Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
compute-runtime/shared/source/direct_submission/direct_submission_hw.inl

835 lines
39 KiB
C++
Raw Blame History

/*
* Copyright (C) 2020-2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/command_container/command_encoder.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/command_stream/submissions_aggregator.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/device/device.h"
#include "shared/source/direct_submission/direct_submission_hw.h"
#include "shared/source/direct_submission/relaxed_ordering_helper.h"
#include "shared/source/execution_environment/execution_environment.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/gmm_helper/gmm_helper.h"
#include "shared/source/gmm_helper/gmm_lib.h"
#include "shared/source/helpers/aligned_memory.h"
#include "shared/source/helpers/compiler_product_helper.h"
#include "shared/source/helpers/definitions/command_encoder_args.h"
#include "shared/source/helpers/flush_stamp.h"
#include "shared/source/helpers/gfx_core_helper.h"
#include "shared/source/helpers/hw_info.h"
#include "shared/source/helpers/ptr_math.h"
#include "shared/source/memory_manager/allocation_properties.h"
#include "shared/source/memory_manager/graphics_allocation.h"
#include "shared/source/memory_manager/memory_manager.h"
#include "shared/source/memory_manager/memory_operations_handler.h"
#include "shared/source/os_interface/os_context.h"
#include "shared/source/os_interface/product_helper.h"
#include "shared/source/utilities/cpu_info.h"
#include "create_direct_submission_hw.inl"
#include <algorithm>
#include <cstring>
namespace NEO {
template <typename GfxFamily, typename Dispatcher>
DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmissionInputParams &inputParams)
: ringBuffers(RingBufferUse::initialRingBufferCount), osContext(inputParams.osContext), rootDeviceIndex(inputParams.rootDeviceIndex), rootDeviceEnvironment(inputParams.rootDeviceEnvironment) {
memoryManager = inputParams.memoryManager;
globalFenceAllocation = inputParams.globalFenceAllocation;
hwInfo = inputParams.rootDeviceEnvironment.getHardwareInfo();
memoryOperationHandler = inputParams.rootDeviceEnvironment.memoryOperationsInterface.get();
auto &productHelper = inputParams.rootDeviceEnvironment.getHelper<ProductHelper>();
auto &compilerProductHelper = inputParams.rootDeviceEnvironment.getHelper<CompilerProductHelper>();
disableCacheFlush = UllsDefaults::defaultDisableCacheFlush;
disableMonitorFence = UllsDefaults::defaultDisableMonitorFence;
if (debugManager.flags.DirectSubmissionDisableMonitorFence.get() != -1) {
this->disableMonitorFence = debugManager.flags.DirectSubmissionDisableMonitorFence.get();
}
if (debugManager.flags.DirectSubmissionMaxRingBuffers.get() != -1) {
this->maxRingBufferCount = debugManager.flags.DirectSubmissionMaxRingBuffers.get();
}
if (debugManager.flags.DirectSubmissionDisableCacheFlush.get() != -1) {
disableCacheFlush = !!debugManager.flags.DirectSubmissionDisableCacheFlush.get();
}
if (debugManager.flags.DirectSubmissionDetectGpuHang.get() != -1) {
detectGpuHang = !!debugManager.flags.DirectSubmissionDetectGpuHang.get();
}
miMemFenceRequired = productHelper.isGlobalFenceInDirectSubmissionRequired(*hwInfo);
if (debugManager.flags.DirectSubmissionInsertExtraMiMemFenceCommands.get() != -1) {
miMemFenceRequired = debugManager.flags.DirectSubmissionInsertExtraMiMemFenceCommands.get();
}
if (miMemFenceRequired && compilerProductHelper.isHeaplessStateInitEnabled(compilerProductHelper.isHeaplessModeEnabled(*hwInfo))) {
this->systemMemoryFenceAddressSet = true;
}
if (debugManager.flags.DirectSubmissionInsertSfenceInstructionPriorToSubmission.get() != -1) {
sfenceMode = static_cast<DirectSubmissionSfenceMode>(debugManager.flags.DirectSubmissionInsertSfenceInstructionPriorToSubmission.get());
}
if (debugManager.flags.DirectSubmissionMonitorFenceInputPolicy.get() != -1) {
this->inputMonitorFenceDispatchRequirement = !!(debugManager.flags.DirectSubmissionMonitorFenceInputPolicy.get());
}
int32_t disableCacheFlushKey = debugManager.flags.DirectSubmissionDisableCpuCacheFlush.get();
if (disableCacheFlushKey != -1) {
disableCpuCacheFlush = (disableCacheFlushKey == 1);
}
isDisablePrefetcherRequired = productHelper.isPrefetcherDisablingInDirectSubmissionRequired();
if (debugManager.flags.DirectSubmissionDisablePrefetcher.get() != -1) {
isDisablePrefetcherRequired = !!debugManager.flags.DirectSubmissionDisablePrefetcher.get();
}
UNRECOVERABLE_IF(!CpuInfo::getInstance().isFeatureSupported(CpuInfo::featureClflush) && !disableCpuCacheFlush);
setImmWritePostSyncOffset();
dcFlushRequired = MemorySynchronizationCommands<GfxFamily>::getDcFlushEnable(true, inputParams.rootDeviceEnvironment);
auto &gfxCoreHelper = inputParams.rootDeviceEnvironment.getHelper<GfxCoreHelper>();
relaxedOrderingEnabled = gfxCoreHelper.isRelaxedOrderingSupported();
this->currentRelaxedOrderingQueueSize = RelaxedOrderingHelper::queueSizeMultiplier;
if (debugManager.flags.DirectSubmissionRelaxedOrdering.get() != -1) {
relaxedOrderingEnabled = (debugManager.flags.DirectSubmissionRelaxedOrdering.get() == 1);
}
if (Dispatcher::isCopy() && relaxedOrderingEnabled) {
relaxedOrderingEnabled = (debugManager.flags.DirectSubmissionRelaxedOrderingForBcs.get() != 0);
}
}
template <typename GfxFamily, typename Dispatcher>
DirectSubmissionHw<GfxFamily, Dispatcher>::~DirectSubmissionHw() = default;
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::allocateResources() {
DirectSubmissionAllocations allocations;
bool isMultiOsContextCapable = osContext.getNumSupportedDevices() > 1u;
constexpr size_t minimumRequiredSize = 256 * MemoryConstants::kiloByte;
constexpr size_t additionalAllocationSize = MemoryConstants::pageSize;
const auto allocationSize = alignUp(minimumRequiredSize + additionalAllocationSize, MemoryConstants::pageSize64k);
const AllocationProperties commandStreamAllocationProperties{rootDeviceIndex,
true, allocationSize,
AllocationType::ringBuffer,
isMultiOsContextCapable, false, osContext.getDeviceBitfield()};
for (uint32_t ringBufferIndex = 0; ringBufferIndex < RingBufferUse::initialRingBufferCount; ringBufferIndex++) {
auto ringBuffer = memoryManager->allocateGraphicsMemoryWithProperties(commandStreamAllocationProperties);
this->ringBuffers[ringBufferIndex].ringBuffer = ringBuffer;
UNRECOVERABLE_IF(ringBuffer == nullptr);
allocations.push_back(ringBuffer);
memset(ringBuffer->getUnderlyingBuffer(), 0, allocationSize);
}
const AllocationProperties semaphoreAllocationProperties{rootDeviceIndex,
true, MemoryConstants::pageSize,
AllocationType::semaphoreBuffer,
isMultiOsContextCapable, false, osContext.getDeviceBitfield()};
semaphores = memoryManager->allocateGraphicsMemoryWithProperties(semaphoreAllocationProperties);
UNRECOVERABLE_IF(semaphores == nullptr);
allocations.push_back(semaphores);
if (this->workPartitionAllocation != nullptr) {
allocations.push_back(workPartitionAllocation);
}
if (completionFenceAllocation != nullptr) {
allocations.push_back(completionFenceAllocation);
}
if (this->relaxedOrderingEnabled) {
const AllocationProperties allocationProperties(rootDeviceIndex,
true, MemoryConstants::pageSize64k,
AllocationType::deferredTasksList,
isMultiOsContextCapable, false, osContext.getDeviceBitfield());
deferredTasksListAllocation = memoryManager->allocateGraphicsMemoryWithProperties(allocationProperties);
UNRECOVERABLE_IF(deferredTasksListAllocation == nullptr);
allocations.push_back(deferredTasksListAllocation);
AllocationProperties relaxedOrderingSchedulerAllocationProperties(rootDeviceIndex,
true, MemoryConstants::pageSize64k,
AllocationType::commandBuffer,
isMultiOsContextCapable, false, osContext.getDeviceBitfield());
relaxedOrderingSchedulerAllocationProperties.flags.cantBeReadOnly = true;
relaxedOrderingSchedulerAllocation = memoryManager->allocateGraphicsMemoryWithProperties(relaxedOrderingSchedulerAllocationProperties);
UNRECOVERABLE_IF(relaxedOrderingSchedulerAllocation == nullptr);
allocations.push_back(relaxedOrderingSchedulerAllocation);
}
if (debugManager.flags.DirectSubmissionPrintBuffers.get()) {
for (uint32_t ringBufferIndex = 0; ringBufferIndex < RingBufferUse::initialRingBufferCount; ringBufferIndex++) {
const auto ringBuffer = this->ringBuffers[ringBufferIndex].ringBuffer;
printf("Ring buffer %u - gpu address: %" PRIx64 " - %" PRIx64 ", cpu address: %p - %p, size: %zu \n",
ringBufferIndex,
ringBuffer->getGpuAddress(),
ptrOffset(ringBuffer->getGpuAddress(), ringBuffer->getUnderlyingBufferSize()),
ringBuffer->getUnderlyingBuffer(),
ptrOffset(ringBuffer->getUnderlyingBuffer(), ringBuffer->getUnderlyingBufferSize()),
ringBuffer->getUnderlyingBufferSize());
}
}
handleResidency();
ringCommandStream.replaceBuffer(this->ringBuffers[0u].ringBuffer->getUnderlyingBuffer(), minimumRequiredSize);
ringCommandStream.replaceGraphicsAllocation(this->ringBuffers[0].ringBuffer);
semaphorePtr = semaphores->getUnderlyingBuffer();
semaphoreGpuVa = semaphores->getGpuAddress();
semaphoreData = static_cast<volatile RingSemaphoreData *>(semaphorePtr);
memset(semaphorePtr, 0, sizeof(RingSemaphoreData));
semaphoreData->queueWorkCount = 0;
cpuCachelineFlush(semaphorePtr, MemoryConstants::cacheLineSize);
this->gpuVaForMiFlush = this->semaphoreGpuVa + offsetof(RingSemaphoreData, miFlushSpace);
this->gpuVaForPagingFenceSemaphore = this->semaphoreGpuVa + offsetof(RingSemaphoreData, pagingFenceCounter);
auto ret = makeResourcesResident(allocations);
return ret && allocateOsResources();
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::makeResourcesResident(DirectSubmissionAllocations &allocations) {
auto ret = memoryOperationHandler->makeResidentWithinOsContext(&this->osContext, ArrayRef<GraphicsAllocation *>(allocations), false, false, true) == MemoryOperationsStatus::success;
return ret;
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::allocateOsResources() {
if (this->semaphorePtr != nullptr) {
this->tagAddress = reinterpret_cast<volatile TagAddressType *>(reinterpret_cast<uint8_t *>(this->semaphorePtr) + offsetof(RingSemaphoreData, tagAllocation));
} else {
this->tagAddress = nullptr;
}
return true;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::unblockGpu() {
SemaphoreFenceHelper fence(*this);
if (this->pciBarrierPtr) {
*this->pciBarrierPtr = 0u;
}
PRINT_DEBUG_STRING(debugManager.flags.DirectSubmissionPrintSemaphoreUsage.get() == 1, stdout, "DirectSubmission semaphore %" PRIx64 " unlocked with value: %u\n", semaphoreGpuVa, currentQueueWorkCount);
semaphoreData->queueWorkCount = currentQueueWorkCount;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::cpuCachelineFlush(void *ptr, size_t size) {
if (disableCpuCacheFlush) {
return;
}
constexpr size_t cachlineBit = 6;
static_assert(MemoryConstants::cacheLineSize == 1 << cachlineBit, "cachlineBit has invalid value");
char *flushPtr = reinterpret_cast<char *>(ptr);
char *flushEndPtr = reinterpret_cast<char *>(ptr) + size;
flushPtr = alignDown(flushPtr, MemoryConstants::cacheLineSize);
flushEndPtr = alignUp(flushEndPtr, MemoryConstants::cacheLineSize);
size_t cachelines = (flushEndPtr - flushPtr) >> cachlineBit;
for (size_t i = 0; i < cachelines; i++) {
CpuIntrinsics::clFlush(flushPtr);
flushPtr += MemoryConstants::cacheLineSize;
}
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::initialize(bool submitOnInit) {
bool ret = allocateResources();
if (ret && submitOnInit) {
size_t startBufferSize = Dispatcher::getSizePreemption() +
getSizeSemaphoreSection(false);
Dispatcher::dispatchPreemption(ringCommandStream);
if (this->partitionedMode) {
startBufferSize += getSizePartitionRegisterConfigurationSection();
dispatchPartitionRegisterConfiguration();
this->partitionConfigSet = true;
}
if (this->globalFenceAllocation && !this->systemMemoryFenceAddressSet) {
startBufferSize += getSizeSystemMemoryFenceAddress();
dispatchSystemMemoryFenceAddress();
this->systemMemoryFenceAddressSet = true;
}
if (this->relaxedOrderingEnabled) {
preinitializeRelaxedOrderingSections();
initRelaxedOrderingRegisters();
dispatchStaticRelaxedOrderingScheduler();
startBufferSize += RelaxedOrderingHelper::getSizeRegistersInit<GfxFamily>();
this->relaxedOrderingInitialized = true;
}
dispatchSemaphoreSection(currentQueueWorkCount);
ringStart = submit(ringCommandStream.getGraphicsAllocation()->getGpuAddress(), startBufferSize, nullptr);
return ringStart;
}
return ret;
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::stopRingBuffer(bool blocking) {
if (!ringStart) {
if (blocking) {
this->ensureRingCompletion();
}
return true;
}
bool relaxedOrderingSchedulerWasRequired = this->relaxedOrderingSchedulerRequired;
if (this->relaxedOrderingEnabled && this->relaxedOrderingSchedulerRequired) {
dispatchRelaxedOrderingQueueStall();
}
void *flushPtr = ringCommandStream.getSpace(0);
Dispatcher::dispatchCacheFlush(ringCommandStream, this->rootDeviceEnvironment, gpuVaForMiFlush);
if (disableMonitorFence) {
dispatchStopRingBufferSection();
}
Dispatcher::dispatchStopCommandBuffer(ringCommandStream);
auto bytesToPad = Dispatcher::getSizeStartCommandBuffer() - Dispatcher::getSizeStopCommandBuffer();
EncodeNoop<GfxFamily>::emitNoop(ringCommandStream, bytesToPad);
EncodeNoop<GfxFamily>::alignToCacheLine(ringCommandStream);
cpuCachelineFlush(flushPtr, getSizeEnd(relaxedOrderingSchedulerWasRequired));
this->unblockGpu();
cpuCachelineFlush(semaphorePtr, MemoryConstants::cacheLineSize);
this->handleStopRingBuffer();
this->ringStart = false;
if (blocking) {
this->ensureRingCompletion();
}
return true;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSemaphoreSection(uint32_t value) {
using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;
if (debugManager.flags.DirectSubmissionPrintSemaphoreUsage.get() == 1) {
printf("DirectSubmission semaphore %" PRIx64 " programmed with value: %u\n", semaphoreGpuVa, value);
}
dispatchDisablePrefetcher(true);
if (this->relaxedOrderingEnabled && this->relaxedOrderingSchedulerRequired) {
dispatchRelaxedOrderingSchedulerSection(value);
} else {
bool switchOnUnsuccessful = false;
if (debugManager.flags.DirectSubmissionSwitchSemaphoreMode.get() != -1) {
switchOnUnsuccessful = !!debugManager.flags.DirectSubmissionSwitchSemaphoreMode.get();
}
EncodeSemaphore<GfxFamily>::addMiSemaphoreWaitCommand(ringCommandStream,
semaphoreGpuVa,
value,
COMPARE_OPERATION::COMPARE_OPERATION_SAD_GREATER_THAN_OR_EQUAL_SDD, false, false, false, switchOnUnsuccessful, nullptr);
}
if (miMemFenceRequired) {
MemorySynchronizationCommands<GfxFamily>::addAdditionalSynchronizationForDirectSubmission(ringCommandStream, this->gpuVaForAdditionalSynchronizationWA, true, rootDeviceEnvironment);
}
dispatchPrefetchMitigation();
dispatchDisablePrefetcher(false);
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeSemaphoreSection(bool relaxedOrderingSchedulerRequired) {
size_t semaphoreSize = (this->relaxedOrderingEnabled && relaxedOrderingSchedulerRequired) ? RelaxedOrderingHelper::DynamicSchedulerSizeAndOffsetSection<GfxFamily>::totalSize
: EncodeSemaphore<GfxFamily>::getSizeMiSemaphoreWait();
semaphoreSize += getSizePrefetchMitigation();
if (isDisablePrefetcherRequired) {
semaphoreSize += 2 * getSizeDisablePrefetcher();
}
if (miMemFenceRequired) {
semaphoreSize += MemorySynchronizationCommands<GfxFamily>::getSizeForSingleAdditionalSynchronizationForDirectSubmission(rootDeviceEnvironment);
}
return semaphoreSize;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchStartSection(uint64_t gpuStartAddress) {
Dispatcher::dispatchStartCommandBuffer(ringCommandStream, gpuStartAddress);
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeStartSection() {
return Dispatcher::getSizeStartCommandBuffer();
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSwitchRingBufferSection(uint64_t nextBufferGpuAddress) {
if (disableMonitorFence) {
TagData currentTagData = {};
getTagAddressValueForRingSwitch(currentTagData);
Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, this->rootDeviceEnvironment, this->partitionedMode, this->dcFlushRequired, this->notifyKmdDuringMonitorFence);
}
Dispatcher::dispatchStartCommandBuffer(ringCommandStream, nextBufferGpuAddress);
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeSwitchRingBufferSection() {
size_t size = Dispatcher::getSizeStartCommandBuffer();
if (disableMonitorFence) {
size += Dispatcher::getSizeMonitorFence(rootDeviceEnvironment);
}
return size;
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeEnd(bool relaxedOrderingSchedulerRequired) {
size_t size = Dispatcher::getSizeStopCommandBuffer() +
Dispatcher::getSizeCacheFlush(rootDeviceEnvironment) +
(Dispatcher::getSizeStartCommandBuffer() - Dispatcher::getSizeStopCommandBuffer()) +
MemoryConstants::cacheLineSize;
if (disableMonitorFence) {
size += dispatchStopRingBufferSectionSize();
}
if (this->relaxedOrderingEnabled && relaxedOrderingSchedulerRequired) {
size += getSizeDispatchRelaxedOrderingQueueStall();
}
return size;
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeDispatch(bool relaxedOrderingSchedulerRequired, bool returnPtrsRequired, bool dispatchMonitorFence) {
size_t size = getSizeSemaphoreSection(relaxedOrderingSchedulerRequired) + getSizeStartSection();
if (this->relaxedOrderingEnabled && returnPtrsRequired) {
size += RelaxedOrderingHelper::getSizeReturnPtrRegs<GfxFamily>();
}
if (!disableCacheFlush) {
size += Dispatcher::getSizeCacheFlush(rootDeviceEnvironment);
}
if (dispatchMonitorFence) {
size += Dispatcher::getSizeMonitorFence(rootDeviceEnvironment);
}
size += getSizeNewResourceHandler();
return size;
}
template <typename GfxFamily, typename Dispatcher>
void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBuffer &batchBuffer, bool dispatchMonitorFence) {
void *currentPosition = ringCommandStream.getSpace(0);
auto copyCmdBuffer = this->copyCommandBufferIntoRing(batchBuffer);
if (debugManager.flags.DirectSubmissionPrintBuffers.get()) {
printf("Client buffer:\n");
printf("Command buffer allocation - gpu address: %" PRIx64 " - %" PRIx64 ", cpu address: %p - %p, size: %zu \n",
batchBuffer.commandBufferAllocation->getGpuAddress(),
ptrOffset(batchBuffer.commandBufferAllocation->getGpuAddress(), batchBuffer.commandBufferAllocation->getUnderlyingBufferSize()),
batchBuffer.commandBufferAllocation->getUnderlyingBuffer(),
ptrOffset(batchBuffer.commandBufferAllocation->getUnderlyingBuffer(), batchBuffer.commandBufferAllocation->getUnderlyingBufferSize()),
batchBuffer.commandBufferAllocation->getUnderlyingBufferSize());
printf("Command buffer - start gpu address: %" PRIx64 " - %" PRIx64 ", start cpu address: %p - %p, start offset: %zu, used size: %zu \n",
ptrOffset(batchBuffer.commandBufferAllocation->getGpuAddress(), batchBuffer.startOffset),
ptrOffset(batchBuffer.commandBufferAllocation->getGpuAddress(), batchBuffer.usedSize),
ptrOffset(batchBuffer.commandBufferAllocation->getUnderlyingBuffer(), batchBuffer.startOffset),
ptrOffset(batchBuffer.commandBufferAllocation->getUnderlyingBuffer(), batchBuffer.usedSize),
batchBuffer.startOffset,
batchBuffer.usedSize);
printf("Ring buffer for submission - start gpu address: %" PRIx64 " - %" PRIx64 ", start cpu address: %p - %p, size: %zu, submission address: %" PRIx64 ", used size: %zu, copyCmdBuffer: %d \n",
ringCommandStream.getGraphicsAllocation()->getGpuAddress(),
ptrOffset(ringCommandStream.getGraphicsAllocation()->getGpuAddress(), ringCommandStream.getGraphicsAllocation()->getUnderlyingBufferSize()),
ringCommandStream.getGraphicsAllocation()->getUnderlyingBuffer(),
ptrOffset(ringCommandStream.getGraphicsAllocation()->getUnderlyingBuffer(), ringCommandStream.getGraphicsAllocation()->getUnderlyingBufferSize()),
ringCommandStream.getGraphicsAllocation()->getUnderlyingBufferSize(),
ptrOffset(ringCommandStream.getGraphicsAllocation()->getGpuAddress(), ringCommandStream.getUsed()),
ringCommandStream.getUsed(),
copyCmdBuffer);
}
if (batchBuffer.pagingFenceSemInfo.requiresProgrammingSemaphore()) {
dispatchSemaphoreForPagingFence(batchBuffer.pagingFenceSemInfo.pagingFenceValue);
}
auto commandStreamAddress = ptrOffset(batchBuffer.commandBufferAllocation->getGpuAddress(), batchBuffer.startOffset);
void *returnCmd = batchBuffer.endCmdPtr;
LinearStream relaxedOrderingReturnPtrCmdStream;
if (this->relaxedOrderingEnabled && batchBuffer.hasRelaxedOrderingDependencies) {
// preallocate and patch after start section
auto relaxedOrderingReturnPtrCmds = ringCommandStream.getSpace(RelaxedOrderingHelper::getSizeReturnPtrRegs<GfxFamily>());
relaxedOrderingReturnPtrCmdStream.replaceBuffer(relaxedOrderingReturnPtrCmds, RelaxedOrderingHelper::getSizeReturnPtrRegs<GfxFamily>());
}
if (copyCmdBuffer) {
auto cmdStreamTaskPtr = ptrOffset(batchBuffer.stream->getCpuBase(), batchBuffer.startOffset);
auto sizeToCopy = ptrDiff(returnCmd, cmdStreamTaskPtr);
auto ringPtr = ringCommandStream.getSpace(sizeToCopy);
memcpy(ringPtr, cmdStreamTaskPtr, sizeToCopy);
} else {
dispatchStartSection(commandStreamAddress);
}
uint64_t returnGpuPointer = ringCommandStream.getCurrentGpuAddressPosition();
if (this->relaxedOrderingEnabled && batchBuffer.hasRelaxedOrderingDependencies) {
dispatchRelaxedOrderingReturnPtrRegs(relaxedOrderingReturnPtrCmdStream, returnGpuPointer);
} else if (!copyCmdBuffer) {
setReturnAddress(returnCmd, returnGpuPointer);
}
if (this->relaxedOrderingEnabled && batchBuffer.hasRelaxedOrderingDependencies) {
dispatchTaskStoreSection(batchBuffer.taskStartAddress);
uint32_t expectedQueueSize = batchBuffer.numCsrClients * RelaxedOrderingHelper::queueSizeMultiplier;
expectedQueueSize = std::min(expectedQueueSize, RelaxedOrderingHelper::maxQueueSize);
if (expectedQueueSize > this->currentRelaxedOrderingQueueSize && debugManager.flags.DirectSubmissionRelaxedOrderingQueueSizeLimit.get() == -1) {
updateRelaxedOrderingQueueSize(expectedQueueSize);
}
}
if (!disableCacheFlush) {
Dispatcher::dispatchCacheFlush(ringCommandStream, this->rootDeviceEnvironment, gpuVaForMiFlush);
}
if (dispatchMonitorFence) {
TagData currentTagData = {};
getTagAddressValue(currentTagData);
Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, this->rootDeviceEnvironment, this->partitionedMode, this->dcFlushRequired, this->notifyKmdDuringMonitorFence);
}
dispatchSemaphoreSection(currentQueueWorkCount + 1);
return currentPosition;
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::copyCommandBufferIntoRing(BatchBuffer &batchBuffer) {
/* Command buffer can't be copied into ring if implicit scaling or metrics are enabled,
because those features uses GPU VAs of command buffer which would be invalid after copy. */
auto ret = !batchBuffer.disableFlatRingBuffer &&
this->osContext.getNumSupportedDevices() == 1u &&
!this->rootDeviceEnvironment.executionEnvironment.areMetricsEnabled() &&
!batchBuffer.chainedBatchBuffer &&
batchBuffer.commandBufferAllocation &&
MemoryPoolHelper::isSystemMemoryPool(batchBuffer.commandBufferAllocation->getMemoryPool()) &&
!batchBuffer.hasRelaxedOrderingDependencies;
if (debugManager.flags.DirectSubmissionFlatRingBuffer.get() != -1) {
ret &= !!debugManager.flags.DirectSubmissionFlatRingBuffer.get();
}
return ret;
}
template <typename GfxFamily, typename Dispatcher>
size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getUllsStateSize() {
size_t startSize = 0u;
if (!this->partitionConfigSet) {
startSize += getSizePartitionRegisterConfigurationSection();
}
if (this->miMemFenceRequired && !this->systemMemoryFenceAddressSet) {
startSize += getSizeSystemMemoryFenceAddress();
}
if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
startSize += RelaxedOrderingHelper::getSizeRegistersInit<GfxFamily>();
}
return startSize;
}
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchUllsState() {
if (!this->partitionConfigSet) {
dispatchPartitionRegisterConfiguration();
this->partitionConfigSet = true;
}
if (this->globalFenceAllocation && !this->systemMemoryFenceAddressSet) {
dispatchSystemMemoryFenceAddress();
this->systemMemoryFenceAddressSet = true;
}
if (this->relaxedOrderingEnabled && !this->relaxedOrderingInitialized) {
preinitializeRelaxedOrderingSections();
dispatchStaticRelaxedOrderingScheduler();
initRelaxedOrderingRegisters();
this->relaxedOrderingInitialized = true;
}
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchCommandBuffer(BatchBuffer &batchBuffer, FlushStampTracker &flushStamp) {
this->handleRingRestartForUllsLightResidency(batchBuffer.allocationsForResidency);
lastSubmittedThrottle = batchBuffer.throttle;
bool relaxedOrderingSchedulerWillBeNeeded = (this->relaxedOrderingSchedulerRequired || batchBuffer.hasRelaxedOrderingDependencies);
bool inputRequiredMonitorFence = false;
if (this->inputMonitorFenceDispatchRequirement) {
inputRequiredMonitorFence = batchBuffer.dispatchMonitorFence;
} else {
inputRequiredMonitorFence = batchBuffer.hasStallingCmds;
}
bool dispatchMonitorFence = this->dispatchMonitorFenceRequired(inputRequiredMonitorFence);
size_t dispatchSize = this->getUllsStateSize() + getSizeDispatch(relaxedOrderingSchedulerWillBeNeeded, batchBuffer.hasRelaxedOrderingDependencies, dispatchMonitorFence);
if (this->copyCommandBufferIntoRing(batchBuffer)) {
dispatchSize += (batchBuffer.stream->getUsed() - batchBuffer.startOffset) - 2 * getSizeStartSection();
}
if (batchBuffer.pagingFenceSemInfo.requiresProgrammingSemaphore()) {
dispatchSize += getSizeSemaphoreForPagingFence();
}
size_t cycleSize = getSizeSwitchRingBufferSection();
size_t requiredMinimalSize = dispatchSize + cycleSize + getSizeEnd(relaxedOrderingSchedulerWillBeNeeded);
if (this->relaxedOrderingEnabled) {
requiredMinimalSize += +RelaxedOrderingHelper::getSizeReturnPtrRegs<GfxFamily>();
if (batchBuffer.hasStallingCmds && this->relaxedOrderingSchedulerRequired) {
requiredMinimalSize += getSizeDispatchRelaxedOrderingQueueStall();
}
if (batchBuffer.hasRelaxedOrderingDependencies) {
requiredMinimalSize += RelaxedOrderingHelper::getSizeTaskStoreSection<GfxFamily>() + sizeof(typename GfxFamily::MI_STORE_DATA_IMM);
}
}
auto needStart = !this->ringStart;
this->switchRingBuffersNeeded(requiredMinimalSize, batchBuffer.allocationsForResidency);
auto startVA = ringCommandStream.getCurrentGpuAddressPosition();
this->dispatchUllsState();
if (this->relaxedOrderingEnabled && batchBuffer.hasStallingCmds && this->relaxedOrderingSchedulerRequired) {
dispatchRelaxedOrderingQueueStall();
}
this->relaxedOrderingSchedulerRequired |= batchBuffer.hasRelaxedOrderingDependencies;
handleNewResourcesSubmission();
void *currentPosition = dispatchWorkloadSection(batchBuffer, dispatchMonitorFence);
cpuCachelineFlush(currentPosition, dispatchSize);
auto requiresBlockingResidencyHandling = batchBuffer.pagingFenceSemInfo.requiresBlockingResidencyHandling;
if (!this->submitCommandBufferToGpu(needStart, startVA, requiredMinimalSize, requiresBlockingResidencyHandling, batchBuffer.allocationsForResidency)) {
return false;
}
cpuCachelineFlush(semaphorePtr, MemoryConstants::cacheLineSize);
currentQueueWorkCount++;
uint64_t flushValue = updateTagValue(dispatchMonitorFence);
if (flushValue == DirectSubmissionHw<GfxFamily, Dispatcher>::updateTagValueFail) {
return false;
}
flushStamp.setStamp(flushValue);
return this->ringStart;
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::submitCommandBufferToGpu(bool needStart, uint64_t gpuAddress, size_t size, bool needWait, const ResidencyContainer *allocationsForResidency) {
if (needStart) {
this->ringStart = this->submit(gpuAddress, size, allocationsForResidency);
return this->ringStart;
} else {
if (needWait) {
handleResidency();
}
this->unblockGpu();
return true;
}
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::setReturnAddress(void *returnCmd, uint64_t returnAddress) {
using MI_BATCH_BUFFER_START = typename GfxFamily::MI_BATCH_BUFFER_START;
MI_BATCH_BUFFER_START cmd = GfxFamily::cmdInitBatchBufferStart;
cmd.setBatchBufferStartAddress(returnAddress);
cmd.setAddressSpaceIndicator(MI_BATCH_BUFFER_START::ADDRESS_SPACE_INDICATOR_PPGTT);
MI_BATCH_BUFFER_START *returnBBStart = static_cast<MI_BATCH_BUFFER_START *>(returnCmd);
*returnBBStart = cmd;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::handleNewResourcesSubmission() {
if (isNewResourceHandleNeeded()) {
auto tlbFlushCounter = this->osContext.peekTlbFlushCounter();
Dispatcher::dispatchTlbFlush(this->ringCommandStream, this->gpuVaForMiFlush, this->rootDeviceEnvironment);
this->osContext.setTlbFlushed(tlbFlushCounter);
}
}
template <typename GfxFamily, typename Dispatcher>
size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeNewResourceHandler() {
// Overestimate to avoid race
return Dispatcher::getSizeTlbFlush(this->rootDeviceEnvironment);
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::isNewResourceHandleNeeded() {
auto newResourcesBound = this->osContext.isTlbFlushRequired();
if (debugManager.flags.DirectSubmissionNewResourceTlbFlush.get() != -1) {
newResourcesBound = debugManager.flags.DirectSubmissionNewResourceTlbFlush.get();
}
return newResourcesBound;
}
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffersNeeded(size_t size, ResidencyContainer *allocationsForResidency) {
if (this->ringCommandStream.getAvailableSpace() < size) {
this->switchRingBuffers(allocationsForResidency);
}
}
template <typename GfxFamily, typename Dispatcher>
inline uint64_t DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffers(ResidencyContainer *allocationsForResidency) {
GraphicsAllocation *nextRingBuffer = switchRingBuffersAllocations(allocationsForResidency);
this->handleRingRestartForUllsLightResidency(allocationsForResidency);
void *flushPtr = ringCommandStream.getSpace(0);
uint64_t currentBufferGpuVa = ringCommandStream.getCurrentGpuAddressPosition();
if (ringStart) {
dispatchSwitchRingBufferSection(nextRingBuffer->getGpuAddress());
cpuCachelineFlush(flushPtr, getSizeSwitchRingBufferSection());
}
ringCommandStream.replaceBuffer(nextRingBuffer->getUnderlyingBuffer(), ringCommandStream.getMaxAvailableSpace());
ringCommandStream.replaceGraphicsAllocation(nextRingBuffer);
handleSwitchRingBuffers(allocationsForResidency);
return currentBufferGpuVa;
}
template <typename GfxFamily, typename Dispatcher>
inline GraphicsAllocation *DirectSubmissionHw<GfxFamily, Dispatcher>::switchRingBuffersAllocations(ResidencyContainer *allocationsForResidency) {
this->previousRingBuffer = this->currentRingBuffer;
GraphicsAllocation *nextAllocation = nullptr;
for (uint32_t ringBufferIndex = 0; ringBufferIndex < this->ringBuffers.size(); ringBufferIndex++) {
if (ringBufferIndex != this->currentRingBuffer && this->isCompleted(ringBufferIndex)) {
this->currentRingBuffer = ringBufferIndex;
nextAllocation = this->ringBuffers[ringBufferIndex].ringBuffer;
break;
}
}
if (nextAllocation == nullptr) {
if (this->ringBuffers.size() == this->maxRingBufferCount) {
this->currentRingBuffer = (this->currentRingBuffer + 1) % this->ringBuffers.size();
nextAllocation = this->ringBuffers[this->currentRingBuffer].ringBuffer;
} else {
bool isMultiOsContextCapable = osContext.getNumSupportedDevices() > 1u;
constexpr size_t minimumRequiredSize = 256 * MemoryConstants::kiloByte;
constexpr size_t additionalAllocationSize = MemoryConstants::pageSize;
const auto allocationSize = alignUp(minimumRequiredSize + additionalAllocationSize, MemoryConstants::pageSize64k);
const AllocationProperties commandStreamAllocationProperties{rootDeviceIndex,
true, allocationSize,
AllocationType::ringBuffer,
isMultiOsContextCapable, false, osContext.getDeviceBitfield()};
nextAllocation = memoryManager->allocateGraphicsMemoryWithProperties(commandStreamAllocationProperties);
this->currentRingBuffer = static_cast<uint32_t>(this->ringBuffers.size());
this->ringBuffers.emplace_back(0ull, 0ull, nextAllocation);
auto ret = memoryOperationHandler->makeResidentWithinOsContext(&this->osContext, ArrayRef<GraphicsAllocation *>(&nextAllocation, 1u), false, false, false) == MemoryOperationsStatus::success;
UNRECOVERABLE_IF(!ret);
this->handleResidencyContainerForUllsLightNewRingAllocation(allocationsForResidency);
}
}
UNRECOVERABLE_IF(this->currentRingBuffer == this->previousRingBuffer);
return nextAllocation;
}
template <typename GfxFamily, typename Dispatcher>
bool DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchMonitorFenceRequired(bool requireMonitorFence) {
return !this->disableMonitorFence;
}
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::deallocateResources() {
for (uint32_t ringBufferIndex = 0; ringBufferIndex < this->ringBuffers.size(); ringBufferIndex++) {
memoryManager->freeGraphicsMemory(this->ringBuffers[ringBufferIndex].ringBuffer);
}
this->ringBuffers.clear();
if (semaphores) {
memoryManager->freeGraphicsMemory(semaphores);
semaphores = nullptr;
}
memoryManager->freeGraphicsMemory(deferredTasksListAllocation);
memoryManager->freeGraphicsMemory(relaxedOrderingSchedulerAllocation);
}
template <typename GfxFamily, typename Dispatcher>
void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSystemMemoryFenceAddress() {
this->makeGlobalFenceAlwaysResident();
EncodeMemoryFence<GfxFamily>::encodeSystemMemoryFence(ringCommandStream, this->globalFenceAllocation);
}
template <typename GfxFamily, typename Dispatcher>
size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeSystemMemoryFenceAddress() {
return EncodeMemoryFence<GfxFamily>::getSystemMemoryFenceSize();
}
template <typename GfxFamily, typename Dispatcher>
uint32_t DirectSubmissionHw<GfxFamily, Dispatcher>::getDispatchErrorCode() {
return dispatchErrorCode;
}
template <typename GfxFamily, typename Dispatcher>
inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSemaphoreForPagingFence(uint64_t value) {
using COMPARE_OPERATION = typename GfxFamily::MI_SEMAPHORE_WAIT::COMPARE_OPERATION;
EncodeSemaphore<GfxFamily>::addMiSemaphoreWaitCommand(ringCommandStream,
this->gpuVaForPagingFenceSemaphore,
value,
COMPARE_OPERATION::COMPARE_OPERATION_SAD_GREATER_THAN_OR_EQUAL_SDD, false, false, false, false, nullptr);
}
template <typename GfxFamily, typename Dispatcher>
inline size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getSizeSemaphoreForPagingFence() {
return EncodeSemaphore<GfxFamily>::getSizeMiSemaphoreWait();
}
} // namespace NEO
Reference in New Issue View Git Blame Copy Permalink