intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-04 15:53:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add multiple root devices support to memory manager

Browse Source 
Related-To: NEO-2941

Change-Id: If4977c19dbd083f2d1559a40f9b4e911eab4859c
Signed-off-by: Igor Venevtsev <igor.venevtsev@intel.com>
This commit is contained in:
Igor Venevtsev
2019-10-22 10:26:23 +02:00
committed by sys_ocldev
parent a6f8a19290
commit fed673861f
66 changed files with 445 additions and 324 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
runtime/os_interface/windows/wddm_memory_manager.cpp
View File

@@ -46,7 +46,10 @@ WddmMemoryManager::WddmMemoryManager(ExecutionEnvironment &executionEnvironment)
if (asyncDeleterEnabled)
deferredDeleter = createDeferredDeleter();
mallocRestrictions.minAddress = wddm->getWddmMinAddress();
wddm->initGfxPartition(*gfxPartition);
for (uint32_t rootDeviceIndex = 0; rootDeviceIndex < gfxPartitions.size(); ++rootDeviceIndex) {
wddm->initGfxPartition(*getGfxPartition(rootDeviceIndex));
}
}
GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForImageImpl(const AllocationData &allocationData, std::unique_ptr<Gmm> gmm) {
@@ -54,7 +57,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForImageImpl(const
return allocateGraphicsMemoryWithAlignment(allocationData);
}
auto allocation = std::make_unique<WddmAllocation>(allocationData.type, nullptr, allocationData.imgInfo->size, nullptr, MemoryPool::SystemCpuInaccessible);
auto allocation = std::make_unique<WddmAllocation>(allocationData.type, nullptr, allocationData.imgInfo->size, nullptr, MemoryPool::SystemCpuInaccessible, allocationData.rootDeviceIndex);
allocation->setDefaultGmm(gmm.get());
if (!createWddmAllocation(allocation.get(), nullptr)) {
return nullptr;
@@ -68,7 +71,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForImageImpl(const
GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemory64kb(const AllocationData &allocationData) {
size_t sizeAligned = alignUp(allocationData.size, MemoryConstants::pageSize64k);
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, nullptr, sizeAligned, nullptr, MemoryPool::System64KBPages);
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, nullptr, sizeAligned, nullptr, MemoryPool::System64KBPages, allocationData.rootDeviceIndex);
auto gmm = new Gmm(nullptr, sizeAligned, false, allocationData.flags.preferRenderCompressed, true, {});
wddmAllocation->setDefaultGmm(gmm);
@@ -98,7 +101,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryWithAlignment(const
return nullptr;
}
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, pSysMem, sizeAligned, nullptr, MemoryPool::System4KBPages);
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, pSysMem, sizeAligned, nullptr, MemoryPool::System4KBPages, allocationData.rootDeviceIndex);
wddmAllocation->setDriverAllocatedCpuPtr(pSysMem);
gmm = new Gmm(pSysMem, sizeAligned, allocationData.flags.uncacheable);
@@ -133,7 +136,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(co
auto alignedSize = alignSizeWholePage(allocationData.hostPtr, allocationData.size);
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, const_cast<void *>(allocationData.hostPtr),
allocationData.size, nullptr, MemoryPool::System4KBPages);
allocationData.size, nullptr, MemoryPool::System4KBPages, allocationData.rootDeviceIndex);
wddmAllocation->setAllocationOffset(offsetInPage);
auto gmm = new Gmm(alignedPtr, alignedSize, false);
@@ -160,7 +163,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryWithHostPtr(const A
return nullptr;
}
auto allocation = new WddmAllocation(allocationData.type, const_cast<void *>(inputPtr), allocationData.size, reserve, MemoryPool::System4KBPages);
auto allocation = new WddmAllocation(allocationData.type, const_cast<void *>(inputPtr), allocationData.size, reserve, MemoryPool::System4KBPages, allocationData.rootDeviceIndex);
allocation->setAllocationOffset(offset);
Gmm *gmm = new Gmm(ptrAligned, sizeAligned, false);
@@ -195,7 +198,7 @@ GraphicsAllocation *WddmMemoryManager::allocate32BitGraphicsMemoryImpl(const All
}
auto wddmAllocation = std::make_unique<WddmAllocation>(allocationData.type, const_cast<void *>(ptrAligned), sizeAligned, nullptr,
MemoryPool::System4KBPagesWith32BitGpuAddressing);
MemoryPool::System4KBPagesWith32BitGpuAddressing, allocationData.rootDeviceIndex);
wddmAllocation->setDriverAllocatedCpuPtr(pSysMem);
wddmAllocation->set32BitAllocation(true);
wddmAllocation->setAllocationOffset(offset);
@@ -209,14 +212,14 @@ GraphicsAllocation *WddmMemoryManager::allocate32BitGraphicsMemoryImpl(const All
return nullptr;
}
auto baseAddress = useInternal32BitAllocator(allocationData.type) ? getInternalHeapBaseAddress() : getExternalHeapBaseAddress();
auto baseAddress = useInternal32BitAllocator(allocationData.type) ? getInternalHeapBaseAddress(allocationData.rootDeviceIndex) : getExternalHeapBaseAddress(allocationData.rootDeviceIndex);
wddmAllocation->setGpuBaseAddress(GmmHelper::canonize(baseAddress));
return wddmAllocation.release();
}
GraphicsAllocation *WddmMemoryManager::createAllocationFromHandle(osHandle handle, bool requireSpecificBitness, bool ntHandle, GraphicsAllocation::AllocationType allocationType) {
auto allocation = std::make_unique<WddmAllocation>(allocationType, nullptr, 0, handle, MemoryPool::SystemCpuInaccessible);
GraphicsAllocation *WddmMemoryManager::createAllocationFromHandle(osHandle handle, bool requireSpecificBitness, bool ntHandle, GraphicsAllocation::AllocationType allocationType, uint32_t rootDeviceIndex) {
auto allocation = std::make_unique<WddmAllocation>(allocationType, nullptr, 0, handle, MemoryPool::SystemCpuInaccessible, rootDeviceIndex);
bool status = ntHandle ? wddm->openNTHandle(reinterpret_cast<HANDLE>(static_cast<uintptr_t>(handle)), allocation.get())
: wddm->openSharedHandle(handle, allocation.get());
@@ -237,7 +240,7 @@ GraphicsAllocation *WddmMemoryManager::createAllocationFromHandle(osHandle handl
allocation->setReservedAddressRange(ptr, size);
} else if (requireSpecificBitness && this->force32bitAllocations) {
allocation->set32BitAllocation(true);
allocation->setGpuBaseAddress(GmmHelper::canonize(getExternalHeapBaseAddress()));
allocation->setGpuBaseAddress(GmmHelper::canonize(getExternalHeapBaseAddress(allocation->getRootDeviceIndex())));
}
status = mapGpuVirtualAddress(allocation.get(), allocation->getReservedAddressPtr());
DEBUG_BREAK_IF(!status);
@@ -251,11 +254,12 @@ GraphicsAllocation *WddmMemoryManager::createAllocationFromHandle(osHandle handl
}
GraphicsAllocation *WddmMemoryManager::createGraphicsAllocationFromSharedHandle(osHandle handle, const AllocationProperties &properties, bool requireSpecificBitness) {
return createAllocationFromHandle(handle, requireSpecificBitness, false, properties.allocationType);
uint32_t rootDeviceIndex = properties.rootDeviceIndex == AllocationProperties::noDeviceSpecified ? 0 : properties.rootDeviceIndex;
return createAllocationFromHandle(handle, requireSpecificBitness, false, properties.allocationType, rootDeviceIndex);
}
GraphicsAllocation *WddmMemoryManager::createGraphicsAllocationFromNTHandle(void *handle) {
return createAllocationFromHandle((osHandle)((UINT_PTR)handle), false, true, GraphicsAllocation::AllocationType::SHARED_IMAGE);
GraphicsAllocation *WddmMemoryManager::createGraphicsAllocationFromNTHandle(void *handle, uint32_t rootDeviceIndex) {
return createAllocationFromHandle((osHandle)((UINT_PTR)handle), false, true, GraphicsAllocation::AllocationType::SHARED_IMAGE, rootDeviceIndex);
}
void WddmMemoryManager::addAllocationToHostPtrManager(GraphicsAllocation *gfxAllocation) {
@@ -469,7 +473,7 @@ void WddmMemoryManager::obtainGpuAddressFromFragments(WddmAllocation *allocation
}
GraphicsAllocation *WddmMemoryManager::createGraphicsAllocation(OsHandleStorage &handleStorage, const AllocationData &allocationData) {
auto allocation = new WddmAllocation(allocationData.type, const_cast<void *>(allocationData.hostPtr), allocationData.size, nullptr, MemoryPool::System4KBPages);
auto allocation = new WddmAllocation(allocationData.type, const_cast<void *>(allocationData.hostPtr), allocationData.size, nullptr, MemoryPool::System4KBPages, allocationData.rootDeviceIndex);
allocation->fragmentsStorage = handleStorage;
obtainGpuAddressFromFragments(allocation, handleStorage);
return allocation;
@@ -504,6 +508,7 @@ bool WddmMemoryManager::mapGpuVaForOneHandleAllocation(WddmAllocation *allocatio
if (allocation->reservedGpuVirtualAddress) {
addressToMap = allocation->reservedGpuVirtualAddress;
}
auto gfxPartition = getGfxPartition(allocation->getRootDeviceIndex());
auto status = wddm->mapGpuVirtualAddress(allocation->getDefaultGmm(), allocation->getDefaultHandle(),
gfxPartition->getHeapMinimalAddress(heapIndex), gfxPartition->getHeapLimit(heapIndex),
addressToMap, allocation->getGpuAddressToModify());
@@ -543,6 +548,7 @@ void WddmMemoryManager::obtainGpuAddressIfNeeded(WddmAllocation *allocation) {
if (allocation->getNumHandles() > 1u) {
auto heapIndex = selectHeap(allocation, false, executionEnvironment.isFullRangeSvm());
allocation->reservedSizeForGpuVirtualAddress = allocation->getAlignedSize();
auto gfxPartition = getGfxPartition(allocation->getRootDeviceIndex());
allocation->reservedGpuVirtualAddress = wddm->reserveGpuVirtualAddress(gfxPartition->getHeapMinimalAddress(heapIndex),
gfxPartition->getHeapLimit(heapIndex),
allocation->reservedSizeForGpuVirtualAddress);