intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 09:09:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Pass canonized gpuAddress to GraphicsAllocation

Browse Source 
Related-To: NEO-6523

Signed-off-by: Krzysztof Gibala <krzysztof.gibala@intel.com>
This commit is contained in:
Krzysztof Gibala
2022-05-30 14:18:50 +00:00
committed by Compute-Runtime-Automation
parent d7420f1786
commit ae56d50b4f
13 changed files with 145 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
shared/source/os_interface/linux/drm_allocation.h
View File

@@ -41,19 +41,19 @@ class DrmAllocation : public GraphicsAllocation {
bufferObjects[0] = bo;
}
DrmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, BufferObject *bo, void *ptrIn, uint64_t gpuAddress, size_t sizeIn, MemoryPool::Type pool)
: DrmAllocation(rootDeviceIndex, 1, allocationType, bo, ptrIn, gpuAddress, sizeIn, pool) {}
DrmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, BufferObject *bo, void *ptrIn, uint64_t canonizedGpuAddress, size_t sizeIn, MemoryPool::Type pool)
: DrmAllocation(rootDeviceIndex, 1, allocationType, bo, ptrIn, canonizedGpuAddress, sizeIn, pool) {}
DrmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, BufferObject *bo, void *ptrIn, uint64_t gpuAddress, size_t sizeIn, MemoryPool::Type pool)
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, ptrIn, gpuAddress, 0, sizeIn, pool, MemoryManager::maxOsContextCount), bufferObjects(EngineLimits::maxHandleCount) {
DrmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, BufferObject *bo, void *ptrIn, uint64_t canonizedGpuAddress, size_t sizeIn, MemoryPool::Type pool)
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, ptrIn, canonizedGpuAddress, 0, sizeIn, pool, MemoryManager::maxOsContextCount), bufferObjects(EngineLimits::maxHandleCount) {
bufferObjects[0] = bo;
}
DrmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, BufferObjects &bos, void *ptrIn, uint64_t gpuAddress, size_t sizeIn, MemoryPool::Type pool)
: DrmAllocation(rootDeviceIndex, 1, allocationType, bos, ptrIn, gpuAddress, sizeIn, pool) {}
DrmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, BufferObjects &bos, void *ptrIn, uint64_t canonizedGpuAddress, size_t sizeIn, MemoryPool::Type pool)
: DrmAllocation(rootDeviceIndex, 1, allocationType, bos, ptrIn, canonizedGpuAddress, sizeIn, pool) {}
DrmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, BufferObjects &bos, void *ptrIn, uint64_t gpuAddress, size_t sizeIn, MemoryPool::Type pool)
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, ptrIn, gpuAddress, 0, sizeIn, pool, MemoryManager::maxOsContextCount),
DrmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, BufferObjects &bos, void *ptrIn, uint64_t canonizedGpuAddress, size_t sizeIn, MemoryPool::Type pool)
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, ptrIn, canonizedGpuAddress, 0, sizeIn, pool, MemoryManager::maxOsContextCount),
bufferObjects(bos) {
}

29
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@@ -274,7 +274,9 @@ void DrmMemoryManager::emitPinningRequest(BufferObject *bo, const AllocationData
DrmAllocation *DrmMemoryManager::createGraphicsAllocation(OsHandleStorage &handleStorage, const AllocationData &allocationData) {
auto hostPtr = const_cast<void *>(allocationData.hostPtr);
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, nullptr, hostPtr, castToUint64(hostPtr), allocationData.size, MemoryPool::System4KBPages);
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(castToUint64(hostPtr));
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, nullptr, hostPtr, canonizedGpuAddress, allocationData.size, MemoryPool::System4KBPages);
allocation->fragmentsStorage = handleStorage;
if (!allocation->setCacheRegion(&this->getDrm(allocationData.rootDeviceIndex), static_cast<CacheRegion>(allocationData.cacheRegion))) {
return nullptr;
@@ -347,7 +349,9 @@ DrmAllocation *DrmMemoryManager::createAllocWithAlignmentFromUserptr(const Alloc
obtainGpuAddress(allocationData, bo.get(), gpuAddress);
emitPinningRequest(bo.get(), allocationData);
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), res, bo->peekAddress(), size, MemoryPool::System4KBPages);
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(bo->peekAddress());
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), res, canonizedGpuAddress, size, MemoryPool::System4KBPages);
allocation->setDriverAllocatedCpuPtr(res);
allocation->setReservedAddressRange(reinterpret_cast<void *>(gpuAddress), alignedSVMSize);
if (!allocation->setCacheRegion(&this->getDrm(allocationData.rootDeviceIndex), static_cast<CacheRegion>(allocationData.cacheRegion))) {
@@ -597,8 +601,9 @@ DrmAllocation *DrmMemoryManager::allocate32BitGraphicsMemoryImpl(const Allocatio
bo->setAddress(gpuVirtualAddress);
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(ptrOffset(gpuVirtualAddress, inputPointerOffset));
auto allocation = new DrmAllocation(allocationData.rootDeviceIndex, allocationData.type, bo.get(), const_cast<void *>(allocationData.hostPtr),
gmmHelper->canonize(ptrOffset(gpuVirtualAddress, inputPointerOffset)),
canonizedGpuAddress,
allocationSize, MemoryPool::System4KBPagesWith32BitGpuAddressing);
allocation->set32BitAllocation(true);
allocation->setGpuBaseAddress(gmmHelper->canonize(gfxPartition->getHeapBase(allocatorToUse)));
@@ -635,8 +640,9 @@ DrmAllocation *DrmMemoryManager::allocate32BitGraphicsMemoryImpl(const Allocatio
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
// softpin to the GPU address, res if it uses limitedRange Allocation
auto canonizedGpuAddress = gmmHelper->canonize(gpuVA);
auto allocation = new DrmAllocation(allocationData.rootDeviceIndex, allocationData.type, bo.get(), ptrAlloc,
gmmHelper->canonize(gpuVA), alignedAllocationSize,
canonizedGpuAddress, alignedAllocationSize,
MemoryPool::System4KBPagesWith32BitGpuAddressing);
allocation->set32BitAllocation(true);
@@ -875,8 +881,9 @@ GraphicsAllocation *DrmMemoryManager::createPaddedAllocation(GraphicsAllocation
}
bo->setAddress(gpuRange);
auto gmmHelper = getGmmHelper(rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(ptrOffset(gpuRange, offset));
auto allocation = new DrmAllocation(rootDeviceIndex, inputGraphicsAllocation->getAllocationType(), bo.get(), srcPtr,
gmmHelper->canonize(ptrOffset(gpuRange, offset)), sizeWithPadding,
canonizedGpuAddress, sizeWithPadding,
inputGraphicsAllocation->getMemoryPool());
allocation->setReservedAddressRange(reinterpret_cast<void *>(gpuRange), sizeWithPadding);
@@ -1471,8 +1478,8 @@ GraphicsAllocation *DrmMemoryManager::allocateGraphicsMemoryInDevicePool(const A
auto gpuAddress = getGpuAddress(this->alignmentSelector, this->heapAssigner, *hwInfo,
allocationData.type, gfxPartition, sizeAllocated,
allocationData.hostPtr, allocationData.flags.resource48Bit, allocationData.flags.use32BitFrontWindow, *gmmHelper);
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, numHandles, allocationData.type, nullptr, nullptr, gpuAddress, sizeAligned, MemoryPool::LocalMemory);
auto canonizedGpuAddress = gmmHelper->canonize(gpuAddress);
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, numHandles, allocationData.type, nullptr, nullptr, canonizedGpuAddress, sizeAligned, MemoryPool::LocalMemory);
DrmAllocation *drmAllocation = static_cast<DrmAllocation *>(allocation.get());
GraphicsAllocation *graphicsAllocation = static_cast<GraphicsAllocation *>(allocation.get());
@@ -1728,7 +1735,9 @@ DrmAllocation *DrmMemoryManager::createAllocWithAlignment(const AllocationData &
obtainGpuAddress(allocationData, bo.get(), gpuAddress);
emitPinningRequest(bo.get(), allocationData);
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), cpuPointer, bo->peekAddress(), alignedSize, MemoryPool::System4KBPages);
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(bo->peekAddress());
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), cpuPointer, canonizedGpuAddress, alignedSize, MemoryPool::System4KBPages);
allocation->setMmapPtr(cpuPointer);
allocation->setMmapSize(alignedSize);
if (pointerDiff != 0) {
@@ -1864,7 +1873,9 @@ GraphicsAllocation *DrmMemoryManager::createSharedUnifiedMemoryAllocation(const
bo->setAddress(reinterpret_cast<uintptr_t>(cpuPointer));
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), cpuPointer, bo->peekAddress(), size, MemoryPool::System4KBPages);
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(bo->peekAddress());
auto allocation = std::make_unique<DrmAllocation>(allocationData.rootDeviceIndex, allocationData.type, bo.get(), cpuPointer, canonizedGpuAddress, size, MemoryPool::System4KBPages);
allocation->setMmapPtr(cpuBasePointer);
allocation->setMmapSize(totalSizeToAlloc);
if (!allocation->setCacheRegion(&this->getDrm(allocationData.rootDeviceIndex), static_cast<CacheRegion>(allocationData.cacheRegion))) {

4
shared/source/os_interface/linux/drm_memory_manager_create_multi_host_allocation.cpp
View File

@@ -36,8 +36,10 @@ DrmAllocation *DrmMemoryManager::createMultiHostAllocation(const AllocationData
gpuAddress = allocationData.gpuAddress;
}
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedGpuAddress = gmmHelper->canonize(gpuAddress);
auto allocation = new DrmAllocation(allocationData.rootDeviceIndex, numTiles, allocationData.type,
nullptr /*bo*/, cpuBasePointer, gpuAddress, sizePerTile, MemoryPool::System4KBPages);
nullptr /*bo*/, cpuBasePointer, canonizedGpuAddress, sizePerTile, MemoryPool::System4KBPages);
allocation->storageInfo = allocationData.storageInfo;
allocation->setFlushL3Required(true);

8
shared/source/os_interface/windows/wddm_allocation.h
View File

@@ -24,12 +24,12 @@ constexpr size_t trimListUnusedPosition = std::numeric_limits<size_t>::max();
class WddmAllocation : public GraphicsAllocation {
public:
WddmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, void *cpuPtrIn, size_t sizeIn, void *reservedAddr, MemoryPool::Type pool, uint32_t shareable, size_t maxOsContextCount)
: WddmAllocation(rootDeviceIndex, 1, allocationType, cpuPtrIn, sizeIn, reservedAddr, pool, shareable, maxOsContextCount) {}
WddmAllocation(uint32_t rootDeviceIndex, AllocationType allocationType, void *cpuPtrIn, uint64_t canonizedAddress, size_t sizeIn, void *reservedAddr, MemoryPool::Type pool, uint32_t shareable, size_t maxOsContextCount)
: WddmAllocation(rootDeviceIndex, 1, allocationType, cpuPtrIn, canonizedAddress, sizeIn, reservedAddr, pool, shareable, maxOsContextCount) {}
WddmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, void *cpuPtrIn, size_t sizeIn,
WddmAllocation(uint32_t rootDeviceIndex, size_t numGmms, AllocationType allocationType, void *cpuPtrIn, uint64_t canonizedAddress, size_t sizeIn,
void *reservedAddr, MemoryPool::Type pool, uint32_t shareable, size_t maxOsContextCount)
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, cpuPtrIn, castToUint64(cpuPtrIn), 0llu, sizeIn, pool, maxOsContextCount),
: GraphicsAllocation(rootDeviceIndex, numGmms, allocationType, cpuPtrIn, canonizedAddress, 0llu, sizeIn, pool, maxOsContextCount),
shareable(shareable), trimCandidateListPositions(maxOsContextCount, trimListUnusedPosition) {
reservedAddressRangeInfo.addressPtr = reservedAddr;
reservedAddressRangeInfo.rangeSize = sizeIn;

39
shared/source/os_interface/windows/wddm_memory_manager.cpp
View File

@@ -78,7 +78,7 @@ GraphicsAllocation *WddmMemoryManager::allocateMemoryByKMD(const AllocationData
CacheSettingsHelper::getGmmUsageType(allocationData.type, !!allocationData.flags.uncacheable, *hwInfo), false, systemMemoryStorageInfo, true);
auto allocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, nullptr, allocationData.size, nullptr,
allocationData.type, nullptr, 0, allocationData.size, nullptr,
MemoryPool::SystemCpuInaccessible, allocationData.flags.shareable, maxOsContextCount);
allocation->setDefaultGmm(gmm.get());
if (!createWddmAllocation(allocation.get(), nullptr)) {
@@ -93,7 +93,7 @@ GraphicsAllocation *WddmMemoryManager::allocateMemoryByKMD(const AllocationData
GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForImageImpl(const AllocationData &allocationData, std::unique_ptr<Gmm> gmm) {
auto allocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, nullptr, allocationData.imgInfo->size,
allocationData.type, nullptr, 0, allocationData.imgInfo->size,
nullptr, MemoryPool::SystemCpuInaccessible,
0u, // shareable
maxOsContextCount);
@@ -119,7 +119,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryUsingKmdAndMapItToC
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, nullptr,
allocationData.type, nullptr, 0,
sizeAligned, nullptr, allowLargePages ? MemoryPool::System64KBPages : MemoryPool::System4KBPages,
0u, // shareable
maxOsContextCount);
@@ -177,8 +177,10 @@ GraphicsAllocation *WddmMemoryManager::allocateHugeGraphicsMemory(const Allocati
auto chunkSize = getHugeGfxMemoryChunkSize(GfxMemoryAllocationMethod::UseUmdSystemPtr);
auto numGmms = (alignedSize + chunkSize - 1) / chunkSize;
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(const_cast<void *>(hostPtr)));
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex, numGmms,
allocationData.type, hostPtr, allocationData.size,
allocationData.type, hostPtr, canonizedAddress, allocationData.size,
nullptr, memoryPool,
0u, // shareable
maxOsContextCount);
@@ -241,9 +243,11 @@ GraphicsAllocation *WddmMemoryManager::allocateSystemMemoryAndCreateGraphicsAllo
return nullptr;
}
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(pSysMem));
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, pSysMem, sizeAligned,
allocationData.type, pSysMem, canonizedAddress, sizeAligned,
nullptr, MemoryPool::System4KBPages,
0u, // shareable
maxOsContextCount);
@@ -286,9 +290,11 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(co
return allocateHugeGraphicsMemory(allocationData, false);
}
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(const_cast<void *>(allocationData.hostPtr)));
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, const_cast<void *>(allocationData.hostPtr),
allocationData.type, const_cast<void *>(allocationData.hostPtr), canonizedAddress,
allocationData.size, nullptr, MemoryPool::System4KBPages,
0u, // shareable
maxOsContextCount);
@@ -328,9 +334,14 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryWithHostPtr(const A
return nullptr;
}
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(const_cast<void *>(inputPtr)));
auto allocation = new WddmAllocation(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, const_cast<void *>(inputPtr), allocationData.size,
allocationData.type,
const_cast<void *>(inputPtr),
canonizedAddress,
allocationData.size,
reserve, MemoryPool::System4KBPages,
0u, // shareable
maxOsContextCount);
@@ -372,9 +383,12 @@ GraphicsAllocation *WddmMemoryManager::allocate32BitGraphicsMemoryImpl(const All
sizeAligned = alignUp(sizeAligned, MemoryConstants::allocationAlignment);
}
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(const_cast<void *>(ptrAligned)));
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, const_cast<void *>(ptrAligned), sizeAligned, nullptr,
allocationData.type, const_cast<void *>(ptrAligned), canonizedAddress,
sizeAligned, nullptr,
MemoryPool::System4KBPagesWith32BitGpuAddressing,
0u, // shareable
maxOsContextCount);
@@ -395,7 +409,6 @@ GraphicsAllocation *WddmMemoryManager::allocate32BitGraphicsMemoryImpl(const All
return nullptr;
}
auto baseAddress = getGfxPartition(allocationData.rootDeviceIndex)->getHeapBase(heapAssigner.get32BitHeapIndex(allocationData.type, useLocalMemory, *hwInfo, allocationData.flags.use32BitFrontWindow));
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
wddmAllocation->setGpuBaseAddress(gmmHelper->canonize(baseAddress));
if (preferredAllocationMethod != GfxMemoryAllocationMethod::UseUmdSystemPtr) {
@@ -696,9 +709,13 @@ void WddmMemoryManager::obtainGpuAddressFromFragments(WddmAllocation *allocation
}
GraphicsAllocation *WddmMemoryManager::createGraphicsAllocation(OsHandleStorage &handleStorage, const AllocationData &allocationData) {
auto gmmHelper = getGmmHelper(allocationData.rootDeviceIndex);
auto canonizedAddress = gmmHelper->canonize(castToUint64(const_cast<void *>(allocationData.hostPtr)));
auto allocation = new WddmAllocation(allocationData.rootDeviceIndex,
1u, // numGmms
allocationData.type, const_cast<void *>(allocationData.hostPtr),
allocationData.type,
const_cast<void *>(allocationData.hostPtr),
canonizedAddress,
allocationData.size, nullptr, MemoryPool::System4KBPages,
0u, // shareable
maxOsContextCount);
@@ -1035,7 +1052,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryInDevicePool(const
const size_t numGmms = (static_cast<uint64_t>(sizeAligned) + chunkSize - 1) / chunkSize;
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.rootDeviceIndex, singleBankAllocation ? numGmms : numBanks,
allocationData.type, nullptr, sizeAligned, nullptr, MemoryPool::LocalMemory, allocationData.flags.shareable, maxOsContextCount);
allocationData.type, nullptr, 0, sizeAligned, nullptr, MemoryPool::LocalMemory, allocationData.flags.shareable, maxOsContextCount);
if (singleBankAllocation) {
if (numGmms > 1) {
splitGmmsInAllocation(gmmHelper, wddmAllocation.get(), alignment, chunkSize, const_cast<StorageInfo &>(allocationData.storageInfo));