compute-runtime/shared/source/memory_manager/unified_memory_manager.cpp

825 lines
38 KiB
C++

/*
* Copyright (C) 2019-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/memory_manager/unified_memory_manager.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/device/sub_device.h"
#include "shared/source/execution_environment/execution_environment.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/api_specific_config.h"
#include "shared/source/helpers/basic_math.h"
#include "shared/source/helpers/memory_properties_helpers.h"
#include "shared/source/memory_manager/allocation_properties.h"
#include "shared/source/memory_manager/compression_selector.h"
#include "shared/source/memory_manager/memory_manager.h"
#include "shared/source/os_interface/os_context.h"
#include "shared/source/os_interface/product_helper.h"
#include "shared/source/page_fault_manager/cpu_page_fault_manager.h"
namespace NEO {
uint32_t SVMAllocsManager::UnifiedMemoryProperties::getRootDeviceIndex() const {
if (device) {
return device->getRootDeviceIndex();
}
UNRECOVERABLE_IF(rootDeviceIndices.begin() == nullptr);
return *rootDeviceIndices.begin();
}
void SVMAllocsManager::MapBasedAllocationTracker::insert(const SvmAllocationData &allocationsPair) {
allocations.insert(std::make_pair(reinterpret_cast<void *>(allocationsPair.gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress()), allocationsPair));
}
void SVMAllocsManager::MapBasedAllocationTracker::remove(const SvmAllocationData &allocationsPair) {
SvmAllocationContainer::iterator iter;
iter = allocations.find(reinterpret_cast<void *>(allocationsPair.gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress()));
allocations.erase(iter);
}
void SVMAllocsManager::SvmAllocationCache::insert(size_t size, void *ptr) {
std::lock_guard<std::mutex> lock(this->mtx);
allocations.emplace(std::lower_bound(allocations.begin(), allocations.end(), size), size, ptr);
}
void *SVMAllocsManager::SvmAllocationCache::get(size_t size, const UnifiedMemoryProperties &unifiedMemoryProperties, SVMAllocsManager *svmAllocsManager) {
std::lock_guard<std::mutex> lock(this->mtx);
for (auto allocationIter = std::lower_bound(allocations.begin(), allocations.end(), size);
allocationIter != allocations.end();
++allocationIter) {
void *allocationPtr = allocationIter->allocation;
SvmAllocationData *svmAllocData = svmAllocsManager->getSVMAlloc(allocationPtr);
UNRECOVERABLE_IF(!svmAllocData);
if (svmAllocData->device == unifiedMemoryProperties.device &&
svmAllocData->allocationFlagsProperty.allFlags == unifiedMemoryProperties.allocationFlags.allFlags &&
svmAllocData->allocationFlagsProperty.allAllocFlags == unifiedMemoryProperties.allocationFlags.allAllocFlags) {
allocations.erase(allocationIter);
return allocationPtr;
}
}
return nullptr;
}
void SVMAllocsManager::SvmAllocationCache::trim(SVMAllocsManager *svmAllocsManager) {
std::lock_guard<std::mutex> lock(this->mtx);
for (auto &cachedAllocationInfo : this->allocations) {
SvmAllocationData *svmData = svmAllocsManager->getSVMAlloc(cachedAllocationInfo.allocation);
DEBUG_BREAK_IF(nullptr == svmData);
svmAllocsManager->freeSVMAllocImpl(cachedAllocationInfo.allocation, FreePolicyType::POLICY_NONE, svmData);
}
this->allocations.clear();
}
SvmAllocationData *SVMAllocsManager::MapBasedAllocationTracker::get(const void *ptr) {
if (allocations.size() == 0) {
return nullptr;
}
if (!ptr) {
return nullptr;
}
SvmAllocationContainer::iterator iter;
const SvmAllocationContainer::iterator end = allocations.end();
SvmAllocationData *svmAllocData;
// try faster find lookup if pointer is aligned to page
if (isAligned<MemoryConstants::pageSize>(ptr)) {
iter = allocations.find(ptr);
if (iter != end) {
return &iter->second;
}
}
// do additional check with lower bound as we may deal with pointer offset
iter = allocations.lower_bound(ptr);
if (((iter != end) && (iter->first != ptr)) ||
(iter == end)) {
if (iter == allocations.begin()) {
iter = end;
} else {
iter--;
}
}
if (iter != end) {
svmAllocData = &iter->second;
char *charPtr = reinterpret_cast<char *>(svmAllocData->gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress());
if (ptr < (charPtr + svmAllocData->size)) {
return svmAllocData;
}
}
return nullptr;
}
void SVMAllocsManager::MapOperationsTracker::insert(SvmMapOperation mapOperation) {
operations.insert(std::make_pair(mapOperation.regionSvmPtr, mapOperation));
}
void SVMAllocsManager::MapOperationsTracker::remove(const void *regionPtr) {
SvmMapOperationsContainer::iterator iter;
iter = operations.find(regionPtr);
operations.erase(iter);
}
SvmMapOperation *SVMAllocsManager::MapOperationsTracker::get(const void *regionPtr) {
SvmMapOperationsContainer::iterator iter;
iter = operations.find(regionPtr);
if (iter == operations.end()) {
return nullptr;
}
return &iter->second;
}
void SVMAllocsManager::addInternalAllocationsToResidencyContainer(uint32_t rootDeviceIndex,
ResidencyContainer &residencyContainer,
uint32_t requestedTypesMask) {
std::shared_lock<std::shared_mutex> lock(mtx);
for (auto &allocation : this->svmAllocs.allocations) {
if (rootDeviceIndex >= allocation.second->gpuAllocations.getGraphicsAllocations().size()) {
continue;
}
if (!(static_cast<uint32_t>(allocation.second->memoryType) & requestedTypesMask) ||
(nullptr == allocation.second->gpuAllocations.getGraphicsAllocation(rootDeviceIndex))) {
continue;
}
auto alloc = allocation.second->gpuAllocations.getGraphicsAllocation(rootDeviceIndex);
residencyContainer.push_back(alloc);
}
}
void SVMAllocsManager::makeInternalAllocationsResident(CommandStreamReceiver &commandStreamReceiver, uint32_t requestedTypesMask) {
std::shared_lock<std::shared_mutex> lock(mtx);
for (auto &allocation : this->svmAllocs.allocations) {
if (static_cast<uint32_t>(allocation.second->memoryType) & requestedTypesMask) {
auto gpuAllocation = allocation.second->gpuAllocations.getGraphicsAllocation(commandStreamReceiver.getRootDeviceIndex());
if (gpuAllocation == nullptr) {
continue;
}
commandStreamReceiver.makeResident(*gpuAllocation);
}
}
}
SVMAllocsManager::SVMAllocsManager(MemoryManager *memoryManager, bool multiOsContextSupport)
: memoryManager(memoryManager), multiOsContextSupport(multiOsContextSupport) {
this->usmDeviceAllocationsCacheEnabled = NEO::ApiSpecificConfig::isDeviceAllocationCacheEnabled();
if (debugManager.flags.ExperimentalEnableDeviceAllocationCache.get() != -1) {
this->usmDeviceAllocationsCacheEnabled = !!debugManager.flags.ExperimentalEnableDeviceAllocationCache.get();
}
if (this->usmDeviceAllocationsCacheEnabled) {
this->initUsmDeviceAllocationsCache();
}
}
SVMAllocsManager::~SVMAllocsManager() = default;
void *SVMAllocsManager::createSVMAlloc(size_t size, const SvmAllocationProperties svmProperties,
const RootDeviceIndicesContainer &rootDeviceIndices,
const std::map<uint32_t, DeviceBitfield> &subdeviceBitfields) {
if (size == 0)
return nullptr;
if (rootDeviceIndices.size() > 1) {
return createZeroCopySvmAllocation(size, svmProperties, rootDeviceIndices, subdeviceBitfields);
}
if (!memoryManager->isLocalMemorySupported(*rootDeviceIndices.begin())) {
return createZeroCopySvmAllocation(size, svmProperties, rootDeviceIndices, subdeviceBitfields);
} else {
UnifiedMemoryProperties unifiedMemoryProperties(InternalMemoryType::notSpecified, 1, rootDeviceIndices, subdeviceBitfields);
return createUnifiedAllocationWithDeviceStorage(size, svmProperties, unifiedMemoryProperties);
}
}
void *SVMAllocsManager::createHostUnifiedMemoryAllocation(size_t size,
const UnifiedMemoryProperties &memoryProperties) {
size_t pageSizeForAlignment = alignUpNonZero<size_t>(memoryProperties.alignment, MemoryConstants::pageSize);
size_t alignedSize = alignUp<size_t>(size, MemoryConstants::pageSize);
bool compressionEnabled = false;
AllocationType allocationType = getGraphicsAllocationTypeAndCompressionPreference(memoryProperties, compressionEnabled);
RootDeviceIndicesContainer rootDeviceIndicesVector(memoryProperties.rootDeviceIndices);
uint32_t rootDeviceIndex = rootDeviceIndicesVector.at(0);
auto &deviceBitfield = memoryProperties.subdeviceBitfields.at(rootDeviceIndex);
AllocationProperties unifiedMemoryProperties{rootDeviceIndex,
true,
alignedSize,
allocationType,
false,
(deviceBitfield.count() > 1) && multiOsContextSupport,
deviceBitfield};
unifiedMemoryProperties.alignment = pageSizeForAlignment;
unifiedMemoryProperties.flags.preferCompressed = compressionEnabled;
unifiedMemoryProperties.flags.shareable = memoryProperties.allocationFlags.flags.shareable;
unifiedMemoryProperties.flags.isUSMHostAllocation = true;
unifiedMemoryProperties.flags.isUSMDeviceAllocation = false;
unifiedMemoryProperties.cacheRegion = MemoryPropertiesHelper::getCacheRegion(memoryProperties.allocationFlags);
auto maxRootDeviceIndex = *std::max_element(rootDeviceIndicesVector.begin(), rootDeviceIndicesVector.end(), std::less<uint32_t const>());
SvmAllocationData allocData(maxRootDeviceIndex);
void *externalHostPointer = reinterpret_cast<void *>(memoryProperties.allocationFlags.hostptr);
void *usmPtr = memoryManager->createMultiGraphicsAllocationInSystemMemoryPool(rootDeviceIndicesVector, unifiedMemoryProperties, allocData.gpuAllocations, externalHostPointer);
if (!usmPtr) {
return nullptr;
}
allocData.cpuAllocation = nullptr;
allocData.size = size;
allocData.memoryType = memoryProperties.memoryType;
allocData.allocationFlagsProperty = memoryProperties.allocationFlags;
allocData.device = nullptr;
allocData.pageSizeForAlignment = pageSizeForAlignment;
allocData.setAllocId(this->allocationsCounter++);
std::unique_lock<std::shared_mutex> lock(mtx);
this->svmAllocs.insert(usmPtr, allocData);
return usmPtr;
}
void *SVMAllocsManager::createUnifiedMemoryAllocation(size_t size,
const UnifiedMemoryProperties &memoryProperties) {
auto rootDeviceIndex = memoryProperties.getRootDeviceIndex();
DeviceBitfield deviceBitfield = memoryProperties.subdeviceBitfields.at(rootDeviceIndex);
size_t pageSizeForAlignment = alignUpNonZero<size_t>(memoryProperties.alignment, MemoryConstants::pageSize64k);
size_t alignedSize = alignUp<size_t>(size, MemoryConstants::pageSize64k);
auto externalPtr = reinterpret_cast<void *>(memoryProperties.allocationFlags.hostptr);
bool useExternalHostPtrForCpu = externalPtr != nullptr;
bool compressionEnabled = false;
AllocationType allocationType = getGraphicsAllocationTypeAndCompressionPreference(memoryProperties, compressionEnabled);
bool multiStorageAllocation = (deviceBitfield.count() > 1) && multiOsContextSupport;
if ((deviceBitfield.count() > 1) && !multiOsContextSupport) {
for (uint32_t i = 0;; i++) {
if (deviceBitfield.test(i)) {
deviceBitfield.reset();
deviceBitfield.set(i);
break;
}
}
}
AllocationProperties unifiedMemoryProperties{rootDeviceIndex,
!useExternalHostPtrForCpu, // allocateMemory
alignedSize,
allocationType,
false,
multiStorageAllocation,
deviceBitfield};
unifiedMemoryProperties.alignment = pageSizeForAlignment;
unifiedMemoryProperties.flags.isUSMDeviceAllocation = false;
unifiedMemoryProperties.flags.shareable = memoryProperties.allocationFlags.flags.shareable;
unifiedMemoryProperties.cacheRegion = MemoryPropertiesHelper::getCacheRegion(memoryProperties.allocationFlags);
unifiedMemoryProperties.flags.uncacheable = memoryProperties.allocationFlags.flags.locallyUncachedResource;
unifiedMemoryProperties.flags.preferCompressed = compressionEnabled || memoryProperties.allocationFlags.flags.compressedHint;
unifiedMemoryProperties.flags.resource48Bit = memoryProperties.allocationFlags.flags.resource48Bit;
if (memoryProperties.memoryType == InternalMemoryType::deviceUnifiedMemory) {
unifiedMemoryProperties.flags.isUSMDeviceAllocation = true;
if (this->usmDeviceAllocationsCacheEnabled) {
void *allocationFromCache = this->usmDeviceAllocationsCache.get(size, memoryProperties, this);
if (allocationFromCache) {
return allocationFromCache;
}
}
} else if (memoryProperties.memoryType == InternalMemoryType::hostUnifiedMemory) {
unifiedMemoryProperties.flags.isUSMHostAllocation = true;
} else {
unifiedMemoryProperties.flags.isUSMHostAllocation = useExternalHostPtrForCpu;
}
GraphicsAllocation *unifiedMemoryAllocation = memoryManager->allocateGraphicsMemoryWithProperties(unifiedMemoryProperties, externalPtr);
if (!unifiedMemoryAllocation) {
if (memoryProperties.memoryType == InternalMemoryType::deviceUnifiedMemory &&
this->usmDeviceAllocationsCacheEnabled) {
this->trimUSMDeviceAllocCache();
unifiedMemoryAllocation = memoryManager->allocateGraphicsMemoryWithProperties(unifiedMemoryProperties, externalPtr);
}
if (!unifiedMemoryAllocation) {
return nullptr;
}
}
setUnifiedAllocationProperties(unifiedMemoryAllocation, {});
SvmAllocationData allocData(rootDeviceIndex);
allocData.gpuAllocations.addAllocation(unifiedMemoryAllocation);
allocData.cpuAllocation = nullptr;
allocData.size = size;
allocData.pageSizeForAlignment = pageSizeForAlignment;
allocData.memoryType = memoryProperties.memoryType;
allocData.allocationFlagsProperty = memoryProperties.allocationFlags;
allocData.device = memoryProperties.device;
allocData.setAllocId(this->allocationsCounter++);
std::unique_lock<std::shared_mutex> lock(mtx);
auto retPtr = reinterpret_cast<void *>(unifiedMemoryAllocation->getGpuAddress());
this->svmAllocs.insert(retPtr, allocData);
UNRECOVERABLE_IF(useExternalHostPtrForCpu && (externalPtr != retPtr));
return retPtr;
}
void *SVMAllocsManager::createSharedUnifiedMemoryAllocation(size_t size,
const UnifiedMemoryProperties &memoryProperties,
void *cmdQ) {
if (memoryProperties.rootDeviceIndices.size() > 1 && memoryProperties.device == nullptr) {
return createHostUnifiedMemoryAllocation(size, memoryProperties);
}
auto rootDeviceIndex = memoryProperties.getRootDeviceIndex();
auto supportDualStorageSharedMemory = memoryManager->isLocalMemorySupported(rootDeviceIndex);
if (debugManager.flags.AllocateSharedAllocationsWithCpuAndGpuStorage.get() != -1) {
supportDualStorageSharedMemory = !!debugManager.flags.AllocateSharedAllocationsWithCpuAndGpuStorage.get();
}
if (supportDualStorageSharedMemory) {
bool useKmdMigration = memoryManager->isKmdMigrationAvailable(rootDeviceIndex);
void *unifiedMemoryPointer = nullptr;
if (useKmdMigration) {
unifiedMemoryPointer = createUnifiedKmdMigratedAllocation(size, {}, memoryProperties);
if (!unifiedMemoryPointer) {
return nullptr;
}
} else {
unifiedMemoryPointer = createUnifiedAllocationWithDeviceStorage(size, {}, memoryProperties);
if (!unifiedMemoryPointer) {
return nullptr;
}
UNRECOVERABLE_IF(cmdQ == nullptr);
auto pageFaultManager = this->memoryManager->getPageFaultManager();
pageFaultManager->insertAllocation(unifiedMemoryPointer, size, this, cmdQ, memoryProperties.allocationFlags);
}
auto unifiedMemoryAllocation = this->getSVMAlloc(unifiedMemoryPointer);
unifiedMemoryAllocation->memoryType = memoryProperties.memoryType;
unifiedMemoryAllocation->allocationFlagsProperty = memoryProperties.allocationFlags;
return unifiedMemoryPointer;
}
return createUnifiedMemoryAllocation(size, memoryProperties);
}
void *SVMAllocsManager::createUnifiedKmdMigratedAllocation(size_t size, const SvmAllocationProperties &svmProperties, const UnifiedMemoryProperties &unifiedMemoryProperties) {
auto rootDeviceIndex = unifiedMemoryProperties.getRootDeviceIndex();
auto &deviceBitfield = unifiedMemoryProperties.subdeviceBitfields.at(rootDeviceIndex);
size_t pageSizeForAlignment = std::max(alignUpNonZero<size_t>(unifiedMemoryProperties.alignment, MemoryConstants::pageSize2M), MemoryConstants::pageSize2M);
size_t alignedSize = alignUp<size_t>(size, MemoryConstants::pageSize2M);
AllocationProperties gpuProperties{rootDeviceIndex,
true,
alignedSize,
AllocationType::unifiedSharedMemory,
false,
false,
deviceBitfield};
gpuProperties.alignment = pageSizeForAlignment;
gpuProperties.flags.resource48Bit = unifiedMemoryProperties.allocationFlags.flags.resource48Bit;
auto cacheRegion = MemoryPropertiesHelper::getCacheRegion(unifiedMemoryProperties.allocationFlags);
MemoryPropertiesHelper::fillCachePolicyInProperties(gpuProperties, false, svmProperties.readOnly, false, cacheRegion);
auto initialPlacement = MemoryPropertiesHelper::getUSMInitialPlacement(unifiedMemoryProperties.allocationFlags);
MemoryPropertiesHelper::setUSMInitialPlacement(gpuProperties, initialPlacement);
GraphicsAllocation *allocationGpu = memoryManager->allocateGraphicsMemoryWithProperties(gpuProperties);
if (!allocationGpu) {
return nullptr;
}
setUnifiedAllocationProperties(allocationGpu, svmProperties);
SvmAllocationData allocData(rootDeviceIndex);
allocData.gpuAllocations.addAllocation(allocationGpu);
allocData.cpuAllocation = nullptr;
allocData.device = unifiedMemoryProperties.device;
allocData.size = size;
allocData.pageSizeForAlignment = pageSizeForAlignment;
allocData.setAllocId(this->allocationsCounter++);
std::unique_lock<std::shared_mutex> lock(mtx);
auto retPtr = allocationGpu->getUnderlyingBuffer();
this->svmAllocs.insert(retPtr, allocData);
return retPtr;
}
void SVMAllocsManager::setUnifiedAllocationProperties(GraphicsAllocation *allocation, const SvmAllocationProperties &svmProperties) {
allocation->setMemObjectsAllocationWithWritableFlags(!svmProperties.readOnly && !svmProperties.hostPtrReadOnly);
allocation->setCoherent(svmProperties.coherent);
}
void SVMAllocsManager::insertSVMAlloc(const SvmAllocationData &svmAllocData) {
std::unique_lock<std::shared_mutex> lock(mtx);
svmAllocs.insert(reinterpret_cast<void *>(svmAllocData.gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress()), svmAllocData);
}
void SVMAllocsManager::removeSVMAlloc(const SvmAllocationData &svmAllocData) {
std::unique_lock<std::shared_mutex> lock(mtx);
svmAllocs.remove(reinterpret_cast<void *>(svmAllocData.gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress()));
}
bool SVMAllocsManager::freeSVMAlloc(void *ptr, bool blocking) {
if (svmDeferFreeAllocs.allocations.size() > 0) {
this->freeSVMAllocDeferImpl();
}
SvmAllocationData *svmData = getSVMAlloc(ptr);
if (svmData) {
if (InternalMemoryType::deviceUnifiedMemory == svmData->memoryType &&
this->usmDeviceAllocationsCacheEnabled) {
this->usmDeviceAllocationsCache.insert(svmData->size, ptr);
return true;
}
if (blocking) {
this->freeSVMAllocImpl(ptr, FreePolicyType::POLICY_BLOCKING, svmData);
} else {
this->freeSVMAllocImpl(ptr, FreePolicyType::POLICY_NONE, svmData);
}
return true;
}
return false;
}
bool SVMAllocsManager::freeSVMAllocDefer(void *ptr) {
if (svmDeferFreeAllocs.allocations.size() > 0) {
this->freeSVMAllocDeferImpl();
}
SvmAllocationData *svmData = getSVMAlloc(ptr);
if (svmData) {
if (InternalMemoryType::deviceUnifiedMemory == svmData->memoryType &&
this->usmDeviceAllocationsCacheEnabled) {
this->usmDeviceAllocationsCache.insert(svmData->size, ptr);
return true;
}
this->freeSVMAllocImpl(ptr, FreePolicyType::POLICY_DEFER, svmData);
return true;
}
return false;
}
void SVMAllocsManager::freeSVMAllocImpl(void *ptr, FreePolicyType policy, SvmAllocationData *svmData) {
this->prepareIndirectAllocationForDestruction(svmData);
if (policy == FreePolicyType::POLICY_BLOCKING) {
if (svmData->cpuAllocation) {
this->memoryManager->waitForEnginesCompletion(*svmData->cpuAllocation);
}
for (auto &gpuAllocation : svmData->gpuAllocations.getGraphicsAllocations()) {
if (gpuAllocation) {
this->memoryManager->waitForEnginesCompletion(*gpuAllocation);
}
}
} else if (policy == FreePolicyType::POLICY_DEFER) {
if (svmData->cpuAllocation) {
if (this->memoryManager->allocInUse(*svmData->cpuAllocation)) {
if (getSVMDeferFreeAlloc(ptr) == nullptr) {
this->svmDeferFreeAllocs.insert(*svmData);
}
return;
}
}
for (auto &gpuAllocation : svmData->gpuAllocations.getGraphicsAllocations()) {
if (gpuAllocation) {
if (this->memoryManager->allocInUse(*gpuAllocation)) {
if (getSVMDeferFreeAlloc(ptr) == nullptr) {
this->svmDeferFreeAllocs.insert(*svmData);
}
return;
}
}
}
}
auto pageFaultManager = this->memoryManager->getPageFaultManager();
if (svmData->cpuAllocation && pageFaultManager) {
pageFaultManager->removeAllocation(svmData->cpuAllocation->getUnderlyingBuffer());
}
if (svmData->gpuAllocations.getAllocationType() == AllocationType::svmZeroCopy) {
freeZeroCopySvmAllocation(svmData);
} else {
freeSvmAllocationWithDeviceStorage(svmData);
}
}
void SVMAllocsManager::freeSVMAllocDeferImpl() {
std::vector<void *> freedPtr;
for (auto iter = svmDeferFreeAllocs.allocations.begin(); iter != svmDeferFreeAllocs.allocations.end(); ++iter) {
void *ptr = reinterpret_cast<void *>(iter->second.gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress());
this->freeSVMAllocImpl(ptr, FreePolicyType::POLICY_DEFER, this->getSVMAlloc(ptr));
if (this->getSVMAlloc(ptr) == nullptr) {
freedPtr.push_back(ptr);
}
}
for (uint32_t i = 0; i < freedPtr.size(); ++i) {
svmDeferFreeAllocs.allocations.erase(freedPtr[i]);
}
}
void SVMAllocsManager::trimUSMDeviceAllocCache() {
this->usmDeviceAllocationsCache.trim(this);
}
void *SVMAllocsManager::createZeroCopySvmAllocation(size_t size, const SvmAllocationProperties &svmProperties,
const RootDeviceIndicesContainer &rootDeviceIndices,
const std::map<uint32_t, DeviceBitfield> &subdeviceBitfields) {
auto rootDeviceIndex = *rootDeviceIndices.begin();
auto &deviceBitfield = subdeviceBitfields.at(rootDeviceIndex);
AllocationProperties properties{rootDeviceIndex,
true, // allocateMemory
size,
AllocationType::svmZeroCopy,
false, // isMultiStorageAllocation
deviceBitfield};
MemoryPropertiesHelper::fillCachePolicyInProperties(properties, false, svmProperties.readOnly, false, properties.cacheRegion);
RootDeviceIndicesContainer rootDeviceIndicesVector(rootDeviceIndices);
auto maxRootDeviceIndex = *std::max_element(rootDeviceIndices.begin(), rootDeviceIndices.end(), std::less<uint32_t const>());
SvmAllocationData allocData(maxRootDeviceIndex);
void *usmPtr = memoryManager->createMultiGraphicsAllocationInSystemMemoryPool(rootDeviceIndicesVector, properties, allocData.gpuAllocations);
if (!usmPtr) {
return nullptr;
}
for (const auto &rootDeviceIndex : rootDeviceIndices) {
auto allocation = allocData.gpuAllocations.getGraphicsAllocation(rootDeviceIndex);
allocation->setMemObjectsAllocationWithWritableFlags(!svmProperties.readOnly && !svmProperties.hostPtrReadOnly);
allocation->setCoherent(svmProperties.coherent);
}
allocData.size = size;
std::unique_lock<std::shared_mutex> lock(mtx);
this->svmAllocs.insert(usmPtr, allocData);
return usmPtr;
}
void *SVMAllocsManager::createUnifiedAllocationWithDeviceStorage(size_t size, const SvmAllocationProperties &svmProperties, const UnifiedMemoryProperties &unifiedMemoryProperties) {
auto rootDeviceIndex = unifiedMemoryProperties.getRootDeviceIndex();
auto externalPtr = reinterpret_cast<void *>(unifiedMemoryProperties.allocationFlags.hostptr);
bool useExternalHostPtrForCpu = externalPtr != nullptr;
const auto pageSizeForAlignment = alignUpNonZero<size_t>(unifiedMemoryProperties.alignment, MemoryConstants::pageSize64k);
size_t alignedSize = alignUp<size_t>(size, MemoryConstants::pageSize64k);
DeviceBitfield subDevices = unifiedMemoryProperties.subdeviceBitfields.at(rootDeviceIndex);
auto cpuAlignment = std::max(pageSizeForAlignment, memoryManager->peekExecutionEnvironment().rootDeviceEnvironments[rootDeviceIndex]->getProductHelper().getSvmCpuAlignment());
AllocationProperties cpuProperties{rootDeviceIndex,
!useExternalHostPtrForCpu, // allocateMemory
alignUp(alignedSize, cpuAlignment), AllocationType::svmCpu,
false, // isMultiStorageAllocation
subDevices};
cpuProperties.alignment = cpuAlignment;
cpuProperties.flags.isUSMHostAllocation = useExternalHostPtrForCpu;
cpuProperties.forceKMDAllocation = true;
cpuProperties.makeGPUVaDifferentThanCPUPtr = true;
auto cacheRegion = MemoryPropertiesHelper::getCacheRegion(unifiedMemoryProperties.allocationFlags);
MemoryPropertiesHelper::fillCachePolicyInProperties(cpuProperties, false, svmProperties.readOnly, false, cacheRegion);
GraphicsAllocation *allocationCpu = memoryManager->allocateGraphicsMemoryWithProperties(cpuProperties, externalPtr);
if (!allocationCpu) {
return nullptr;
}
setUnifiedAllocationProperties(allocationCpu, svmProperties);
void *svmPtr = allocationCpu->getUnderlyingBuffer();
UNRECOVERABLE_IF(useExternalHostPtrForCpu && (externalPtr != svmPtr));
bool multiStorageAllocation = (subDevices.count() > 1) && multiOsContextSupport;
if ((subDevices.count() > 1) && !multiOsContextSupport) {
for (uint32_t i = 0;; i++) {
if (subDevices.test(i)) {
subDevices.reset();
subDevices.set(i);
break;
}
}
}
AllocationProperties gpuProperties{rootDeviceIndex,
false,
alignedSize,
AllocationType::svmGpu,
false,
multiStorageAllocation,
subDevices};
gpuProperties.alignment = pageSizeForAlignment;
MemoryPropertiesHelper::fillCachePolicyInProperties(gpuProperties, false, svmProperties.readOnly, false, cacheRegion);
GraphicsAllocation *allocationGpu = memoryManager->allocateGraphicsMemoryWithProperties(gpuProperties, svmPtr);
if (!allocationGpu) {
memoryManager->freeGraphicsMemory(allocationCpu);
return nullptr;
}
setUnifiedAllocationProperties(allocationGpu, svmProperties);
SvmAllocationData allocData(rootDeviceIndex);
allocData.gpuAllocations.addAllocation(allocationGpu);
allocData.cpuAllocation = allocationCpu;
allocData.device = unifiedMemoryProperties.device;
allocData.pageSizeForAlignment = cpuAlignment;
allocData.size = size;
allocData.setAllocId(this->allocationsCounter++);
std::unique_lock<std::shared_mutex> lock(mtx);
this->svmAllocs.insert(svmPtr, allocData);
return svmPtr;
}
void SVMAllocsManager::freeSVMData(SvmAllocationData *svmData) {
std::unique_lock<std::mutex> lockForIndirect(mtxForIndirectAccess);
std::unique_lock<std::shared_mutex> lock(mtx);
svmAllocs.remove(reinterpret_cast<void *>(svmData->gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress()));
}
void SVMAllocsManager::freeZeroCopySvmAllocation(SvmAllocationData *svmData) {
auto gpuAllocations = svmData->gpuAllocations;
freeSVMData(svmData);
for (const auto &graphicsAllocation : gpuAllocations.getGraphicsAllocations()) {
memoryManager->freeGraphicsMemory(graphicsAllocation);
}
}
void SVMAllocsManager::initUsmDeviceAllocationsCache() {
this->usmDeviceAllocationsCache.allocations.reserve(128u);
}
void SVMAllocsManager::freeSvmAllocationWithDeviceStorage(SvmAllocationData *svmData) {
auto graphicsAllocations = svmData->gpuAllocations.getGraphicsAllocations();
GraphicsAllocation *cpuAllocation = svmData->cpuAllocation;
bool isImportedAllocation = svmData->isImportedAllocation;
freeSVMData(svmData);
for (auto gpuAllocation : graphicsAllocations) {
memoryManager->freeGraphicsMemory(gpuAllocation, isImportedAllocation);
}
memoryManager->freeGraphicsMemory(cpuAllocation, isImportedAllocation);
}
bool SVMAllocsManager::hasHostAllocations() {
std::shared_lock<std::shared_mutex> lock(mtx);
for (auto &allocation : this->svmAllocs.allocations) {
if (allocation.second->memoryType == InternalMemoryType::hostUnifiedMemory) {
return true;
}
}
return false;
}
void SVMAllocsManager::makeIndirectAllocationsResident(CommandStreamReceiver &commandStreamReceiver, TaskCountType taskCount) {
std::unique_lock<std::shared_mutex> lock(mtx);
bool parseAllAllocations = false;
auto entry = indirectAllocationsResidency.find(&commandStreamReceiver);
if (entry == indirectAllocationsResidency.end()) {
parseAllAllocations = true;
InternalAllocationsTracker tracker = {};
tracker.latestResidentObjectId = this->allocationsCounter;
tracker.latestSentTaskCount = taskCount;
this->indirectAllocationsResidency.insert(std::make_pair(&commandStreamReceiver, tracker));
} else {
if (this->allocationsCounter > entry->second.latestResidentObjectId) {
parseAllAllocations = true;
entry->second.latestResidentObjectId = this->allocationsCounter;
}
entry->second.latestSentTaskCount = taskCount;
}
if (parseAllAllocations) {
for (auto &allocation : this->svmAllocs.allocations) {
auto gpuAllocation = allocation.second->gpuAllocations.getGraphicsAllocation(commandStreamReceiver.getRootDeviceIndex());
if (gpuAllocation == nullptr) {
continue;
}
commandStreamReceiver.makeResident(*gpuAllocation);
gpuAllocation->updateResidencyTaskCount(GraphicsAllocation::objectAlwaysResident, commandStreamReceiver.getOsContext().getContextId());
gpuAllocation->setEvictable(false);
}
}
}
void SVMAllocsManager::prepareIndirectAllocationForDestruction(SvmAllocationData *allocationData) {
std::unique_lock<std::shared_mutex> lock(mtx);
if (this->indirectAllocationsResidency.size() > 0u) {
for (auto &internalAllocationsHandling : this->indirectAllocationsResidency) {
auto commandStreamReceiver = internalAllocationsHandling.first;
auto gpuAllocation = allocationData->gpuAllocations.getGraphicsAllocation(commandStreamReceiver->getRootDeviceIndex());
if (gpuAllocation == nullptr) {
continue;
}
auto desiredTaskCount = std::max(internalAllocationsHandling.second.latestSentTaskCount, gpuAllocation->getTaskCount(commandStreamReceiver->getOsContext().getContextId()));
if (gpuAllocation->isAlwaysResident(commandStreamReceiver->getOsContext().getContextId())) {
gpuAllocation->updateResidencyTaskCount(GraphicsAllocation::objectNotResident, commandStreamReceiver->getOsContext().getContextId());
gpuAllocation->updateResidencyTaskCount(desiredTaskCount, commandStreamReceiver->getOsContext().getContextId());
gpuAllocation->updateTaskCount(desiredTaskCount, commandStreamReceiver->getOsContext().getContextId());
}
}
}
}
SvmMapOperation *SVMAllocsManager::getSvmMapOperation(const void *ptr) {
std::shared_lock<std::shared_mutex> lock(mtx);
return svmMapOperations.get(ptr);
}
void SVMAllocsManager::insertSvmMapOperation(void *regionSvmPtr, size_t regionSize, void *baseSvmPtr, size_t offset, bool readOnlyMap) {
SvmMapOperation svmMapOperation;
svmMapOperation.regionSvmPtr = regionSvmPtr;
svmMapOperation.baseSvmPtr = baseSvmPtr;
svmMapOperation.offset = offset;
svmMapOperation.regionSize = regionSize;
svmMapOperation.readOnlyMap = readOnlyMap;
std::unique_lock<std::shared_mutex> lock(mtx);
svmMapOperations.insert(svmMapOperation);
}
void SVMAllocsManager::removeSvmMapOperation(const void *regionSvmPtr) {
std::unique_lock<std::shared_mutex> lock(mtx);
svmMapOperations.remove(regionSvmPtr);
}
AllocationType SVMAllocsManager::getGraphicsAllocationTypeAndCompressionPreference(const UnifiedMemoryProperties &unifiedMemoryProperties, bool &compressionEnabled) const {
compressionEnabled = false;
AllocationType allocationType = AllocationType::bufferHostMemory;
if (unifiedMemoryProperties.memoryType == InternalMemoryType::deviceUnifiedMemory) {
if (unifiedMemoryProperties.allocationFlags.allocFlags.allocWriteCombined) {
allocationType = AllocationType::writeCombined;
} else {
UNRECOVERABLE_IF(nullptr == unifiedMemoryProperties.device);
if (CompressionSelector::allowStatelessCompression()) {
compressionEnabled = true;
}
if (unifiedMemoryProperties.requestedAllocationType != AllocationType::unknown) {
allocationType = unifiedMemoryProperties.requestedAllocationType;
} else {
allocationType = AllocationType::buffer;
}
}
}
return allocationType;
}
void SVMAllocsManager::prefetchMemory(Device &device, CommandStreamReceiver &commandStreamReceiver, SvmAllocationData &svmData) {
auto getSubDeviceId = [](Device &device) {
if (!device.isSubDevice()) {
uint32_t deviceBitField = static_cast<uint32_t>(device.getDeviceBitfield().to_ulong());
if (device.getDeviceBitfield().count() > 1) {
deviceBitField &= ~deviceBitField + 1;
}
return Math::log2(deviceBitField);
}
return static_cast<NEO::SubDevice *>(&device)->getSubDeviceIndex();
};
auto getSubDeviceIds = [](CommandStreamReceiver &csr) {
SubDeviceIdsVec subDeviceIds;
for (auto subDeviceId = 0u; subDeviceId < csr.getOsContext().getDeviceBitfield().size(); subDeviceId++) {
if (csr.getOsContext().getDeviceBitfield().test(subDeviceId)) {
subDeviceIds.push_back(subDeviceId);
}
}
return subDeviceIds;
};
// Perform prefetch for chunks if EnableBOChunkingPrefetch is 1
// and if KMD migration is set, as current target is to use
// chunking only with KMD migration
bool isChunkingNeededForDeviceAllocations = false;
if (NEO::debugManager.flags.EnableBOChunkingDevMemPrefetch.get() &&
memoryManager->isKmdMigrationAvailable(device.getRootDeviceIndex()) &&
(svmData.memoryType == InternalMemoryType::deviceUnifiedMemory)) {
isChunkingNeededForDeviceAllocations = true;
}
if ((memoryManager->isKmdMigrationAvailable(device.getRootDeviceIndex()) &&
(svmData.memoryType == InternalMemoryType::sharedUnifiedMemory)) ||
isChunkingNeededForDeviceAllocations) {
auto gfxAllocation = svmData.gpuAllocations.getGraphicsAllocation(device.getRootDeviceIndex());
auto subDeviceIds = commandStreamReceiver.getActivePartitions() > 1 ? getSubDeviceIds(commandStreamReceiver) : SubDeviceIdsVec{getSubDeviceId(device)};
memoryManager->setMemPrefetch(gfxAllocation, subDeviceIds, device.getRootDeviceIndex());
}
}
void SVMAllocsManager::prefetchSVMAllocs(Device &device, CommandStreamReceiver &commandStreamReceiver) {
std::shared_lock<std::shared_mutex> lock(mtx);
for (auto &allocation : this->svmAllocs.allocations) {
NEO::SvmAllocationData allocData = *allocation.second;
this->prefetchMemory(device, commandStreamReceiver, allocData);
}
}
std::unique_lock<std::mutex> SVMAllocsManager::obtainOwnership() {
return std::unique_lock<std::mutex>(mtxForIndirectAccess);
}
} // namespace NEO