compute-runtime/opencl/source/kernel/kernel.h

558 lines
22 KiB
C++

/*
* Copyright (C) 2018-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "shared/source/command_stream/csr_properties_flags.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/device/device.h"
#include "shared/source/helpers/aux_translation.h"
#include "shared/source/helpers/vec.h"
#include "shared/source/kernel/implicit_args.h"
#include "shared/source/kernel/kernel_execution_type.h"
#include "shared/source/program/kernel_info.h"
#include "shared/source/unified_memory/unified_memory.h"
#include "shared/source/utilities/logger.h"
#include "opencl/extensions/public/cl_ext_private.h"
#include "opencl/source/cl_device/cl_device.h"
#include "opencl/source/kernel/kernel_objects_for_aux_translation.h"
#include <map>
#include <vector>
namespace NEO {
class MemObj;
class TimestampPacketContainer;
class Context;
class Program;
struct ImplicitArgs;
enum class AllocationType;
struct PatchInfoData;
struct CompletionStamp;
class Buffer;
class CommandQueue;
class CommandStreamReceiver;
class GraphicsAllocation;
class ImageTransformer;
class Surface;
class PrintfHandler;
class MultiDeviceKernel;
class LocalIdsCache;
class Kernel : public ReferenceTrackedObject<Kernel> {
public:
static const uint32_t kernelBinaryAlignment = 64;
enum kernelArgType {
NONE_OBJ,
IMAGE_OBJ,
BUFFER_OBJ,
PIPE_OBJ,
SVM_OBJ,
SVM_ALLOC_OBJ,
SAMPLER_OBJ,
ACCELERATOR_OBJ,
DEVICE_QUEUE_OBJ,
SLM_OBJ
};
struct SimpleKernelArgInfo {
cl_mem_flags svmFlags;
void *object;
const void *value;
size_t size;
GraphicsAllocation *svmAllocation;
kernelArgType type;
uint32_t allocId;
uint32_t allocIdMemoryManagerCounter;
bool isPatched = false;
bool isStatelessUncacheable = false;
bool isSetToNullptr = false;
};
enum class TunningStatus {
STANDARD_TUNNING_IN_PROGRESS,
SUBDEVICE_TUNNING_IN_PROGRESS,
TUNNING_DONE
};
enum class TunningType {
DISABLED,
SIMPLE,
FULL
};
typedef int32_t (Kernel::*KernelArgHandler)(uint32_t argIndex,
size_t argSize,
const void *argVal);
template <typename kernel_t = Kernel, typename program_t = Program>
static kernel_t *create(program_t *program, const KernelInfo &kernelInfo, ClDevice &clDevice, cl_int *errcodeRet) {
cl_int retVal;
kernel_t *pKernel = nullptr;
pKernel = new kernel_t(program, kernelInfo, clDevice);
retVal = pKernel->initialize();
if (retVal != CL_SUCCESS) {
delete pKernel;
pKernel = nullptr;
}
if (pKernel) {
auto localMemSize = static_cast<uint32_t>(clDevice.getDevice().getDeviceInfo().localMemSize);
auto slmTotalSize = pKernel->getSlmTotalSize();
if (slmTotalSize > 0 && localMemSize < slmTotalSize) {
PRINT_DEBUG_STRING(NEO::DebugManager.flags.PrintDebugMessages.get(), stderr, "Size of SLM (%u) larger than available (%u)\n", slmTotalSize, localMemSize);
retVal = CL_OUT_OF_RESOURCES;
}
}
if (errcodeRet) {
*errcodeRet = retVal;
}
if (fileLoggerInstance().enabled()) {
std::string source;
program->getSource(source);
fileLoggerInstance().dumpKernel(kernelInfo.kernelDescriptor.kernelMetadata.kernelName, source);
}
return pKernel;
}
Kernel &operator=(const Kernel &) = delete;
Kernel(const Kernel &) = delete;
~Kernel() override;
static bool isMemObj(kernelArgType kernelArg) {
return kernelArg == BUFFER_OBJ || kernelArg == IMAGE_OBJ || kernelArg == PIPE_OBJ;
}
bool isAuxTranslationRequired() const { return auxTranslationRequired; }
void setAuxTranslationRequired(bool onOff) { auxTranslationRequired = onOff; }
void updateAuxTranslationRequired();
ArrayRef<uint8_t> getCrossThreadDataRef() {
return ArrayRef<uint8_t>(reinterpret_cast<uint8_t *>(crossThreadData), crossThreadDataSize);
}
char *getCrossThreadData() const {
return crossThreadData;
}
uint32_t getCrossThreadDataSize() const {
return crossThreadDataSize;
}
cl_int initialize();
MOCKABLE_VIRTUAL cl_int cloneKernel(Kernel *pSourceKernel);
MOCKABLE_VIRTUAL bool canTransformImages() const;
MOCKABLE_VIRTUAL bool isPatched() const;
// API entry points
cl_int setArgument(uint32_t argIndex, size_t argSize, const void *argVal) { return setArg(argIndex, argSize, argVal); }
cl_int setArgSvm(uint32_t argIndex, size_t svmAllocSize, void *svmPtr, GraphicsAllocation *svmAlloc, cl_mem_flags svmFlags);
MOCKABLE_VIRTUAL cl_int setArgSvmAlloc(uint32_t argIndex, void *svmPtr, GraphicsAllocation *svmAlloc, uint32_t allocId);
void setSvmKernelExecInfo(GraphicsAllocation *argValue);
void clearSvmKernelExecInfo();
cl_int getInfo(cl_kernel_info paramName, size_t paramValueSize,
void *paramValue, size_t *paramValueSizeRet) const;
cl_int getArgInfo(cl_uint argIndx, cl_kernel_arg_info paramName,
size_t paramValueSize, void *paramValue, size_t *paramValueSizeRet) const;
cl_int getWorkGroupInfo(cl_kernel_work_group_info paramName,
size_t paramValueSize, void *paramValue, size_t *paramValueSizeRet) const;
cl_int getSubGroupInfo(cl_kernel_sub_group_info paramName,
size_t inputValueSize, const void *inputValue,
size_t paramValueSize, void *paramValue,
size_t *paramValueSizeRet) const;
const void *getKernelHeap() const;
void *getSurfaceStateHeap() const;
const void *getDynamicStateHeap() const;
size_t getKernelHeapSize() const;
size_t getSurfaceStateHeapSize() const;
size_t getDynamicStateHeapSize() const;
size_t getNumberOfBindingTableStates() const;
size_t getBindingTableOffset() const {
return localBindingTableOffset;
}
void resizeSurfaceStateHeap(void *pNewSsh, size_t newSshSize, size_t newBindingTableCount, size_t newBindingTableOffset);
void substituteKernelHeap(void *newKernelHeap, size_t newKernelHeapSize);
bool isKernelHeapSubstituted() const;
uint64_t getKernelId() const;
void setKernelId(uint64_t newKernelId);
uint32_t getStartOffset() const;
void setStartOffset(uint32_t offset);
const std::vector<SimpleKernelArgInfo> &getKernelArguments() const {
return kernelArguments;
}
size_t getKernelArgsNumber() const {
return kernelArguments.size();
}
bool usesBindfulAddressingForBuffers() const {
return KernelDescriptor::BindfulAndStateless == kernelInfo.kernelDescriptor.kernelAttributes.bufferAddressingMode;
}
inline const KernelDescriptor &getDescriptor() const {
return kernelInfo.kernelDescriptor;
}
inline const KernelInfo &getKernelInfo() const {
return kernelInfo;
}
Context &getContext() const;
Program *getProgram() const { return program; }
uint32_t getScratchSize() {
return kernelInfo.kernelDescriptor.kernelAttributes.perThreadScratchSize[0];
}
uint32_t getPrivateScratchSize() {
return kernelInfo.kernelDescriptor.kernelAttributes.perThreadScratchSize[1];
}
bool usesSyncBuffer() const;
void patchSyncBuffer(GraphicsAllocation *gfxAllocation, size_t bufferOffset);
void *patchBindlessSurfaceState(NEO::GraphicsAllocation *alloc, uint32_t bindless);
// Helpers
cl_int setArg(uint32_t argIndex, uint32_t argValue);
cl_int setArg(uint32_t argIndex, uint64_t argValue);
cl_int setArg(uint32_t argIndex, cl_mem argValue);
cl_int setArg(uint32_t argIndex, cl_mem argValue, uint32_t mipLevel);
cl_int setArg(uint32_t argIndex, size_t argSize, const void *argVal);
// Handlers
void setKernelArgHandler(uint32_t argIndex, KernelArgHandler handler);
void unsetArg(uint32_t argIndex);
cl_int setArgImmediate(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgBuffer(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgPipe(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgImage(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgImageWithMipLevel(uint32_t argIndex,
size_t argSize,
const void *argVal, uint32_t mipLevel);
cl_int setArgLocal(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgSampler(uint32_t argIndex,
size_t argSize,
const void *argVal);
cl_int setArgAccelerator(uint32_t argIndex,
size_t argSize,
const void *argVal);
void storeKernelArg(uint32_t argIndex,
kernelArgType argType,
void *argObject,
const void *argValue,
size_t argSize,
GraphicsAllocation *argSvmAlloc = nullptr,
cl_mem_flags argSvmFlags = 0);
void storeKernelArgAllocIdMemoryManagerCounter(uint32_t argIndex, uint32_t allocIdMemoryManagerCounter);
const void *getKernelArg(uint32_t argIndex) const;
const SimpleKernelArgInfo &getKernelArgInfo(uint32_t argIndex) const;
bool getAllowNonUniform() const;
bool isVmeKernel() const { return kernelInfo.kernelDescriptor.kernelAttributes.flags.usesVme; }
bool requiresSystolicPipelineSelectMode() const { return systolicPipelineSelectMode; }
void performKernelTuning(CommandStreamReceiver &commandStreamReceiver, const Vec3<size_t> &lws, const Vec3<size_t> &gws, const Vec3<size_t> &offsets, TimestampPacketContainer *timestampContainer);
MOCKABLE_VIRTUAL bool isSingleSubdevicePreferred() const;
void setInlineSamplers();
// residency for kernel surfaces
MOCKABLE_VIRTUAL void makeResident(CommandStreamReceiver &commandStreamReceiver);
MOCKABLE_VIRTUAL void getResidency(std::vector<Surface *> &dst);
bool requiresCoherency();
void resetSharedObjectsPatchAddresses();
bool isUsingSharedObjArgs() const { return usingSharedObjArgs; }
bool hasUncacheableStatelessArgs() const { return statelessUncacheableArgsCount > 0; }
bool hasPrintfOutput() const;
cl_int checkCorrectImageAccessQualifier(cl_uint argIndex,
size_t argSize,
const void *argValue) const;
static uint32_t dummyPatchLocation;
uint32_t allBufferArgsStateful = CL_TRUE;
bool isBuiltIn = false;
KernelExecutionType getExecutionType() const {
return executionType;
}
bool is32Bit() const {
return kernelInfo.kernelDescriptor.kernelAttributes.gpuPointerSize == 4;
}
size_t getPerThreadSystemThreadSurfaceSize() const {
return kernelInfo.kernelDescriptor.kernelAttributes.perThreadSystemThreadSurfaceSize;
}
std::vector<PatchInfoData> &getPatchInfoDataList() { return patchInfoDataList; };
bool usesImages() const {
return usingImages;
}
bool usesOnlyImages() const {
return usingImagesOnly;
}
std::unique_ptr<KernelObjsForAuxTranslation> fillWithKernelObjsForAuxTranslation();
MOCKABLE_VIRTUAL bool requiresCacheFlushCommand(const CommandQueue &commandQueue) const;
using CacheFlushAllocationsVec = StackVec<GraphicsAllocation *, 32>;
void getAllocationsForCacheFlush(CacheFlushAllocationsVec &out) const;
void setAuxTranslationDirection(AuxTranslationDirection auxTranslationDirection) {
this->auxTranslationDirection = auxTranslationDirection;
}
void setUnifiedMemorySyncRequirement(bool isUnifiedMemorySyncRequired) {
this->isUnifiedMemorySyncRequired = isUnifiedMemorySyncRequired;
}
void setUnifiedMemoryProperty(cl_kernel_exec_info infoType, bool infoValue);
void setUnifiedMemoryExecInfo(GraphicsAllocation *argValue);
void clearUnifiedMemoryExecInfo();
bool areStatelessWritesUsed() { return containsStatelessWrites; }
int setKernelThreadArbitrationPolicy(uint32_t propertyValue);
cl_int setKernelExecutionType(cl_execution_info_kernel_type_intel executionType);
void getSuggestedLocalWorkSize(const cl_uint workDim, const size_t *globalWorkSize, const size_t *globalWorkOffset,
size_t *localWorkSize);
uint32_t getMaxWorkGroupCount(const cl_uint workDim, const size_t *localWorkSize, const CommandQueue *commandQueue) const;
uint64_t getKernelStartAddress(const bool localIdsGenerationByRuntime, const bool kernelUsesLocalIds, const bool isCssUsed, const bool returnFullAddress) const;
bool isKernelDebugEnabled() const { return debugEnabled; }
void setAdditionalKernelExecInfo(uint32_t additionalKernelExecInfo);
uint32_t getAdditionalKernelExecInfo() const;
MOCKABLE_VIRTUAL bool requiresWaDisableRccRhwoOptimization() const;
// dispatch traits
void setGlobalWorkOffsetValues(uint32_t globalWorkOffsetX, uint32_t globalWorkOffsetY, uint32_t globalWorkOffsetZ);
void setGlobalWorkSizeValues(uint32_t globalWorkSizeX, uint32_t globalWorkSizeY, uint32_t globalWorkSizeZ);
void setLocalWorkSizeValues(uint32_t localWorkSizeX, uint32_t localWorkSizeY, uint32_t localWorkSizeZ);
void setLocalWorkSize2Values(uint32_t localWorkSizeX, uint32_t localWorkSizeY, uint32_t localWorkSizeZ);
void setEnqueuedLocalWorkSizeValues(uint32_t localWorkSizeX, uint32_t localWorkSizeY, uint32_t localWorkSizeZ);
void setNumWorkGroupsValues(uint32_t numWorkGroupsX, uint32_t numWorkGroupsY, uint32_t numWorkGroupsZ);
void setWorkDim(uint32_t workDim);
const uint32_t *getDispatchTrait(const CrossThreadDataOffset offset) const {
return isValidOffset(offset) ? reinterpret_cast<uint32_t *>(getCrossThreadData() + offset)
: &Kernel::dummyPatchLocation;
}
const uint32_t *getWorkDim() const { return getDispatchTrait(getDescriptor().payloadMappings.dispatchTraits.workDim); }
std::array<const uint32_t *, 3> getDispatchTraitArray(const CrossThreadDataOffset dispatchTrait[3]) const { return {getDispatchTrait(dispatchTrait[0]), getDispatchTrait(dispatchTrait[1]), getDispatchTrait(dispatchTrait[2])}; }
std::array<const uint32_t *, 3> getGlobalWorkOffsetValues() const { return getDispatchTraitArray(getDescriptor().payloadMappings.dispatchTraits.globalWorkOffset); }
std::array<const uint32_t *, 3> getLocalWorkSizeValues() const { return getDispatchTraitArray(getDescriptor().payloadMappings.dispatchTraits.localWorkSize); }
std::array<const uint32_t *, 3> getLocalWorkSize2Values() const { return getDispatchTraitArray(getDescriptor().payloadMappings.dispatchTraits.localWorkSize2); }
std::array<const uint32_t *, 3> getEnqueuedLocalWorkSizeValues() const { return getDispatchTraitArray(getDescriptor().payloadMappings.dispatchTraits.enqueuedLocalWorkSize); }
std::array<const uint32_t *, 3> getNumWorkGroupsValues() const { return getDispatchTraitArray(getDescriptor().payloadMappings.dispatchTraits.numWorkGroups); }
bool isLocalWorkSize2Patchable();
uint32_t getMaxKernelWorkGroupSize() const;
uint32_t getSlmTotalSize() const;
bool getHasIndirectAccess() const {
return this->kernelHasIndirectAccess;
}
MultiDeviceKernel *getMultiDeviceKernel() const { return pMultiDeviceKernel; }
void setMultiDeviceKernel(MultiDeviceKernel *pMultiDeviceKernelToSet) { pMultiDeviceKernel = pMultiDeviceKernelToSet; }
bool areMultipleSubDevicesInContext() const;
bool requiresMemoryMigration() const { return migratableArgsMap.size() > 0; }
const std::map<uint32_t, MemObj *> &getMemObjectsToMigrate() const { return migratableArgsMap; }
ImplicitArgs *getImplicitArgs() const { return pImplicitArgs.get(); }
const HardwareInfo &getHardwareInfo() const;
bool isAnyKernelArgumentUsingSystemMemory() const {
return anyKernelArgumentUsingSystemMemory;
}
static bool graphicsAllocationTypeUseSystemMemory(AllocationType type);
void setDestinationAllocationInSystemMemory(bool value) {
isDestinationAllocationInSystemMemory = value;
}
bool getDestinationAllocationInSystemMemory() const {
return isDestinationAllocationInSystemMemory;
}
void setLocalIdsForGroup(const Vec3<uint16_t> &groupSize, void *destination) const;
size_t getLocalIdsSizeForGroup(const Vec3<uint16_t> &groupSize) const;
size_t getLocalIdsSizePerThread() const;
protected:
struct KernelConfig {
Vec3<size_t> gws;
Vec3<size_t> lws;
Vec3<size_t> offsets;
bool operator==(const KernelConfig &other) const { return this->gws == other.gws && this->lws == other.lws && this->offsets == other.offsets; }
};
struct KernelConfigHash {
size_t operator()(KernelConfig const &config) const {
auto hash = std::hash<size_t>{};
size_t gwsHashX = hash(config.gws.x);
size_t gwsHashY = hash(config.gws.y);
size_t gwsHashZ = hash(config.gws.z);
size_t gwsHash = hashCombine(gwsHashX, gwsHashY, gwsHashZ);
size_t lwsHashX = hash(config.lws.x);
size_t lwsHashY = hash(config.lws.y);
size_t lwsHashZ = hash(config.lws.z);
size_t lwsHash = hashCombine(lwsHashX, lwsHashY, lwsHashZ);
size_t offsetsHashX = hash(config.offsets.x);
size_t offsetsHashY = hash(config.offsets.y);
size_t offsetsHashZ = hash(config.offsets.z);
size_t offsetsHash = hashCombine(offsetsHashX, offsetsHashY, offsetsHashZ);
return hashCombine(gwsHash, lwsHash, offsetsHash);
}
size_t hashCombine(size_t hash1, size_t hash2, size_t hash3) const {
return (hash1 ^ (hash2 << 1u)) ^ (hash3 << 2u);
}
};
struct KernelSubmissionData {
std::unique_ptr<TimestampPacketContainer> kernelStandardTimestamps;
std::unique_ptr<TimestampPacketContainer> kernelSubdeviceTimestamps;
TunningStatus status;
bool singleSubdevicePreferred = false;
};
Kernel(Program *programArg, const KernelInfo &kernelInfo, ClDevice &clDevice);
void makeArgsResident(CommandStreamReceiver &commandStreamReceiver);
void *patchBufferOffset(const ArgDescPointer &argAsPtr, void *svmPtr, GraphicsAllocation *svmAlloc);
void patchWithImplicitSurface(void *ptrToPatchInCrossThreadData, GraphicsAllocation &allocation, const ArgDescPointer &arg);
void provideInitializationHints();
void markArgPatchedAndResolveArgs(uint32_t argIndex);
void resolveArgs();
void reconfigureKernel();
bool hasDirectStatelessAccessToSharedBuffer() const;
bool hasDirectStatelessAccessToHostMemory() const;
bool hasIndirectStatelessAccessToHostMemory() const;
void addAllocationToCacheFlushVector(uint32_t argIndex, GraphicsAllocation *argAllocation);
bool allocationForCacheFlush(GraphicsAllocation *argAllocation) const;
const ClDevice &getDevice() const {
return clDevice;
}
cl_int patchPrivateSurface();
bool hasTunningFinished(KernelSubmissionData &submissionData);
bool hasRunFinished(TimestampPacketContainer *timestampContainer);
void initializeLocalIdsCache();
std::unique_ptr<LocalIdsCache> localIdsCache;
UnifiedMemoryControls unifiedMemoryControls{};
std::map<uint32_t, MemObj *> migratableArgsMap{};
std::unordered_map<KernelConfig, KernelSubmissionData, KernelConfigHash> kernelSubmissionMap;
std::vector<SimpleKernelArgInfo> kernelArguments;
std::vector<KernelArgHandler> kernelArgHandlers;
std::vector<GraphicsAllocation *> kernelSvmGfxAllocations;
std::vector<GraphicsAllocation *> kernelUnifiedMemoryGfxAllocations;
std::vector<PatchInfoData> patchInfoDataList;
std::vector<GraphicsAllocation *> kernelArgRequiresCacheFlush;
std::vector<size_t> slmSizes;
std::unique_ptr<ImageTransformer> imageTransformer;
std::unique_ptr<char[]> pSshLocal;
std::unique_ptr<ImplicitArgs> pImplicitArgs = nullptr;
uint64_t privateSurfaceSize = 0u;
size_t numberOfBindingTableStates = 0u;
size_t localBindingTableOffset = 0u;
const ExecutionEnvironment &executionEnvironment;
Program *program;
ClDevice &clDevice;
const KernelInfo &kernelInfo;
GraphicsAllocation *privateSurface = nullptr;
MultiDeviceKernel *pMultiDeviceKernel = nullptr;
uint32_t *maxWorkGroupSizeForCrossThreadData = &Kernel::dummyPatchLocation;
uint32_t *dataParameterSimdSize = &Kernel::dummyPatchLocation;
uint32_t *parentEventOffset = &Kernel::dummyPatchLocation;
uint32_t *preferredWkgMultipleOffset = &Kernel::dummyPatchLocation;
char *crossThreadData = nullptr;
AuxTranslationDirection auxTranslationDirection = AuxTranslationDirection::None;
KernelExecutionType executionType = KernelExecutionType::Default;
uint32_t patchedArgumentsNum = 0;
uint32_t startOffset = 0;
uint32_t statelessUncacheableArgsCount = 0;
uint32_t additionalKernelExecInfo = AdditionalKernelExecInfo::DisableOverdispatch;
uint32_t maxKernelWorkGroupSize = 0;
uint32_t slmTotalSize = 0u;
uint32_t sshLocalSize = 0u;
uint32_t crossThreadDataSize = 0u;
bool containsStatelessWrites = true;
bool usingSharedObjArgs = false;
bool usingImages = false;
bool usingImagesOnly = false;
bool auxTranslationRequired = false;
bool systolicPipelineSelectMode = false;
bool svmAllocationsRequireCacheFlush = false;
bool isUnifiedMemorySyncRequired = true;
bool debugEnabled = false;
bool singleSubdevicePreferredInCurrentEnqueue = false;
bool kernelHasIndirectAccess = true;
bool anyKernelArgumentUsingSystemMemory = false;
bool isDestinationAllocationInSystemMemory = false;
};
} // namespace NEO