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/runtime/built_ins/built_ins.cpp

856 lines
39 KiB
C++
Raw Blame History

/*
* Copyright (c) 2017 - 2018, Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included
* in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/
#include <cstdint>
#include "runtime/built_ins/built_ins.h"
#include "runtime/built_ins/vme_dispatch_builder.h"
#include "runtime/built_ins/sip.h"
#include "runtime/compiler_interface/compiler_interface.h"
#include "runtime/program/program.h"
#include "runtime/mem_obj/image.h"
#include "runtime/kernel/kernel.h"
#include "runtime/helpers/basic_math.h"
#include "runtime/helpers/convert_color.h"
#include "runtime/helpers/dispatch_info_builder.h"
#include "runtime/helpers/debug_helpers.h"
#include <sstream>
namespace OCLRT {
BuiltIns *BuiltIns::pInstance = nullptr;
const char *mediaKernelsBuildOptions = {
"-D cl_intel_device_side_advanced_vme_enable "
"-D cl_intel_device_side_avc_vme_enable "
"-D cl_intel_device_side_vme_enable "
"-D cl_intel_media_block_io "
"-cl-fast-relaxed-math "};
BuiltIns::BuiltIns() {
builtinsLib.reset(new BuiltinsLib());
}
BuiltIns::~BuiltIns() {
delete static_cast<SchedulerKernel *>(schedulerBuiltIn.pKernel);
delete schedulerBuiltIn.pProgram;
schedulerBuiltIn.pKernel = nullptr;
schedulerBuiltIn.pProgram = nullptr;
}
BuiltIns &BuiltIns::getInstance() {
static std::mutex initMutex;
std::lock_guard<std::mutex> autolock(initMutex);
if (pInstance == nullptr) {
pInstance = new BuiltIns();
}
return *pInstance;
}
void BuiltIns::shutDown() {
if (pInstance) {
auto inst = pInstance;
pInstance = nullptr;
delete inst;
}
}
SchedulerKernel &BuiltIns::getSchedulerKernel(Context &context) {
if (schedulerBuiltIn.pKernel) {
return *static_cast<SchedulerKernel *>(schedulerBuiltIn.pKernel);
}
auto initializeSchedulerProgramAndKernel = [&] {
cl_int retVal = CL_SUCCESS;
auto src = getInstance().builtinsLib->getBuiltinCode(EBuiltInOps::Scheduler, BuiltinCode::ECodeType::Any, *context.getDevice(0));
auto program = Program::createFromGenBinary(&context,
src.resource.data(),
src.resource.size(),
true,
&retVal);
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
DEBUG_BREAK_IF(!program);
retVal = program->processGenBinary();
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
schedulerBuiltIn.pProgram = program;
auto kernelInfo = schedulerBuiltIn.pProgram->getKernelInfo(SchedulerKernel::schedulerName);
DEBUG_BREAK_IF(!kernelInfo);
schedulerBuiltIn.pKernel = Kernel::create<SchedulerKernel>(
schedulerBuiltIn.pProgram,
*kernelInfo,
&retVal);
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
};
std::call_once(schedulerBuiltIn.programIsInitialized, initializeSchedulerProgramAndKernel);
return *static_cast<SchedulerKernel *>(schedulerBuiltIn.pKernel);
}
const SipKernel &BuiltIns::getSipKernel(SipKernelType type, const Device &device) {
uint32_t kernelId = static_cast<uint32_t>(type);
UNRECOVERABLE_IF(kernelId >= static_cast<uint32_t>(SipKernelType::COUNT));
auto &sipBuiltIn = this->sipKernels[kernelId];
auto initializer = [&] {
cl_int retVal = CL_SUCCESS;
std::vector<char> sipBinary;
auto compilerInteface = CompilerInterface::getInstance();
UNRECOVERABLE_IF(compilerInteface == nullptr);
auto ret = compilerInteface->getSipKernelBinary(type, device, sipBinary);
UNRECOVERABLE_IF(ret != CL_SUCCESS);
UNRECOVERABLE_IF(sipBinary.size() == 0);
auto program = Program::createFromGenBinary(nullptr,
sipBinary.data(),
sipBinary.size(),
true,
&retVal);
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
UNRECOVERABLE_IF(program == nullptr);
retVal = program->processGenBinary();
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
auto kernelInfo = program->getKernelInfo(size_t{0});
UNRECOVERABLE_IF(kernelInfo == nullptr);
uint32_t sipOffset = kernelInfo->systemKernelOffset;
UNRECOVERABLE_IF(sipOffset >= kernelInfo->heapInfo.pKernelHeader->KernelHeapSize)
sipBuiltIn.first.reset(new SipKernel(type, ptrOffset(kernelInfo->heapInfo.pKernelHeap, sipOffset),
kernelInfo->heapInfo.pKernelHeader->KernelHeapSize - sipOffset));
DEBUG_BREAK_IF(retVal != CL_SUCCESS);
program->release();
};
std::call_once(sipBuiltIn.second, initializer);
UNRECOVERABLE_IF(sipBuiltIn.first == nullptr);
return *sipBuiltIn.first;
}
// VME:
static const char *blockMotionEstimateIntelSrc = {
#include "kernels/vme_block_motion_estimate_intel_frontend.igdrcl_built_in"
};
static const char *blockAdvancedMotionEstimateCheckIntelSrc = {
#include "kernels/vme_block_advanced_motion_estimate_check_intel_frontend.igdrcl_built_in"
};
static const char *blockAdvancedMotionEstimateBidirectionalCheckIntelSrc = {
#include "kernels/vme_block_advanced_motion_estimate_bidirectional_check_intel_frontend.igdrcl_built_in"
};
// VEBOX:
static const char *veEnhanceIntelSrc = {
#include "kernels/vebox_ve_enhance_intel.igdrcl_built_in"
};
static const char *veDnEnhanceIntelSrc = {
#include "kernels/vebox_ve_dn_enhance_intel.igdrcl_built_in"
};
static const char *veDnDiEnhanceIntelSrc = {
#include "kernels/vebox_ve_dn_di_enhance_intel.igdrcl_built_in"
};
static const std::tuple<const char *, const char *> mediaBuiltIns[] = {
std::make_tuple("block_motion_estimate_intel", blockMotionEstimateIntelSrc),
std::make_tuple("block_advanced_motion_estimate_check_intel", blockAdvancedMotionEstimateCheckIntelSrc),
std::make_tuple("block_advanced_motion_estimate_bidirectional_check_intel", blockAdvancedMotionEstimateBidirectionalCheckIntelSrc),
std::make_tuple("ve_enhance_intel", veEnhanceIntelSrc),
std::make_tuple("ve_dn_enhance_intel", veDnEnhanceIntelSrc),
std::make_tuple("ve_dn_di_enhance_intel", veDnDiEnhanceIntelSrc),
};
// Unlike other built-ins media kernels are not stored in BuiltIns object.
// Pointer to program with built in kernels is returned to the user through API
// call and user is responsible for releasing it by calling clReleaseProgram.
Program *BuiltIns::createBuiltInProgram(
Context &context,
Device &device,
const char *kernelNames,
int &errcodeRet) {
std::string programSourceStr = "";
std::istringstream ss(kernelNames);
std::string currentKernelName;
while (std::getline(ss, currentKernelName, ';')) {
bool found = false;
for (auto &builtInTuple : mediaBuiltIns) {
if (currentKernelName == std::get<0>(builtInTuple)) {
programSourceStr += std::get<1>(builtInTuple);
found = true;
break;
}
}
if (!found) {
errcodeRet = CL_INVALID_VALUE;
return nullptr;
}
}
if (programSourceStr.empty() == true) {
errcodeRet = CL_INVALID_VALUE;
return nullptr;
}
Program *pBuiltInProgram = nullptr;
pBuiltInProgram = Program::create(programSourceStr.c_str(), &context, device, true, nullptr);
if (pBuiltInProgram) {
std::unordered_map<std::string, BuiltinDispatchInfoBuilder *> builtinsBuilders;
builtinsBuilders["block_motion_estimate_intel"] =
&BuiltIns::getInstance().getBuiltinDispatchInfoBuilder(EBuiltInOps::VmeBlockMotionEstimateIntel, context, device);
builtinsBuilders["block_advanced_motion_estimate_check_intel"] =
&BuiltIns::getInstance().getBuiltinDispatchInfoBuilder(EBuiltInOps::VmeBlockAdvancedMotionEstimateCheckIntel, context, device);
builtinsBuilders["block_advanced_motion_estimate_bidirectional_check_intel"] =
&BuiltIns::getInstance().getBuiltinDispatchInfoBuilder(EBuiltInOps::VmeBlockAdvancedMotionEstimateBidirectionalCheckIntel, context, device);
const cl_device_id clDevice = &device;
errcodeRet = pBuiltInProgram->build(
clDevice,
mediaKernelsBuildOptions,
enableCacheing,
builtinsBuilders);
} else {
errcodeRet = CL_INVALID_VALUE;
}
return pBuiltInProgram;
}
void BuiltinDispatchInfoBuilder::takeOwnership(Context *context) {
for (auto &k : usedKernels) {
k->takeOwnership(true);
k->setContext(context);
}
}
void BuiltinDispatchInfoBuilder::releaseOwnership() {
for (auto &k : usedKernels) {
k->setContext(nullptr);
k->releaseOwnership();
}
}
template <typename... KernelsDescArgsT>
void BuiltinDispatchInfoBuilder::populate(Context &context, Device &device, EBuiltInOps op, const char *options, KernelsDescArgsT &&... desc) {
auto src = kernelsLib.getBuiltinsLib().getBuiltinCode(op, BuiltinCode::ECodeType::Any, device);
prog.reset(BuiltinsLib::createProgramFromCode(src, context, device).release());
prog->build(0, nullptr, options, nullptr, nullptr, kernelsLib.isCacheingEnabled());
grabKernels(std::forward<KernelsDescArgsT>(desc)...);
}
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::CopyBufferToBuffer> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernLeftLeftover(nullptr), kernMiddle(nullptr), kernRightLeftover(nullptr) {
populate(context, device,
EBuiltInOps::CopyBufferToBuffer,
"",
"CopyBufferToBufferLeftLeftover", kernLeftLeftover,
"CopyBufferToBufferMiddle", kernMiddle,
"CopyBufferToBufferRightLeftover", kernRightLeftover);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d1D, SplitDispatch::SplitMode::KernelSplit> kernelSplit1DBuilder;
uintptr_t start = reinterpret_cast<uintptr_t>(operationParams.dstPtr) + operationParams.dstOffset.x;
size_t middleAlignment = MemoryConstants::cacheLineSize;
size_t middleElSize = sizeof(uint32_t) * 4;
uintptr_t leftSize = start % middleAlignment;
leftSize = (leftSize > 0) ? (middleAlignment - leftSize) : 0; // calc left leftover size
leftSize = std::min(leftSize, operationParams.size.x); // clamp left leftover size to requested size
uintptr_t rightSize = (start + operationParams.size.x) % middleAlignment; // calc right leftover size
rightSize = std::min(rightSize, operationParams.size.x - leftSize); // clamp
uintptr_t middleSizeBytes = operationParams.size.x - leftSize - rightSize; // calc middle size
if (!isAligned<4>(reinterpret_cast<uintptr_t>(operationParams.srcPtr) + operationParams.srcOffset.x + leftSize)) {
//corner case - src relative to dst does not have DWORD alignment
leftSize += middleSizeBytes;
middleSizeBytes = 0;
}
auto middleSizeEls = middleSizeBytes / middleElSize; // num work items in middle walker
// Set-up ISA
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Left, kernLeftLeftover);
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Middle, kernMiddle);
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Right, kernRightLeftover);
// Set-up common kernel args
if (operationParams.srcSvmAlloc) {
kernelSplit1DBuilder.setArgSvmAlloc(0, operationParams.srcPtr, operationParams.srcSvmAlloc);
} else if (operationParams.srcMemObj) {
kernelSplit1DBuilder.setArg(0, operationParams.srcMemObj);
} else {
kernelSplit1DBuilder.setArgSvm(0, operationParams.size.x, operationParams.srcPtr, nullptr, CL_MEM_READ_ONLY);
}
if (operationParams.dstSvmAlloc) {
kernelSplit1DBuilder.setArgSvmAlloc(1, operationParams.dstPtr, operationParams.dstSvmAlloc);
} else if (operationParams.dstMemObj) {
kernelSplit1DBuilder.setArg(1, operationParams.dstMemObj);
} else {
kernelSplit1DBuilder.setArgSvm(1, operationParams.size.x, operationParams.dstPtr);
}
// Set-up srcOffset
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Left, 2, static_cast<uint32_t>(operationParams.srcOffset.x));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Middle, 2, static_cast<uint32_t>(operationParams.srcOffset.x + leftSize));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Right, 2, static_cast<uint32_t>(operationParams.srcOffset.x + leftSize + middleSizeBytes));
// Set-up dstOffset
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Left, 3, static_cast<uint32_t>(operationParams.dstOffset.x));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Middle, 3, static_cast<uint32_t>(operationParams.dstOffset.x + leftSize));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Right, 3, static_cast<uint32_t>(operationParams.dstOffset.x + leftSize + middleSizeBytes));
// Set-up work sizes
// Note for split walker, it would be just builder.SetDipatchGeometry(GWS, ELWS, OFFSET)
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Left, Vec3<size_t>{leftSize, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Middle, Vec3<size_t>{middleSizeEls, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Right, Vec3<size_t>{rightSize, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernLeftLeftover;
Kernel *kernMiddle;
Kernel *kernRightLeftover;
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::CopyBufferRect> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernelBytes{nullptr} {
populate(context, device,
EBuiltInOps::CopyBufferRect,
"",
"CopyBufferRectBytes2d", kernelBytes[0],
"CopyBufferRectBytes2d", kernelBytes[1],
"CopyBufferRectBytes3d", kernelBytes[2]);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d3D, SplitDispatch::SplitMode::NoSplit> kernelNoSplit3DBuilder;
size_t hostPtrSize = 0;
bool is3D = false;
if (operationParams.srcMemObj && operationParams.dstMemObj) {
DEBUG_BREAK_IF(!((operationParams.srcPtr == nullptr) && (operationParams.dstPtr == nullptr)));
is3D = (operationParams.size.z > 1) || (operationParams.srcOffset.z > 0) || (operationParams.dstOffset.z > 0);
} else {
if (operationParams.srcPtr) {
size_t origin[] = {operationParams.srcOffset.x, operationParams.srcOffset.y, operationParams.srcOffset.z};
size_t region[] = {operationParams.size.x, operationParams.size.y, operationParams.size.z};
hostPtrSize = Buffer::calculateHostPtrSize(origin, region, operationParams.srcRowPitch, operationParams.srcSlicePitch);
is3D = (operationParams.size.z > 1) || (operationParams.dstOffset.z > 0);
} else if (operationParams.dstPtr) {
size_t origin[] = {operationParams.dstOffset.x, operationParams.dstOffset.y, operationParams.dstOffset.z};
size_t region[] = {operationParams.size.x, operationParams.size.y, operationParams.size.z};
hostPtrSize = Buffer::calculateHostPtrSize(origin, region, operationParams.dstRowPitch, operationParams.dstSlicePitch);
is3D = (operationParams.size.z > 1) || (operationParams.srcOffset.z > 0);
} else {
DEBUG_BREAK_IF(!false);
}
}
// Set-up ISA
int dimensions = is3D ? 3 : 2;
kernelNoSplit3DBuilder.setKernel(kernelBytes[dimensions - 1]);
// arg0 = src
if (operationParams.srcMemObj) {
kernelNoSplit3DBuilder.setArg(0, operationParams.srcMemObj);
} else {
kernelNoSplit3DBuilder.setArgSvm(0, hostPtrSize, is3D ? operationParams.srcPtr : ptrOffset(operationParams.srcPtr, operationParams.srcOffset.z * operationParams.srcSlicePitch));
}
// arg1 = dst
if (operationParams.dstMemObj) {
kernelNoSplit3DBuilder.setArg(1, operationParams.dstMemObj);
} else {
kernelNoSplit3DBuilder.setArgSvm(1, hostPtrSize, is3D ? operationParams.dstPtr : ptrOffset(operationParams.dstPtr, operationParams.dstOffset.z * operationParams.dstSlicePitch));
}
// arg2 = srcOrigin
uint32_t kSrcOrigin[4] = {(uint32_t)operationParams.srcOffset.x, (uint32_t)operationParams.srcOffset.y, (uint32_t)operationParams.srcOffset.z, 0};
kernelNoSplit3DBuilder.setArg(2, sizeof(uint32_t) * 4, kSrcOrigin);
// arg3 = dstOrigin
uint32_t kDstOrigin[4] = {(uint32_t)operationParams.dstOffset.x, (uint32_t)operationParams.dstOffset.y, (uint32_t)operationParams.dstOffset.z, 0};
kernelNoSplit3DBuilder.setArg(3, sizeof(uint32_t) * 4, kDstOrigin);
// arg4 = srcPitch
uint32_t kSrcPitch[2] = {(uint32_t)operationParams.srcRowPitch, (uint32_t)operationParams.srcSlicePitch};
kernelNoSplit3DBuilder.setArg(4, sizeof(uint32_t) * 2, kSrcPitch);
// arg5 = dstPitch
uint32_t kDstPitch[2] = {(uint32_t)operationParams.dstRowPitch, (uint32_t)operationParams.dstSlicePitch};
kernelNoSplit3DBuilder.setArg(5, sizeof(uint32_t) * 2, kDstPitch);
// Set-up work sizes
kernelNoSplit3DBuilder.setDispatchGeometry(operationParams.size, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelNoSplit3DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernelBytes[3];
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::FillBuffer> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernLeftLeftover(nullptr), kernMiddle(nullptr), kernRightLeftover(nullptr) {
populate(context, device,
EBuiltInOps::FillBuffer,
"",
"FillBufferLeftLeftover", kernLeftLeftover,
"FillBufferMiddle", kernMiddle,
"FillBufferRightLeftover", kernRightLeftover);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d1D, SplitDispatch::SplitMode::KernelSplit> kernelSplit1DBuilder;
uintptr_t start = reinterpret_cast<uintptr_t>(operationParams.dstPtr) + operationParams.dstOffset.x;
size_t middleAlignment = MemoryConstants::cacheLineSize;
size_t middleElSize = sizeof(uint32_t);
uintptr_t leftSize = start % middleAlignment;
leftSize = (leftSize > 0) ? (middleAlignment - leftSize) : 0; // calc left leftover size
leftSize = std::min(leftSize, operationParams.size.x); // clamp left leftover size to requested size
uintptr_t rightSize = (start + operationParams.size.x) % middleAlignment; // calc right leftover size
rightSize = std::min(rightSize, operationParams.size.x - leftSize); // clamp
uintptr_t middleSizeBytes = operationParams.size.x - leftSize - rightSize; // calc middle size
auto middleSizeEls = middleSizeBytes / middleElSize; // num work items in middle walker
// Set-up ISA
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Left, kernLeftLeftover);
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Middle, kernMiddle);
kernelSplit1DBuilder.setKernel(SplitDispatch::RegionCoordX::Right, kernRightLeftover);
DEBUG_BREAK_IF((operationParams.srcMemObj == nullptr) || (operationParams.srcOffset != 0));
DEBUG_BREAK_IF((operationParams.dstMemObj == nullptr) && (operationParams.dstSvmAlloc == nullptr));
// Set-up dstMemObj with buffer
if (operationParams.dstSvmAlloc) {
kernelSplit1DBuilder.setArgSvmAlloc(0, operationParams.dstPtr, operationParams.dstSvmAlloc);
} else {
kernelSplit1DBuilder.setArg(0, operationParams.dstMemObj);
}
// Set-up dstOffset
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Left, 1, static_cast<uint32_t>(operationParams.dstOffset.x));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Middle, 1, static_cast<uint32_t>(operationParams.dstOffset.x + leftSize));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Right, 1, static_cast<uint32_t>(operationParams.dstOffset.x + leftSize + middleSizeBytes));
// Set-up srcMemObj with pattern
kernelSplit1DBuilder.setArgSvm(2, operationParams.srcMemObj->getSize(), operationParams.srcMemObj->getGraphicsAllocation()->getUnderlyingBuffer(), operationParams.srcMemObj->getGraphicsAllocation());
// Set-up patternSizeInEls
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Left, 3, static_cast<uint32_t>(operationParams.srcMemObj->getSize()));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Middle, 3, static_cast<uint32_t>(operationParams.srcMemObj->getSize() / middleElSize));
kernelSplit1DBuilder.setArg(SplitDispatch::RegionCoordX::Right, 3, static_cast<uint32_t>(operationParams.srcMemObj->getSize()));
// Set-up work sizes
// Note for split walker, it would be just builder.SetDipatchGeomtry(GWS, ELWS, OFFSET)
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Left, Vec3<size_t>{leftSize, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Middle, Vec3<size_t>{middleSizeEls, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.setDispatchGeometry(SplitDispatch::RegionCoordX::Right, Vec3<size_t>{rightSize, 0, 0}, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelSplit1DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernLeftLeftover;
Kernel *kernMiddle;
Kernel *kernRightLeftover;
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::CopyBufferToImage3d> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernelBytes{nullptr} {
populate(context, device,
EBuiltInOps::CopyBufferToImage3d,
"",
"CopyBufferToImage3dBytes", kernelBytes[0],
"CopyBufferToImage3d2Bytes", kernelBytes[1],
"CopyBufferToImage3d4Bytes", kernelBytes[2],
"CopyBufferToImage3d8Bytes", kernelBytes[3],
"CopyBufferToImage3d16Bytes", kernelBytes[4]);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d3D, SplitDispatch::SplitMode::NoSplit> kernelNoSplit3DBuilder;
DEBUG_BREAK_IF(!(((operationParams.srcPtr != nullptr) || (operationParams.srcMemObj != nullptr)) && (operationParams.dstPtr == nullptr)));
auto dstImage = castToObjectOrAbort<Image>(operationParams.dstMemObj);
// Redescribe image to be byte-copy
auto dstImageRedescribed = dstImage->redescribe();
multiDispatchInfo.pushRedescribedMemObj(std::unique_ptr<MemObj>(dstImageRedescribed)); // life range same as mdi's
// Calculate srcRowPitch and srcSlicePitch
auto bytesPerPixel = dstImage->getSurfaceFormatInfo().ImageElementSizeInBytes;
size_t region[] = {operationParams.size.x, operationParams.size.y, operationParams.size.z};
auto srcRowPitch = static_cast<uint32_t>(operationParams.dstRowPitch ? operationParams.dstRowPitch : region[0] * bytesPerPixel);
auto srcSlicePitch = static_cast<uint32_t>(
operationParams.dstSlicePitch ? operationParams.dstSlicePitch : ((dstImage->getImageDesc().image_type == CL_MEM_OBJECT_IMAGE1D_ARRAY ? 1 : region[1]) * srcRowPitch));
// Determine size of host ptr surface for residency purposes
size_t hostPtrSize = operationParams.srcPtr ? Image::calculateHostPtrSize(region, srcRowPitch, srcSlicePitch, bytesPerPixel, dstImage->getImageDesc().image_type) : 0;
// Set-up kernel
auto bytesExponent = Math::log2(bytesPerPixel);
DEBUG_BREAK_IF(bytesExponent >= 5);
kernelNoSplit3DBuilder.setKernel(kernelBytes[bytesExponent]);
// Set-up source host ptr / buffer
if (operationParams.srcPtr) {
kernelNoSplit3DBuilder.setArgSvm(0, hostPtrSize, operationParams.srcPtr);
} else {
kernelNoSplit3DBuilder.setArg(0, operationParams.srcMemObj);
}
// Set-up destination image
kernelNoSplit3DBuilder.setArg(1, dstImageRedescribed);
// Set-up srcOffset
kernelNoSplit3DBuilder.setArg(2, static_cast<uint32_t>(operationParams.srcOffset.x));
// Set-up dstOrigin
{
uint32_t origin[] = {
static_cast<uint32_t>(operationParams.dstOffset.x),
static_cast<uint32_t>(operationParams.dstOffset.y),
static_cast<uint32_t>(operationParams.dstOffset.z),
0};
kernelNoSplit3DBuilder.setArg(3, sizeof(origin), origin);
}
// Set-up srcRowPitch
{
uint32_t pitch[] = {
static_cast<uint32_t>(srcRowPitch),
static_cast<uint32_t>(srcSlicePitch)};
kernelNoSplit3DBuilder.setArg(4, sizeof(pitch), pitch);
}
// Set-up work sizes
kernelNoSplit3DBuilder.setDispatchGeometry(operationParams.size, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelNoSplit3DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernelBytes[5];
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::CopyImage3dToBuffer> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernelBytes{nullptr} {
populate(context, device,
EBuiltInOps::CopyImage3dToBuffer,
"",
"CopyImage3dToBufferBytes", kernelBytes[0],
"CopyImage3dToBuffer2Bytes", kernelBytes[1],
"CopyImage3dToBuffer4Bytes", kernelBytes[2],
"CopyImage3dToBuffer8Bytes", kernelBytes[3],
"CopyImage3dToBuffer16Bytes", kernelBytes[4]);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d3D, SplitDispatch::SplitMode::NoSplit> kernelNoSplit3DBuilder;
DEBUG_BREAK_IF(!((operationParams.srcPtr == nullptr) && ((operationParams.dstPtr != nullptr) || (operationParams.dstMemObj != nullptr))));
auto srcImage = castToObjectOrAbort<Image>(operationParams.srcMemObj);
// Redescribe image to be byte-copy
auto srcImageRedescribed = srcImage->redescribe();
multiDispatchInfo.pushRedescribedMemObj(std::unique_ptr<MemObj>(srcImageRedescribed)); // life range same as mdi's
// Calculate dstRowPitch and dstSlicePitch
auto bytesPerPixel = srcImage->getSurfaceFormatInfo().ImageElementSizeInBytes;
size_t region[] = {operationParams.size.x, operationParams.size.y, operationParams.size.z};
auto dstRowPitch = static_cast<uint32_t>(operationParams.srcRowPitch ? operationParams.srcRowPitch : region[0] * bytesPerPixel);
auto dstSlicePitch = static_cast<uint32_t>(
operationParams.srcSlicePitch ? operationParams.srcSlicePitch : ((srcImage->getImageDesc().image_type == CL_MEM_OBJECT_IMAGE1D_ARRAY ? 1 : region[1]) * dstRowPitch));
// Determine size of host ptr surface for residency purposes
size_t hostPtrSize = operationParams.dstPtr ? Image::calculateHostPtrSize(region, dstRowPitch, dstSlicePitch, bytesPerPixel, srcImage->getImageDesc().image_type) : 0;
// Set-up ISA
auto bytesExponent = Math::log2(bytesPerPixel);
DEBUG_BREAK_IF(bytesExponent >= 5);
kernelNoSplit3DBuilder.setKernel(kernelBytes[bytesExponent]);
// Set-up source image
kernelNoSplit3DBuilder.setArg(0, srcImageRedescribed);
// Set-up destination host ptr / buffer
if (operationParams.dstPtr) {
kernelNoSplit3DBuilder.setArgSvm(1, hostPtrSize, operationParams.dstPtr);
} else {
kernelNoSplit3DBuilder.setArg(1, operationParams.dstMemObj);
}
// Set-up srcOrigin
{
uint32_t origin[] = {
static_cast<uint32_t>(operationParams.srcOffset.x),
static_cast<uint32_t>(operationParams.srcOffset.y),
static_cast<uint32_t>(operationParams.srcOffset.z),
0};
kernelNoSplit3DBuilder.setArg(2, sizeof(origin), origin);
}
// Set-up dstOffset
kernelNoSplit3DBuilder.setArg(3, static_cast<uint32_t>(operationParams.dstOffset.x));
// Set-up dstRowPitch
{
uint32_t pitch[] = {
static_cast<uint32_t>(dstRowPitch),
static_cast<uint32_t>(dstSlicePitch)};
kernelNoSplit3DBuilder.setArg(4, sizeof(pitch), pitch);
}
// Set-up work sizes
kernelNoSplit3DBuilder.setDispatchGeometry(operationParams.size, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelNoSplit3DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernelBytes[5];
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::CopyImageToImage3d> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernel(nullptr) {
populate(context, device,
EBuiltInOps::CopyImageToImage3d,
"",
"CopyImageToImage3d", kernel);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d3D, SplitDispatch::SplitMode::NoSplit> kernelNoSplit3DBuilder;
DEBUG_BREAK_IF(!((operationParams.srcPtr == nullptr) && (operationParams.dstPtr == nullptr)));
auto srcImage = castToObjectOrAbort<Image>(operationParams.srcMemObj);
auto dstImage = castToObjectOrAbort<Image>(operationParams.dstMemObj);
// Redescribe images to be byte-copies
auto srcImageRedescribed = srcImage->redescribe();
auto dstImageRedescribed = dstImage->redescribe();
multiDispatchInfo.pushRedescribedMemObj(std::unique_ptr<MemObj>(srcImageRedescribed)); // life range same as mdi's
multiDispatchInfo.pushRedescribedMemObj(std::unique_ptr<MemObj>(dstImageRedescribed)); // life range same as mdi's
// Set-up kernel
kernelNoSplit3DBuilder.setKernel(kernel);
// Set-up source image
kernelNoSplit3DBuilder.setArg(0, srcImageRedescribed);
// Set-up destination image
kernelNoSplit3DBuilder.setArg(1, dstImageRedescribed);
// Set-up srcOrigin
{
uint32_t origin[] = {
static_cast<uint32_t>(operationParams.srcOffset.x),
static_cast<uint32_t>(operationParams.srcOffset.y),
static_cast<uint32_t>(operationParams.srcOffset.z),
0};
kernelNoSplit3DBuilder.setArg(2, sizeof(origin), origin);
}
// Set-up dstOrigin
{
uint32_t origin[] = {
static_cast<uint32_t>(operationParams.dstOffset.x),
static_cast<uint32_t>(operationParams.dstOffset.y),
static_cast<uint32_t>(operationParams.dstOffset.z),
0};
kernelNoSplit3DBuilder.setArg(3, sizeof(origin), origin);
}
// Set-up work sizes
kernelNoSplit3DBuilder.setDispatchGeometry(operationParams.size, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelNoSplit3DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernel;
};
template <typename HWFamily>
class BuiltInOp<HWFamily, EBuiltInOps::FillImage3d> : public BuiltinDispatchInfoBuilder {
public:
BuiltInOp(BuiltIns &kernelsLib, Context &context, Device &device)
: BuiltinDispatchInfoBuilder(kernelsLib), kernel(nullptr) {
populate(context, device,
EBuiltInOps::FillImage3d,
"",
"FillImage3d", kernel);
}
bool buildDispatchInfos(MultiDispatchInfo &multiDispatchInfo, const BuiltinOpParams &operationParams) const override {
DispatchInfoBuilder<SplitDispatch::Dim::d3D, SplitDispatch::SplitMode::NoSplit> kernelNoSplit3DBuilder;
DEBUG_BREAK_IF(!((operationParams.srcMemObj == nullptr) && (operationParams.srcPtr != nullptr) && (operationParams.dstPtr == nullptr)));
auto image = castToObjectOrAbort<Image>(operationParams.dstMemObj);
// Redescribe image to be byte-copy
auto imageRedescribed = image->redescribeFillImage();
multiDispatchInfo.pushRedescribedMemObj(std::unique_ptr<MemObj>(imageRedescribed));
// Set-up kernel
kernelNoSplit3DBuilder.setKernel(kernel);
// Set-up destination image
kernelNoSplit3DBuilder.setArg(0, imageRedescribed);
// Set-up fill color
int iFillColor[4] = {0};
const void *fillColor = operationParams.srcPtr;
convertFillColor(fillColor,
iFillColor,
image->getSurfaceFormatInfo().OCLImageFormat,
imageRedescribed->getSurfaceFormatInfo().OCLImageFormat);
kernelNoSplit3DBuilder.setArg(1, 4 * sizeof(int32_t), iFillColor);
// Set-up dstOffset
{
uint32_t offset[] = {
static_cast<uint32_t>(operationParams.dstOffset.x),
static_cast<uint32_t>(operationParams.dstOffset.y),
static_cast<uint32_t>(operationParams.dstOffset.z),
0};
kernelNoSplit3DBuilder.setArg(2, sizeof(offset), offset);
}
// Set-up work sizes
kernelNoSplit3DBuilder.setDispatchGeometry(operationParams.size, Vec3<size_t>{0, 0, 0}, Vec3<size_t>{0, 0, 0});
kernelNoSplit3DBuilder.bake(multiDispatchInfo);
return true;
}
protected:
Kernel *kernel;
};
BuiltinDispatchInfoBuilder &BuiltIns::getBuiltinDispatchInfoBuilder(EBuiltInOps operation, Context &context, Device &device) {
uint32_t operationId = static_cast<uint32_t>(operation);
auto &operationBuilder = BuiltinOpsBuilders[operationId];
switch (operation) {
default:
throw std::runtime_error("getBuiltinDispatchInfoBuilder failed");
case EBuiltInOps::CopyBufferToBuffer:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::CopyBufferToBuffer>(*this, context, device)); });
break;
case EBuiltInOps::CopyBufferRect:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::CopyBufferRect>(*this, context, device)); });
break;
case EBuiltInOps::FillBuffer:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::FillBuffer>(*this, context, device)); });
break;
case EBuiltInOps::CopyBufferToImage3d:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::CopyBufferToImage3d>(*this, context, device)); });
break;
case EBuiltInOps::CopyImage3dToBuffer:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::CopyImage3dToBuffer>(*this, context, device)); });
break;
case EBuiltInOps::CopyImageToImage3d:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::CopyImageToImage3d>(*this, context, device)); });
break;
case EBuiltInOps::FillImage3d:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::FillImage3d>(*this, context, device)); });
break;
case EBuiltInOps::VmeBlockMotionEstimateIntel:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::VmeBlockMotionEstimateIntel>(*this, context, device)); });
break;
case EBuiltInOps::VmeBlockAdvancedMotionEstimateCheckIntel:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::VmeBlockAdvancedMotionEstimateCheckIntel>(*this, context, device)); });
break;
case EBuiltInOps::VmeBlockAdvancedMotionEstimateBidirectionalCheckIntel:
std::call_once(operationBuilder.second, [&] { operationBuilder.first.reset(new BuiltInOp<HWFamily, EBuiltInOps::VmeBlockAdvancedMotionEstimateBidirectionalCheckIntel>(*this, context, device)); });
break;
}
return *operationBuilder.first;
}
std::unique_ptr<BuiltinDispatchInfoBuilder> BuiltIns::setBuiltinDispatchInfoBuilder(EBuiltInOps operation, Context &context, Device &device, std::unique_ptr<BuiltinDispatchInfoBuilder> builder) {
uint32_t operationId = static_cast<uint32_t>(operation);
auto &operationBuilder = BuiltinOpsBuilders[operationId];
operationBuilder.first.swap(builder);
return builder;
}
} // namespace OCLRT
Reference in New Issue View Git Blame Copy Permalink