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/pal/palkernel.cpp
+22 -35
View File
@@ -934,6 +934,8 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
const amd::KernelSignature& signature = kernel.signature();
const amd::KernelParameters& kernelParams = kernel.parameters();
amd::Memory* const* memories =
reinterpret_cast<amd::Memory* const*>(parameters + kernelParams.memoryObjOffset());
// Find all parameters for the current kernel
for (auto arg : arguments_) {
@@ -1005,45 +1007,24 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
// If it is a global pointer
Memory* gpuMem = nullptr;
amd::Memory* mem = nullptr;
if (kernelParams.boundToSvmPointer(dev(), parameters, arg->index_)) {
WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr), sizeof(paramaddr));
mem = amd::SvmManager::FindSvmBuffer(*reinterpret_cast<void* const*>(paramaddr));
if (mem != nullptr) {
gpuMem = dev().getGpuMemory(mem);
gpuMem->wait(gpu, WaitOnBusyEngine);
if ((mem->getMemFlags() & CL_MEM_READ_ONLY) == 0) {
mem->signalWrite(&dev());
}
gpu.addVmMemory(gpuMem);
}
// If finegrainsystem is present then the pointer can be malloced by the app and
// passed to kernel directly. If so copy the pointer location to aqlArgBuf
else if (!dev().isFineGrainedSystem(true)) {
return nullptr;
}
break;
}
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
if (nativeMem) {
gpuMem = *reinterpret_cast<Memory* const*>(paramaddr);
gpuMem = reinterpret_cast<Memory* const*>(memories)[index];
if (nullptr != gpuMem) {
mem = gpuMem->owner();
}
} else {
mem = *reinterpret_cast<amd::Memory* const*>(paramaddr);
mem = memories[index];
if (mem != nullptr) {
gpuMem = dev().getGpuMemory(mem);
}
}
WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr), sizeof(paramaddr));
if (gpuMem == nullptr) {
WriteAqlArg(&aqlArgBuf, &gpuMem, arg->size_, arg->alignment_);
break;
}
//! 64 bit isn't supported with 32 bit binary
uint64_t globalAddress = gpuMem->vmAddress();
WriteAqlArg(&aqlArgBuf, &globalAddress, arg->size_, arg->alignment_);
// Wait for resource if it was used on an inactive engine
//! \note syncCache may call DRM transfer
gpuMem->wait(gpu, WaitOnBusyEngine);
@@ -1083,15 +1064,17 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
case HSAIL_ARGTYPE_IMAGE: {
Image* image = nullptr;
amd::Memory* mem = nullptr;
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
if (nativeMem) {
image = static_cast<Image*>(*reinterpret_cast<Memory* const*>(paramaddr));
} else {
mem = *reinterpret_cast<amd::Memory* const*>(paramaddr);
if (mem == nullptr) {
LogError("The kernel image argument isn't an image object!");
return nullptr;
image = reinterpret_cast<Image* const*>(memories)[index];
if (nullptr != image) {
mem = image->owner();
}
} else {
mem = memories[index];
if (mem != nullptr) {
image = static_cast<Image*>(dev().getGpuMemory(mem));
}
image = static_cast<Image*>(dev().getGpuMemory(mem));
}
// Wait for resource if it was used on an inactive engine
@@ -1127,7 +1110,9 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
break;
}
case HSAIL_ARGTYPE_SAMPLER: {
const amd::Sampler* sampler = *reinterpret_cast<amd::Sampler* const*>(paramaddr);
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
const amd::Sampler* sampler = reinterpret_cast<amd::Sampler* const*>(parameters +
kernelParams.samplerObjOffset())[index];
const Sampler* gpuSampler = static_cast<Sampler*>(sampler->getDeviceSampler(dev()));
uint64_t srd = gpuSampler->hwSrd();
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
@@ -1135,7 +1120,9 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
break;
}
case HSAIL_ARGTYPE_QUEUE: {
const amd::DeviceQueue* queue = *reinterpret_cast<amd::DeviceQueue* const*>(paramaddr);
uint32_t index = signature.at(arg->index_).info_.arrayIndex_;
const amd::DeviceQueue* queue = reinterpret_cast<amd::DeviceQueue* const*>(
parameters + kernelParams.queueObjOffset())[index];
VirtualGPU* gpuQueue = static_cast<VirtualGPU*>(queue->vDev());
uint64_t vmQueue;
if (dev().settings().useDeviceQueue_) {