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

P4 to Git Change 1987582 by cpaquot@cpaquot-ocl-lc-lnx on 2019/08/22 13:20:28

Browse Source 
SWDEV-185742 - [HIP] Add a svm path to avoid analyzing the mem objs
	and just memcpy the kernel args

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#82 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#96 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.hpp#94 edit
This commit is contained in:
foreman
2019-08-22 13:26:57 -04:00
parent d34d591676
commit b02f98556a
4 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/runtime/device/rocm/rocvirtual.cpp
+4 -5
View File
@@ -576,7 +576,7 @@ VirtualGPU::VirtualGPU(Device& device)
kernarg_pool_base_ = nullptr;
kernarg_pool_size_ = 0;
kernarg_pool_cur_offset_ = 0;
aqlHeader_ = kDispatchPacketHeaderNoSync;
aqlHeader_ = kDispatchPacketHeader;
barrier_signal_.handle = 0;
// Note: Virtual GPU device creation must be a thread safe operation
@@ -1981,13 +1981,13 @@ bool VirtualGPU::createVirtualQueue(uint deviceQueueSize)
}
bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const amd::Kernel& kernel,
const_address parameters, void* eventHandle, uint32_t sharedMemBytes, bool cooperativeGroups) {
const_address parameters, void* eventHandle, uint32_t sharedMemBytes, bool cooperativeGroups, bool isSvm) {
device::Kernel* devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(dev()));
Kernel& gpuKernel = static_cast<Kernel&>(*devKernel);
size_t ldsUsage = gpuKernel.WorkgroupGroupSegmentByteSize();
// Check memory dependency and SVM objects
if (!processMemObjects(kernel, parameters, ldsUsage, cooperativeGroups)) {
if (!isSvm && !processMemObjects(kernel, parameters, ldsUsage, cooperativeGroups)) {
LogError("Wrong memory objects!");
return false;
}
@@ -2113,7 +2113,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(reinterpret_cast<uint64_t>(schedulerParam_->getHostMem()) + sizeof(SchedulerParam));
memset(wrap, 0, sizeof(AmdAqlWrap));
wrap->state = AQL_WRAP_DONE;
spVA = reinterpret_cast<uint64_t>(schedulerMem->getDeviceMemory()) + sizeof(SchedulerParam);
}
WriteAqlArgAt(const_cast<address>(parameters), &spVA, it.size_, it.offset_);
@@ -2225,7 +2224,7 @@ void VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd) {
// Submit kernel to HW
if (!queue->submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(),
static_cast<void*>(as_cl(&vcmd.event())), vcmd.sharedMemBytes(), vcmd.cooperativeGroups())) {
static_cast<void*>(as_cl(&vcmd.event())), vcmd.sharedMemBytes(), vcmd.cooperativeGroups()), vcmd.isSvm()) {
LogError("AQL dispatch failed!");
vcmd.setStatus(CL_INVALID_OPERATION);
}