MNN/source/backend/cpu/CPUWhere.cpp

61 lines
1.8 KiB
C++

//
// CPUWhere.cpp
// MNN
//
// Created by MNN on 2018/08/31.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "backend/cpu/CPUWhere.hpp"
#include "backend/cpu/CPUBackend.hpp"
namespace MNN {
template <typename T>
std::vector<int32_t> _collect(Tensor* t) {
const T* ptr = t->host<T>();
std::vector<int32_t> collect;
for (int i = 0; i < t->elementSize(); i++) {
if (ptr[i] > 0) {
collect.push_back(i);
}
}
return collect;
}
ErrorCode CPUWhere::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
auto& ib = inputs[0]->buffer();
auto outputData = outputs[0]->host<int32_t>();
std::vector<int32_t> collect;
if (ib.type == halide_type_of<float>()) {
collect = _collect<float>(inputs[0]);
} else if (ib.type == halide_type_of<int32_t>()) {
collect = _collect<int32_t>(inputs[0]);
} else if (ib.type == halide_type_of<uint8_t>()) {
collect = _collect<uint8_t>(inputs[0]);
}
//MNN_ASSERT(outputs[0]->batch() == trueVec.size());
for (int i = 0; i < collect.size(); i++) {
int index = collect[i];
for (int j = 0; j < ib.dimensions; j++) {
int result = ib.dim[j].stride == 0 ? index : index / ib.dim[j].stride;
index = index - result * ib.dim[j].stride;
outputData[i * ib.dimensions + j] = result;
}
}
return NO_ERROR;
}
class CPUWhereCreator : public CPUBackend::Creator {
public:
virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
const MNN::Op* op, Backend* backend) const override {
return new CPUWhere(backend);
}
};
REGISTER_CPU_OP_CREATOR(CPUWhereCreator, OpType_Where);
} // namespace MNN