Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-15 13:01:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

[1/n] Improve indirect allocations handling.

Browse Source 
Add new functions that would treat all indirect allocations as single pack.

Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek
2022-01-24 15:13:30 +00:00
committed by Compute-Runtime-Automation
parent 22bbce42dd
commit 52d636394c
3 changed files with 200 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

142
opencl/test/unit_test/memory_manager/unified_memory_manager_tests.cpp
View File

@ -723,6 +723,148 @@ TEST(UnifiedMemoryTest, givenDeviceBitfieldWithSingleBitsSetWhenMultiOsContextFl
svmManager->freeSVMAlloc(ptr);
}
TEST(UnifiedMemoryTest, givenInternalAllocationWhenItIsMadeResidentThenNewTrackingEntryIsCreated) {
MockCommandQueue cmdQ;
MockDevice device;
MockExecutionEnvironment executionEnvironment;
auto memoryManager = std::make_unique<MemoryManagerPropertiesCheck>(false, true, executionEnvironment);
auto unifiedMemoryManager = std::make_unique<MockSVMAllocsManager>(memoryManager.get(), false);
memoryManager->pageFaultManager = std::make_unique<MockPageFaultManager>();
std::set<uint32_t> rootDeviceIndices{mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, DeviceBitfield(0x1)}};
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY, rootDeviceIndices, deviceBitfields);
auto ptr = unifiedMemoryManager->createSharedUnifiedMemoryAllocation(4096u, unifiedMemoryProperties, &cmdQ);
ASSERT_NE(nullptr, ptr);
auto graphicsAllocation = unifiedMemoryManager->getSVMAlloc(ptr);
auto &commandStreamReceiver = device.getGpgpuCommandStreamReceiver();
EXPECT_FALSE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_EQ(0u, unifiedMemoryManager->indirectAllocationsResidency.size());
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 1u);
EXPECT_TRUE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_EQ(GraphicsAllocation::objectAlwaysResident, graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->getResidencyTaskCount(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_EQ(1u, unifiedMemoryManager->indirectAllocationsResidency.size());
auto internalEntry = unifiedMemoryManager->indirectAllocationsResidency.find(&commandStreamReceiver)->second;
EXPECT_EQ(1u, internalEntry.latestSentTaskCount);
EXPECT_EQ(1u, internalEntry.latestResidentObjectId);
unifiedMemoryManager->freeSVMAlloc(ptr);
}
TEST(UnifiedMemoryTest, givenInternalAllocationWhenItIsMadeResidentThenSubsequentCallsDoNotCallResidency) {
MockCommandQueue cmdQ;
MockDevice device;
MockExecutionEnvironment executionEnvironment;
auto memoryManager = std::make_unique<MemoryManagerPropertiesCheck>(false, true, executionEnvironment);
auto unifiedMemoryManager = std::make_unique<MockSVMAllocsManager>(memoryManager.get(), false);
memoryManager->pageFaultManager = std::make_unique<MockPageFaultManager>();
std::set<uint32_t> rootDeviceIndices{mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, DeviceBitfield(0x1)}};
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY, rootDeviceIndices, deviceBitfields);
auto ptr = unifiedMemoryManager->createSharedUnifiedMemoryAllocation(4096u, unifiedMemoryProperties, &cmdQ);
ASSERT_NE(nullptr, ptr);
auto graphicsAllocation = unifiedMemoryManager->getSVMAlloc(ptr);
auto &commandStreamReceiver = device.getGpgpuCommandStreamReceiver();
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 1u);
EXPECT_TRUE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
//now call with task count 2 , allocations shouldn't change
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 2u);
auto internalEntry = unifiedMemoryManager->indirectAllocationsResidency.find(&commandStreamReceiver)->second;
EXPECT_EQ(2u, internalEntry.latestSentTaskCount);
EXPECT_TRUE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
//force Graphics Allocation to be non resident
graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->updateResidencyTaskCount(GraphicsAllocation::objectNotResident, commandStreamReceiver.getOsContext().getContextId());
EXPECT_FALSE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
//now call with task count 3 , allocations shouldn't change
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 2u);
EXPECT_FALSE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
unifiedMemoryManager->freeSVMAlloc(ptr);
}
TEST(UnifiedMemoryTest, givenInternalAllocationWhenNewAllocationIsCreatedThenItIsMadeResident) {
MockCommandQueue cmdQ;
MockDevice device;
MockExecutionEnvironment executionEnvironment;
auto memoryManager = std::make_unique<MemoryManagerPropertiesCheck>(false, true, executionEnvironment);
auto unifiedMemoryManager = std::make_unique<MockSVMAllocsManager>(memoryManager.get(), false);
memoryManager->pageFaultManager = std::make_unique<MockPageFaultManager>();
std::set<uint32_t> rootDeviceIndices{mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, DeviceBitfield(0x1)}};
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY, rootDeviceIndices, deviceBitfields);
auto ptr = unifiedMemoryManager->createSharedUnifiedMemoryAllocation(4096u, unifiedMemoryProperties, &cmdQ);
ASSERT_NE(nullptr, ptr);
auto graphicsAllocation = unifiedMemoryManager->getSVMAlloc(ptr);
auto &commandStreamReceiver = device.getGpgpuCommandStreamReceiver();
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 1u);
EXPECT_TRUE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
//force to non resident
graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->updateResidencyTaskCount(GraphicsAllocation::objectNotResident, commandStreamReceiver.getOsContext().getContextId());
auto ptr2 = unifiedMemoryManager->createSharedUnifiedMemoryAllocation(4096u, unifiedMemoryProperties, &cmdQ);
auto graphicsAllocation2 = unifiedMemoryManager->getSVMAlloc(ptr);
EXPECT_FALSE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_FALSE(graphicsAllocation2->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
//now call with task count 2, both allocations needs to be made resident
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 2u);
EXPECT_TRUE(graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_TRUE(graphicsAllocation2->gpuAllocations.getDefaultGraphicsAllocation()->isResident(commandStreamReceiver.getOsContext().getContextId()));
unifiedMemoryManager->freeSVMAlloc(ptr);
unifiedMemoryManager->freeSVMAlloc(ptr2);
}
TEST(UnifiedMemoryTest, givenInternalAllocationsWhenTheyArePreparedForFreeingThenProperTaskCountIsAssigned) {
MockCommandQueue cmdQ;
MockDevice device;
MockExecutionEnvironment executionEnvironment;
auto memoryManager = std::make_unique<MemoryManagerPropertiesCheck>(false, true, executionEnvironment);
auto unifiedMemoryManager = std::make_unique<MockSVMAllocsManager>(memoryManager.get(), false);
memoryManager->pageFaultManager = std::make_unique<MockPageFaultManager>();
std::set<uint32_t> rootDeviceIndices{mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, DeviceBitfield(0x1)}};
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::SHARED_UNIFIED_MEMORY, rootDeviceIndices, deviceBitfields);
auto ptr = unifiedMemoryManager->createSharedUnifiedMemoryAllocation(4096u, unifiedMemoryProperties, &cmdQ);
ASSERT_NE(nullptr, ptr);
auto graphicsAllocation = unifiedMemoryManager->getSVMAlloc(ptr);
auto &commandStreamReceiver = device.getGpgpuCommandStreamReceiver();
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 1u);
unifiedMemoryManager->makeIndirectAllocationsResident(commandStreamReceiver, 124u);
EXPECT_EQ(1u, graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->getTaskCount(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_EQ(GraphicsAllocation::objectAlwaysResident, graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->getResidencyTaskCount(commandStreamReceiver.getOsContext().getContextId()));
auto allocationData = unifiedMemoryManager->getSVMAlloc(ptr);
unifiedMemoryManager->prepareIndirectAllocationForDestruction(allocationData);
EXPECT_EQ(124u, graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->getTaskCount(commandStreamReceiver.getOsContext().getContextId()));
EXPECT_EQ(124u, graphicsAllocation->gpuAllocations.getDefaultGraphicsAllocation()->getResidencyTaskCount(commandStreamReceiver.getOsContext().getContextId()));
unifiedMemoryManager->freeSVMAlloc(ptr);
}
TEST_F(UnifiedMemoryManagerPropertiesTest, givenDeviceBitfieldWithSingleBitSetWhenDeviceUnifiedMemoryAllocationIsCreatedThenProperPropertiesArePassedToMemoryManager) {
MockCommandQueue cmdQ;
std::set<uint32_t> rootDeviceIndices{mockRootDeviceIndex};