MNN/source/backend/opencl/execution/buffer/Interp3DBufExecution.cpp

138 lines
5.3 KiB
C++

//
// Interp3DBufExecution.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/Interp3DBufExecution.hpp"
#include "core/TensorUtils.hpp"
namespace MNN {
namespace OpenCL {
Interp3DBufExecution::Interp3DBufExecution(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend) : Execution(backend) {
mOpenCLBackend = static_cast<OpenCLBackend *>(backend);
auto runtime = mOpenCLBackend->getOpenCLRuntime();
auto interp3DParam = op->main_as_Interp();
mCordTransform[0] = interp3DParam->widthScale();
mCordTransform[1] = interp3DParam->widthOffset();
mCordTransform[2] = interp3DParam->heightScale();
mCordTransform[3] = interp3DParam->heightOffset();
mCordTransform[4] = interp3DParam->depthScale();
mCordTransform[5] = interp3DParam->depthOffset();
std::set<std::string> buildOptions;
if (op->main_as_Interp()->resizeType() == 1) {
mKernelName = "nearest3D_buf";
mKernel = runtime->buildKernel("interp_buf", mKernelName, buildOptions);
} else {
MNN_ERROR("Resize type other than nearest is not supported in Interp3DBuf, change to nearest!");
mKernelName = "nearest3D_buf";
mKernel = runtime->buildKernel("interp_buf", mKernelName, buildOptions);
}
mMaxWorkGroupSize = static_cast<uint32_t>(runtime->getMaxWorkGroupSize(mKernel));
}
ErrorCode Interp3DBufExecution::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
Tensor *input = inputs[0];
Tensor *output = outputs[0];
auto runtime = ((OpenCLBackend *)backend())->getOpenCLRuntime();
std::vector<int> inputImageShape = tensorShapeFormat(input); // {C/4 * H * W, N * D} for 5-D Tensor
std::vector<int> outputImageShape = tensorShapeFormat(output);
auto inputShape = input->shape();
auto outputShape = output->shape();
const int inputBatch = inputShape[0];
const int inputChannels = inputShape[1];
const int inputDepth = inputShape[2];
const int inputHeight = inputShape[3];
const int inputWidth = inputShape[4];
const int channelBlocks = UP_DIV(inputChannels, 4);
const int outputDepth = outputShape[2];
const int outputHeight = outputShape[3];
const int outputWidth = outputShape[4];
mGWS = {static_cast<uint32_t>(channelBlocks),
static_cast<uint32_t>(outputHeight * outputWidth),
static_cast<uint32_t>(outputDepth * inputBatch)};
MNN_ASSERT(outputDepth > 0 && outputHeight > 0 && outputWidth > 0);
uint32_t idx = 0;
cl_int ret = CL_SUCCESS;
ret |= mKernel.setArg(idx++, mGWS[0]);
ret |= mKernel.setArg(idx++, mGWS[1]);
ret |= mKernel.setArg(idx++, mGWS[2]);
ret |= mKernel.setArg(idx++, openCLBuffer(input));
ret |= mKernel.setArg(idx++, openCLBuffer(output));
ret |= mKernel.setArg(idx++, mCordTransform[4]);
ret |= mKernel.setArg(idx++, mCordTransform[2]);
ret |= mKernel.setArg(idx++, mCordTransform[0]);
ret |= mKernel.setArg(idx++, mCordTransform[5]);
ret |= mKernel.setArg(idx++, mCordTransform[3]);
ret |= mKernel.setArg(idx++, mCordTransform[1]);
ret |= mKernel.setArg(idx++, static_cast<int32_t>(inputDepth));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(inputHeight));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(inputWidth));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(outputDepth));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(outputHeight));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(outputWidth));
ret |= mKernel.setArg(idx++, static_cast<int32_t>(channelBlocks));
MNN_CHECK_CL_SUCCESS(ret, "setArg Interp3DBufExecution");
mLWS = localWS3DDefault(mGWS, mMaxWorkGroupSize, runtime, mKernelName, mKernel).first;
return NO_ERROR;
}
ErrorCode Interp3DBufExecution::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
#ifdef LOG_VERBOSE
MNN_PRINT("Start InterpBufExecution onExecute... \n");
#endif
#ifdef ENABLE_OPENCL_TIME_PROFILER
cl::Event event;
run3DKernelDefault(mKernel, mGWS, mLWS,
mOpenCLBackend->getOpenCLRuntime(), &event);
mOpenCLBackend->getOpenCLRuntime()->pushEvent({"Interp", event});
#else
run3DKernelDefault(mKernel, mGWS, mLWS, mOpenCLBackend->getOpenCLRuntime());
#endif
#ifdef LOG_VERBOSE
MNN_PRINT("end InterpBufExecution onExecute... \n");
#endif
return NO_ERROR;
}
class Interp3DBufCreator : public OpenCLBackend::Creator {
public:
virtual ~Interp3DBufCreator() = default;
virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
const MNN::Op *op, Backend *backend) const override {
for (int i = 0; i < inputs.size(); ++i) {
TensorUtils::setTensorSupportPack(inputs[i], false);
}
for (int i = 0; i < outputs.size(); ++i) {
TensorUtils::setTensorSupportPack(outputs[i], false);
}
return new Interp3DBufExecution(inputs, op, backend);
;
}
};
OpenCLCreatorRegister<Interp3DBufCreator> __Interp3DBuf_op_(OpType_Interp3D, BUFFER);
} // namespace OpenCL
} // namespace MNN
#endif /* MNN_OPENCL_BUFFER_CLOSED */