AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4 to Git Change 1558704 by gandryey@gera-w8 on 2018/05/23 17:20:01

Browse Source 
SWDEV-79445 - OCL generic changes and code clean-up
	- ABI clean-up. Stage 1: Separate kernel arguments and OCL objects. OCL objects will be passed in the new arrays of mem objects, samplers and device queue objects. The kernel arguments will contain GPU virtual addresses.

	http://ocltc.amd.com/reviews/r/14881/

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_program.cpp#48 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_svm.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#302 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#129 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#323 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.hpp#128 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpumemory.hpp#51 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#417 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#23 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#50 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.hpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#97 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.cpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.hpp#9 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#28 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocmemory.hpp#12 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#51 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#86 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.cpp#26 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.hpp#20 edit


[ROCm/clr commit: 2176dc3b19]
This commit is contained in:
foreman
2018-05-23 17:24:32 -04:00
parent 4d8bd812ec
commit abfbedddc3
18 changed files with 467 additions and 310 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/runtime/device/rocm/rocvirtual.cpp
+32 -41
View File
@@ -236,49 +236,42 @@ bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address para
}
}
amd::Memory* const* memories =
reinterpret_cast<amd::Memory* const*>(params + kernelParams.memoryObjOffset());
// Check all parameters for the current kernel
for (size_t i = 0; i < signature.numParameters(); ++i) {
const amd::KernelParameterDescriptor& desc = signature.at(i);
const Kernel::Argument* arg = hsaKernel.hsailArgAt(i);
Memory* memory = nullptr;
Memory* gpuMem = nullptr;
bool readOnly = false;
amd::Memory* svmMem = nullptr;
amd::Memory* mem = nullptr;
// Find if current argument is a buffer
if ((desc.type_ == T_POINTER) && (arg->addrQual_ != ROC_ADDRESS_LOCAL)) {
if (kernelParams.boundToSvmPointer(dev(), params, i)) {
svmMem =
amd::SvmManager::FindSvmBuffer(*reinterpret_cast<void* const*>(params + desc.offset_));
if (!svmMem) {
// Sync AQL packets
setAqlHeader(kDispatchPacketHeader);
// Clear memory dependency state
const static bool All = true;
memoryDependency().clear(!All);
continue;
}
}
if (*reinterpret_cast<amd::Memory* const*>(params + desc.offset_) != nullptr) {
if (nullptr == svmMem) {
memory =
static_cast<Memory*>((*reinterpret_cast<amd::Memory* const*>(params + desc.offset_))
->getDeviceMemory(dev()));
} else {
memory = static_cast<Memory*>(svmMem->getDeviceMemory(dev()));
}
uint32_t index = desc.info_.arrayIndex_;
mem = memories[index];
if (mem != nullptr) {
gpuMem = static_cast<Memory*>(mem->getDeviceMemory(dev()));
// Don't sync for internal objects,
// since they are not shared between devices
if (memory->owner()->getVirtualDevice() == nullptr) {
if (gpuMem->owner()->getVirtualDevice() == nullptr) {
// Synchronize data with other memory instances if necessary
memory->syncCacheFromHost(*this);
gpuMem->syncCacheFromHost(*this);
}
}
if (memory != nullptr) {
//! This condition is for SVM fine-grain
if ((gpuMem == nullptr) && dev().isFineGrainedSystem(true)) {
// Sync AQL packets
setAqlHeader(kDispatchPacketHeader);
// Clear memory dependency state
const static bool All = true;
memoryDependency().clear(!All);
continue;
} else if (gpuMem != nullptr) {
readOnly |= (arg->access_ == ROC_ACCESS_TYPE_RO);
// Validate memory for a dependency in the queue
memoryDependency().validate(*this, memory, readOnly);
memoryDependency().validate(*this, gpuMem, readOnly);
}
}
}
@@ -1601,6 +1594,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
}
}
amd::Memory* const* memories =
reinterpret_cast<amd::Memory* const*>(parameters + kernelParams.memoryObjOffset());
for (int j = 0; j < iteration; j++) {
// Reset global size for dimension dim if split is needed
if (dim != -1) {
@@ -1675,21 +1671,13 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
}
assert((arg->addrQual_ == ROC_ADDRESS_GLOBAL || arg->addrQual_ == ROC_ADDRESS_CONSTANT) &&
"Unsupported address qualifier");
if (kernelParams.boundToSvmPointer(dev(), parameters, arg->index_)) {
argPtr = addArg(argPtr, srcArgPtr, arg->size_, arg->alignment_);
break;
}
amd::Memory* mem = *reinterpret_cast<amd::Memory* const*>(srcArgPtr);
argPtr = addArg(argPtr, srcArgPtr, arg->size_, arg->alignment_);
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
amd::Memory* mem = memories[index];
if (mem == nullptr) {
argPtr = addArg(argPtr, srcArgPtr, arg->size_, arg->alignment_);
break;
}
Memory* devMem = static_cast<Memory*>(mem->getDeviceMemory(dev()));
//! @todo add multi-devices synchronization when supported.
void* globalAddress = devMem->getDeviceMemory();
argPtr = addArg(argPtr, &globalAddress, arg->size_, arg->alignment_);
const bool readOnly =
#if defined(WITH_LIGHTNING_COMPILER)
signature.at(arg->index_).typeQualifier_ == CL_KERNEL_ARG_TYPE_CONST ||
@@ -1715,7 +1703,8 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
argPtr = addArg(argPtr, srcArgPtr, arg->size_, arg->alignment_);
break;
case ROC_ARGTYPE_IMAGE: {
amd::Memory* mem = *reinterpret_cast<amd::Memory* const*>(srcArgPtr);
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
amd::Memory* mem = memories[index];
Image* image = static_cast<Image*>(mem->getDeviceMemory(dev()));
if (image == nullptr) {
LogError("Kernel image argument is not an image object");
@@ -1744,7 +1733,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
break;
}
case ROC_ARGTYPE_SAMPLER: {
amd::Sampler* sampler = *reinterpret_cast<amd::Sampler* const*>(srcArgPtr);
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
const amd::Sampler* sampler = reinterpret_cast<amd::Sampler* const*>(parameters +
kernelParams.samplerObjOffset())[index];
if (sampler == nullptr) {
LogError("Kernel sampler argument is not an sampler object");
return false;