intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-22 10:17:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Introduced standard heap with 2MB alignment on Linux

Browse Source 
Related-To: NEO-5507

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
This commit is contained in:
Slawomir Milczarek
2021-03-01 12:11:41 +00:00
committed by Compute-Runtime-Automation
parent c1cabb505a
commit 049166688c
7 changed files with 47 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
opencl/test/unit_test/memory_manager/gfx_partition_tests.inl
View File

@@ -80,7 +80,8 @@ void testGfxPartition(MockGfxPartition &gfxPartition, uint64_t gfxBase, uint64_t
gfxBase += sizeHeap32;
}
uint64_t sizeStandard = (gfxTop - gfxBase) >> 1;
uint32_t numStandardHeaps = static_cast<uint32_t>(HeapIndex::HEAP_STANDARD2MB) - static_cast<uint32_t>(HeapIndex::HEAP_STANDARD) + 1;
uint64_t sizeStandard = alignDown((gfxTop - gfxBase) / numStandardHeaps, GfxPartition::heapGranularity);
EXPECT_TRUE(gfxPartition.heapInitialized(HeapIndex::HEAP_STANDARD));
auto heapStandardBase = gfxPartition.getHeapBase(HeapIndex::HEAP_STANDARD);
@@ -94,11 +95,19 @@ void testGfxPartition(MockGfxPartition &gfxPartition, uint64_t gfxBase, uint64_t
auto heapStandard64KbBase = gfxPartition.getHeapBase(HeapIndex::HEAP_STANDARD64KB);
auto heapStandard64KbSize = gfxPartition.getHeapSize(HeapIndex::HEAP_STANDARD64KB);
EXPECT_TRUE(isAligned<GfxPartition::heapGranularity>(heapStandard64KbBase));
EXPECT_EQ(heapStandard64KbBase, heapStandardBase + heapStandardSize);
EXPECT_EQ(heapStandard64KbSize, heapStandardSize);
EXPECT_EQ(heapStandard64KbBase + heapStandard64KbSize, gfxTop);
EXPECT_EQ(gfxBase + sizeStandard, gfxTop);
gfxBase += sizeStandard;
EXPECT_TRUE(gfxPartition.heapInitialized(HeapIndex::HEAP_STANDARD2MB));
auto heapStandard2MbBase = gfxPartition.getHeapBase(HeapIndex::HEAP_STANDARD2MB);
auto heapStandard2MbSize = gfxPartition.getHeapSize(HeapIndex::HEAP_STANDARD2MB);
EXPECT_TRUE(isAligned<GfxPartition::heapGranularity>(heapStandard2MbBase));
EXPECT_EQ(heapStandard2MbBase, heapStandard64KbBase + heapStandard64KbSize);
EXPECT_EQ(heapStandard2MbSize, heapStandard64KbSize);
EXPECT_LE(heapStandard2MbBase + heapStandard2MbSize, gfxTop);
EXPECT_LE(gfxBase + sizeStandard, gfxTop);
EXPECT_EQ(gfxPartition.getHeapMinimalAddress(HeapIndex::HEAP_INTERNAL_FRONT_WINDOW), gfxPartition.getHeapBase(HeapIndex::HEAP_INTERNAL_FRONT_WINDOW));
EXPECT_EQ(gfxPartition.getHeapMinimalAddress(HeapIndex::HEAP_INTERNAL_DEVICE_FRONT_WINDOW), gfxPartition.getHeapBase(HeapIndex::HEAP_INTERNAL_DEVICE_FRONT_WINDOW));
@@ -113,7 +122,9 @@ void testGfxPartition(MockGfxPartition &gfxPartition, uint64_t gfxBase, uint64_t
const bool isInternalHeapType = heap == HeapIndex::HEAP_INTERNAL || heap == HeapIndex::HEAP_INTERNAL_DEVICE_MEMORY;
if (isInternalHeapType) {
if (heap == HeapIndex::HEAP_SVM) {
EXPECT_EQ(gfxPartition.getHeapMinimalAddress(heap), gfxPartition.getHeapBase(heap));
} else if (isInternalHeapType) {
EXPECT_EQ(gfxPartition.getHeapMinimalAddress(heap), gfxPartition.getHeapBase(heap) + GfxPartition::internalFrontWindowPoolSize);
} else {
EXPECT_GT(gfxPartition.getHeapMinimalAddress(heap), gfxPartition.getHeapBase(heap));
@@ -202,7 +213,8 @@ TEST(GfxPartitionTest, GivenFullRange48BitSvmHeap64KbSplitWhenTestingGfxPartitio
uint64_t gfxBase = is32bit ? MemoryConstants::maxSvmAddress + 1 : maxNBitValue(48 - 1) + 1;
uint64_t gfxTop = maxNBitValue(48) + 1;
auto heapStandardSize = (gfxTop - gfxBase - 4 * sizeHeap32) / 2;
auto numStandardHeaps = static_cast<uint32_t>(HeapIndex::HEAP_STANDARD2MB) - static_cast<uint32_t>(HeapIndex::HEAP_STANDARD) + 1;
auto heapStandardSize = alignDown((gfxTop - gfxBase - 4 * sizeHeap32) / numStandardHeaps, GfxPartition::heapGranularity);
auto heapStandard64KBSize = alignDown(heapStandardSize / numRootDevices, GfxPartition::heapGranularity);
EXPECT_EQ(heapStandard64KBSize, gfxPartition.getHeapSize(HeapIndex::HEAP_STANDARD64KB));
@@ -219,7 +231,8 @@ TEST(GfxPartitionTest, GivenFullRange47BitSvmHeap64KbSplitWhenTestingGfxPartitio
uint64_t gfxBase = is32bit ? MemoryConstants::maxSvmAddress + 1 : (uint64_t)gfxPartition.getReservedCpuAddressRange();
uint64_t gfxTop = is32bit ? maxNBitValue(47) + 1 : gfxBase + gfxPartition.getReservedCpuAddressRangeSize();
auto heapStandardSize = ((gfxTop - gfxBase) - 4 * sizeHeap32) / 2;
auto numStandardHeaps = static_cast<uint32_t>(HeapIndex::HEAP_STANDARD2MB) - static_cast<uint32_t>(HeapIndex::HEAP_STANDARD) + 1;
auto heapStandardSize = alignDown((gfxTop - gfxBase - 4 * sizeHeap32) / numStandardHeaps, GfxPartition::heapGranularity);
auto heapStandard64KBSize = alignDown(heapStandardSize / numRootDevices, GfxPartition::heapGranularity);
EXPECT_EQ(heapStandard64KBSize, gfxPartition.getHeapSize(HeapIndex::HEAP_STANDARD64KB));

8
opencl/test/unit_test/os_interface/linux/drm_memory_manager_allocate_in_device_pool_tests_dg1.cpp
View File

@@ -581,17 +581,15 @@ TEST_F(DrmMemoryManagerLocalMemoryTest, givenNotSetUseSystemMemoryWhenGraphicsAl
EXPECT_NE(0u, gpuAddress);
auto heap = HeapIndex::HEAP_STANDARD64KB;
auto sizeReserved = alignUp(sizeAligned, 2 * MemoryConstants::megaByte);
if (memoryManager->getGfxPartition(0)->getHeapLimit(HeapIndex::HEAP_EXTENDED)) {
heap = HeapIndex::HEAP_EXTENDED;
sizeReserved = alignUp(sizeAligned, MemoryConstants::pageSize);
}
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(0)->getHeapBase(heap)), gpuAddress);
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(0)->getHeapLimit(heap)), gpuAddress);
EXPECT_EQ(0u, allocation->getGpuBaseAddress());
EXPECT_EQ(sizeAligned, allocation->getUnderlyingBufferSize());
EXPECT_EQ(gpuAddress, reinterpret_cast<uint64_t>(allocation->getReservedAddressPtr()));
EXPECT_EQ(sizeReserved, allocation->getReservedAddressSize());
EXPECT_EQ(sizeAligned, allocation->getReservedAddressSize());
EXPECT_EQ(1u, allocation->storageInfo.getNumBanks());
EXPECT_EQ(allocData.storageInfo.getMemoryBanks(), allocation->storageInfo.getMemoryBanks());
@@ -642,7 +640,7 @@ TEST_F(DrmMemoryManagerLocalMemoryTest, givenNotSetUseSystemMemoryWhenGraphicsAl
EXPECT_EQ(0u, allocation->getGpuBaseAddress());
EXPECT_EQ(sizeAligned, allocation->getUnderlyingBufferSize());
EXPECT_EQ(gpuAddress, reinterpret_cast<uint64_t>(allocation->getReservedAddressPtr()));
EXPECT_EQ(alignUp(sizeAligned, 2 * MemoryConstants::megaByte), allocation->getReservedAddressSize());
EXPECT_EQ(sizeAligned, allocation->getReservedAddressSize());
EXPECT_EQ(1u, allocation->storageInfo.getNumBanks());
EXPECT_EQ(allocData.storageInfo.getMemoryBanks(), allocation->storageInfo.getMemoryBanks());
@@ -1254,7 +1252,7 @@ TEST_F(DrmMemoryManagerLocalMemoryTest, givenGraphicsAllocationInDevicePoolIsAll
EXPECT_NE(0u, gpuAddress);
EXPECT_EQ(sizeAlignedTo64KB, allocation->getUnderlyingBufferSize());
EXPECT_EQ(gpuAddress, reinterpret_cast<uint64_t>(allocation->getReservedAddressPtr()));
EXPECT_EQ(alignUp(sizeAlignedTo64KB, 2 * MemoryConstants::megaByte), allocation->getReservedAddressSize());
EXPECT_EQ(sizeAlignedTo64KB, allocation->getReservedAddressSize());
auto drmAllocation = static_cast<DrmAllocation *>(allocation);
auto bo = drmAllocation->getBO();

16
opencl/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp
View File

@@ -1061,7 +1061,7 @@ TEST_F(DrmMemoryManagerTest, GivenMemoryManagerWhenCreatingGraphicsAllocation64k
EXPECT_EQ(nullptr, allocation);
}
TEST_F(DrmMemoryManagerTest, givenRequiresStandard64KBHeapSetToFalseThenStandardHeapIsUsed) {
TEST_F(DrmMemoryManagerTest, givenRequiresStandard2MBHeapSetToFalseThenStandardHeapIsUsed) {
const uint32_t rootDeviceIndex = 0;
size_t bufferSize = 4096u;
uint64_t range = memoryManager->acquireGpuRange(bufferSize, false, rootDeviceIndex, false);
@@ -1070,13 +1070,13 @@ TEST_F(DrmMemoryManagerTest, givenRequiresStandard64KBHeapSetToFalseThenStandard
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_STANDARD)), range);
}
TEST_F(DrmMemoryManagerTest, givenRequiresStandard64KBHeapSetToTrueThenStandard64KBHeapIsUsed) {
TEST_F(DrmMemoryManagerTest, givenRequiresStandard2MBHeapSetToTrueThenStandard2MBHeapIsUsed) {
const uint32_t rootDeviceIndex = 0;
size_t bufferSize = 4096u;
uint64_t range = memoryManager->acquireGpuRange(bufferSize, false, rootDeviceIndex, true);
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapBase(HeapIndex::HEAP_STANDARD64KB)), range);
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_STANDARD64KB)), range);
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapBase(HeapIndex::HEAP_STANDARD2MB)), range);
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_STANDARD2MB)), range);
}
TEST_F(DrmMemoryManagerTest, GivenShareableEnabledWhenAskedToCreateGraphicsAllocationThenValidAllocationIsReturnedAndStandard64KBHeapIsUsed) {
@@ -1090,8 +1090,8 @@ TEST_F(DrmMemoryManagerTest, GivenShareableEnabledWhenAskedToCreateGraphicsAlloc
EXPECT_NE(nullptr, allocation);
EXPECT_NE(0u, allocation->getGpuAddress());
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(allocation->getRootDeviceIndex())->getHeapBase(HeapIndex::HEAP_STANDARD64KB)), allocation->getGpuAddress());
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(allocation->getRootDeviceIndex())->getHeapLimit(HeapIndex::HEAP_STANDARD64KB)), allocation->getGpuAddress());
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(allocation->getRootDeviceIndex())->getHeapBase(HeapIndex::HEAP_STANDARD2MB)), allocation->getGpuAddress());
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(allocation->getRootDeviceIndex())->getHeapLimit(HeapIndex::HEAP_STANDARD2MB)), allocation->getGpuAddress());
memoryManager->freeGraphicsMemory(allocation);
}
@@ -2008,8 +2008,8 @@ TEST_F(DrmMemoryManagerWithLocalMemoryTest, givenDrmMemoryManagerWithLocalMemory
EXPECT_EQ(this->mock->inputFd, static_cast<int32_t>(handle));
auto gpuAddress = graphicsAllocation->getGpuAddress();
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapBase(HeapIndex::HEAP_STANDARD64KB)), gpuAddress);
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_STANDARD64KB)), gpuAddress);
EXPECT_LT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapBase(HeapIndex::HEAP_STANDARD2MB)), gpuAddress);
EXPECT_GT(GmmHelper::canonize(memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_STANDARD2MB)), gpuAddress);
DrmAllocation *drmAllocation = static_cast<DrmAllocation *>(graphicsAllocation);
auto bo = drmAllocation->getBO();

10
shared/source/memory_manager/gfx_partition.cpp
View File

@@ -184,14 +184,20 @@ bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToRe
gfxBase += gfxHeap32Size;
}
uint64_t gfxStandardSize = alignDown((gfxTop - gfxBase) >> 1, heapGranularity);
uint32_t numStandardHeaps = static_cast<uint32_t>(HeapIndex::HEAP_STANDARD2MB) - static_cast<uint32_t>(HeapIndex::HEAP_STANDARD) + 1;
uint64_t gfxStandardSize = alignDown((gfxTop - gfxBase) / numStandardHeaps, heapGranularity);
heapInit(HeapIndex::HEAP_STANDARD, gfxBase, gfxStandardSize);
gfxBase += gfxStandardSize;
// Split HEAP_STANDARD64K among root devices
auto gfxStandard64KBSize = alignDown(gfxStandardSize / numRootDevices, GfxPartition::heapGranularity);
heapInitWithAllocationAlignment(HeapIndex::HEAP_STANDARD64KB, gfxBase + rootDeviceIndex * gfxStandard64KBSize, gfxStandard64KBSize, 2 * MemoryConstants::megaByte);
heapInitWithAllocationAlignment(HeapIndex::HEAP_STANDARD64KB, gfxBase + rootDeviceIndex * gfxStandard64KBSize, gfxStandard64KBSize, MemoryConstants::pageSize64k);
gfxBase += gfxStandardSize;
// Split HEAP_STANDARD2MB among root devices
auto gfxStandard2MBSize = alignDown(gfxStandardSize / numRootDevices, GfxPartition::heapGranularity);
heapInitWithAllocationAlignment(HeapIndex::HEAP_STANDARD2MB, gfxBase + rootDeviceIndex * gfxStandard2MBSize, gfxStandard2MBSize, 2 * MemoryConstants::megaByte);
return true;
}

6
shared/source/memory_manager/gfx_partition.h
View File

@@ -23,6 +23,7 @@ enum class HeapIndex : uint32_t {
HEAP_EXTERNAL = 3u,
HEAP_STANDARD,
HEAP_STANDARD64KB,
HEAP_STANDARD2MB,
HEAP_SVM,
HEAP_EXTENDED,
HEAP_EXTERNAL_FRONT_WINDOW,
@@ -83,7 +84,8 @@ class GfxPartition {
}
uint64_t getHeapMinimalAddress(HeapIndex heapIndex) {
if (heapIndex == HeapIndex::HEAP_EXTERNAL_DEVICE_FRONT_WINDOW ||
if (heapIndex == HeapIndex::HEAP_SVM ||
heapIndex == HeapIndex::HEAP_EXTERNAL_DEVICE_FRONT_WINDOW ||
heapIndex == HeapIndex::HEAP_EXTERNAL_FRONT_WINDOW ||
heapIndex == HeapIndex::HEAP_INTERNAL_DEVICE_FRONT_WINDOW ||
heapIndex == HeapIndex::HEAP_INTERNAL_FRONT_WINDOW) {
@@ -103,7 +105,7 @@ class GfxPartition {
bool isLimitedRange() { return getHeap(HeapIndex::HEAP_SVM).getSize() == 0ull; }
static const uint64_t heapGranularity = MemoryConstants::pageSize64k;
static const uint64_t heapGranularity = 2 * MemoryConstants::megaByte;
static constexpr size_t externalFrontWindowPoolSize = 16 * MemoryConstants::megaByte;
static constexpr size_t internalFrontWindowPoolSize = 1 * MemoryConstants::megaByte;

6
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@@ -161,13 +161,13 @@ uint32_t DrmMemoryManager::unreference(NEO::BufferObject *bo, bool synchronousDe
return r;
}
uint64_t DrmMemoryManager::acquireGpuRange(size_t &size, bool specificBitness, uint32_t rootDeviceIndex, bool requiresStandard64KBHeap) {
uint64_t DrmMemoryManager::acquireGpuRange(size_t &size, bool specificBitness, uint32_t rootDeviceIndex, bool requiresStandard2MBHeap) {
auto gfxPartition = getGfxPartition(rootDeviceIndex);
if (specificBitness && this->force32bitAllocations) {
return GmmHelper::canonize(gfxPartition->heapAllocate(HeapIndex::HEAP_EXTERNAL, size));
}
if (requiresStandard64KBHeap) {
return GmmHelper::canonize(gfxPartition->heapAllocate(HeapIndex::HEAP_STANDARD64KB, size));
if (requiresStandard2MBHeap) {
return GmmHelper::canonize(gfxPartition->heapAllocate(HeapIndex::HEAP_STANDARD2MB, size));
}
return GmmHelper::canonize(gfxPartition->heapAllocate(HeapIndex::HEAP_STANDARD, size));
}

2
shared/source/os_interface/linux/drm_memory_manager.h
View File

@@ -78,7 +78,7 @@ class DrmMemoryManager : public MemoryManager {
void pushSharedBufferObject(BufferObject *bo);
BufferObject *allocUserptr(uintptr_t address, size_t size, uint64_t flags, uint32_t rootDeviceIndex);
bool setDomainCpu(GraphicsAllocation &graphicsAllocation, bool writeEnable);
uint64_t acquireGpuRange(size_t &size, bool requireSpecificBitness, uint32_t rootDeviceIndex, bool requiresStandard64KBHeap);
uint64_t acquireGpuRange(size_t &size, bool requireSpecificBitness, uint32_t rootDeviceIndex, bool requiresStandard2MBHeap);
MOCKABLE_VIRTUAL void releaseGpuRange(void *address, size_t size, uint32_t rootDeviceIndex);
void emitPinningRequest(BufferObject *bo, const AllocationData &allocationData) const;
uint32_t getDefaultDrmContextId() const;