intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-18 22:08:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: set misaligned source memory 1-way coherent

Browse Source 
Related-To: NEO-14443

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
This commit is contained in:
Szymon Morek
2025-03-28 12:24:28 +00:00
committed by Compute-Runtime-Automation
parent 98b50661a5
commit 3fff3dd77b
8 changed files with 47 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
level_zero/core/source/device/device_imp.cpp
View File

@@ -1750,7 +1750,8 @@ NEO::GraphicsAllocation *DeviceImp::allocateMemoryFromHostPtr(const void *buffer
NEO::AllocationProperties properties = {getRootDeviceIndex(), false, size,
NEO::AllocationType::externalHostPtr,
false, neoDevice->getDeviceBitfield()};
properties.flags.flushL3RequiredForRead = properties.flags.flushL3RequiredForWrite = true;
// L3 must be flushed only if host memory is a transfer destination.
properties.flags.flushL3RequiredForRead = properties.flags.flushL3RequiredForWrite = !hostCopyAllowed;
auto allocation = neoDevice->getMemoryManager()->allocateGraphicsMemoryWithProperties(properties,
buffer);
if (allocation == nullptr && hostCopyAllowed) {

1
level_zero/core/test/unit_tests/sources/device/test_l0_device.cpp
View File

@@ -917,6 +917,7 @@ TEST_F(DeviceHostPointerTest, givenHostPointerNotAcceptedByKernelThenNewAllocati
EXPECT_NE(allocation->getUnderlyingBuffer(), reinterpret_cast<void *>(buffer));
EXPECT_EQ(alignUp(size, MemoryConstants::pageSize), allocation->getUnderlyingBufferSize());
EXPECT_EQ(0, memcmp(buffer, allocation->getUnderlyingBuffer(), size));
EXPECT_FALSE(allocation->isFlushL3Required());
neoDevice->getMemoryManager()->freeGraphicsMemory(allocation);
delete[] buffer;

4
shared/source/gmm_helper/cache_settings_helper.cpp
View File

@@ -126,8 +126,8 @@ GMM_RESOURCE_USAGE_TYPE_ENUM CacheSettingsHelper::getDefaultUsageTypeWithCaching
}
// Set 2-way coherency for allocations which are not aligned to cacheline
GMM_RESOURCE_USAGE_TYPE_ENUM CacheSettingsHelper::getGmmUsageTypeForUserPtr(const void *userPtr, size_t size, const ProductHelper &productHelper) {
if (!isL3Capable(userPtr, size) && productHelper.isMisalignedUserPtr2WayCoherent()) {
GMM_RESOURCE_USAGE_TYPE_ENUM CacheSettingsHelper::getGmmUsageTypeForUserPtr(bool isCacheFlushRequired, const void *userPtr, size_t size, const ProductHelper &productHelper) {
if (isCacheFlushRequired && !isL3Capable(userPtr, size) && productHelper.isMisalignedUserPtr2WayCoherent()) {
return GMM_RESOURCE_USAGE_HW_CONTEXT;
} else {
return GMM_RESOURCE_USAGE_OCL_SYSTEM_MEMORY_BUFFER;

2
shared/source/gmm_helper/cache_settings_helper.h
View File

@@ -20,7 +20,7 @@ struct RootDeviceEnvironment;
struct CacheSettingsHelper {
static GMM_RESOURCE_USAGE_TYPE_ENUM getGmmUsageType(AllocationType allocationType, bool forceUncached, const ProductHelper &productHelper);
static GMM_RESOURCE_USAGE_TYPE_ENUM getGmmUsageTypeForUserPtr(const void *userPtr, size_t size, const ProductHelper &productHelper);
static GMM_RESOURCE_USAGE_TYPE_ENUM getGmmUsageTypeForUserPtr(bool isCacheFlushRequired, const void *userPtr, size_t size, const ProductHelper &productHelper);
static constexpr bool isUncachedType(GMM_RESOURCE_USAGE_TYPE_ENUM gmmResourceUsageType) {
return ((gmmResourceUsageType == GMM_RESOURCE_USAGE_OCL_BUFFER_CSR_UC) ||

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

@@ -633,7 +633,7 @@ GraphicsAllocation *DrmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(con
bo->setAddress(gpuVirtualAddress);
auto usageType = CacheSettingsHelper::getGmmUsageTypeForUserPtr(allocationData.hostPtr, allocationData.size, productHelper);
auto usageType = CacheSettingsHelper::getGmmUsageTypeForUserPtr(allocationData.flags.flushL3, allocationData.hostPtr, allocationData.size, productHelper);
auto patIndex = rootDeviceEnvironment->getGmmClientContext()->cachePolicyGetPATIndex(nullptr, usageType, false, true);
bo->setPatIndex(patIndex);

2
shared/source/os_interface/windows/wddm_memory_manager.cpp
View File

@@ -474,7 +474,7 @@ GraphicsAllocation *WddmMemoryManager::allocateGraphicsMemoryForNonSvmHostPtr(co
}
auto gmm = new Gmm(executionEnvironment.rootDeviceEnvironments[allocationData.rootDeviceIndex]->getGmmHelper(), alignedPtr, alignedSize, 0u,
CacheSettingsHelper::getGmmUsageTypeForUserPtr(allocationData.hostPtr, allocationData.size, productHelper), {}, gmmRequirements);
CacheSettingsHelper::getGmmUsageTypeForUserPtr(allocationData.flags.flushL3, allocationData.hostPtr, allocationData.size, productHelper), {}, gmmRequirements);
wddmAllocation->setDefaultGmm(gmm);

21
shared/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp
View File

@@ -3803,7 +3803,7 @@ TEST_F(DrmMemoryManagerBasic, givenSpecificAddressSpaceWhenInitializingMemoryMan
EXPECT_EQ(maxNBitValue(48 - 1), limit);
}
TEST_F(DrmMemoryManagerBasic, givenUnalignedHostPtrWhenAllocateGraphicsMemoryThenSetCorrectPatIndex) {
TEST_F(DrmMemoryManagerBasic, givenUnalignedHostPtrWithFlushL3RequiredWhenAllocateGraphicsMemoryThenSetCorrectPatIndex) {
AllocationData allocationData;
std::unique_ptr<TestedDrmMemoryManager> memoryManager(new (std::nothrow) TestedDrmMemoryManager(false, false, false, executionEnvironment));
@@ -3812,6 +3812,7 @@ TEST_F(DrmMemoryManagerBasic, givenUnalignedHostPtrWhenAllocateGraphicsMemoryThe
allocationData.size = 13;
allocationData.hostPtr = reinterpret_cast<const void *>(0x5001);
allocationData.rootDeviceIndex = rootDeviceIndex;
allocationData.flags.flushL3 = true;
auto allocation = static_cast<DrmAllocation *>(memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData));
EXPECT_NE(nullptr, allocation);
@@ -3828,6 +3829,23 @@ TEST_F(DrmMemoryManagerBasic, givenUnalignedHostPtrWhenAllocateGraphicsMemoryThe
memoryManager->freeGraphicsMemory(allocation);
}
TEST_F(DrmMemoryManagerBasic, givenUnalignedHostPtrWithFlushL3NotRequiredWhenAllocateGraphicsMemoryThenSetCorrectPatIndex) {
AllocationData allocationData;
std::unique_ptr<TestedDrmMemoryManager> memoryManager(new (std::nothrow) TestedDrmMemoryManager(false, false, false, executionEnvironment));
memoryManager->forceLimitedRangeAllocator(MemoryConstants::max48BitAddress);
allocationData.size = 13;
allocationData.hostPtr = reinterpret_cast<const void *>(0x5001);
allocationData.rootDeviceIndex = rootDeviceIndex;
allocationData.flags.flushL3 = false;
auto allocation = static_cast<DrmAllocation *>(memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData));
EXPECT_NE(nullptr, allocation);
EXPECT_EQ(MockGmmClientContextBase::MockPatIndex::cached, allocation->getBO()->peekPatIndex());
memoryManager->freeGraphicsMemory(allocation);
}
TEST_F(DrmMemoryManagerBasic, givenAlignedHostPtrWhenAllocateGraphicsMemoryThenSetCorrectPatIndex) {
AllocationData allocationData;
std::unique_ptr<TestedDrmMemoryManager> memoryManager(new (std::nothrow) TestedDrmMemoryManager(false, false, false, executionEnvironment));
@@ -3837,6 +3855,7 @@ TEST_F(DrmMemoryManagerBasic, givenAlignedHostPtrWhenAllocateGraphicsMemoryThenS
allocationData.size = MemoryConstants::cacheLineSize;
allocationData.hostPtr = reinterpret_cast<const void *>(MemoryConstants::pageSize);
allocationData.rootDeviceIndex = rootDeviceIndex;
allocationData.flags.flushL3 = true;
auto allocation = static_cast<DrmAllocation *>(memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData));
EXPECT_NE(nullptr, allocation);

20
shared/test/unit_test/os_interface/windows/wddm_memory_manager_tests.cpp
View File

@@ -1050,7 +1050,7 @@ TEST_F(WddmMemoryManagerSimpleTest, whenCreateAllocationFromHandleAndMapCallFail
EXPECT_EQ(1u, memoryManager->freeGraphicsMemoryImplCalled);
}
TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrIsCalledWhenNotAlignedPtrIsPassedThenAlignedGraphicsAllocationIsCreatedWithCorrectGmmResource) {
TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrIsCalledWhenNotAlignedPtrIsPassedAndFlushL3RequiredThenSetCorrectGmmResource) {
memoryManager.reset(new MockWddmMemoryManager(false, false, executionEnvironment));
auto size = 13u;
auto hostPtr = reinterpret_cast<const void *>(0x10001);
@@ -1058,6 +1058,7 @@ TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrI
AllocationData allocationData;
allocationData.size = size;
allocationData.hostPtr = hostPtr;
allocationData.flags.flushL3 = true;
auto allocation = memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData);
EXPECT_NE(nullptr, allocation);
EXPECT_EQ(hostPtr, allocation->getUnderlyingBuffer());
@@ -1074,6 +1075,22 @@ TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrI
memoryManager->freeGraphicsMemory(allocation);
}
TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrIsCalledWhenNotAlignedPtrIsPassedAndFlushL3NotRequiredThenSetCorrectGmmResource) {
memoryManager.reset(new MockWddmMemoryManager(false, false, executionEnvironment));
auto size = 13u;
auto hostPtr = reinterpret_cast<const void *>(0x10001);
AllocationData allocationData;
allocationData.size = size;
allocationData.hostPtr = hostPtr;
allocationData.flags.flushL3 = false;
auto allocation = memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData);
EXPECT_NE(nullptr, allocation);
EXPECT_EQ(GMM_RESOURCE_USAGE_OCL_SYSTEM_MEMORY_BUFFER, allocation->getGmm(0)->resourceParams.Usage);
memoryManager->freeGraphicsMemory(allocation);
}
TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrIsCalledWhenNotAlignedPtrIsPassedAndImportedAllocationIsFalseThenAlignedGraphicsAllocationIsFreed) {
memoryManager.reset(new MockWddmMemoryManager(false, false, executionEnvironment));
auto size = 13u;
@@ -1082,6 +1099,7 @@ TEST_F(WddmMemoryManagerSimpleTest, givenAllocateGraphicsMemoryForNonSvmHostPtrI
AllocationData allocationData;
allocationData.size = size;
allocationData.hostPtr = hostPtr;
allocationData.flags.flushL3 = true;
auto allocation = memoryManager->allocateGraphicsMemoryForNonSvmHostPtr(allocationData);
EXPECT_NE(nullptr, allocation);
EXPECT_EQ(hostPtr, allocation->getUnderlyingBuffer());