mirror of https://github.com/alibaba/MNN.git
110 lines
3.8 KiB
C++
110 lines
3.8 KiB
C++
//
|
|
// LrnExecution.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/02/28.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "LrnExecution.hpp"
|
|
#include <Macro.h>
|
|
#include "TensorUtils.hpp"
|
|
|
|
namespace MNN {
|
|
namespace OpenCL {
|
|
|
|
LrnExecution::LrnExecution(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend)
|
|
: Execution(backend) {
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("start LrnExecution init !\n");
|
|
#endif
|
|
mOpenCLBackend = static_cast<OpenCLBackend *>(backend);
|
|
auto lrn = op->main_as_LRN();
|
|
mRegionType = lrn->regionType();
|
|
mLocalSize = lrn->localSize();
|
|
mAlpha = lrn->alpha() / (float)mLocalSize;
|
|
mBeta = lrn->beta();
|
|
auto runtime = mOpenCLBackend->getOpenCLRuntime();
|
|
std::set<std::string> buildOptions;
|
|
std::string kernelName = "lrn_buffer";
|
|
mKernel = runtime->buildKernel("lrn", kernelName, buildOptions);
|
|
mMaxWorkGroupSize = static_cast<uint32_t>(runtime->getMaxWorkGroupSize(mKernel));
|
|
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("end LrnExecution init !\n");
|
|
#endif
|
|
}
|
|
|
|
ErrorCode LrnExecution::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
|
auto bufferPool = mOpenCLBackend->getBufferPool();
|
|
mInputTemp.reset(Tensor::createDevice<float>(tensorShapeFormat(inputs[0])));
|
|
mOutputTemp.reset(Tensor::createDevice<float>(tensorShapeFormat(outputs[0])));
|
|
auto inputBuffer = bufferPool->alloc(mInputTemp->size());
|
|
auto outputBuffer = bufferPool->alloc(mOutputTemp->size());
|
|
mInputTemp->buffer().device = (uint64_t)inputBuffer;
|
|
mOutputTemp->buffer().device = (uint64_t)outputBuffer;
|
|
|
|
bufferPool->recycle(inputBuffer);
|
|
bufferPool->recycle(outputBuffer);
|
|
|
|
return NO_ERROR;
|
|
}
|
|
|
|
ErrorCode LrnExecution::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("start LrnExecution onExecute !\n");
|
|
#endif
|
|
|
|
Tensor *input = inputs[0];
|
|
Tensor *output = outputs[0];
|
|
|
|
std::vector<int> inputShape = tensorShapeFormat(input);
|
|
std::vector<int> outputShape = tensorShapeFormat(output);
|
|
|
|
int oN = outputShape.at(0);
|
|
int oH = outputShape.at(1);
|
|
int oW = outputShape.at(2);
|
|
int oC = outputShape.at(3);
|
|
|
|
convertImageToNCHWBuffer(input, mInputTemp.get(), mImageToBufferKernel, mOpenCLBackend->getOpenCLRuntime());
|
|
{
|
|
std::vector<uint32_t> gws = {static_cast<uint32_t>(oW), static_cast<uint32_t>(oH), static_cast<uint32_t>(oC)};
|
|
const std::vector<uint32_t> lws = {16, 16, 1};
|
|
int32_t shape[4] = {oW, oH, oC, oN};
|
|
{
|
|
uint32_t idx = 0;
|
|
mKernel.setArg(idx++, openCLBuffer(mInputTemp.get()));
|
|
mKernel.setArg(idx++, openCLBuffer(mOutputTemp.get()));
|
|
mKernel.setArg(idx++, shape);
|
|
mKernel.setArg(idx++, mLocalSize);
|
|
mKernel.setArg(idx++, mAlpha);
|
|
mKernel.setArg(idx++, mBeta);
|
|
}
|
|
run3DKernelDefault(mKernel, gws, lws, mOpenCLBackend->getOpenCLRuntime());
|
|
}
|
|
convertNCHWBufferToImage(mOutputTemp.get(), output, mBufferToImageKernel, mOpenCLBackend->getOpenCLRuntime());
|
|
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("end LrnExecution onExecute !\n");
|
|
#endif
|
|
return NO_ERROR;
|
|
}
|
|
|
|
class LRNCreator : 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 {
|
|
auto lrn = op->main_as_LRN();
|
|
if (lrn->regionType() != 0) {
|
|
return nullptr;
|
|
}
|
|
|
|
return new LrnExecution(inputs, op, backend);
|
|
}
|
|
};
|
|
|
|
OpenCLCreatorRegister<LRNCreator> __lrn_op(OpType_LRN);
|
|
|
|
} // namespace OpenCL
|
|
} // namespace MNN
|