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

P4 to Git Change 1537540 by gandryey@gera-w8 on 2018/04/06 12:30:49

Browse Source 
SWDEV-79445 - OCL generic changes and code clean-up
	- Reduce the managed buffer alignment to 128 bytes
	- Add direct upload function to the constant buffer
	- Add heap preference print into the active allocations

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.hpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#87 edit


[ROCm/clr commit: c006b7b262]
This commit is contained in:
foreman
2018-04-06 12:39:00 -04:00
parent df1d5e9de8
commit c507deb37f
4 changed files with 52 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/runtime/device/pal/palkernel.cpp
+7 -11
View File
@@ -921,16 +921,14 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
static const bool WaitOnBusyEngine = true;
uint64_t ldsAddress = ldsSize();
address aqlArgBuf = gpu.cb(0)->SysMemCopy();
address aqlStruct = gpu.cb(1)->SysMemCopy();
bool srdResource = false;
if (dynamicParallelism()) {
// Provide the host parent AQL wrap object to the kernel
AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(aqlStruct);
memset(wrap, 0, sizeof(AmdAqlWrap));
wrap->state = AQL_WRAP_BUSY;
AmdAqlWrap wrap = {};
wrap.state = AQL_WRAP_BUSY;
const ConstantBuffer* cb = gpu.cb(1);
*vmParentWrap = cb->UploadDataToHw(sizeof(AmdAqlWrap));
*vmParentWrap = cb->UploadDataToHw(&wrap, sizeof(AmdAqlWrap));
gpu.addVmMemory(cb->ActiveMemory());
}
@@ -1063,11 +1061,10 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
break;
}
case HSAIL_ARGTYPE_REFERENCE: {
// Copy the current structure into CB1
memcpy(aqlStruct, paramaddr, arg->size_);
const ConstantBuffer* cb = gpu.cb(1);
// Copy the current structure into CB1
size_t gpuPtr = static_cast<size_t>(cb->UploadDataToHw(paramaddr, arg->size_));
// Then use a pointer in aqlArgBuffer to CB1
size_t gpuPtr = static_cast<size_t>(cb->UploadDataToHw(arg->size_));
WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t), sizeof(size_t));
gpu.addVmMemory(cb->ActiveMemory());
break;
@@ -1105,11 +1102,10 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
//! Copy SRD to CB1, so blit manager will be able to release
//! this view without a wait for SRD resource.
if (image->memoryType() == Resource::ImageView) {
// Copy the current structre into CB1
memcpy(aqlStruct, image->hwState(), HsaImageObjectSize);
// Copy the current image SRD into CB1
const ConstantBuffer* cb = gpu.cb(1);
uint64_t srd = cb->UploadDataToHw(image->hwState(), HsaImageObjectSize);
// Then use a pointer in aqlArgBuffer to CB1
uint64_t srd = cb->UploadDataToHw(HsaImageObjectSize);
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
gpu.addVmMemory(cb->ActiveMemory());
} else {