intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-21 09:14:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "fix: limit usm device reuse based on used memory"

Browse Source 
This reverts commit 1252b10ba9.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
This commit is contained in:
Compute-Runtime-Validation
2024-12-05 22:46:55 +01:00
committed by Compute-Runtime-Automation
parent ffec97acc5
commit 484210d656
6 changed files with 12 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
shared/source/device/device.h
View File

@@ -181,7 +181,7 @@ class Device : public ReferenceTrackedObject<Device> {
void initializeRayTracing(uint32_t maxBvhLevels);
void allocateRTDispatchGlobals(uint32_t maxBvhLevels);
MOCKABLE_VIRTUAL uint64_t getGlobalMemorySize(uint32_t deviceBitfield) const;
uint64_t getGlobalMemorySize(uint32_t deviceBitfield) const;
const std::vector<SubDevice *> &getSubDevices() const { return subdevices; }
bool getUuid(std::array<uint8_t, ProductHelper::uuidSize> &uuid);
void generateUuid(std::array<uint8_t, ProductHelper::uuidSize> &uuid);

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

@@ -62,13 +62,7 @@ bool SVMAllocsManager::SvmAllocationCache::insert(size_t size, void *ptr, SvmAll
if (auto device = svmData->device) {
auto lock = device->obtainAllocationsReuseLock();
const auto usedSize = device->getAllocationsSavedForReuseSize();
uint64_t availableMemory = device->getGlobalMemorySize(static_cast<uint32_t>(device->getDeviceBitfield().to_ulong()));
availableMemory -= memoryManager->getUsedLocalMemorySize(device->getRootDeviceIndex());
if (!localMemorySupported) {
availableMemory -= memoryManager->getUsedSystemMemorySize();
}
const auto availableMemoryForReuse = static_cast<uint64_t>(availableMemory * fractionOfAvailableMemoryForRecycling);
if (size + usedSize > availableMemoryForReuse) {
if (size + usedSize > this->maxSize) {
return false;
}
device->recordAllocationSaveForReuse(size);
@@ -762,8 +756,6 @@ void SVMAllocsManager::initUsmDeviceAllocationsCache(Device &device) {
if (this->usmDeviceAllocationsCache.maxSize > 0u) {
this->usmDeviceAllocationsCache.allocations.reserve(128u);
}
this->usmDeviceAllocationsCache.fractionOfAvailableMemoryForRecycling = fractionOfTotalMemoryForRecycling;
this->usmDeviceAllocationsCache.localMemorySupported = memoryManager->isLocalMemorySupported(device.getRootDeviceIndex());
}
void SVMAllocsManager::initUsmHostAllocationsCache() {

2
shared/source/memory_manager/unified_memory_manager.h
View File

@@ -172,8 +172,6 @@ class SVMAllocsManager {
std::vector<SvmCacheAllocationInfo> allocations;
std::mutex mtx;
bool localMemorySupported = true;
double fractionOfAvailableMemoryForRecycling = 0.0;
size_t maxSize = 0;
};

10
shared/test/common/mocks/mock_device.h
View File

@@ -63,6 +63,7 @@ class MockDevice : public RootDevice {
using Device::deviceInfo;
using Device::executionEnvironment;
using Device::generateUuidFromPciBusInfo;
using Device::getGlobalMemorySize;
using Device::initializeCaps;
using Device::preemptionMode;
using Device::regularEngineGroups;
@@ -98,13 +99,6 @@ class MockDevice : public RootDevice {
performanceCounters = std::move(perfCounters);
}
uint64_t getGlobalMemorySize(uint32_t deviceBitfield) const override {
if (callBaseGetGlobalMemorySize) {
return Device::getGlobalMemorySize(deviceBitfield);
}
return getGlobalMemorySizeReturn;
}
size_t getMaxParameterSizeFromIGC() const override {
if (callBaseGetMaxParameterSizeFromIGC) {
return Device::getMaxParameterSizeFromIGC();
@@ -179,8 +173,6 @@ class MockDevice : public RootDevice {
bool callBaseGetMaxParameterSizeFromIGC = false;
bool callBaseVerifyAdapterLuid = true;
bool callBaseGetGlobalMemorySize = true;
uint64_t getGlobalMemorySizeReturn = 0u;
bool verifyAdapterLuidReturnValue = true;
size_t maxParameterSizeFromIGC = 0u;
bool rtDispatchGlobalsForceAllocation = true;

1
shared/test/common/mocks/mock_memory_manager.h
View File

@@ -64,7 +64,6 @@ class MockMemoryManager : public MemoryManagerCreate<OsAgnosticMemoryManager> {
using MemoryManager::localMemorySupported;
using MemoryManager::reservedMemory;
using MemoryManager::secondaryEngines;
using MemoryManager::sysMemAllocsSize;
static constexpr osHandle invalidSharedHandle = -1;
static const unsigned int moduleId;

119
shared/test/unit_test/memory_manager/unified_memory_manager_cache_tests.cpp
View File

@@ -231,34 +231,25 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledWhenInitializedT
auto expectedMaxSize = static_cast<size_t>(device->getGlobalMemorySize(static_cast<uint32_t>(mockDeviceBitfield.to_ulong())) * 0.02);
EXPECT_EQ(expectedMaxSize, svmManager->usmDeviceAllocationsCache.maxSize);
EXPECT_EQ(0.02, svmManager->usmDeviceAllocationsCache.fractionOfAvailableMemoryForRecycling);
EXPECT_EQ(device->getMemoryManager()->isLocalMemorySupported(device->getRootDeviceIndex()), svmManager->usmDeviceAllocationsCache.localMemorySupported);
}
TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndLocalMemorySupportedWhenFreeingDeviceAllocationThenItIsPutIntoCacheOnlyIfMaxSizeWillNotBeExceeded) {
TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledWhenFreeingDeviceAllocationThenItIsPutIntoCacheOnlyIfMaxSizeWillNotBeExceeded) {
std::unique_ptr<UltDeviceFactory> deviceFactory(new UltDeviceFactory(1, 1));
RootDeviceIndicesContainer rootDeviceIndices = {mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, mockDeviceBitfield}};
DebugManagerStateRestore restore;
debugManager.flags.ExperimentalEnableDeviceAllocationCache.set(2);
debugManager.flags.ExperimentalEnableDeviceAllocationCache.set(1);
auto device = deviceFactory->rootDevices[0];
auto mockMemoryManager = reinterpret_cast<MockMemoryManager *>(device->getMemoryManager());
mockMemoryManager->localMemorySupported[device->getRootDeviceIndex()] = true;
constexpr auto allocationSize = MemoryConstants::pageSize64k;
device->callBaseGetGlobalMemorySize = false;
device->getGlobalMemorySizeReturn = allocationSize * 100;
auto svmManager = std::make_unique<MockSVMAllocsManager>(device->getMemoryManager(), false);
svmManager->initUsmAllocationsCaches(*device);
ASSERT_TRUE(svmManager->usmDeviceAllocationsCacheEnabled);
constexpr auto allocationSize = MemoryConstants::pageSize64k;
svmManager->usmDeviceAllocationsCache.maxSize = allocationSize;
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::deviceUnifiedMemory, 1, rootDeviceIndices, deviceBitfields);
unifiedMemoryProperties.device = device;
{
mockMemoryManager->sysMemAllocsSize = allocationSize * 50;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = 0u;
auto allocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation, nullptr);
auto allocation2 = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
@@ -270,8 +261,6 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndLocalMemorySu
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = allocationSize * 50;
svmManager->freeSVMAlloc(allocation2);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
@@ -288,8 +277,6 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndLocalMemorySu
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
}
{
mockMemoryManager->sysMemAllocsSize = allocationSize * 50;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = 0u;
auto allocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation, nullptr);
auto allocation2 = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
@@ -301,94 +288,6 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndLocalMemorySu
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = allocationSize * 50;
svmManager->freeSVMAllocDefer(allocation2);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
auto recycledAllocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
EXPECT_EQ(recycledAllocation, allocation);
EXPECT_EQ(svmManager->usmDeviceAllocationsCache.allocations.size(), 0u);
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
svmManager->freeSVMAllocDefer(recycledAllocation);
svmManager->trimUSMDeviceAllocCache();
EXPECT_EQ(svmManager->usmDeviceAllocationsCache.allocations.size(), 0u);
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
}
}
TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndLocalMemoryNotSupportedWhenFreeingDeviceAllocationThenItIsPutIntoCacheOnlyIfMaxSizeWillNotBeExceeded) {
std::unique_ptr<UltDeviceFactory> deviceFactory(new UltDeviceFactory(1, 1));
RootDeviceIndicesContainer rootDeviceIndices = {mockRootDeviceIndex};
std::map<uint32_t, DeviceBitfield> deviceBitfields{{mockRootDeviceIndex, mockDeviceBitfield}};
DebugManagerStateRestore restore;
debugManager.flags.ExperimentalEnableDeviceAllocationCache.set(2);
auto device = deviceFactory->rootDevices[0];
auto mockMemoryManager = reinterpret_cast<MockMemoryManager *>(device->getMemoryManager());
mockMemoryManager->localMemorySupported[device->getRootDeviceIndex()] = false;
constexpr auto allocationSize = MemoryConstants::pageSize64k;
device->callBaseGetGlobalMemorySize = false;
device->getGlobalMemorySizeReturn = allocationSize * 100;
auto svmManager = std::make_unique<MockSVMAllocsManager>(device->getMemoryManager(), false);
svmManager->initUsmAllocationsCaches(*device);
ASSERT_TRUE(svmManager->usmDeviceAllocationsCacheEnabled);
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::deviceUnifiedMemory, 1, rootDeviceIndices, deviceBitfields);
unifiedMemoryProperties.device = device;
{
mockMemoryManager->sysMemAllocsSize = 0u;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = 0u;
auto allocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation, nullptr);
auto allocation2 = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation2, nullptr);
EXPECT_EQ(0u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
svmManager->freeSVMAlloc(allocation);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
mockMemoryManager->sysMemAllocsSize = allocationSize * 25;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = allocationSize * 25;
svmManager->freeSVMAlloc(allocation2);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
auto recycledAllocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
EXPECT_EQ(recycledAllocation, allocation);
EXPECT_EQ(svmManager->usmDeviceAllocationsCache.allocations.size(), 0u);
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
svmManager->freeSVMAlloc(recycledAllocation);
svmManager->trimUSMDeviceAllocCache();
EXPECT_EQ(svmManager->usmDeviceAllocationsCache.allocations.size(), 0u);
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
}
{
mockMemoryManager->sysMemAllocsSize = 0u;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = 0u;
auto allocation = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation, nullptr);
auto allocation2 = svmManager->createUnifiedMemoryAllocation(allocationSize, unifiedMemoryProperties);
ASSERT_NE(allocation2, nullptr);
EXPECT_EQ(0u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(0u, device->getAllocationsSavedForReuseSize());
svmManager->freeSVMAllocDefer(allocation);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
mockMemoryManager->sysMemAllocsSize = allocationSize * 25;
mockMemoryManager->localMemAllocsSize[device->getRootDeviceIndex()] = allocationSize * 25;
svmManager->freeSVMAllocDefer(allocation2);
EXPECT_EQ(1u, svmManager->usmDeviceAllocationsCache.allocations.size());
EXPECT_EQ(allocationSize, device->getAllocationsSavedForReuseSize());
@@ -413,10 +312,6 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndMultipleSVMMa
DebugManagerStateRestore restore;
debugManager.flags.ExperimentalEnableDeviceAllocationCache.set(1);
auto device = deviceFactory->rootDevices[0];
constexpr auto allocationSize = MemoryConstants::pageSize64k;
device->callBaseGetGlobalMemorySize = false;
device->getGlobalMemorySizeReturn = allocationSize * 100;
auto svmManager = std::make_unique<MockSVMAllocsManager>(device->getMemoryManager(), false);
auto secondSvmManager = std::make_unique<MockSVMAllocsManager>(device->getMemoryManager(), false);
svmManager->initUsmAllocationsCaches(*device);
@@ -424,6 +319,10 @@ TEST_F(SvmDeviceAllocationCacheTest, givenAllocationCacheEnabledAndMultipleSVMMa
ASSERT_TRUE(svmManager->usmDeviceAllocationsCacheEnabled);
ASSERT_TRUE(secondSvmManager->usmDeviceAllocationsCacheEnabled);
constexpr auto allocationSize = MemoryConstants::pageSize64k;
svmManager->usmDeviceAllocationsCache.maxSize = allocationSize;
secondSvmManager->usmDeviceAllocationsCache.maxSize = allocationSize;
SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::deviceUnifiedMemory, 1, rootDeviceIndices, deviceBitfields);
unifiedMemoryProperties.device = device;
{