Merge "Replace parentheses with curly braces to fix vector doc bug."
diff --git a/cpp/rsDispatch.cpp b/cpp/rsDispatch.cpp
index f09f539..10adcf2 100644
--- a/cpp/rsDispatch.cpp
+++ b/cpp/rsDispatch.cpp
@@ -411,6 +411,11 @@
LOG_API("Couldn't initialize dispatchTab.Allocation3DRead");
return false;
}
+ dispatchTab.ScriptForEachMulti = (ScriptForEachMultiFnPtr)dlsym(handle, "rsScriptForEachMulti");
+ if (dispatchTab.ScriptForEachMulti == NULL) {
+ LOG_API("Couldn't initialize dispatchTab.ScriptForEachMulti");
+ return false;
+ }
}
return true;
diff --git a/cpp/rsDispatch.h b/cpp/rsDispatch.h
index d77a4d9..740457a 100644
--- a/cpp/rsDispatch.h
+++ b/cpp/rsDispatch.h
@@ -93,6 +93,7 @@
typedef void (*ScriptGroupSetOutputFnPtr) (RsContext, RsScriptGroup, RsScriptKernelID, RsAllocation);
typedef void (*ScriptGroupSetInputFnPtr) (RsContext, RsScriptGroup, RsScriptKernelID, RsAllocation);
typedef void (*ScriptGroupExecuteFnPtr) (RsContext, RsScriptGroup);
+typedef void (*ScriptForEachMultiFnPtr) (RsContext, RsScript, uint32_t, RsAllocation *, size_t, RsAllocation, const void *, size_t, const RsScriptCall *, size_t);
typedef void (*AllocationIoSendFnPtr) (RsContext, RsAllocation);
typedef void (*AllocationIoReceiveFnPtr) (RsContext, RsAllocation);
typedef void * (*AllocationGetPointerFnPtr) (RsContext, RsAllocation, uint32_t lod, RsAllocationCubemapFace face, uint32_t z, uint32_t array, size_t *stride, size_t stride_len);
@@ -176,6 +177,7 @@
ScriptGroupSetOutputFnPtr ScriptGroupSetOutput;
ScriptGroupSetInputFnPtr ScriptGroupSetInput;
ScriptGroupExecuteFnPtr ScriptGroupExecute;
+ ScriptForEachMultiFnPtr ScriptForEachMulti;
AllocationIoSendFnPtr AllocationIoSend;
AllocationIoReceiveFnPtr AllocationIoReceive;
AllocationGetPointerFnPtr AllocationGetPointer;
diff --git a/cpu_ref/rsCpuCore.cpp b/cpu_ref/rsCpuCore.cpp
index 5fe7801..911a0f0 100644
--- a/cpu_ref/rsCpuCore.cpp
+++ b/cpu_ref/rsCpuCore.cpp
@@ -646,11 +646,9 @@
case RS_SCRIPT_INTRINSIC_ID_RESIZE:
i = rsdIntrinsic_Resize(this, s, e);
break;
-#if !defined(RS_COMPATIBILITY_LIB)
case RS_SCRIPT_INTRINSIC_ID_BLAS:
i = rsdIntrinsic_BLAS(this, s, e);
break;
-#endif
default:
rsAssert(0);
diff --git a/cpu_ref/rsCpuExecutable.cpp b/cpu_ref/rsCpuExecutable.cpp
index e2c27b5..867a2cd 100644
--- a/cpu_ref/rsCpuExecutable.cpp
+++ b/cpu_ref/rsCpuExecutable.cpp
@@ -20,27 +20,6 @@
namespace {
-// Create a len length string containing random characters from [A-Za-z0-9].
-static std::string getRandomString(size_t len) {
- char buf[len + 1];
- for (size_t i = 0; i < len; i++) {
- uint32_t r = arc4random() & 0xffff;
- r %= 62;
- if (r < 26) {
- // lowercase
- buf[i] = 'a' + r;
- } else if (r < 52) {
- // uppercase
- buf[i] = 'A' + (r - 26);
- } else {
- // Use a number
- buf[i] = '0' + (r - 52);
- }
- }
- buf[len] = '\0';
- return std::string(buf);
-}
-
// Check if a path exists and attempt to create it if it doesn't.
static bool ensureCacheDirExists(const char *path) {
if (access(path, R_OK | W_OK | X_OK) == 0) {
@@ -151,7 +130,8 @@
void* SharedLibraryUtils::loadSharedLibrary(const char *cacheDir,
const char *resName,
- const char *nativeLibDir) {
+ const char *nativeLibDir,
+ bool* alreadyLoaded) {
void *loaded = nullptr;
#if defined(RS_COMPATIBILITY_LIB) && defined(__LP64__)
@@ -162,7 +142,7 @@
// We should check if we can load the library from the standard app
// location for shared libraries first.
- loaded = loadSOHelper(scriptSOName.c_str(), cacheDir, resName);
+ loaded = loadSOHelper(scriptSOName.c_str(), cacheDir, resName, alreadyLoaded);
if (loaded == nullptr) {
ALOGE("Unable to open shared library (%s): %s",
@@ -189,8 +169,28 @@
return loaded;
}
+String8 SharedLibraryUtils::getRandomString(size_t len) {
+ char buf[len + 1];
+ for (size_t i = 0; i < len; i++) {
+ uint32_t r = arc4random() & 0xffff;
+ r %= 62;
+ if (r < 26) {
+ // lowercase
+ buf[i] = 'a' + r;
+ } else if (r < 52) {
+ // uppercase
+ buf[i] = 'A' + (r - 26);
+ } else {
+ // Use a number
+ buf[i] = '0' + (r - 52);
+ }
+ }
+ buf[len] = '\0';
+ return String8(buf);
+}
+
void* SharedLibraryUtils::loadSOHelper(const char *origName, const char *cacheDir,
- const char *resName) {
+ const char *resName, bool *alreadyLoaded) {
// Keep track of which .so libraries have been loaded. Once a library is
// in the set (per-process granularity), we must instead make a copy of
// the original shared object (randomly named .so file) and load that one
@@ -208,6 +208,9 @@
// Common path is that we have not loaded this Script/library before.
if (LoadedLibraries.find(origName) == LoadedLibraries.end()) {
+ if (alreadyLoaded != nullptr) {
+ *alreadyLoaded = false;
+ }
loaded = dlopen(origName, RTLD_NOW | RTLD_LOCAL);
if (loaded) {
LoadedLibraries.insert(origName);
@@ -215,6 +218,10 @@
return loaded;
}
+ if (alreadyLoaded != nullptr) {
+ *alreadyLoaded = true;
+ }
+
std::string newName(cacheDir);
// Append RS_CACHE_DIR only if it is not found in cacheDir
@@ -234,7 +241,7 @@
newName.append("librs.");
newName.append(resName);
newName.append("#");
- newName.append(getRandomString(6)); // 62^6 potential filename variants.
+ newName.append(getRandomString(6).string()); // 62^6 potential filename variants.
newName.append(".so");
int r = copyFile(newName.c_str(), origName);
diff --git a/cpu_ref/rsCpuExecutable.h b/cpu_ref/rsCpuExecutable.h
index 0464dac..6880970 100644
--- a/cpu_ref/rsCpuExecutable.h
+++ b/cpu_ref/rsCpuExecutable.h
@@ -43,14 +43,18 @@
// For 64bit RS Support Lib, the shared lib path cannot be constructed from
// cacheDir, so nativeLibDir is needed to load shared libs.
static void* loadSharedLibrary(const char *cacheDir, const char *resName,
- const char *nativeLibDir = nullptr);
+ const char *nativeLibDir = nullptr,
+ bool *alreadyLoaded = nullptr);
+
+ // Create a len length string containing random characters from [A-Za-z0-9].
+ static String8 getRandomString(size_t len);
private:
// Attempt to load the shared library from origName, but then fall back to
// creating a copy of the shared library if necessary (to ensure instancing).
// This function returns the dlopen()-ed handle if successful.
static void *loadSOHelper(const char *origName, const char *cacheDir,
- const char *resName);
+ const char *resName, bool* alreadyLoaded = nullptr);
static const char* LD_EXE_PATH;
static const char* RS_CACHE_DIR;
diff --git a/cpu_ref/rsCpuIntrinsicBLAS.cpp b/cpu_ref/rsCpuIntrinsicBLAS.cpp
index 3034454..a7705b4 100644
--- a/cpu_ref/rsCpuIntrinsicBLAS.cpp
+++ b/cpu_ref/rsCpuIntrinsicBLAS.cpp
@@ -47,10 +47,10 @@
uint8_t c_offset = 0;
static void kernelBNNM(size_t m, size_t n, size_t k,
- const uint8_t* a, uint32_t a_offset, size_t lda,
- const uint8_t* b, uint32_t b_offset, size_t ldb,
- uint8_t* c, uint32_t c_offset, size_t ldc,
- uint32_t c_mult_int);
+ const uint8_t* a, uint8_t a_offset, size_t lda,
+ const uint8_t* b, uint8_t b_offset, size_t ldb,
+ uint8_t* c, int32_t c_offset, size_t ldc,
+ int32_t c_mult_int);
@@ -653,10 +653,10 @@
}
void RsdCpuScriptIntrinsicBLAS::kernelBNNM(size_t m, size_t n, size_t k,
- const uint8_t* a, uint32_t a_offset, size_t lda,
- const uint8_t* b, uint32_t b_offset, size_t ldb,
- uint8_t* c, uint32_t c_offset, size_t ldc,
- uint32_t c_mult_int) {
+ const uint8_t* a, uint8_t a_offset, size_t lda,
+ const uint8_t* b, uint8_t b_offset, size_t ldb,
+ uint8_t* c, int32_t c_offset, size_t ldc,
+ int32_t c_mult_int) {
// Calculations are done in 1.10.21 fixed-point format for the final output,
// just before there's a shift down to drop the fractional parts. The output
// values are gated to 0 to 255 to fit in a byte, but the 10-bit format
diff --git a/cpu_ref/rsCpuIntrinsicBlend.cpp b/cpu_ref/rsCpuIntrinsicBlend.cpp
index d546bbd..ca40c5e 100644
--- a/cpu_ref/rsCpuIntrinsicBlend.cpp
+++ b/cpu_ref/rsCpuIntrinsicBlend.cpp
@@ -121,7 +121,9 @@
uint32_t x2 = xend;
#if defined(ARCH_ARM_USE_INTRINSICS) && !defined(ARCH_ARM64_USE_INTRINSICS)
- if (gArchUseSIMD) {
+ // Bug: 22047392 - Skip optimized version for BLEND_DST_ATOP until this
+ // been fixed.
+ if (gArchUseSIMD && info->slot != BLEND_DST_ATOP) {
if (rsdIntrinsicBlend_K(out, in, info->slot, x1, x2) >= 0)
return;
}
@@ -271,6 +273,9 @@
}
break;
case BLEND_DST_ATOP:
+ // Bug: 22047392 - We need to make sure that "out->w = in->w;" in all
+ // accelerated versions before re-enabling optimizations.
+ #if false // Bug: 22047392
#if defined(ARCH_X86_HAVE_SSSE3)
if (gArchUseSIMD) {
if ((x1 + 8) < x2) {
@@ -282,11 +287,13 @@
}
}
#endif
+ #endif // false for Bug: 22047392
for (;x1 < x2; x1++, out++, in++) {
short4 in_s = convert_short4(*in);
short4 out_s = convert_short4(*out);
out_s.xyz = ((out_s.xyz * in_s.w) +
(in_s.xyz * ((short3)255 - (short3)out_s.w))) >> (short3)8;
+ out_s.w = in_s.w;
*out = convert_uchar4(out_s);
}
break;
diff --git a/cpu_ref/rsCpuScript.cpp b/cpu_ref/rsCpuScript.cpp
index b95d8f7..1909e13 100644
--- a/cpu_ref/rsCpuScript.cpp
+++ b/cpu_ref/rsCpuScript.cpp
@@ -862,10 +862,8 @@
}
RsdCpuScriptImpl::~RsdCpuScriptImpl() {
- if (mScriptExec != nullptr) {
- delete mScriptExec;
- }
- if (mBoundAllocs) delete[] mBoundAllocs;
+ delete mScriptExec;
+ delete[] mBoundAllocs;
if (mScriptSO) {
dlclose(mScriptSO);
}
diff --git a/cpu_ref/rsCpuScriptGroup2.cpp b/cpu_ref/rsCpuScriptGroup2.cpp
index 8923c65..bf01403 100644
--- a/cpu_ref/rsCpuScriptGroup2.cpp
+++ b/cpu_ref/rsCpuScriptGroup2.cpp
@@ -156,6 +156,7 @@
mExecutable(nullptr), mScriptObj(nullptr) {
rsAssert(!mGroup->mClosures.empty());
+ mCpuRefImpl->lockMutex();
Batch* batch = new Batch(this, "Batch0");
int i = 0;
for (Closure* closure: mGroup->mClosures) {
@@ -192,6 +193,7 @@
}
}
#endif // RS_COMPATIBILITY_LIB
+ mCpuRefImpl->unlockMutex();
}
void Batch::resolveFuncPtr(void* sharedObj) {
@@ -279,6 +281,9 @@
}
args->push_back("-output_path");
args->push_back(outputDir);
+
+ // The output filename has to be the last, in case we need to pop it out and
+ // replace with a different name.
args->push_back("-o");
args->push_back(outputFileName);
}
@@ -396,15 +401,41 @@
// Try to load a shared lib from code cache matching filename and checksum
//===--------------------------------------------------------------------===//
- mScriptObj = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
+ bool alreadyLoaded = false;
+ std::string cloneName;
+
+ mScriptObj = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName, nullptr,
+ &alreadyLoaded);
if (mScriptObj != nullptr) {
+ // A shared library named resName is found in code cache directory
+ // cacheDir, and loaded with the handle stored in mScriptObj.
+
mExecutable = ScriptExecutable::createFromSharedObject(
getCpuRefImpl()->getContext(), mScriptObj, checksum);
+
if (mExecutable != nullptr) {
+ // The loaded shared library in mScriptObj has a matching checksum.
+ // An executable object has been created.
return;
- } else {
- ALOGE("Failed to create an executable object from so file");
}
+
+ ALOGV("Failed to create an executable object from so file due to "
+ "mismatching checksum");
+
+ if (alreadyLoaded) {
+ // The shared object found in code cache has already been loaded.
+ // A different file name is needed for the new shared library, to
+ // avoid corrupting the currently loaded instance.
+
+ cloneName.append(resName);
+ cloneName.append("#");
+ cloneName.append(SharedLibraryUtils::getRandomString(6).string());
+
+ // The last element in arguments is the output filename.
+ arguments.pop_back();
+ arguments.push_back(cloneName.c_str());
+ }
+
dlclose(mScriptObj);
mScriptObj = nullptr;
}
@@ -443,6 +474,16 @@
return;
}
+ if (alreadyLoaded) {
+ // Delete the temporary, random-named file that we created to avoid
+ // interfering with an already loaded shared library.
+ string cloneFilePath(cacheDir);
+ cloneFilePath.append("/");
+ cloneFilePath.append(cloneName.c_str());
+ cloneFilePath.append(".so");
+ unlink(cloneFilePath.c_str());
+ }
+
mExecutable = ScriptExecutable::createFromSharedObject(
getCpuRefImpl()->getContext(),
mScriptObj);
diff --git a/rsDefines.h b/rsDefines.h
index 839520f..6da672f 100644
--- a/rsDefines.h
+++ b/rsDefines.h
@@ -458,10 +458,10 @@
int incY;
int KL;
int KU;
- uint32_t a_offset;
- uint32_t b_offset;
- uint32_t c_offset;
- uint32_t c_mult_int;
+ uint8_t a_offset;
+ uint8_t b_offset;
+ int32_t c_offset;
+ int32_t c_mult_int;
} RsBlasCall;
enum RsGlobalProperty {
diff --git a/rsFileA3D.cpp b/rsFileA3D.cpp
index a81d0f9..a7b1e27 100644
--- a/rsFileA3D.cpp
+++ b/rsFileA3D.cpp
@@ -54,7 +54,7 @@
delete mWriteStream;
}
if (mReadStream) {
- delete mWriteStream;
+ delete mReadStream;
}
if (mAlloc) {
free(mAlloc);
