compute-runtime/shared/source/helpers/kernel_helpers.cpp

152 lines
7.3 KiB
C++

/*
* Copyright (C) 2019-2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/helpers/kernel_helpers.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/device/device.h"
#include "shared/source/execution_environment/root_device_environment.h"
#include "shared/source/helpers/basic_math.h"
#include "shared/source/helpers/debug_helpers.h"
#include "shared/source/helpers/gfx_core_helper.h"
#include "shared/source/helpers/hw_info.h"
#include "shared/source/program/sync_buffer_handler.h"
namespace NEO {
uint32_t KernelHelper::getMaxWorkGroupCount(Device &device, uint16_t numGrfRequired, uint8_t simdSize, uint8_t barrierCount, uint32_t alignedSlmSize, uint32_t workDim, const size_t *localWorkSize,
EngineGroupType engineGroupType, bool implicitScalingEnabled, bool forceSingleTileQuery) {
uint32_t numSubDevicesForExecution = 1;
auto deviceBitfield = device.getDeviceBitfield();
if (!forceSingleTileQuery && implicitScalingEnabled) {
numSubDevicesForExecution = static_cast<uint32_t>(deviceBitfield.count());
}
return KernelHelper::getMaxWorkGroupCount(device.getRootDeviceEnvironment(), numGrfRequired, simdSize, barrierCount, numSubDevicesForExecution, alignedSlmSize, workDim, localWorkSize, engineGroupType);
}
uint32_t KernelHelper::getMaxWorkGroupCount(const RootDeviceEnvironment &rootDeviceEnvironment, uint16_t numGrfRequired, uint8_t simdSize, uint8_t barrierCount,
uint32_t numSubDevices, uint32_t usedSlmSize, uint32_t workDim, const size_t *localWorkSize, EngineGroupType engineGroupType) {
if (debugManager.flags.OverrideMaxWorkGroupCount.get() != -1) {
return static_cast<uint32_t>(debugManager.flags.OverrideMaxWorkGroupCount.get());
}
auto &helper = rootDeviceEnvironment.getHelper<NEO::GfxCoreHelper>();
auto &hwInfo = *rootDeviceEnvironment.getHardwareInfo();
auto dssCount = hwInfo.gtSystemInfo.DualSubSliceCount;
if (dssCount == 0) {
dssCount = hwInfo.gtSystemInfo.SubSliceCount;
}
auto availableThreadCount = helper.calculateAvailableThreadCount(hwInfo, numGrfRequired);
auto availableSlmSize = static_cast<uint32_t>(dssCount * MemoryConstants::kiloByte * hwInfo.capabilityTable.maxProgrammableSlmSize);
auto maxBarrierCount = static_cast<uint32_t>(helper.getMaxBarrierRegisterPerSlice());
UNRECOVERABLE_IF((workDim == 0) || (workDim > 3));
UNRECOVERABLE_IF(localWorkSize == nullptr);
size_t workGroupSize = localWorkSize[0];
for (uint32_t i = 1; i < workDim; i++) {
workGroupSize *= localWorkSize[i];
}
UNRECOVERABLE_IF(workGroupSize == 0);
auto numThreadsPerThreadGroup = static_cast<uint32_t>(Math::divideAndRoundUp(workGroupSize, simdSize));
auto maxWorkGroupsCount = availableThreadCount / numThreadsPerThreadGroup;
if (barrierCount > 0 || usedSlmSize > 0) {
helper.alignThreadGroupCountToDssSize(maxWorkGroupsCount, dssCount, availableThreadCount / dssCount, numThreadsPerThreadGroup);
if (barrierCount > 0) {
auto maxWorkGroupsCountDueToBarrierUsage = dssCount * (maxBarrierCount / barrierCount);
maxWorkGroupsCount = std::min(maxWorkGroupsCount, maxWorkGroupsCountDueToBarrierUsage);
}
if (usedSlmSize > 0) {
auto maxWorkGroupsCountDueToSlm = availableSlmSize / usedSlmSize;
maxWorkGroupsCount = std::min(maxWorkGroupsCount, maxWorkGroupsCountDueToSlm);
}
}
maxWorkGroupsCount = helper.adjustMaxWorkGroupCount(maxWorkGroupsCount, engineGroupType, rootDeviceEnvironment);
if (!helper.singleTileExecImplicitScalingRequired(true)) {
maxWorkGroupsCount *= numSubDevices;
}
return maxWorkGroupsCount;
}
KernelHelper::ErrorCode KernelHelper::checkIfThereIsSpaceForScratchOrPrivate(KernelDescriptor::KernelAttributes attributes, Device *device) {
auto &gfxCoreHelper = device->getRootDeviceEnvironment().getHelper<NEO::GfxCoreHelper>();
auto &productHelper = device->getRootDeviceEnvironment().getHelper<NEO::ProductHelper>();
uint32_t maxScratchSize = gfxCoreHelper.getMaxScratchSize(productHelper);
if ((attributes.perThreadScratchSize[0] > maxScratchSize) || (attributes.perThreadScratchSize[1] > maxScratchSize)) {
return KernelHelper::ErrorCode::invalidKernel;
}
auto globalMemorySize = device->getDeviceInfo().globalMemSize;
auto computeUnitsForScratch = device->getDeviceInfo().computeUnitsUsedForScratch;
auto totalPrivateMemorySize = KernelHelper::getPrivateSurfaceSize(attributes.perHwThreadPrivateMemorySize, computeUnitsForScratch);
auto totalScratchSize = KernelHelper::getScratchSize(attributes.perThreadScratchSize[0], computeUnitsForScratch);
auto totalPrivateScratchSize = KernelHelper::getPrivateScratchSize(attributes.perThreadScratchSize[1], computeUnitsForScratch);
PRINT_DEBUG_STRING(debugManager.flags.PrintDebugMessages.get(), stderr,
"computeUnits for each thread: %u\n", computeUnitsForScratch);
PRINT_DEBUG_STRING(debugManager.flags.PrintDebugMessages.get(), stderr,
"global memory size: %llu\n", globalMemorySize);
PRINT_DEBUG_STRING(debugManager.flags.PrintDebugMessages.get(), stderr,
"perHwThreadPrivateMemorySize: %u\t totalPrivateMemorySize: %lu\n",
attributes.perHwThreadPrivateMemorySize, totalPrivateMemorySize);
PRINT_DEBUG_STRING(debugManager.flags.PrintDebugMessages.get(), stderr,
"perHwThreadScratchSize: %u\t totalScratchSize: %lu\n",
attributes.perThreadScratchSize[0], totalScratchSize);
PRINT_DEBUG_STRING(debugManager.flags.PrintDebugMessages.get(), stderr,
"perHwThreadPrivateScratchSize: %u\t totalPrivateScratchSize: %lu\n",
attributes.perThreadScratchSize[1], totalPrivateScratchSize);
if (totalPrivateMemorySize > globalMemorySize ||
totalScratchSize > globalMemorySize ||
totalPrivateScratchSize > globalMemorySize) {
return KernelHelper::ErrorCode::outOfDeviceMemory;
}
return KernelHelper::ErrorCode::success;
}
bool KernelHelper::isAnyArgumentPtrByValue(const KernelDescriptor &kernelDescriptor) {
for (auto &argDescriptor : kernelDescriptor.payloadMappings.explicitArgs) {
if (argDescriptor.type == NEO::ArgDescriptor::argTValue) {
for (auto &element : argDescriptor.as<NEO::ArgDescValue>().elements) {
if (element.isPtr) {
return true;
}
}
}
}
return false;
}
std::pair<GraphicsAllocation *, size_t> KernelHelper::getRegionGroupBarrierAllocationOffset(Device &device, const size_t threadGroupCount, const size_t localRegionSize) {
device.allocateSyncBufferHandler();
size_t size = KernelHelper::getRegionGroupBarrierSize(threadGroupCount, localRegionSize);
return device.syncBufferHandler->obtainAllocationAndOffset(size);
}
std::pair<GraphicsAllocation *, size_t> KernelHelper::getSyncBufferAllocationOffset(Device &device, const size_t requestedNumberOfWorkgroups) {
device.allocateSyncBufferHandler();
size_t requiredSize = KernelHelper::getSyncBufferSize(requestedNumberOfWorkgroups);
return device.syncBufferHandler->obtainAllocationAndOffset(requiredSize);
}
} // namespace NEO