intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-21 09:14:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: check indirect access for kernel

Browse Source 
Do not make indirect allocations resident if kernel does not use
indirect access.
For both level zero and opencl.
Currently disabled by default, enable with debug flag
DetectIndirectAccessInKernel

Related-To: NEO-7712

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2023-03-07 15:22:10 +00:00
committed by Compute-Runtime-Automation
parent c7311bc140
commit 69a16fd3ed
20 changed files with 354 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
opencl/source/kernel/kernel.cpp
View File

@@ -148,13 +148,13 @@ void Kernel::patchWithImplicitSurface(uint64_t ptrToPatchInCrossThreadData, Grap
}
cl_int Kernel::initialize() {
this->kernelHasIndirectAccess = false;
auto pClDevice = &getDevice();
auto rootDeviceIndex = pClDevice->getRootDeviceIndex();
reconfigureKernel();
auto &hwInfo = pClDevice->getHardwareInfo();
auto &rootDeviceEnvironment = pClDevice->getRootDeviceEnvironment();
auto &gfxCoreHelper = rootDeviceEnvironment.getHelper<GfxCoreHelper>();
auto &productHelper = rootDeviceEnvironment.getHelper<ProductHelper>();
auto &kernelDescriptor = kernelInfo.kernelDescriptor;
const auto &implicitArgs = kernelDescriptor.payloadMappings.implicitArgs;
const auto &explicitArgs = kernelDescriptor.payloadMappings.explicitArgs;
@@ -281,11 +281,19 @@ cl_int Kernel::initialize() {
slmSizes.resize(numArgs);
this->setInlineSamplers();
this->kernelHasIndirectAccess |= kernelInfo.kernelDescriptor.kernelAttributes.hasNonKernelArgLoad ||
kernelInfo.kernelDescriptor.kernelAttributes.hasNonKernelArgStore ||
kernelInfo.kernelDescriptor.kernelAttributes.hasNonKernelArgAtomic;
bool detectIndirectAccessInKernel = productHelper.isDetectIndirectAccessInKernelSupported(kernelDescriptor);
if (DebugManager.flags.DetectIndirectAccessInKernel.get() != -1) {
detectIndirectAccessInKernel = DebugManager.flags.DetectIndirectAccessInKernel.get() == 1;
}
if (detectIndirectAccessInKernel) {
this->kernelHasIndirectAccess = kernelDescriptor.kernelAttributes.hasNonKernelArgLoad ||
kernelDescriptor.kernelAttributes.hasNonKernelArgStore ||
kernelDescriptor.kernelAttributes.hasNonKernelArgAtomic ||
kernelDescriptor.kernelAttributes.hasIndirectStatelessAccess ||
NEO::KernelHelper::isAnyArgumentPtrByValue(kernelDescriptor);
} else {
this->kernelHasIndirectAccess = true;
}
provideInitializationHints();
// resolve the new kernel info to account for kernel handlers
// I think by this time we have decoded the binary and know the number of args etc.
@@ -1311,7 +1319,7 @@ void Kernel::makeResident(CommandStreamReceiver &commandStreamReceiver) {
}
}
if (unifiedMemoryControls.indirectSharedAllocationsAllowed && pageFaultManager) {
if (getHasIndirectAccess() && unifiedMemoryControls.indirectSharedAllocationsAllowed && pageFaultManager) {
pageFaultManager->moveAllocationsWithinUMAllocsManagerToGpuDomain(this->getContext().getSVMAllocsManager());
}
makeArgsResident(commandStreamReceiver);
@@ -1323,9 +1331,9 @@ void Kernel::makeResident(CommandStreamReceiver &commandStreamReceiver) {
gtpinNotifyMakeResident(this, &commandStreamReceiver);
if (unifiedMemoryControls.indirectDeviceAllocationsAllowed ||
unifiedMemoryControls.indirectHostAllocationsAllowed ||
unifiedMemoryControls.indirectSharedAllocationsAllowed) {
if (getHasIndirectAccess() && (unifiedMemoryControls.indirectDeviceAllocationsAllowed ||
unifiedMemoryControls.indirectHostAllocationsAllowed ||
unifiedMemoryControls.indirectSharedAllocationsAllowed)) {
this->getContext().getSVMAllocsManager()->makeInternalAllocationsResident(commandStreamReceiver, unifiedMemoryControls.generateMask());
}
}