AI/rocm-systems
2
0
포크 0
코드 이슈 풀 리퀘스트 액션 패키지 프로젝트 릴리즈 위키 활동

P4 to Git Change 1501660 by gandryey@gera-w8 on 2018/01/09 14:04:56

소스 검색 
SWDEV-79445 - OCL generic changes and code clean-up
	- Code style clean-up. No functional changes.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#127 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#164 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#321 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#236 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#412 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.hpp#142 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcompiler.cpp#20 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#70 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevicegl.cpp#9 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#18 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprintf.cpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#53 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#45 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#69 edit


[ROCm/clr commit: e5ee74a086]
이 커밋은 다음에 포함됨:
foreman
2018-01-09 14:09:12 -05:00
부모 c29b086c23
커밋 6111e38160
17개의 변경된 파일과 82개의 추가작업 그리고 285개의 파일을 삭제
패치 파일 다운로드 Diff 파일 다운로드 모든 파일 펼침 모든 파일 접기
projects/clr/rocclr/runtime/device/gpu/gpublit.cpp
+8 -14
파일 보기
@@ -261,7 +261,6 @@ bool DmaBlitManager::writeMemoryStaged(const void* srcHost, Memory& dstMemory, M
size_t origin, size_t& offset, size_t& totalSize,
size_t xferSize) const {
amd::Coord3D src(0, 0, 0);
size_t tmpSize;
size_t chunkSize;
static const bool CopyRect = false;
// Flush DMA for ASYNC copy
@@ -278,7 +277,7 @@ bool DmaBlitManager::writeMemoryStaged(const void* srcHost, Memory& dstMemory, M
while (xferSize != 0) {
// Find the partial transfer size
tmpSize = std::min(chunkSize, xferSize);
size_t tmpSize = std::min(chunkSize, xferSize);
amd::Coord3D dst(origin + offset, 0, 0);
amd::Coord3D copySize(tmpSize, 0, 0);
@@ -1717,11 +1716,10 @@ bool KernelBlitManager::copyBufferRect(device::Memory& srcMemory, device::Memory
const static uint CopyRectAlignment[3] = {16, 4, 1};
bool aligned;
uint i;
for (i = 0; i < sizeof(CopyRectAlignment) / sizeof(uint); i++) {
// Check source alignments
aligned = ((srcRectIn.rowPitch_ % CopyRectAlignment[i]) == 0);
bool aligned = ((srcRectIn.rowPitch_ % CopyRectAlignment[i]) == 0);
aligned &= ((srcRectIn.slicePitch_ % CopyRectAlignment[i]) == 0);
aligned &= ((srcRectIn.start_ % CopyRectAlignment[i]) == 0);
@@ -2310,9 +2308,7 @@ bool KernelBlitManager::fillImage(device::Memory& memory, const void* pattern,
bool KernelBlitManager::runScheduler(device::Memory& vqueue, device::Memory& params, uint paramIdx,
uint threads) const {
amd::ScopedLock k(lockXferOps_);
bool result = false;
size_t dim = 1;
size_t globalWorkOffset[1] = {0};
size_t globalWorkSize[1] = {threads};
size_t localWorkSize[1] = {1};
@@ -2329,7 +2325,7 @@ bool KernelBlitManager::runScheduler(device::Memory& vqueue, device::Memory& par
// Execute the blit
address parameters = kernels_[Scheduler]->parameters().values();
result = gpu().submitKernelInternal(ndrange, *kernels_[Scheduler], parameters);
bool result = gpu().submitKernelInternal(ndrange, *kernels_[Scheduler], parameters);
synchronize();
@@ -2385,15 +2381,13 @@ amd::Memory* DmaBlitManager::pinHostMemory(const void* hostMem, size_t pinSize,
Memory* KernelBlitManager::createView(const Memory& parent, const CalFormat& format) const {
assert(!parent.cal()->buffer_ && "View supports images only");
gpu::Memory* gpuImage = NULL;
gpuImage = new gpu::Image(dev(), parent.size(), parent.cal()->width_, parent.cal()->height_,
parent.cal()->depth_, format.type_, format.channelOrder_,
parent.cal()->imageType_, 1);
gpu::Memory* gpuImage = new gpu::Image(
dev(), parent.size(), parent.cal()->width_, parent.cal()->height_,
parent.cal()->depth_, format.type_, format.channelOrder_,
parent.cal()->imageType_, 1);
// Create resource
if (NULL != gpuImage) {
bool result = false;
Resource::ImageViewParams params;
const Memory& gpuMem = static_cast<const Memory&>(parent);
@@ -2405,7 +2399,7 @@ Memory* KernelBlitManager::createView(const Memory& parent, const CalFormat& for
params.gpu_ = &gpu();
// Create memory object
result = gpuImage->create(Resource::ImageView, &params);
bool result = gpuImage->create(Resource::ImageView, &params);
if (!result) {
delete gpuImage;
return NULL;
projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp
+2 -2
파일 보기
@@ -300,7 +300,7 @@ class Device : public NullDevice, public CALGSLDevice {
size_t bufSize_; //!< Staged buffer size
std::list<Memory*> freeBuffers_; //!< The list of free buffers
amd::Atomic<uint> acquiredCnt_; //!< The total number of acquired buffers
amd::Monitor lock_; //!< Stgaed buffer acquire/release lock
amd::Monitor lock_; //!< Staged buffer acquire/release lock
const Device& gpuDevice_; //!< GPU device object
};
@@ -311,7 +311,7 @@ class Device : public NullDevice, public CALGSLDevice {
uint64_t size_; //!< Scratch buffer size on this queue
//! Default constructor
ScratchBuffer() : regNum_(0), memObj_(NULL), offset_(0) {}
ScratchBuffer() : regNum_(0), memObj_(NULL), offset_(0), size_(0) {}
//! Default constructor
~ScratchBuffer();
projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp
+2 -6
파일 보기
@@ -1031,7 +1031,6 @@ void Kernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkSize,
if (workGroupInfo()->compileSize_[0] == 0) {
// Find the default local workgroup size, if it wasn't specified
if (lclWorkSize[0] == 0) {
size_t thrPerGrp;
bool b1DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE);
bool b2DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_X) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_Y);
@@ -1043,7 +1042,7 @@ void Kernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkSize,
((workDim == 3) && b3DOverrideSet);
if (!overrideSet) {
// Find threads per group
thrPerGrp = workGroupInfo()->size_;
size_t thrPerGrp = workGroupInfo()->size_;
// Check if kernel uses images
if ((flags() & ImageEnable) &&
@@ -1376,8 +1375,6 @@ bool Kernel::bindConstantBuffers(VirtualGPU& gpu) const {
void Kernel::processMemObjects(VirtualGPU& gpu, const amd::Kernel& kernel, const_address params,
bool nativeMem) const {
VirtualGPU::MemoryDependency& dependecy = gpu.memoryDependency();
// Mark the tracker with a new kernel,
// so we can avoid checks of the aliased objects
gpu.memoryDependency().newKernel();
@@ -3323,7 +3320,6 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
if (workGroupInfo()->compileSize_[0] == 0) {
// Find the default local workgroup size, if it wasn't specified
if (lclWorkSize[0] == 0) {
size_t thrPerGrp;
bool b1DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE);
bool b2DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_X) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_Y);
@@ -3335,7 +3331,7 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
((workDim == 3) && b3DOverrideSet);
if (!overrideSet) {
// Find threads per group
thrPerGrp = workGroupInfo()->size_;
size_t thrPerGrp = workGroupInfo()->size_;
// Check if kernel uses images
if (flags_.imageEna_ &&
projects/clr/rocclr/runtime/device/gpu/gpuprogram.cpp
+2 -2
파일 보기
@@ -2032,11 +2032,11 @@ bool HSAILProgram::linkImpl(amd::option::Options* options) {
&kernelNamesSize);
if (errorCode != ACL_SUCCESS) {
buildLog_ += "Error: Querying of kernel names from the binary failed.\n";
delete kernelNames;
delete [] kernelNames;
return false;
}
std::vector<std::string> vKernels = splitSpaceSeparatedString(kernelNames);
delete kernelNames;
delete [] kernelNames;
std::vector<std::string>::iterator it = vKernels.begin();
bool dynamicParallelism = false;
aclMetadata md;
projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp
+4 -4
파일 보기
@@ -1366,7 +1366,7 @@ void VirtualGPU::submitMigrateMemObjects(amd::MigrateMemObjectsCommand& vcmd) {
profilingBegin(vcmd, true);
std::vector<amd::Memory*>::const_iterator itr;
for (itr = vcmd.memObjects().begin(); itr != vcmd.memObjects().end(); itr++) {
for (itr = vcmd.memObjects().begin(); itr != vcmd.memObjects().end(); ++itr) {
// Find device memory
gpu::Memory* memory = dev().getGpuMemory(*itr);
@@ -1395,7 +1395,7 @@ void VirtualGPU::submitSvmFreeMemory(amd::SvmFreeMemoryCommand& vcmd) {
std::vector<void*>& svmPointers = vcmd.svmPointers();
if (vcmd.pfnFreeFunc() == NULL) {
// pointers allocated using clSVMAlloc
for (cl_uint i = 0; i < svmPointers.size(); i++) {
for (cl_uint i = 0; i < svmPointers.size(); ++i) {
dev().svmFree(svmPointers[i]);
}
} else {
@@ -2298,7 +2298,7 @@ void VirtualGPU::submitAcquireExtObjects(amd::AcquireExtObjectsCommand& vcmd) {
profilingBegin(vcmd);
for (std::vector<amd::Memory*>::const_iterator it = vcmd.getMemList().begin();
it != vcmd.getMemList().end(); it++) {
it != vcmd.getMemList().end(); ++it) {
// amd::Memory object should never be NULL
assert(*it && "Memory object for interop is NULL");
gpu::Memory* memory = dev().getGpuMemory(*it);
@@ -2337,7 +2337,7 @@ void VirtualGPU::submitReleaseExtObjects(amd::ReleaseExtObjectsCommand& vcmd) {
profilingBegin(vcmd);
for (std::vector<amd::Memory*>::const_iterator it = vcmd.getMemList().begin();
it != vcmd.getMemList().end(); it++) {
it != vcmd.getMemList().end(); ++it) {
// amd::Memory object should never be NULL
assert(*it && "Memory object for interop is NULL");
gpu::Memory* memory = dev().getGpuMemory(*it);
projects/clr/rocclr/runtime/device/gpu/gpuvirtual.hpp
+2 -2
파일 보기
@@ -120,7 +120,7 @@ class VirtualGPU : public device::VirtualDevice, public CALGSLContext {
public:
//! Default constructor
MemoryDependency()
: memObjectsInQueue_(NULL), numMemObjectsInQueue_(0), maxMemObjectsInQueue_(0) {}
: memObjectsInQueue_(NULL), endMemObjectsInQueue_(0), numMemObjectsInQueue_(0), maxMemObjectsInQueue_(0) {}
~MemoryDependency() { delete[] memObjectsInQueue_; }
@@ -183,7 +183,7 @@ class VirtualGPU : public device::VirtualDevice, public CALGSLContext {
typedef std::vector<ResourceSlot> ResourceSlots;
public:
VirtualGPU(Device& device);
explicit VirtualGPU(Device& device);
bool create(bool profiling, uint rtCUs = amd::CommandQueue::RealTimeDisabled,
uint deviceQueueSize = 0,
amd::CommandQueue::Priority priority = amd::CommandQueue::Priority::Normal);
projects/clr/rocclr/runtime/device/pal/palblit.cpp
+8 -17
파일 보기
@@ -197,7 +197,6 @@ bool DmaBlitManager::readBufferRect(device::Memory& srcMemory, void* dstHost,
Memory& xferBuf = dev().xferRead().acquire();
amd::Coord3D dst(0, 0, 0);
size_t tmpSize = 0;
size_t bufOffset;
size_t hostOffset;
size_t srcSize;
@@ -210,7 +209,7 @@ bool DmaBlitManager::readBufferRect(device::Memory& srcMemory, void* dstHost,
while (srcSize != 0) {
// Find the partial transfer size
tmpSize = std::min(dev().xferRead().bufSize(), srcSize);
size_t tmpSize = std::min(dev().xferRead().bufSize(), srcSize);
amd::Coord3D src(bufOffset, 0, 0);
amd::Coord3D copySize(tmpSize, 0, 0);
@@ -258,7 +257,6 @@ bool DmaBlitManager::writeMemoryStaged(const void* srcHost, Memory& dstMemory, M
size_t origin, size_t& offset, size_t& totalSize,
size_t xferSize) const {
amd::Coord3D src(0, 0, 0);
size_t tmpSize;
size_t chunkSize;
static const bool CopyRect = false;
// Flush DMA for ASYNC copy
@@ -275,7 +273,7 @@ bool DmaBlitManager::writeMemoryStaged(const void* srcHost, Memory& dstMemory, M
while (xferSize != 0) {
// Find the partial transfer size
tmpSize = std::min(chunkSize, xferSize);
size_t tmpSize = std::min(chunkSize, xferSize);
amd::Coord3D dst(origin + offset, 0, 0);
amd::Coord3D copySize(tmpSize, 0, 0);
@@ -1711,11 +1709,10 @@ bool KernelBlitManager::copyBufferRect(device::Memory& srcMemory, device::Memory
const static uint CopyRectAlignment[3] = {16, 4, 1};
bool aligned;
uint i;
for (i = 0; i < sizeof(CopyRectAlignment) / sizeof(uint); i++) {
// Check source alignments
aligned = ((srcRectIn.rowPitch_ % CopyRectAlignment[i]) == 0);
bool aligned = ((srcRectIn.rowPitch_ % CopyRectAlignment[i]) == 0);
aligned &= ((srcRectIn.slicePitch_ % CopyRectAlignment[i]) == 0);
aligned &= ((srcRectIn.start_ % CopyRectAlignment[i]) == 0);
@@ -2064,11 +2061,10 @@ bool KernelBlitManager::copyBuffer(device::Memory& srcMemory, device::Memory& ds
const static uint CopyBuffAlignment[3] = {16, 4, 1};
amd::Coord3D size(sizeIn[0], sizeIn[1], sizeIn[2]);
bool aligned;
uint i;
for (i = 0; i < sizeof(CopyBuffAlignment) / sizeof(uint); i++) {
// Check source alignments
aligned = ((srcOrigin[0] % CopyBuffAlignment[i]) == 0);
bool aligned = ((srcOrigin[0] % CopyBuffAlignment[i]) == 0);
// Check destination alignments
aligned &= ((dstOrigin[0] % CopyBuffAlignment[i]) == 0);
// Check copy size alignment in the first dimension
@@ -2283,9 +2279,7 @@ bool KernelBlitManager::fillImage(device::Memory& memory, const void* pattern,
bool KernelBlitManager::runScheduler(device::Memory& vqueue, device::Memory& params, uint paramIdx,
uint threads) const {
amd::ScopedLock k(lockXferOps_);
bool result = false;
size_t dim = 1;
size_t globalWorkOffset[1] = {0};
size_t globalWorkSize[1] = {threads};
size_t localWorkSize[1] = {1};
@@ -2302,7 +2296,7 @@ bool KernelBlitManager::runScheduler(device::Memory& vqueue, device::Memory& par
// Execute the blit
address parameters = kernels_[Scheduler]->parameters().values();
result = gpu().submitKernelInternal(ndrange, *kernels_[Scheduler], parameters);
bool result = gpu().submitKernelInternal(ndrange, *kernels_[Scheduler], parameters);
synchronize();
@@ -2366,14 +2360,11 @@ amd::Memory* DmaBlitManager::pinHostMemory(const void* hostMem, size_t pinSize,
Memory* KernelBlitManager::createView(const Memory& parent, const cl_image_format format) const {
assert(!parent.desc().buffer_ && "View supports images only");
Memory* gpuImage = NULL;
gpuImage = new Image(dev(), parent.size(), parent.desc().width_, parent.desc().height_,
parent.desc().depth_, format, parent.desc().topology_, 1);
Memory* gpuImage = new Image(dev(), parent.size(), parent.desc().width_, parent.desc().height_,
parent.desc().depth_, format, parent.desc().topology_, 1);
// Create resource
if (NULL != gpuImage) {
bool result = false;
Resource::ImageViewParams params;
const Memory& gpuMem = static_cast<const Memory&>(parent);
@@ -2385,7 +2376,7 @@ Memory* KernelBlitManager::createView(const Memory& parent, const cl_image_forma
params.gpu_ = &gpu();
// Create memory object
result = gpuImage->create(Resource::ImageView, &params);
bool result = gpuImage->create(Resource::ImageView, &params);
if (!result) {
delete gpuImage;
return NULL;
projects/clr/rocclr/runtime/device/pal/palcompiler.cpp
-3
파일 보기
@@ -56,11 +56,9 @@ bool HSAILProgram::compileImpl(const std::string& sourceCode,
// Find the temp folder for the OS
std::string tempFolder = amd::Os::getTempPath();
std::string tempFileName = amd::Os::getTempFileName();
// Iterate through each source code and dump it into tmp
std::fstream f;
std::vector<std::string> headerFileNames(headers.size());
std::vector<std::string> newDirs;
for (size_t i = 0; i < headers.size(); ++i) {
std::string headerPath = tempFolder;
@@ -84,7 +82,6 @@ bool HSAILProgram::compileImpl(const std::string& sourceCode,
newDirs.push_back(headerPath);
}
std::string headerFullName = headerPath + amd::Os::fileSeparator() + headerIncludeName;
headerFileNames[i] = headerFullName;
f.open(headerFullName.c_str(), std::fstream::out);
// Should we allow asserts
assert(!f.fail() && "failed creating header file!");
projects/clr/rocclr/runtime/device/pal/paldevice.cpp
+11 -23
파일 보기
@@ -213,7 +213,7 @@ bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel,
// Report 512MB for all offline devices
Pal::GpuMemoryHeapProperties heaps[Pal::GpuHeapCount];
heaps[Pal::GpuHeapLocal].heapSize =
heaps[Pal::GpuHeapLocal].heapSize =
heaps[Pal::GpuHeapLocal].physicalHeapSize = 512 * Mi;
Pal::WorkStationCaps wscaps = {};
@@ -547,10 +547,9 @@ Device::XferBuffers::~XferBuffers() {
}
bool Device::XferBuffers::create() {
Memory* xferBuf = nullptr;
bool result = false;
// Create a buffer object
xferBuf = new Memory(dev(), bufSize_);
Memory* xferBuf = new Memory(dev(), bufSize_);
// Try to allocate memory for the transfer buffer
if ((nullptr == xferBuf) || !xferBuf->create(type_)) {
@@ -990,14 +989,13 @@ bool Device::initializeHeapResources() {
device::VirtualDevice* Device::createVirtualDevice(amd::CommandQueue* queue) {
bool profiling = false;
bool interopQueue = false;
uint rtCUs = amd::CommandQueue::RealTimeDisabled;
uint deviceQueueSize = 0;
if (queue != nullptr) {
profiling = queue->properties().test(CL_QUEUE_PROFILING_ENABLE);
if (queue->asHostQueue() != nullptr) {
interopQueue = (0 != (queue->context().info().flags_ &
bool interopQueue = (0 != (queue->context().info().flags_ &
(amd::Context::GLDeviceKhr | amd::Context::D3D10DeviceKhr |
amd::Context::D3D11DeviceKhr)));
rtCUs = queue->rtCUs();
@@ -1044,11 +1042,10 @@ typedef std::map<int, bool> requestedDevices_t;
//! Parses the requested list of devices to be exposed to the user.
static void parseRequestedDeviceList(requestedDevices_t& requestedDevices) {
char* pch = nullptr;
int requestedDeviceCount = 0;
const char* requestedDeviceList = GPU_DEVICE_ORDINAL;
pch = strtok(const_cast<char*>(requestedDeviceList), ",");
char* pch = strtok(const_cast<char*>(requestedDeviceList), ",");
while (pch != nullptr) {
bool deviceIdValid = true;
int currentDeviceIndex = atoi(pch);
@@ -1196,10 +1193,8 @@ Pal::ChNumFormat Device::getPalFormat(const amd::Image::Format& format,
// Create buffer without an owner (merge common code with createBuffer() ?)
pal::Memory* Device::createScratchBuffer(size_t size) const {
Memory* gpuMemory = nullptr;
// Create a memory object
gpuMemory = new pal::Memory(*this, size);
Memory* gpuMemory = new pal::Memory(*this, size);
if (nullptr == gpuMemory || !gpuMemory->create(Resource::Local)) {
delete gpuMemory;
gpuMemory = nullptr;
@@ -1381,7 +1376,6 @@ pal::Memory* Device::createBuffer(amd::Memory& owner, bool directAccess) const {
}
pal::Memory* Device::createImage(amd::Memory& owner, bool directAccess) const {
size_t size = owner.getSize();
amd::Image& image = *owner.asImage();
pal::Memory* gpuImage = nullptr;
@@ -1527,8 +1521,6 @@ bool Device::createSampler(const amd::Sampler& owner, device::Sampler** sampler)
//! Otherwise a deadlock in lockVgpus() is possible
bool Device::reallocMemory(amd::Memory& owner) const {
bool directAccess = false;
// For now we have to serialize reallocation code
amd::ScopedLock lk(*lockAsyncOps_);
@@ -1548,7 +1540,7 @@ bool Device::reallocMemory(amd::Memory& owner) const {
}
if (owner.asBuffer()) {
gpuMemory = createBuffer(owner, directAccess);
gpuMemory = createBuffer(owner, false);
} else if (owner.asImage()) {
return true;
} else {
@@ -1584,17 +1576,14 @@ bool Device::reallocMemory(amd::Memory& owner) const {
}
device::Memory* Device::createView(amd::Memory& owner, const device::Memory& parent) const {
size_t size = owner.getSize();
assert((owner.asImage() != nullptr) && "View supports images only");
const amd::Image& image = *owner.asImage();
pal::Memory* gpuImage = nullptr;
gpuImage = new pal::Image(*this, owner, image.getWidth(), image.getHeight(), image.getDepth(),
image.getImageFormat(), image.getType(), image.getMipLevels());
pal::Memory* gpuImage = new pal::Image(
*this, owner, image.getWidth(), image.getHeight(), image.getDepth(),
image.getImageFormat(), image.getType(), image.getMipLevels());
// Create resource
if (nullptr != gpuImage) {
bool result = false;
Resource::ImageViewParams params;
const pal::Memory& gpuMem = static_cast<const pal::Memory&>(parent);
@@ -1606,7 +1595,7 @@ device::Memory* Device::createView(amd::Memory& owner, const device::Memory& par
params.memory_ = &gpuMem;
// Create memory object
result = gpuImage->create(Resource::ImageView, &params);
bool result = gpuImage->create(Resource::ImageView, &params);
if (!result) {
delete gpuImage;
return nullptr;
@@ -2033,8 +2022,7 @@ void Device::svmFree(void* ptr) const {
if (freeCPUMem_) {
amd::Os::alignedFree(ptr);
} else {
amd::Memory* svmMem = nullptr;
svmMem = amd::SvmManager::FindSvmBuffer(ptr);
amd::Memory* svmMem = amd::SvmManager::FindSvmBuffer(ptr);
if (nullptr != svmMem) {
svmMem->release();
amd::SvmManager::RemoveSvmBuffer(ptr);
projects/clr/rocclr/runtime/device/pal/paldevice.hpp
+3 -3
파일 보기
@@ -159,7 +159,7 @@ class ThreadTrace;
class Sampler : public device::Sampler {
public:
//! Constructor
Sampler(const Device& dev) : dev_(dev) {}
Sampler(const Device& dev) : dev_(dev) {}
//! Default destructor for the device memory object
virtual ~Sampler();
@@ -252,7 +252,7 @@ class Device : public NullDevice {
uint64_t size_; //!< Scratch buffer size on this queue
//! Default constructor
ScratchBuffer() : regNum_(0), memObj_(NULL), offset_(0) {}
ScratchBuffer() : regNum_(0), memObj_(NULL), offset_(0), size_(0) {}
//! Default constructor
~ScratchBuffer();
@@ -493,7 +493,7 @@ class Device : public NullDevice {
bool resGLFree(void* GLplatformContext, void* mbResHandle, uint type) const;
//! Adds a resource to the global list
void addResource(GpuMemoryReference* mem) const {
void addResource(GpuMemoryReference* mem) const {
amd::ScopedLock lock(lockResources());
auto findIt = std::find(resourceList_->begin(), resourceList_->end(), mem);
mem->events_.resize(numOfVgpus());
projects/clr/rocclr/runtime/device/pal/paldevicegl.cpp
+3 -5
파일 보기
@@ -639,7 +639,7 @@ bool Device::glCanInterop(void* GLplatformContext, void* GLdeviceContext) const
GLXContext ctx = static_cast<GLXContext>(GLplatformContext);
Display* disp = static_cast<Display*>(GLdeviceContext);
if (glXGetContextMVPUInfoAMD(ctx, &glDeviceId, &glChainMask)) {
mesa_glinterop_device_info info = {};
if (pfnMesaGLInteropGLXQueryDeviceInfo(disp, ctx, &info) == 0) {
@@ -661,7 +661,6 @@ bool Device::glAssociate(void* GLplatformContext, void* GLdeviceContext) const {
return false;
}
int flags = 0;
/*
if (m_adp->pAsicInfo->svmFineGrainSystem)
{
@@ -673,12 +672,11 @@ bool Device::glAssociate(void* GLplatformContext, void* GLdeviceContext) const {
return (glXBeginCLInteropAMD(ctx, 0)) ? true : false;
#else
HGLRC hRC = (HGLRC)GLplatformContext;
return (wglBeginCLInteropAMD(hRC, flags)) ? true : false;
return (wglBeginCLInteropAMD(hRC, 0)) ? true : false;
#endif
}
bool Device::glDissociate(void* GLplatformContext, void* GLdeviceContext) const {
int flags = 0;
/*
if (m_adp->pAsicInfo->svmFineGrainSystem)
{
@@ -690,7 +688,7 @@ bool Device::glDissociate(void* GLplatformContext, void* GLdeviceContext) const
return (glXEndCLInteropAMD(ctx, 0)) ? true : false;
#else
HGLRC hRC = (HGLRC)GLplatformContext;
return (wglEndCLInteropAMD(hRC, flags)) ? true : false;
return (wglEndCLInteropAMD(hRC, 0)) ? true : false;
#endif
}
projects/clr/rocclr/runtime/device/pal/palkernel.cpp
+2 -3
파일 보기
@@ -770,7 +770,6 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
if (workGroupInfo()->compileSize_[0] == 0) {
// Find the default local workgroup size, if it wasn't specified
if (lclWorkSize[0] == 0) {
size_t thrPerGrp;
bool b1DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE);
bool b2DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_X) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_Y);
@@ -782,7 +781,7 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
((workDim == 3) && b3DOverrideSet);
if (!overrideSet) {
// Find threads per group
thrPerGrp = workGroupInfo()->size_;
size_t thrPerGrp = workGroupInfo()->size_;
// Check if kernel uses images
if (flags_.imageEna_ &&
@@ -1378,7 +1377,7 @@ static inline HSAIL_ADDRESS_QUALIFIER GetKernelAddrQual(const KernelArgMD& lcArg
}
LogError("Unsupported address type");
return HSAIL_ADDRESS_ERROR;
} else if (lcArg.mValueKind == ValueKind::Image ||
} else if (lcArg.mValueKind == ValueKind::Image ||
lcArg.mValueKind == ValueKind::Sampler ||
lcArg.mValueKind == ValueKind::Pipe) {
return HSAIL_ADDRESS_GLOBAL;
projects/clr/rocclr/runtime/device/pal/palmemory.cpp
+3 -5
파일 보기
@@ -170,7 +170,7 @@ bool Memory::create(Resource::MemoryType memType, Resource::CreateParams* params
}
if (result) {
if ((params != nullptr) && (memoryType() == Pinned)) {
if ((params != nullptr) && (memoryType() == Pinned)) {
memRef()->gpu_ = params->gpu_;
}
}
@@ -789,7 +789,6 @@ void* Memory::allocMapTarget(const amd::Coord3D& origin, const amd::Coord3D& reg
if (memory == nullptr) {
// for map target of svm buffer , we need use svm host ptr
memory = new (dev().context()) amd::Buffer(dev().context(), flag, owner()->getSize());
Memory* gpuMemory;
do {
if ((memory == nullptr) || !memory->create(initHostPtr, SysMem)) {
@@ -798,7 +797,7 @@ void* Memory::allocMapTarget(const amd::Coord3D& origin, const amd::Coord3D& reg
}
memory->setCacheStatus(canBeCached);
gpuMemory = reinterpret_cast<Memory*>(memory->getDeviceMemory(dev()));
Memory* gpuMemory = reinterpret_cast<Memory*>(memory->getDeviceMemory(dev()));
// Create, Map and get the base pointer for the resource
if ((gpuMemory == nullptr) || (nullptr == gpuMemory->map(nullptr))) {
@@ -1099,14 +1098,13 @@ void* Image::allocMapTarget(const amd::Coord3D& origin, const amd::Coord3D& regi
amd::Buffer(dev().context(), 0, desc().width_ * height * depth * elementSize());
memory->setVirtualDevice(owner()->getVirtualDevice());
Memory* gpuMemory;
do {
if ((memory == nullptr) || !memory->create(nullptr, SysMem)) {
failed = true;
break;
}
gpuMemory = reinterpret_cast<Memory*>(memory->getDeviceMemory(dev()));
Memory* gpuMemory = reinterpret_cast<Memory*>(memory->getDeviceMemory(dev()));
// Create, Map and get the base pointer for the resource
if ((gpuMemory == nullptr) || (nullptr == gpuMemory->map(nullptr))) {
projects/clr/rocclr/runtime/device/pal/palprintf.cpp
+1 -2
파일 보기
@@ -386,7 +386,6 @@ void PrintfDbg::outputDbgBuffer(const PrintfInfo& info, const uint32_t* workitem
if (posStart != std::string::npos) {
bool printFloat = false;
int vectorSize = 0;
size_t length;
size_t idPos = 0;
// Search for PrintfDbg specifier in the format string.
@@ -423,7 +422,7 @@ void PrintfDbg::outputDbgBuffer(const PrintfInfo& info, const uint32_t* workitem
// Is it a scalar value?
if (vectorSize == 0) {
length = outputArgument(fmt, printFloat, info.arguments_[j], &s[i]);
size_t length = outputArgument(fmt, printFloat, info.arguments_[j], &s[i]);
if (0 == length) {
return;
}
projects/clr/rocclr/runtime/device/pal/palprogram.cpp
+4 -6
파일 보기
@@ -93,8 +93,7 @@ void Segment::copy(size_t offset, const void* src, size_t size) {
size_t srcOffs = 0;
while (size != 0) {
xferBuf.hostWrite(&gpu, reinterpret_cast<const_address>(src) + srcOffs, 0, tmpSize);
bool result =
xferBuf.partialMemCopyTo(gpu, 0, (offset + srcOffs), tmpSize, *gpuAccess_, false, true);
xferBuf.partialMemCopyTo(gpu, 0, (offset + srcOffs), tmpSize, *gpuAccess_, false, true);
size -= tmpSize;
srcOffs += tmpSize;
tmpSize = std::min(static_cast<size_t>(xferBuf.vmSize()), size);
@@ -648,11 +647,11 @@ bool HSAILProgram::linkImpl(amd::option::Options* options) {
&kernelNamesSize);
if (errorCode != ACL_SUCCESS) {
buildLog_ += "Error: Querying of kernel names from the binary failed.\n";
delete kernelNames;
delete [] kernelNames;
return false;
}
std::vector<std::string> vKernels = splitSpaceSeparatedString(kernelNames);
delete kernelNames;
delete [] kernelNames;
std::vector<std::string>::iterator it = vKernels.begin();
bool dynamicParallelism = false;
for (it; it != vKernels.end(); ++it) {
@@ -862,11 +861,10 @@ hsa_isa_t PALHSALoaderContext::IsaFromName(const char* name) {
uint32_t gfxip = 0;
std::string gfx_target(name);
uint32_t shift = 1;
size_t first;
size_t last = gfx_target.length();
std::string ver;
do {
first = gfx_target.find_last_of(':', last);
size_t first = gfx_target.find_last_of(':', last);
ver = gfx_target.substr(first + 1, last - first);
last = first - 1;
gfxip += static_cast<uint32_t>(atoi(ver.c_str())) * shift;
projects/clr/rocclr/runtime/device/pal/palresource.cpp
+17 -172
파일 보기
@@ -614,7 +614,7 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
if (memImg != nullptr) {
result = dev().iDev()->CreateImage(imgCreateInfo, memImg, &image_);
if (result != Pal::Result::Success) {
delete memImg;
delete [] memImg;
return false;
}
}
@@ -782,7 +782,6 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
Pal::ImageViewInfo viewInfo = {};
Pal::ImageCreateInfo imgCreateInfo = {};
Pal::GpuMemoryRequirements req = {};
char* memImg;
imgCreateInfo.imageType = Pal::ImageType::Tex2d;
viewInfo.viewType = Pal::ImageViewType::Tex2d;
imgCreateInfo.extent.width = desc_.width_;
@@ -868,11 +867,11 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
return false;
}
memImg = new char[imageSize];
char* memImg = new char[imageSize];
if (memImg != nullptr) {
result = dev().iDev()->CreateImage(imgCreateInfo, memImg, &image_);
if (result != Pal::Result::Success) {
delete memImg;
delete [] memImg;
return false;
}
}
@@ -1208,7 +1207,6 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
Resource& dstResource, bool enableCopyRect, bool flushDMA,
uint bytesPerElement) const {
GpuEvent event;
bool result = true;
EngineType activeEngineID = gpu.engineID_;
static const bool waitOnBusyEngine = true;
@@ -1350,19 +1348,17 @@ bool Resource::partialMemCopyTo(VirtualGPU& gpu, const amd::Coord3D& srcOrigin,
gpu.eventEnd(gpu.engineID_, event);
if (result) {
// Mark source and destination as busy
setBusy(gpu, event);
dstResource.setBusy(gpu, event);
// Mark source and destination as busy
setBusy(gpu, event);
dstResource.setBusy(gpu, event);
// Update the global GPU event
gpu.setGpuEvent(event, flushDMA);
}
// Update the global GPU event
gpu.setGpuEvent(event, flushDMA);
// Restore the original engine
gpu.engineID_ = activeEngineID;
return result;
return true;
}
void Resource::setBusy(VirtualGPU& gpu, GpuEvent gpuEvent) const {
@@ -1380,7 +1376,7 @@ void Resource::wait(VirtualGPU& gpu, bool waitOnBusyEngine) const {
// Check if we have to wait unconditionally
if (!waitOnBusyEngine ||
// or we have to wait only if another engine was used on this resource
(waitOnBusyEngine && (gpuEvent->engineId_ != gpu.engineID_))) {
(gpuEvent->engineId_ != gpu.engineID_)) {
gpu.waitForEvent(gpuEvent);
}
@@ -1418,9 +1414,7 @@ bool Resource::hostWrite(VirtualGPU* gpu, const void* hostPtr, const amd::Coord3
// Copy memory
amd::Os::fastMemcpy(dst, hostPtr, copySize);
} else {
size_t srcOffs = 0;
size_t dstOffsBase = origin[0] * elementSize_;
size_t dstOffs;
// Make sure we use the right pitch if it's not specified
if (rowPitch == 0) {
@@ -1440,8 +1434,8 @@ bool Resource::hostWrite(VirtualGPU* gpu, const void* hostPtr, const amd::Coord3
// Copy memory slice by slice
for (size_t slice = 0; slice < size[2]; ++slice) {
dstOffs = dstOffsBase + slice * desc().slice_ * elementSize_;
srcOffs = slice * slicePitch;
size_t dstOffs = dstOffsBase + slice * desc().slice_ * elementSize_;
size_t srcOffs = slice * slicePitch;
// Copy memory line by line
for (size_t row = 0; row < size[1]; ++row) {
@@ -1490,8 +1484,6 @@ bool Resource::hostRead(VirtualGPU* gpu, void* hostPtr, const amd::Coord3D& orig
amd::Os::fastMemcpy(hostPtr, src, copySize);
} else {
size_t srcOffsBase = origin[0] * elementSize_;
size_t srcOffs;
size_t dstOffs = 0;
// Make sure we use the right pitch if it's not specified
if (rowPitch == 0) {
@@ -1511,8 +1503,8 @@ bool Resource::hostRead(VirtualGPU* gpu, void* hostPtr, const amd::Coord3D& orig
// Copy memory line by line
for (size_t slice = 0; slice < size[2]; ++slice) {
srcOffs = srcOffsBase + slice * desc().slice_ * elementSize_;
dstOffs = slice * slicePitch;
size_t srcOffs = srcOffsBase + slice * desc().slice_ * elementSize_;
size_t dstOffs = slice * slicePitch;
// Copy memory line by line
for (size_t row = 0; row < size[1]; ++row) {
@@ -1701,85 +1693,8 @@ void* Resource::map(VirtualGPU* gpu, uint flags, uint startLayer, uint numLayers
}
void* Resource::mapLayers(VirtualGPU* gpu, uint flags) {
size_t srcOffs = 0;
size_t dstOffs = 0;
Pal::IGpuMemory* sliceResource = 0;
PalGpuMemoryType palDim = PAL_TEXTURE_2D;
size_t layers = desc().depth_;
size_t height = desc().height_;
// Use 1D layers
if (CL_MEM_OBJECT_IMAGE1D_ARRAY == desc().topology_) {
palDim = PAL_TEXTURE_1D;
height = 1;
layers = desc().height_;
}
desc_.pitch_ = desc().width_;
desc_.slice_ = desc().pitch_ * height;
address_ = new char[desc().slice_ * layers * elementSize()];
if (nullptr == address_) {
return nullptr;
}
// Check if map is write only
if (flags & WriteOnly) {
return address_;
}
if (numLayers_ != 0) {
layers = startLayer_ + numLayers_;
}
dstOffs = startLayer_ * desc().slice_ * elementSize();
// Loop through all layers
for (uint i = startLayer_; i < layers; ++i) {
// gslResource3D gslSize;
size_t calOffset;
void* sliceAddr;
size_t pitch;
Unimplemented();
// Allocate a layer from the image
// gslSize.width = desc().width_;
// gslSize.height = height;
// gslSize.depth = 1;
calOffset = 0;
/*
sliceResource = dev().resAllocView(
iMem(), gslSize,
calOffset, desc().format_, desc().channelOrder_, palDim,
0, i, CAL_RESALLOCSLICEVIEW_LEVEL_AND_LAYER);
if (0 == sliceResource) {
LogError("Map layer. resAllocSliceView failed!");
return nullptr;
}
*/
// Map 2D layer
sliceAddr = gpuMemoryMap(&pitch, ReadOnly, sliceResource);
if (sliceAddr == nullptr) {
LogError("Map layer. CalResMap failed!");
return nullptr;
}
srcOffs = 0;
// Copy memory line by line
for (size_t rows = 0; rows < height; ++rows) {
// Copy memory
amd::Os::fastMemcpy((reinterpret_cast<address>(address_) + dstOffs),
(reinterpret_cast<const_address>(sliceAddr) + srcOffs),
desc().width_ * elementSize_);
dstOffs += desc().pitch_ * elementSize();
srcOffs += pitch * elementSize();
}
// Unmap a layer
gpuMemoryUnmap(sliceResource);
// dev().resFree(sliceResource);
}
return address_;
Unimplemented();
return nullptr;
}
void Resource::unmap(VirtualGPU* gpu) {
@@ -1809,77 +1724,7 @@ void Resource::unmap(VirtualGPU* gpu) {
}
void Resource::unmapLayers(VirtualGPU* gpu) {
size_t srcOffs = 0;
size_t dstOffs = 0;
PalGpuMemoryType palDim = PAL_TEXTURE_2D;
Pal::IGpuMemory* sliceResource = nullptr;
uint layers = desc().depth_;
uint height = desc().height_;
// Use 1D layers
if (CL_MEM_OBJECT_IMAGE1D_ARRAY == desc().topology_) {
palDim = PAL_TEXTURE_1D;
height = 1;
layers = desc().height_;
}
if (numLayers_ != 0) {
layers = startLayer_ + numLayers_;
}
srcOffs = startLayer_ * desc().slice_ * elementSize();
// Check if map is write only
if (!(mapFlags_ & ReadOnly)) {
// Loop through all layers
for (uint i = startLayer_; i < layers; ++i) {
Unimplemented();
// gslResource3D gslSize;
size_t calOffset;
void* sliceAddr;
size_t pitch;
// Allocate a layer from the image
// gslSize.width = desc().width_;
// gslSize.height = height;
// gslSize.depth = 1;
calOffset = 0;
/*sliceResource = dev().resAllocView(
iMem(), gslSize,
calOffset, desc().format_, desc().channelOrder_, palDim,
0, i, CAL_RESALLOCSLICEVIEW_LEVEL_AND_LAYER);
if (0 == sliceResource) {
LogError("Unmap layer. resAllocSliceView failed!");
return;
}
*/
// Map a layer
sliceAddr = gpuMemoryMap(&pitch, WriteOnly, sliceResource);
if (sliceAddr == nullptr) {
LogError("Unmap layer. CalResMap failed!");
return;
}
dstOffs = 0;
// Copy memory line by line
for (size_t rows = 0; rows < height; ++rows) {
// Copy memory
amd::Os::fastMemcpy((reinterpret_cast<address>(sliceAddr) + dstOffs),
(reinterpret_cast<const_address>(address_) + srcOffs),
desc().width_ * elementSize_);
dstOffs += pitch * elementSize();
srcOffs += desc().pitch_ * elementSize();
}
// Unmap a layer
gpuMemoryUnmap(sliceResource);
// dev().resFree(sliceResource);
}
}
// Destroy the mapped memory
delete[] reinterpret_cast<char*>(address_);
Unimplemented();
}
void Resource::setActiveRename(VirtualGPU& gpu, GpuMemoryReference* rename) {
projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
+10 -16
파일 보기
@@ -384,7 +384,6 @@ void VirtualGPU::Queue::DumpMemoryReferences() const {
}
if (last_kernel_ != nullptr) {
const amd::KernelSignature& signature = last_kernel_->signature();
const amd::KernelParameters& params = last_kernel_->parameters();
dump << last_kernel_->name() << std::endl;
for (size_t i = 0; i < signature.numParameters(); ++i) {
const amd::KernelParameterDescriptor& desc = signature.at(i);
@@ -1427,7 +1426,6 @@ void VirtualGPU::submitMapMemory(amd::MapMemoryCommand& vcmd) {
vcmd.setStatus(CL_MAP_FAILURE);
}
} else if ((vcmd.memory().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
amd::Memory* bufferFromImage = nullptr;
Memory* memoryBuf = memory;
amd::Coord3D origin(vcmd.origin()[0]);
amd::Coord3D size(vcmd.size()[0]);
@@ -1435,7 +1433,7 @@ void VirtualGPU::submitMapMemory(amd::MapMemoryCommand& vcmd) {
origin.c[0] *= elemSize;
size.c[0] *= elemSize;
bufferFromImage = createBufferFromImage(vcmd.memory());
amd::Memory* bufferFromImage = createBufferFromImage(vcmd.memory());
if (nullptr == bufferFromImage) {
LogError("We should not fail buffer creation from image_buffer!");
} else {
@@ -1531,7 +1529,6 @@ void VirtualGPU::submitUnmapMemory(amd::UnmapMemoryCommand& vcmd) {
vcmd.setStatus(CL_OUT_OF_RESOURCES);
}
} else if ((vcmd.memory().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
amd::Memory* bufferFromImage = nullptr;
Memory* memoryBuf = memory;
amd::Coord3D origin(writeMapInfo->origin_[0]);
amd::Coord3D size(writeMapInfo->region_[0]);
@@ -1539,7 +1536,7 @@ void VirtualGPU::submitUnmapMemory(amd::UnmapMemoryCommand& vcmd) {
origin.c[0] *= elemSize;
size.c[0] *= elemSize;
bufferFromImage = createBufferFromImage(vcmd.memory());
amd::Memory* bufferFromImage = createBufferFromImage(vcmd.memory());
if (nullptr == bufferFromImage) {
LogError("We should not fail buffer creation from image_buffer!");
} else {
@@ -1747,12 +1744,10 @@ void VirtualGPU::submitSvmFillMemory(amd::SvmFillMemoryCommand& vcmd) {
if (!dev().isFineGrainedSystem()) {
size_t patternSize = vcmd.patternSize();
size_t fillSize = patternSize * vcmd.times();
size_t offset = 0;
amd::Memory* dstMemory = amd::SvmManager::FindSvmBuffer(vcmd.dst());
assert(dstMemory && "No svm Buffer to fill with!");
offset = reinterpret_cast<uintptr_t>(vcmd.dst()) -
size_t offset = reinterpret_cast<uintptr_t>(vcmd.dst()) -
reinterpret_cast<uintptr_t>(dstMemory->getSvmPtr());
assert((offset >= 0) && "wrong svm ptr to fill with!");
pal::Memory* memory = dev().getGpuMemory(dstMemory);
@@ -1785,7 +1780,7 @@ void VirtualGPU::submitMigrateMemObjects(amd::MigrateMemObjectsCommand& vcmd) {
profilingBegin(vcmd, true);
std::vector<amd::Memory*>::const_iterator itr;
for (itr = vcmd.memObjects().begin(); itr != vcmd.memObjects().end(); itr++) {
for (itr = vcmd.memObjects().begin(); itr != vcmd.memObjects().end(); ++itr) {
// Find device memory
pal::Memory* memory = dev().getGpuMemory(*itr);
@@ -1814,7 +1809,7 @@ void VirtualGPU::submitSvmFreeMemory(amd::SvmFreeMemoryCommand& vcmd) {
std::vector<void*>& svmPointers = vcmd.svmPointers();
if (vcmd.pfnFreeFunc() == nullptr) {
// pointers allocated using clSVMAlloc
for (cl_uint i = 0; i < svmPointers.size(); i++) {
for (cl_uint i = 0; i < svmPointers.size(); ++i) {
dev().svmFree(svmPointers[i]);
}
} else {
@@ -2428,7 +2423,7 @@ void VirtualGPU::submitAcquireExtObjects(amd::AcquireExtObjectsCommand& vcmd) {
profilingBegin(vcmd);
for (std::vector<amd::Memory*>::const_iterator it = vcmd.getMemList().begin();
it != vcmd.getMemList().end(); it++) {
it != vcmd.getMemList().end(); ++it) {
// amd::Memory object should never be nullptr
assert(*it && "Memory object for interop is nullptr");
pal::Memory* memory = dev().getGpuMemory(*it);
@@ -2467,7 +2462,7 @@ void VirtualGPU::submitReleaseExtObjects(amd::ReleaseExtObjectsCommand& vcmd) {
profilingBegin(vcmd);
for (std::vector<amd::Memory*>::const_iterator it = vcmd.getMemList().begin();
it != vcmd.getMemList().end(); it++) {
it != vcmd.getMemList().end(); ++it) {
// amd::Memory object should never be nullptr
assert(*it && "Memory object for interop is nullptr");
pal::Memory* memory = dev().getGpuMemory(*it);
@@ -2992,7 +2987,6 @@ bool VirtualGPU::processMemObjectsHSA(const amd::Kernel& kernel, const_address p
const amd::KernelParameterDescriptor& desc = signature.at(i);
const HSAILKernel::Argument* arg = hsaKernel.argumentAt(i);
Memory* memory = nullptr;
bool readOnly = false;
amd::Memory* svmMem = nullptr;
// Find if current argument is a buffer
@@ -3024,7 +3018,7 @@ bool VirtualGPU::processMemObjectsHSA(const amd::Kernel& kernel, const_address p
if (memory != nullptr) {
// Check image
readOnly = (desc.accessQualifier_ == CL_KERNEL_ARG_ACCESS_READ_ONLY) ? true : false;
bool readOnly = (desc.accessQualifier_ == CL_KERNEL_ARG_ACCESS_READ_ONLY) ? true : false;
// Check buffer
readOnly |= (arg->access_ == HSAIL_ACCESS_TYPE_RO) ? true : false;
// Validate memory for a dependency in the queue
@@ -3236,7 +3230,7 @@ void VirtualGPU::submitTransferBufferFromFile(amd::TransferBufferFileCommand& cm
}
staging->cpuUnmap(*this);
bool result = blitMgr().copyBuffer(*staging, *mem, 0, dstOffset, dstSize, false);
blitMgr().copyBuffer(*staging, *mem, 0, dstOffset, dstSize, false);
flushDMA(staging->getGpuEvent(*this)->engineId_);
fileOffset += dstSize;
dstOffset += dstSize;
@@ -3248,7 +3242,7 @@ void VirtualGPU::submitTransferBufferFromFile(amd::TransferBufferFileCommand& cm
Memory* staging = dev().getGpuMemory(&cmd.staging(idx));
size_t srcSize = amd::TransferBufferFileCommand::StagingBufferSize;
srcSize = std::min(srcSize, copySize);
bool result = blitMgr().copyBuffer(*mem, *staging, srcOffset, 0, srcSize, false);
blitMgr().copyBuffer(*mem, *staging, srcOffset, 0, srcSize, false);
void* srcBuffer = staging->cpuMap(*this);
if (!cmd.file()->transferBlock(writeBuffer, srcBuffer, staging->size(), fileOffset, 0,