mirror of https://github.com/alibaba/MNN.git
64 lines
1.9 KiB
C++
64 lines
1.9 KiB
C++
//
|
|
// CPUSpatialProduct.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2018/07/19.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "backend/cpu/CPUSpatialProduct.hpp"
|
|
#include "backend/cpu/CPUBackend.hpp"
|
|
#include "core/Macro.h"
|
|
|
|
#ifdef MNN_USE_NEON
|
|
#include <arm_neon.h>
|
|
#endif
|
|
|
|
namespace MNN {
|
|
CPUSpatialProduct::CPUSpatialProduct(Backend *b) : MNN::Execution(b) {
|
|
// nothing to do
|
|
}
|
|
|
|
ErrorCode CPUSpatialProduct::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
|
// Assume
|
|
// bottom[0] dim CxHxW
|
|
// bottom[1] dim 1xHxW
|
|
// top[0] dim CxHxW
|
|
auto inputTensor = inputs[0];
|
|
auto outputTensor = outputs[0];
|
|
int w = inputTensor->width();
|
|
int h = inputTensor->height();
|
|
int channels = UP_DIV(inputTensor->channel(), 4);
|
|
int size = w * h;
|
|
|
|
auto inputT1 = inputs[1];
|
|
|
|
// second, top[0](CxHxW) = bottom[0](CxHxW) * bottom[1](CxHxW)
|
|
for (int q = 0; q < channels; q++) {
|
|
const float *ptr = inputTensor->host<float>() + q * size * 4;
|
|
const float *ptr1 = inputT1->host<float>();
|
|
float *outptr = outputTensor->host<float>() + q * size * 4;
|
|
|
|
for (int v = 0; v < size; ++v) {
|
|
#ifdef MNN_USE_NEON
|
|
vst1q_f32(outptr + 4 * v, vld1q_f32(ptr + 4 * v) * vdupq_n_f32(ptr1[4 * v]));
|
|
#else
|
|
for (int j = 0; j < 4; ++j) {
|
|
outptr[4 * v + j] = ptr[4 * v + j] * ptr1[4 * v + 0];
|
|
}
|
|
#endif
|
|
}
|
|
}
|
|
return NO_ERROR;
|
|
}
|
|
|
|
class CPUSpatialProductCreator : public CPUBackend::Creator {
|
|
public:
|
|
virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
|
|
const MNN::Op *op, Backend *backend) const {
|
|
return new CPUSpatialProduct(backend);
|
|
}
|
|
};
|
|
REGISTER_CPU_OP_CREATOR(CPUSpatialProductCreator, OpType_SpatialProduct);
|
|
} // namespace MNN
|