intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 12:23:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

Improve caching in clSetKernelArgSVMPointer

Browse Source 
Check allocId earlier and also reuse if allocationsCounter did not
change from last call.

Related-To: NEO-6737

Co-authored-by: Michal Mrozek <michal.mrozek@intel.com>

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2022-03-02 11:08:18 +00:00
committed by Compute-Runtime-Automation
parent 385c60948e
commit 7ab86d44d6
4 changed files with 40 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
opencl/source/api/api.cpp
View File

@@ -4866,13 +4866,41 @@ cl_int CL_API_CALL clSetKernelArgSVMPointer(cl_kernel kernel,
auto retVal = validateObjects(WithCastToInternal(kernel, &pMultiDeviceKernel));
API_ENTER(&retVal);
DBG_LOG_INPUTS("kernel", kernel, "argIndex", argIndex, "argValue", argValue);
if (CL_SUCCESS != retVal) {
TRACING_EXIT(clSetKernelArgSVMPointer, &retVal);
return retVal;
}
if (argIndex >= pMultiDeviceKernel->getKernelArgsNumber()) {
retVal = CL_INVALID_ARG_INDEX;
TRACING_EXIT(clSetKernelArgSVMPointer, &retVal);
return retVal;
}
auto svmManager = pMultiDeviceKernel->getContext().getSVMAllocsManager();
if (argValue != nullptr) {
if (pMultiDeviceKernel->getKernelArguments()[argIndex].allocId > 0 &&
pMultiDeviceKernel->getKernelArguments()[argIndex].value == argValue) {
bool reuseFromCache = false;
if (svmManager->allocationsCounter == pMultiDeviceKernel->getKernelArguments()[argIndex].allocIdMemoryManagerCounter) {
reuseFromCache = true;
} else {
auto svmData = svmManager->getSVMAlloc(argValue);
if (pMultiDeviceKernel->getKernelArguments()[argIndex].allocId == svmData->getAllocId()) {
reuseFromCache = true;
}
}
if (reuseFromCache) {
TRACING_EXIT(clSetKernelArgSVMPointer, &retVal);
return CL_SUCCESS;
}
}
}
DBG_LOG_INPUTS("kernel", kernel, "argIndex", argIndex, "argValue", argValue);
for (const auto &pDevice : pMultiDeviceKernel->getDevices()) {
const HardwareInfo &hwInfo = pDevice->getHardwareInfo();
if (!hwInfo.capabilityTable.ftrSvm) {
@@ -4882,12 +4910,6 @@ cl_int CL_API_CALL clSetKernelArgSVMPointer(cl_kernel kernel,
}
}
if (argIndex >= pMultiDeviceKernel->getKernelArgsNumber()) {
retVal = CL_INVALID_ARG_INDEX;
TRACING_EXIT(clSetKernelArgSVMPointer, &retVal);
return retVal;
}
for (const auto &pDevice : pMultiDeviceKernel->getDevices()) {
auto pKernel = pMultiDeviceKernel->getKernel(pDevice->getRootDeviceIndex());
cl_int kernelArgAddressQualifier = asClKernelArgAddressQualifier(pKernel->getKernelInfo()
@@ -4905,7 +4927,6 @@ cl_int CL_API_CALL clSetKernelArgSVMPointer(cl_kernel kernel,
MultiGraphicsAllocation *pSvmAllocs = nullptr;
uint32_t allocId = 0u;
if (argValue != nullptr) {
auto svmManager = pMultiDeviceKernel->getContext().getSVMAllocsManager();
auto svmData = svmManager->getSVMAlloc(argValue);
if (svmData == nullptr) {
for (const auto &pDevice : pMultiDeviceKernel->getDevices()) {
@@ -4916,12 +4937,6 @@ cl_int CL_API_CALL clSetKernelArgSVMPointer(cl_kernel kernel,
}
}
} else {
if (pMultiDeviceKernel->getKernelArguments()[argIndex].allocId > 0 &&
pMultiDeviceKernel->getKernelArguments()[argIndex].allocId == svmData->getAllocId() &&
pMultiDeviceKernel->getKernelArguments()[argIndex].value == argValue) {
TRACING_EXIT(clSetKernelArgSVMPointer, &retVal);
return CL_SUCCESS;
}
pSvmAllocs = &svmData->gpuAllocations;
allocId = svmData->getAllocId();
}