intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 21:18:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: program pat index on mtl linux

Browse Source 
Enable programming pat indexes on mtl linux for device buffers.

Change DrmMemoryManager::allocateMemoryByKMD to use gemCreateExt.

Changes currently disabled, can be enabled with flag
DisableGemCreateExtSetPat=0

Related-To: NEO-7896

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2024-02-15 15:45:12 +00:00
committed by Compute-Runtime-Automation
parent 618f1dd346
commit ed011de03e
14 changed files with 64 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@@ -635,14 +635,25 @@ GraphicsAllocation *DrmMemoryManager::allocateMemoryByKMD(const AllocationData &
auto gmm = std::make_unique<Gmm>(executionEnvironment.rootDeviceEnvironments[allocationData.rootDeviceIndex]->getGmmHelper(), allocationData.hostPtr,
allocationData.size, 0u, CacheSettingsHelper::getGmmUsageType(allocationData.type, allocationData.flags.uncacheable, productHelper), systemMemoryStorageInfo, gmmRequirements);
size_t bufferSize = allocationData.size;
uint64_t gpuRange = acquireGpuRangeWithCustomAlignment(bufferSize, allocationData.rootDeviceIndex, HeapIndex::heapStandard64KB, allocationData.alignment);
auto alignment = allocationData.alignment;
if (bufferSize >= 2 * MemoryConstants::megaByte) {
alignment = MemoryConstants::pageSize2M;
}
uint64_t gpuRange = acquireGpuRangeWithCustomAlignment(bufferSize, allocationData.rootDeviceIndex, HeapIndex::heapStandard64KB, alignment);
auto &drm = getDrm(allocationData.rootDeviceIndex);
auto ioctlHelper = drm.getIoctlHelper();
uint32_t handle = ioctlHelper->createGem(bufferSize, static_cast<uint32_t>(allocationData.storageInfo.memoryBanks.to_ulong()));
int ret = -1;
uint32_t handle;
auto patIndex = drm.getPatIndex(gmm.get(), allocationData.type, CacheRegion::defaultRegion, CachePolicy::writeBack, false, MemoryPoolHelper::isSystemMemoryPool(memoryPool));
const bool tryToUseGemCreateExt = !debugManager.flags.DisableGemCreateExtSetPat.get();
if (tryToUseGemCreateExt && drm.getMemoryInfo()) {
ret = drm.getMemoryInfo()->createGemExtWithSingleRegion(allocationData.storageInfo.getMemoryBanks(), bufferSize, handle, patIndex, -1, allocationData.flags.isUSMHostAllocation);
}
if (0 != ret) {
auto ioctlHelper = drm.getIoctlHelper();
handle = ioctlHelper->createGem(bufferSize, static_cast<uint32_t>(allocationData.storageInfo.memoryBanks.to_ulong()));
}
std::unique_ptr<BufferObject, BufferObject::Deleter> bo(new BufferObject(allocationData.rootDeviceIndex, &drm, patIndex, handle, bufferSize, maxOsContextCount));
bo->setAddress(gpuRange);

2
shared/source/os_interface/linux/drm_neo.cpp
View File

@@ -1231,7 +1231,7 @@ uint64_t Drm::getPatIndex(Gmm *gmm, AllocationType allocationType, CacheRegion c
}
uint64_t patIndex = rootDeviceEnvironment.getGmmClientContext()->cachePolicyGetPATIndex(resourceInfo, usageType, compressed, cachable);
patIndex = productHelper.overridePatIndex(isUncachedType, patIndex);
patIndex = productHelper.overridePatIndex(isUncachedType, patIndex, allocationType);
UNRECOVERABLE_IF(patIndex == static_cast<uint64_t>(GMM_PAT_ERROR));