root
/
MNN
mirror of https://github.com/alibaba/MNN.git
1
0
Fork 0
Code Issues Actions 2 Packages Projects Releases Wiki Activity
MNN/source/backend/opencl/execution/buffer/ReluBufExecution.cpp

297 lines
16 KiB
C++
Raw Permalink Blame History

//
// ReluBufExecution.cpp
// MNN
//
// Created by MNN on 2019/02/28.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef MNN_OPENCL_BUFFER_CLOSED
#include "backend/opencl/execution/buffer/ReluBufExecution.hpp"
#include "backend/opencl/execution/buffer/UnaryBufExecution.hpp"
namespace MNN {
namespace OpenCL {
ReluBufExecution::ReluBufExecution(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend)
: CommonExecution(backend, op) {
mOpenCLBackend = static_cast<OpenCLBackend *>(backend);
auto mPreluParamPtr = op->main_as_PRelu();
int preluSize = mPreluParamPtr->slopeCount();
const float *preluDataPtr = mPreluParamPtr->slope()->data();
int buffer_size = ALIGN_UP4(preluSize);
if (mOpenCLBackend->getPrecision() != BackendConfig::Precision_High) {
buffer_size *= sizeof(half_float::half);
} else {
buffer_size *= sizeof(float);
}
mPreluParam.reset(Tensor::createDevice<float>({1, 1, 1, ALIGN_UP4(preluSize)}));
mOpenCLBackend->onAcquireBuffer(mPreluParam.get(), Backend::STATIC);
cl::Buffer &preluBuffer = openCLBuffer(mPreluParam.get());
cl_int error;
auto preluDataPtrCL = mOpenCLBackend->getOpenCLRuntime()->commandQueue().enqueueMapBuffer(
preluBuffer, true, CL_MAP_WRITE, 0, buffer_size, nullptr, nullptr, &error);
if(preluDataPtrCL != nullptr && error == CL_SUCCESS){
if (mOpenCLBackend->getPrecision() != BackendConfig::Precision_High) {
for(int i=0; i<preluSize; i++) {
((half_float::half*)preluDataPtrCL)[i] = (half_float::half)(preluDataPtr[i]);
}
for(int i=preluSize; i<ALIGN_UP4(preluSize); i++) {
((half_float::half*)preluDataPtrCL)[i] = (half_float::half)(0.0f);
}
}else{
::memset(preluDataPtrCL, 0, buffer_size);
::memcpy(preluDataPtrCL, preluDataPtr, preluSize * sizeof(float));
}
}else{
MNN_ERROR("Map error preluDataPtrCL == nullptr \n");
}
mOpenCLBackend->getOpenCLRuntime()->commandQueue().enqueueUnmapMemObject(preluBuffer, preluDataPtrCL);
}
ReluBufExecution::~ReluBufExecution() {
// Do nothing
}
ErrorCode ReluBufExecution::onEncode(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
mUnits.resize(1);
auto nhwc = tensorShapeFormat(outputs[0]);
int nhwcArray[4] = {nhwc[0], nhwc[1], nhwc[2], UP_DIV(nhwc[3], 4)};
auto imageWidth = nhwc[0] * UP_DIV(nhwc[3], 4);
auto imageHeight = nhwc[1] * nhwc[2];
std::vector<uint32_t> localSize = {1, 1};
std::vector<uint32_t> globalSize = {(uint32_t)imageWidth, (uint32_t)imageHeight};
auto runTime = mOpenCLBackend->getOpenCLRuntime();
#ifdef MNN_SUPPORT_INTEL_SUBGROUP
if (runTime->isSupportedIntelSubgroup()){
return SubgrouponResize(inputs, outputs);
}
#endif /* MNN_SUPPORT_INTEL_SUBGROUP */
std::set<std::string> buildOption;
buildOption.emplace("-DOPERATOR=select(in0*in1,in0,in0>=(float4)0)");
mUnits[0].kernel = runTime->buildKernel("binary_buf", "prelu_buf", buildOption, mOpenCLBackend->getPrecision(), inputs[0], outputs[0]);
mMaxWorkGroupSize = static_cast<uint32_t>(runTime->getMaxWorkGroupSize(mUnits[0].kernel));
int fullCount[2] = {1, 1};
uint32_t index = 0;
cl_int ret = CL_SUCCESS;
ret |= mUnits[0].kernel->get().setArg(index++, globalSize[0]);
ret |= mUnits[0].kernel->get().setArg(index++, globalSize[1]);
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(inputs[0]));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(mPreluParam.get()));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(outputs[0]));
ret |= mUnits[0].kernel->get().setArg(index++, nhwcArray);
MNN_CHECK_CL_SUCCESS(ret, "setArg ReluBufExecution");
std::string name = "prelu_buf";
localSize = localWS2DDefault(globalSize, mMaxWorkGroupSize, mOpenCLBackend->getOpenCLRuntime(), name, mUnits[0].kernel, mOpenCLBackend->getCLTuneLevel(), "binary_buf").first;
mUnits[0].globalWorkSize = {globalSize[0], globalSize[1]};
mUnits[0].localWorkSize = {localSize[0], localSize[1]};
mOpenCLBackend->recordKernel2d(mUnits[0].kernel, globalSize, localSize);
return NO_ERROR;
}
#ifdef MNN_SUPPORT_INTEL_SUBGROUP
ErrorCode ReluBufExecution::SubgrouponResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
mUnits.resize(1);
auto nhwc = tensorShapeFormat(outputs[0]);
int nhwcArray[4] = {nhwc[0], nhwc[1], nhwc[2], nhwc[3]};
auto runTime = mOpenCLBackend->getOpenCLRuntime();
int input_c_pack = TensorUtils::getTensorChannelPack(inputs[0]);
int output_c_pack = TensorUtils::getTensorChannelPack(outputs[0]);
auto inputpad = TensorUtils::getDescribe(inputs[0])->mPads;
auto outputpad = TensorUtils::getDescribe(outputs[0])->mPads;
std::string kernelName = "prelu_buf_c" + std::to_string(input_c_pack) + "_c" + std::to_string(output_c_pack);
auto output = outputs[0];
std::set<std::string> buildOptions;
if (output->getType().code == halide_type_int) {
if (output->getType().bits == 8) {
buildOptions.emplace("-DINTEL_DATA=uchar");
buildOptions.emplace("-DAS_INPUT_DATA=as_char");
buildOptions.emplace("-DAS_INPUT_DATA4=as_char4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_uchar4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read_uc");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read_uc4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write_uc4");
} else if (output->getType().bits == 32) {
buildOptions.emplace("-DINTEL_DATA=uint");
buildOptions.emplace("-DAS_INPUT_DATA=as_int");
buildOptions.emplace("-DAS_INPUT_DATA4=as_int4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_uint4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write4");
}
} else if (output->getType().code == halide_type_uint) {
if (output->getType().bits == 8) {
buildOptions.emplace("-DINTEL_DATA=uchar");
buildOptions.emplace("-DAS_INPUT_DATA=as_uchar");
buildOptions.emplace("-DAS_INPUT_DATA4=as_uchar4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_uchar4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read_uc");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read_uc4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write_uc4");
} else if (output->getType().bits == 32) {
buildOptions.emplace("-DINTEL_DATA=uint");
buildOptions.emplace("-DAS_INPUT_DATA=as_uint");
buildOptions.emplace("-DAS_INPUT_DATA4=as_uint4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_uint4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write4");
}
} else {
if (mOpenCLBackend->getPrecision() != BackendConfig::Precision_High) {
buildOptions.emplace("-DINTEL_DATA=ushort");
buildOptions.emplace("-DAS_INPUT_DATA=as_half");
buildOptions.emplace("-DAS_INPUT_DATA4=as_half4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_ushort4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read_us");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read_us4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write_us4");
} else {
buildOptions.emplace("-DINTEL_DATA=uint");
buildOptions.emplace("-DAS_INPUT_DATA=as_float");
buildOptions.emplace("-DAS_INPUT_DATA4=as_float4");
buildOptions.emplace("-DAS_OUTPUT_DATA4=as_uint4");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ=intel_sub_group_block_read");
buildOptions.emplace("-DINTEL_SUB_GROUP_READ4=intel_sub_group_block_read4");
buildOptions.emplace("-DINTEL_SUB_GROUP_WRITE4=intel_sub_group_block_write4");
}
}
buildOptions.emplace("-DOPERATOR=select(in0*in1,in0,in0>=(float4)0)");
mUnits[0].kernel = runTime->buildKernel("binary_subgroup_buf", kernelName, buildOptions, mOpenCLBackend->getPrecision(), inputs[0], output);
mMaxWorkGroupSize = static_cast<uint32_t>(runTime->getMaxWorkGroupSize(mUnits[0].kernel));
int fullCount[2] = {1, 1};
uint32_t index = 0;
cl_int ret = CL_SUCCESS;
std::vector<uint32_t> gws = {(uint32_t)nhwc[2] * nhwc[1], (uint32_t)UP_DIV(nhwc[3], 4),
(uint32_t)nhwc[0]};
std::vector<uint32_t> lws = {1, 16, 1};
if (input_c_pack == 4) {
mUnits[0].globalWorkSize = {gws[0], gws[1], gws[2]};
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[0]);
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[1]);
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[2]);
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(inputs[0]));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(mPreluParam.get()));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(output));
ret |= mUnits[0].kernel->get().setArg(index++, nhwcArray);
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(inputpad.left));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(inputpad.right));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(outputpad.left));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(outputpad.right));
MNN_CHECK_CL_SUCCESS(ret, "setArg ReluBufExecution SubGroup C4");
lws = localWS3DDefault(gws, mMaxWorkGroupSize, mOpenCLBackend->getOpenCLRuntime(), kernelName, mUnits[0].kernel, mOpenCLBackend->getCLTuneLevel(), "binary_subgroup_buf").first;
mUnits[0].localWorkSize = {lws[0], lws[1], lws[2]};
} else {
gws = {(uint32_t)UP_DIV(nhwc[2], 4) * nhwc[1], (uint32_t)ROUND_UP(nhwc[3], 16),
(uint32_t)nhwc[0]};
mUnits[0].globalWorkSize = {gws[0], gws[1], gws[2]};
mUnits[0].localWorkSize = {lws[0], lws[1], lws[2]};
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[0]);
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[1]);
ret |= mUnits[0].kernel->get().setArg(index++, mUnits[0].globalWorkSize[2]);
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(inputs[0]));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(mPreluParam.get()));
ret |= mUnits[0].kernel->get().setArg(index++, openCLBuffer(outputs[0]));
ret |= mUnits[0].kernel->get().setArg(index++, nhwcArray);
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(inputpad.left));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(inputpad.right));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(outputpad.left));
ret |= mUnits[0].kernel->get().setArg(index++, static_cast<uint32_t>(outputpad.right));
MNN_CHECK_CL_SUCCESS(ret, "setArg ReluBufExecution SubGroup");
}
mOpenCLBackend->recordKernel3d(mUnits[0].kernel, gws, lws);
return NO_ERROR;
}
#endif /* MNN_SUPPORT_INTEL_SUBGROUP */
class ReluBufCreator : public OpenCLBackend::Creator {
public:
virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
const MNN::Op *op, Backend *backend) const override {
// There seems to be a bug on OpenCL compiler of AMD Radeon HD 7000 series.
// When use build option -Dname=definition, definition will be truncated by
// a comma, which violate opencl specification (quote, 'In particular, the definition will
// be truncated by embedded newline characters'.)
// So we use ternary operation (A ? B: C) instead of function call with comma
// (e.g, fmax(in,(float4)(0))), when there is a Radeon GPU.
bool isRadeonGpu = (static_cast<OpenCLBackend*>(backend)->getOpenCLRuntime()->getGpuType() == RADEON);
#ifdef MNN_SUPPORT_INTEL_SUBGROUP
for (int i = 0; i < inputs.size(); ++i) {
int channel = inputs[i]->channel();
if (channel >= 16 && static_cast<OpenCLBackend *>(backend)->getOpenCLRuntime()->isSupportedIntelSubgroup()) {
TensorUtils::setTensorChannelPack(inputs[i], 16);
}
}
#endif /* MNN_SUPPORT_INTEL_SUBGROUP */
if (op->type() == OpType_ReLU6) {
char storage[256];
float minValue = 0.0f;
float maxValue = 6.0f;
if (nullptr != op->main_as_Relu6()) {
minValue = op->main_as_Relu6()->minValue();
maxValue = op->main_as_Relu6()->maxValue();
}
if (isRadeonGpu) {
std::string temp = "(in<=(float4)((float)%f)?(float4)((float)%f):(in>=(float4)((float)%f)?(float4)((float)%f):in))";
sprintf(storage, temp.c_str(), minValue, minValue, maxValue, maxValue);
return new UnaryBufExecution(storage, op, backend);
}
std::string temp = "clamp(in,(float4)((float)%f),(float4)((float)%f))";
sprintf(storage, temp.c_str(), minValue, maxValue);
return new UnaryBufExecution(storage, op, backend);
}
if (op->type() == OpType_ReLU) {
if (op->main_as_Relu()->slope() == 0.0f) {
if (isRadeonGpu) {
return new UnaryBufExecution("(in>(float4)((float)0)?in:(float4)((float)0))", op, backend);
}
return new UnaryBufExecution("fmax(in,(float4)((float)0))", op, backend);
}
auto slope = op->main_as_Relu()->slope();
char slopeCStr[30] = {};
sprintf(slopeCStr, "%.8f", slope);
std::string slopeStr = slopeCStr;
if (isRadeonGpu) {
return new UnaryBufExecution("in<(float4)((float)0)?(float)(" + slopeStr + "f)*in:in", op, backend);
}
return new UnaryBufExecution("select((float)(" + slopeStr + "f)*in,in,in>=(float4)((float)0))", op, backend);
}
if (op->type() == OpType_PReLU) {
if (op->main_as_PRelu()->slopeCount() == 1) {
auto slope = op->main_as_PRelu()->slope()->data()[0];
char slopeCStr[30] = {};
sprintf(slopeCStr, "%.8f", slope);
std::string slopeStr = slopeCStr;
if (isRadeonGpu) {
return new UnaryBufExecution("in<(float4)((float)0)?(float)(" + slopeStr + "f)*in:in", op, backend);
}
return new UnaryBufExecution("select((float)(" + slopeStr + "f)*in,in,in>=(float4)((float)0))", op, backend);
}
return new ReluBufExecution(inputs, op, backend);
}
return nullptr;
}
};
REGISTER_OPENCL_OP_CREATOR(ReluBufCreator, OpType_ReLU, BUFFER);
REGISTER_OPENCL_OP_CREATOR(ReluBufCreator, OpType_PReLU, BUFFER);
REGISTER_OPENCL_OP_CREATOR(ReluBufCreator, OpType_ReLU6, BUFFER);
} // namespace OpenCL
} // namespace MNN
#endif /* MNN_OPENCL_BUFFER_CLOSED */
Reference in New Issue View Git Blame Copy Permalink