intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-19 06:24:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: apply stateful programming only if needed

Browse Source 
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek
2025-10-20 15:09:57 +00:00
committed by Compute-Runtime-Automation
parent 5df120ddbc
commit 6cdb73ccca
1 changed files with 43 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
opencl/source/kernel/kernel.cpp
View File

@@ -984,56 +984,57 @@ cl_int Kernel::setArgSvmAlloc(uint32_t argIndex, void *svmPtr, GraphicsAllocatio
patchWithRequiredSize(patchLocation, argAsPtr.pointerSize, reinterpret_cast<uintptr_t>(svmPtr));
auto &kernelArgInfo = kernelArguments[argIndex];
bool disableL3 = false;
bool forceNonAuxMode = false;
const bool isAuxTranslationKernel = (AuxTranslationDirection::none != auxTranslationDirection);
auto &rootDeviceEnvironment = getDevice().getRootDeviceEnvironment();
auto &clGfxCoreHelper = rootDeviceEnvironment.getHelper<ClGfxCoreHelper>();
if (isAuxTranslationKernel) {
if (((AuxTranslationDirection::auxToNonAux == auxTranslationDirection) && argIndex == 1) ||
((AuxTranslationDirection::nonAuxToAux == auxTranslationDirection) && argIndex == 0)) {
forceNonAuxMode = true;
}
disableL3 = (argIndex == 0);
} else if (svmAlloc && svmAlloc->isCompressionEnabled() && clGfxCoreHelper.requiresNonAuxMode(argAsPtr)) {
forceNonAuxMode = true;
}
const bool argWasUncacheable = kernelArgInfo.isStatelessUncacheable;
const bool argIsUncacheable = svmAlloc ? svmAlloc->isUncacheable() : false;
statelessUncacheableArgsCount += (argIsUncacheable ? 1 : 0) - (argWasUncacheable ? 1 : 0);
void *ptrToPatch = patchBufferOffset(argAsPtr, svmPtr, svmAlloc);
if (isValidOffset(argAsPtr.bindful)) {
auto surfaceState = ptrOffset(getSurfaceStateHeap(), argAsPtr.bindful);
size_t allocSize = 0;
size_t offset = 0;
if (svmAlloc != nullptr) {
allocSize = svmAlloc->getUnderlyingBufferSize();
offset = ptrDiff(ptrToPatch, svmAlloc->getGpuAddressToPatch());
allocSize -= offset;
}
Buffer::setSurfaceState(&getDevice().getDevice(), surfaceState, forceNonAuxMode, disableL3, allocSize, ptrToPatch, offset, svmAlloc, 0, 0,
areMultipleSubDevicesInContext());
} else if (isValidOffset(argAsPtr.bindless)) {
size_t allocSize = 0;
size_t offset = 0;
if (svmAlloc != nullptr) {
allocSize = svmAlloc->getUnderlyingBufferSize();
offset = ptrDiff(ptrToPatch, svmAlloc->getGpuAddressToPatch());
allocSize -= offset;
if (isValidOffset(argAsPtr.bindful) || isValidOffset(argAsPtr.bindless)) {
bool disableL3 = false;
bool forceNonAuxMode = false;
const bool isAuxTranslationKernel = (AuxTranslationDirection::none != auxTranslationDirection);
auto &rootDeviceEnvironment = getDevice().getRootDeviceEnvironment();
auto &clGfxCoreHelper = rootDeviceEnvironment.getHelper<ClGfxCoreHelper>();
if (isAuxTranslationKernel) {
if (((AuxTranslationDirection::auxToNonAux == auxTranslationDirection) && argIndex == 1) ||
((AuxTranslationDirection::nonAuxToAux == auxTranslationDirection) && argIndex == 0)) {
forceNonAuxMode = true;
}
disableL3 = (argIndex == 0);
} else if (svmAlloc && svmAlloc->isCompressionEnabled() && clGfxCoreHelper.requiresNonAuxMode(argAsPtr)) {
forceNonAuxMode = true;
}
auto &gfxCoreHelper = this->getGfxCoreHelper();
auto surfaceStateSize = gfxCoreHelper.getRenderSurfaceStateSize();
auto ssIndex = getSurfaceStateIndexForBindlessOffset(argAsPtr.bindless);
if (ssIndex < std::numeric_limits<uint32_t>::max()) {
auto surfaceState = ptrOffset(getSurfaceStateHeap(), ssIndex * surfaceStateSize);
void *ptrToPatch = patchBufferOffset(argAsPtr, svmPtr, svmAlloc);
if (isValidOffset(argAsPtr.bindful)) {
auto surfaceState = ptrOffset(getSurfaceStateHeap(), argAsPtr.bindful);
size_t allocSize = 0;
size_t offset = 0;
if (svmAlloc != nullptr) {
allocSize = svmAlloc->getUnderlyingBufferSize();
offset = ptrDiff(ptrToPatch, svmAlloc->getGpuAddressToPatch());
allocSize -= offset;
}
Buffer::setSurfaceState(&getDevice().getDevice(), surfaceState, forceNonAuxMode, disableL3, allocSize, ptrToPatch, offset, svmAlloc, 0, 0,
areMultipleSubDevicesInContext());
} else { // bindless
size_t allocSize = 0;
size_t offset = 0;
if (svmAlloc != nullptr) {
allocSize = svmAlloc->getUnderlyingBufferSize();
offset = ptrDiff(ptrToPatch, svmAlloc->getGpuAddressToPatch());
allocSize -= offset;
}
auto &gfxCoreHelper = this->getGfxCoreHelper();
auto surfaceStateSize = gfxCoreHelper.getRenderSurfaceStateSize();
auto ssIndex = getSurfaceStateIndexForBindlessOffset(argAsPtr.bindless);
if (ssIndex < std::numeric_limits<uint32_t>::max()) {
auto surfaceState = ptrOffset(getSurfaceStateHeap(), ssIndex * surfaceStateSize);
Buffer::setSurfaceState(&getDevice().getDevice(), surfaceState, forceNonAuxMode, disableL3, allocSize, ptrToPatch, offset, svmAlloc, 0, 0,
areMultipleSubDevicesInContext());
}
}
}