hdc);
if (t < 0) return stbi__err("bad huffman code","Corrupt JPEG");
// 0 all the ac values now so we can do it 32-bits at a time
memset(data,0,64*sizeof(data[0]));
diff = t ? stbi__extend_receive(j, t) : 0;
dc = j->img_comp[b].dc_pred + diff;
j->img_comp[b].dc_pred = dc;
data[0] = (short) (dc * dequant[0]);
// decode AC components, see JPEG spec
k = 1;
do {
unsigned int zig;
int c,r,s;
if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
r = fac[c];
if (r) { // fast-AC path
k += (r >> 4) & 15; // run
s = r & 15; // combined length
j->code_buffer <<= s;
j->code_bits -= s;
// decode into unzigzag'd location
zig = stbi__jpeg_dezigzag[k++];
data[zig] = (short) ((r >> 8) * dequant[zig]);
} else {
int rs = stbi__jpeg_huff_decode(j, hac);
if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
s = rs & 15;
r = rs >> 4;
if (s == 0) {
if (rs != 0xf0) break; // end block
k += 16;
} else {
k += r;
// decode into unzigzag'd location
zig = stbi__jpeg_dezigzag[k++];
data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]);
}
}
} while (k < 64);
return 1;
}
static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b)
{
int diff,dc;
int t;
if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
if (j->succ_high == 0) {
// first scan for DC coefficient, must be first
memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
t = stbi__jpeg_huff_decode(j, hdc);
diff = t ? stbi__extend_receive(j, t) : 0;
dc = j->img_comp[b].dc_pred + diff;
j->img_comp[b].dc_pred = dc;
data[0] = (short) (dc << j->succ_low);
} else {
// refinement scan for DC coefficient
if (stbi__jpeg_get_bit(j))
data[0] += (short) (1 << j->succ_low);
}
return 1;
}
// @OPTIMIZE: store non-zigzagged during the decode passes,
// and only de-zigzag when dequantizing
static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac)
{
int k;
if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
if (j->succ_high == 0) {
int shift = j->succ_low;
if (j->eob_run) {
--j->eob_run;
return 1;
}
k = j->spec_start;
do {
unsigned int zig;
int c,r,s;
if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
r = fac[c];
if (r) { // fast-AC path
k += (r >> 4) & 15; // run
s = r & 15; // combined length
j->code_buffer <<= s;
j->code_bits -= s;
zig = stbi__jpeg_dezigzag[k++];
data[zig] = (short) ((r >> 8) << shift);
} else {
int rs = stbi__jpeg_huff_decode(j, hac);
if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
s = rs & 15;
r = rs >> 4;
if (s == 0) {
if (r < 15) {
j->eob_run = (1 << r);
if (r)
j->eob_run += stbi__jpeg_get_bits(j, r);
--j->eob_run;
break;
}
k += 16;
} else {
k += r;
zig = stbi__jpeg_dezigzag[k++];
data[zig] = (short) (stbi__extend_receive(j,s) << shift);
}
}
} while (k <= j->spec_end);
} else {
// refinement scan for these AC coefficients
short bit = (short) (1 << j->succ_low);
if (j->eob_run) {
--j->eob_run;
for (k = j->spec_start; k <= j->spec_end; ++k) {
short *p = &data[stbi__jpeg_dezigzag[k]];
if (*p != 0)
if (stbi__jpeg_get_bit(j))
if ((*p & bit)==0) {
if (*p > 0)
*p += bit;
else
*p -= bit;
}
}
} else {
k = j->spec_start;
do {
int r,s;
int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
s = rs & 15;
r = rs >> 4;
if (s == 0) {
if (r < 15) {
j->eob_run = (1 << r) - 1;
if (r)
j->eob_run += stbi__jpeg_get_bits(j, r);
r = 64; // force end of block
} else {
// r=15 s=0 should write 16 0s, so we just do
// a run of 15 0s and then write s (which is 0),
// so we don't have to do anything special here
}
} else {
if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG");
// sign bit
if (stbi__jpeg_get_bit(j))
s = bit;
else
s = -bit;
}
// advance by r
while (k <= j->spec_end) {
short *p = &data[stbi__jpeg_dezigzag[k++]];
if (*p != 0) {
if (stbi__jpeg_get_bit(j))
if ((*p & bit)==0) {
if (*p > 0)
*p += bit;
else
*p -= bit;
}
} else {
if (r == 0) {
*p = (short) s;
break;
}
--r;
}
}
} while (k <= j->spec_end);
}
}
return 1;
}
// take a -128..127 value and stbi__clamp it and convert to 0..255
stbi_inline static stbi_uc stbi__clamp(int x)
{
// trick to use a single test to catch both cases
if ((unsigned int) x > 255) {
if (x < 0) return 0;
if (x > 255) return 255;
}
return (stbi_uc) x;
}
#define stbi__f2f(x) ((int) (((x) * 4096 + 0.5)))
#define stbi__fsh(x) ((x) << 12)
// derived from jidctint -- DCT_ISLOW
#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \
int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \
p2 = s2; \
p3 = s6; \
p1 = (p2+p3) * stbi__f2f(0.5411961f); \
t2 = p1 + p3*stbi__f2f(-1.847759065f); \
t3 = p1 + p2*stbi__f2f( 0.765366865f); \
p2 = s0; \
p3 = s4; \
t0 = stbi__fsh(p2+p3); \
t1 = stbi__fsh(p2-p3); \
x0 = t0+t3; \
x3 = t0-t3; \
x1 = t1+t2; \
x2 = t1-t2; \
t0 = s7; \
t1 = s5; \
t2 = s3; \
t3 = s1; \
p3 = t0+t2; \
p4 = t1+t3; \
p1 = t0+t3; \
p2 = t1+t2; \
p5 = (p3+p4)*stbi__f2f( 1.175875602f); \
t0 = t0*stbi__f2f( 0.298631336f); \
t1 = t1*stbi__f2f( 2.053119869f); \
t2 = t2*stbi__f2f( 3.072711026f); \
t3 = t3*stbi__f2f( 1.501321110f); \
p1 = p5 + p1*stbi__f2f(-0.899976223f); \
p2 = p5 + p2*stbi__f2f(-2.562915447f); \
p3 = p3*stbi__f2f(-1.961570560f); \
p4 = p4*stbi__f2f(-0.390180644f); \
t3 += p1+p4; \
t2 += p2+p3; \
t1 += p2+p4; \
t0 += p1+p3;
static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64])
{
int i,val[64],*v=val;
stbi_uc *o;
short *d = data;
// columns
for (i=0; i < 8; ++i,++d, ++v) {
// if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0
&& d[40]==0 && d[48]==0 && d[56]==0) {
// no shortcut 0 seconds
// (1|2|3|4|5|6|7)==0 0 seconds
// all separate -0.047 seconds
// 1 && 2|3 && 4|5 && 6|7: -0.047 seconds
int dcterm = d[0] << 2;
v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
} else {
STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56])
// constants scaled things up by 1<<12; let's bring them back
// down, but keep 2 extra bits of precision
x0 += 512; x1 += 512; x2 += 512; x3 += 512;
v[ 0] = (x0+t3) >> 10;
v[56] = (x0-t3) >> 10;
v[ 8] = (x1+t2) >> 10;
v[48] = (x1-t2) >> 10;
v[16] = (x2+t1) >> 10;
v[40] = (x2-t1) >> 10;
v[24] = (x3+t0) >> 10;
v[32] = (x3-t0) >> 10;
}
}
for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) {
// no fast case since the first 1D IDCT spread components out
STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7])
// constants scaled things up by 1<<12, plus we had 1<<2 from first
// loop, plus horizontal and vertical each scale by sqrt(8) so together
// we've got an extra 1<<3, so 1<<17 total we need to remove.
// so we want to round that, which means adding 0.5 * 1<<17,
// aka 65536. Also, we'll end up with -128 to 127 that we want
// to encode as 0..255 by adding 128, so we'll add that before the shift
x0 += 65536 + (128<<17);
x1 += 65536 + (128<<17);
x2 += 65536 + (128<<17);
x3 += 65536 + (128<<17);
// tried computing the shifts into temps, or'ing the temps to see
// if any were out of range, but that was slower
o[0] = stbi__clamp((x0+t3) >> 17);
o[7] = stbi__clamp((x0-t3) >> 17);
o[1] = stbi__clamp((x1+t2) >> 17);
o[6] = stbi__clamp((x1-t2) >> 17);
o[2] = stbi__clamp((x2+t1) >> 17);
o[5] = stbi__clamp((x2-t1) >> 17);
o[3] = stbi__clamp((x3+t0) >> 17);
o[4] = stbi__clamp((x3-t0) >> 17);
}
}
#ifdef STBI_SSE2
// sse2 integer IDCT. not the fastest possible implementation but it
// produces bit-identical results to the generic C version so it's
// fully "transparent".
static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
{
// This is constructed to match our regular (generic) integer IDCT exactly.
__m128i row0, row1, row2, row3, row4, row5, row6, row7;
__m128i tmp;
// dot product constant: even elems=x, odd elems=y
#define dct_const(x,y) _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y))
// out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit)
// out(1) = c1[even]*x + c1[odd]*y
#define dct_rot(out0,out1, x,y,c0,c1) \
__m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \
__m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \
__m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \
__m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \
__m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \
__m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
// out = in << 12 (in 16-bit, out 32-bit)
#define dct_widen(out, in) \
__m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \
__m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
// wide add
#define dct_wadd(out, a, b) \
__m128i out##_l = _mm_add_epi32(a##_l, b##_l); \
__m128i out##_h = _mm_add_epi32(a##_h, b##_h)
// wide sub
#define dct_wsub(out, a, b) \
__m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \
__m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
// butterfly a/b, add bias, then shift by "s" and pack
#define dct_bfly32o(out0, out1, a,b,bias,s) \
{ \
__m128i abiased_l = _mm_add_epi32(a##_l, bias); \
__m128i abiased_h = _mm_add_epi32(a##_h, bias); \
dct_wadd(sum, abiased, b); \
dct_wsub(dif, abiased, b); \
out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \
out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \
}
// 8-bit interleave step (for transposes)
#define dct_interleave8(a, b) \
tmp = a; \
a = _mm_unpacklo_epi8(a, b); \
b = _mm_unpackhi_epi8(tmp, b)
// 16-bit interleave step (for transposes)
#define dct_interleave16(a, b) \
tmp = a; \
a = _mm_unpacklo_epi16(a, b); \
b = _mm_unpackhi_epi16(tmp, b)
#define dct_pass(bias,shift) \
{ \
/* even part */ \
dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \
__m128i sum04 = _mm_add_epi16(row0, row4); \
__m128i dif04 = _mm_sub_epi16(row0, row4); \
dct_widen(t0e, sum04); \
dct_widen(t1e, dif04); \
dct_wadd(x0, t0e, t3e); \
dct_wsub(x3, t0e, t3e); \
dct_wadd(x1, t1e, t2e); \
dct_wsub(x2, t1e, t2e); \
/* odd part */ \
dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \
dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \
__m128i sum17 = _mm_add_epi16(row1, row7); \
__m128i sum35 = _mm_add_epi16(row3, row5); \
dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \
dct_wadd(x4, y0o, y4o); \
dct_wadd(x5, y1o, y5o); \
dct_wadd(x6, y2o, y5o); \
dct_wadd(x7, y3o, y4o); \
dct_bfly32o(row0,row7, x0,x7,bias,shift); \
dct_bfly32o(row1,row6, x1,x6,bias,shift); \
dct_bfly32o(row2,row5, x2,x5,bias,shift); \
dct_bfly32o(row3,row4, x3,x4,bias,shift); \
}
__m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
__m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f));
__m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
__m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
__m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f));
__m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f));
__m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f));
__m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f));
// rounding biases in column/row passes, see stbi__idct_block for explanation.
__m128i bias_0 = _mm_set1_epi32(512);
__m128i bias_1 = _mm_set1_epi32(65536 + (128<<17));
// load
row0 = _mm_load_si128((const __m128i *) (data + 0*8));
row1 = _mm_load_si128((const __m128i *) (data + 1*8));
row2 = _mm_load_si128((const __m128i *) (data + 2*8));
row3 = _mm_load_si128((const __m128i *) (data + 3*8));
row4 = _mm_load_si128((const __m128i *) (data + 4*8));
row5 = _mm_load_si128((const __m128i *) (data + 5*8));
row6 = _mm_load_si128((const __m128i *) (data + 6*8));
row7 = _mm_load_si128((const __m128i *) (data + 7*8));
// column pass
dct_pass(bias_0, 10);
{
// 16bit 8x8 transpose pass 1
dct_interleave16(row0, row4);
dct_interleave16(row1, row5);
dct_interleave16(row2, row6);
dct_interleave16(row3, row7);
// transpose pass 2
dct_interleave16(row0, row2);
dct_interleave16(row1, row3);
dct_interleave16(row4, row6);
dct_interleave16(row5, row7);
// transpose pass 3
dct_interleave16(row0, row1);
dct_interleave16(row2, row3);
dct_interleave16(row4, row5);
dct_interleave16(row6, row7);
}
// row pass
dct_pass(bias_1, 17);
{
// pack
__m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
__m128i p1 = _mm_packus_epi16(row2, row3);
__m128i p2 = _mm_packus_epi16(row4, row5);
__m128i p3 = _mm_packus_epi16(row6, row7);
// 8bit 8x8 transpose pass 1
dct_interleave8(p0, p2); // a0e0a1e1...
dct_interleave8(p1, p3); // c0g0c1g1...
// transpose pass 2
dct_interleave8(p0, p1); // a0c0e0g0...
dct_interleave8(p2, p3); // b0d0f0h0...
// transpose pass 3
dct_interleave8(p0, p2); // a0b0c0d0...
dct_interleave8(p1, p3); // a4b4c4d4...
// store
_mm_storel_epi64((__m128i *) out, p0); out += out_stride;
_mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride;
_mm_storel_epi64((__m128i *) out, p2); out += out_stride;
_mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride;
_mm_storel_epi64((__m128i *) out, p1); out += out_stride;
_mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride;
_mm_storel_epi64((__m128i *) out, p3); out += out_stride;
_mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e));
}
#undef dct_const
#undef dct_rot
#undef dct_widen
#undef dct_wadd
#undef dct_wsub
#undef dct_bfly32o
#undef dct_interleave8
#undef dct_interleave16
#undef dct_pass
}
#endif // STBI_SSE2
#ifdef STBI_NEON
// NEON integer IDCT. should produce bit-identical
// results to the generic C version.
static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
{
int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f));
int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f));
int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f));
int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f));
int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f));
int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f));
#define dct_long_mul(out, inq, coeff) \
int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \
int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
#define dct_long_mac(out, acc, inq, coeff) \
int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \
int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
#define dct_widen(out, inq) \
int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \
int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
// wide add
#define dct_wadd(out, a, b) \
int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \
int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
// wide sub
#define dct_wsub(out, a, b) \
int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \
int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
// butterfly a/b, then shift using "shiftop" by "s" and pack
#define dct_bfly32o(out0,out1, a,b,shiftop,s) \
{ \
dct_wadd(sum, a, b); \
dct_wsub(dif, a, b); \
out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \
out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \
}
#define dct_pass(shiftop, shift) \
{ \
/* even part */ \
int16x8_t sum26 = vaddq_s16(row2, row6); \
dct_long_mul(p1e, sum26, rot0_0); \
dct_long_mac(t2e, p1e, row6, rot0_1); \
dct_long_mac(t3e, p1e, row2, rot0_2); \
int16x8_t sum04 = vaddq_s16(row0, row4); \
int16x8_t dif04 = vsubq_s16(row0, row4); \
dct_widen(t0e, sum04); \
dct_widen(t1e, dif04); \
dct_wadd(x0, t0e, t3e); \
dct_wsub(x3, t0e, t3e); \
dct_wadd(x1, t1e, t2e); \
dct_wsub(x2, t1e, t2e); \
/* odd part */ \
int16x8_t sum15 = vaddq_s16(row1, row5); \
int16x8_t sum17 = vaddq_s16(row1, row7); \
int16x8_t sum35 = vaddq_s16(row3, row5); \
int16x8_t sum37 = vaddq_s16(row3, row7); \
int16x8_t sumodd = vaddq_s16(sum17, sum35); \
dct_long_mul(p5o, sumodd, rot1_0); \
dct_long_mac(p1o, p5o, sum17, rot1_1); \
dct_long_mac(p2o, p5o, sum35, rot1_2); \
dct_long_mul(p3o, sum37, rot2_0); \
dct_long_mul(p4o, sum15, rot2_1); \
dct_wadd(sump13o, p1o, p3o); \
dct_wadd(sump24o, p2o, p4o); \
dct_wadd(sump23o, p2o, p3o); \
dct_wadd(sump14o, p1o, p4o); \
dct_long_mac(x4, sump13o, row7, rot3_0); \
dct_long_mac(x5, sump24o, row5, rot3_1); \
dct_long_mac(x6, sump23o, row3, rot3_2); \
dct_long_mac(x7, sump14o, row1, rot3_3); \
dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \
dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \
dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \
dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \
}
// load
row0 = vld1q_s16(data + 0*8);
row1 = vld1q_s16(data + 1*8);
row2 = vld1q_s16(data + 2*8);
row3 = vld1q_s16(data + 3*8);
row4 = vld1q_s16(data + 4*8);
row5 = vld1q_s16(data + 5*8);
row6 = vld1q_s16(data + 6*8);
row7 = vld1q_s16(data + 7*8);
// add DC bias
row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
// column pass
dct_pass(vrshrn_n_s32, 10);
// 16bit 8x8 transpose
{
// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
// whether compilers actually get this is another story, sadly.
#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; }
#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); }
#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); }
// pass 1
dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
dct_trn16(row2, row3);
dct_trn16(row4, row5);
dct_trn16(row6, row7);
// pass 2
dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
dct_trn32(row1, row3);
dct_trn32(row4, row6);
dct_trn32(row5, row7);
// pass 3
dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
dct_trn64(row1, row5);
dct_trn64(row2, row6);
dct_trn64(row3, row7);
#undef dct_trn16
#undef dct_trn32
#undef dct_trn64
}
// row pass
// vrshrn_n_s32 only supports shifts up to 16, we need
// 17. so do a non-rounding shift of 16 first then follow
// up with a rounding shift by 1.
dct_pass(vshrn_n_s32, 16);
{
// pack and round
uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
// again, these can translate into one instruction, but often don't.
#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; }
#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); }
#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); }
// sadly can't use interleaved stores here since we only write
// 8 bytes to each scan line!
// 8x8 8-bit transpose pass 1
dct_trn8_8(p0, p1);
dct_trn8_8(p2, p3);
dct_trn8_8(p4, p5);
dct_trn8_8(p6, p7);
// pass 2
dct_trn8_16(p0, p2);
dct_trn8_16(p1, p3);
dct_trn8_16(p4, p6);
dct_trn8_16(p5, p7);
// pass 3
dct_trn8_32(p0, p4);
dct_trn8_32(p1, p5);
dct_trn8_32(p2, p6);
dct_trn8_32(p3, p7);
// store
vst1_u8(out, p0); out += out_stride;
vst1_u8(out, p1); out += out_stride;
vst1_u8(out, p2); out += out_stride;
vst1_u8(out, p3); out += out_stride;
vst1_u8(out, p4); out += out_stride;
vst1_u8(out, p5); out += out_stride;
vst1_u8(out, p6); out += out_stride;
vst1_u8(out, p7);
#undef dct_trn8_8
#undef dct_trn8_16
#undef dct_trn8_32
}
#undef dct_long_mul
#undef dct_long_mac
#undef dct_widen
#undef dct_wadd
#undef dct_wsub
#undef dct_bfly32o
#undef dct_pass
}
#endif // STBI_NEON
#define STBI__MARKER_none 0xff
// if there's a pending marker from the entropy stream, return that
// otherwise, fetch from the stream and get a marker. if there's no
// marker, return 0xff, which is never a valid marker value
static stbi_uc stbi__get_marker(stbi__jpeg *j)
{
stbi_uc x;
if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; }
x = stbi__get8(j->s);
if (x != 0xff) return STBI__MARKER_none;
while (x == 0xff)
x = stbi__get8(j->s);
return x;
}
// in each scan, we'll have scan_n components, and the order
// of the components is specified by order[]
#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7)
// after a restart interval, stbi__jpeg_reset the entropy decoder and
// the dc prediction
static void stbi__jpeg_reset(stbi__jpeg *j)
{
j->code_bits = 0;
j->code_buffer = 0;
j->nomore = 0;
j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = 0;
j->marker = STBI__MARKER_none;
j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
j->eob_run = 0;
// no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
// since we don't even allow 1<<30 pixels
}
static int stbi__parse_entropy_coded_data(stbi__jpeg *z)
{
stbi__jpeg_reset(z);
if (!z->progressive) {
if (z->scan_n == 1) {
int i,j;
STBI_SIMD_ALIGN(short, data[64]);
int n = z->order[0];
// non-interleaved data, we just need to process one block at a time,
// in trivial scanline order
// number of blocks to do just depends on how many actual "pixels" this
// component has, independent of interleaved MCU blocking and such
int w = (z->img_comp[n].x+7) >> 3;
int h = (z->img_comp[n].y+7) >> 3;
for (j=0; j < h; ++j) {
for (i=0; i < w; ++i) {
int ha = z->img_comp[n].ha;
if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
// every data block is an MCU, so countdown the restart interval
if (--z->todo <= 0) {
if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
// if it's NOT a restart, then just bail, so we get corrupt data
// rather than no data
if (!STBI__RESTART(z->marker)) return 1;
stbi__jpeg_reset(z);
}
}
}
return 1;
} else { // interleaved
int i,j,k,x,y;
STBI_SIMD_ALIGN(short, data[64]);
for (j=0; j < z->img_mcu_y; ++j) {
for (i=0; i < z->img_mcu_x; ++i) {
// scan an interleaved mcu... process scan_n components in order
for (k=0; k < z->scan_n; ++k) {
int n = z->order[k];
// scan out an mcu's worth of this component; that's just determined
// by the basic H and V specified for the component
for (y=0; y < z->img_comp[n].v; ++y) {
for (x=0; x < z->img_comp[n].h; ++x) {
int x2 = (i*z->img_comp[n].h + x)*8;
int y2 = (j*z->img_comp[n].v + y)*8;
int ha = z->img_comp[n].ha;
if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data);
}
}
}
// after all interleaved components, that's an interleaved MCU,
// so now count down the restart interval
if (--z->todo <= 0) {
if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
if (!STBI__RESTART(z->marker)) return 1;
stbi__jpeg_reset(z);
}
}