Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-10 12:53:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add debug flag to control alignment.

Browse Source 
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek
2022-11-08 17:32:18 +00:00
committed by Compute-Runtime-Automation
parent c06ddfc7b8
commit d9762c0337
4 changed files with 45 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
shared/source/debug_settings/debug_variables_base.inl
View File

@ -303,6 +303,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, SplitBcsCopy, -1, "-1: default, 0:disabled, 1: e
DECLARE_DEBUG_VARIABLE(int32_t, SplitBcsMask, 0, "0: default, >0: bitmask: indicates bcs engines for split")
DECLARE_DEBUG_VARIABLE(int32_t, ReuseKernelBinaries, -1, "-1: default, 0:disabled, 1: enabled. If enabled, driver reuses kernel binaries.")
DECLARE_DEBUG_VARIABLE(int32_t, SetAmountOfReusableAllocations, -1, "-1: default, 0:disabled, > 1: enabled. If enabled, driver will fill reusable allocation lists with given amount of command buffers and heaps at initialization of immediate command list.")
DECLARE_DEBUG_VARIABLE(int32_t, UseHighAlignmentForHeapExtended, -1, "-1: default, 0:disabled, > 1: enabled. If enabled, driver aligns HEAP_EXTENDED allocations to GPU VA that is next power of 2 for a given size, if disables GPU VA is using 2MB/64KB alignment.")
/*DIRECT SUBMISSION FLAGS*/
DECLARE_DEBUG_VARIABLE(int32_t, EnableDirectSubmission, -1, "-1: default (disabled), 0: disable, 1:enable. Enables direct submission of command buffers bypassing KMD")

10
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@ -1364,7 +1364,15 @@ uint64_t getGpuAddress(const AlignmentSelector &alignmentSelector, HeapAssigner
default:
AlignmentSelector::CandidateAlignment alignment = alignmentSelector.selectAlignment(sizeAllocated);
if (gfxPartition->getHeapLimit(HeapIndex::HEAP_EXTENDED) > 0 && !resource48Bit) {
alignment.alignment = Math::nextPowerOfTwo(sizeAllocated);
auto alignSize = true;
if (DebugManager.flags.UseHighAlignmentForHeapExtended.get() != -1) {
alignSize = !!DebugManager.flags.UseHighAlignmentForHeapExtended.get();
}
if (alignSize) {
alignment.alignment = Math::nextPowerOfTwo(sizeAllocated);
}
alignment.heap = HeapIndex::HEAP_EXTENDED;
}
gpuAddress = gmmHelper.canonize(gfxPartition->heapAllocateWithCustomAlignment(alignment.heap, sizeAllocated, alignment.alignment));

1
shared/test/common/test_files/igdrcl.config
View File

@ -481,3 +481,4 @@ ExperimentalSmallBufferPoolAllocator = -1
ForceZeDeviceCanAccessPerReturnValue = -1
AdjustThreadGroupDispatchSize = -1
ForceNonblockingExecbufferCalls = -1
UseHighAlignmentForHeapExtended = -1

34
shared/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp
View File

@ -5069,6 +5069,40 @@ TEST_F(DrmMemoryManagerWithLocalMemoryAndExplicitExpectationsTest, givenHeapExte
memoryManager->freeGraphicsMemory(allocation2);
}
TEST_F(DrmMemoryManagerWithLocalMemoryAndExplicitExpectationsTest, givenDebugVariableToDisableAddressAlignmentWhenAllocationsAreMadeTheyAreAlignedTo2MB) {
if (!memoryManager->getGfxPartition(rootDeviceIndex)->getHeapLimit(HeapIndex::HEAP_EXTENDED)) {
GTEST_SKIP();
}
DebugManagerStateRestore restorer;
DebugManager.flags.UseHighAlignmentForHeapExtended.set(0);
auto size = 16 * MemoryConstants::megaByte;
auto status = MemoryManager::AllocationStatus::Error;
AllocationData allocData;
allocData.size = size;
allocData.type = AllocationType::BUFFER;
allocData.rootDeviceIndex = rootDeviceIndex;
auto allocation = memoryManager->allocateGraphicsMemoryInDevicePool(allocData, status);
EXPECT_EQ(MemoryManager::AllocationStatus::Success, status);
ASSERT_NE(nullptr, allocation);
EXPECT_EQ(allocData.size, allocation->getUnderlyingBufferSize());
EXPECT_EQ(allocData.size, static_cast<DrmAllocation *>(allocation)->getBO()->peekSize());
EXPECT_TRUE(allocation->getGpuAddress() % MemoryConstants::pageSize2Mb == 0u);
size = 32 * MemoryConstants::megaByte;
allocData.size = size;
auto allocation2 = memoryManager->allocateGraphicsMemoryInDevicePool(allocData, status);
EXPECT_EQ(MemoryManager::AllocationStatus::Success, status);
ASSERT_NE(nullptr, allocation2);
EXPECT_EQ(allocData.size, allocation2->getUnderlyingBufferSize());
EXPECT_EQ(allocData.size, static_cast<DrmAllocation *>(allocation2)->getBO()->peekSize());
EXPECT_TRUE(allocation2->getGpuAddress() % MemoryConstants::pageSize2Mb == 0u);
memoryManager->freeGraphicsMemory(allocation);
memoryManager->freeGraphicsMemory(allocation2);
}
struct DrmMemoryManagerToTestCopyMemoryToAllocationBanks : public DrmMemoryManager {
DrmMemoryManagerToTestCopyMemoryToAllocationBanks(ExecutionEnvironment &executionEnvironment, size_t lockableLocalMemorySize)
: DrmMemoryManager(gemCloseWorkerMode::gemCloseWorkerInactive, false, false, executionEnvironment) {