MNN/source/backend/cpu/CPUHistogram.cpp

121 lines
3.9 KiB
C++

//
// CPUHistogram.cpp
// MNN
//
// Created by MNN on 2022/07/26.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "backend/cpu/CPUHistogram.hpp"
#include "backend/cpu/CPUBackend.hpp"
#include "core/Concurrency.h"
#include "core/Macro.h"
#include <cmath>
namespace MNN {
CPUHistogram::CPUHistogram(Backend *backend, const Op* op): Execution(backend) {
auto param = op->main_as_ArgMax();
mChannel = param->axis();
mBinNum = param->outMaxVal();
mMin = param->softmaxThreshold();
mMax = param->topK();
mAlpha = static_cast<float>(mBinNum) / (mMax - mMin);
mBeta = mAlpha * mMin;
}
template <typename T>
ErrorCode CPUHistogram::histogram(Tensor* input, Tensor* output) {
auto iptr = input->host<T>() + mChannel;
auto optr = output->host<float>();
memset(optr, 0, mBinNum * sizeof(float));
for (int i = 0; i < mSize; i++) {
T val = iptr[i * mStride];
if (val >= mMin && val <= mMax) {
const int bin = (int)(val * mAlpha - mBeta);
optr[std::min(bin, mBinNum -1)]++;
}
}
return NO_ERROR;
}
template <>
ErrorCode CPUHistogram::histogram<uint8_t>(Tensor* input, Tensor* output) {
auto iptr = input->host<uint8_t>() + mChannel;
auto optr = output->host<float>();
int hist_map[256] = { 0 };
// add hist_ptr to avoid iOS compile error: cannot refer to declaration with an array type inside block
int* hist_ptr = hist_map;
// auto numberThread = ((CPUBackend*)backend())->threadNumber();
// TODO: Support multi thread
int numberThread = 1;
int sizeDivide = mSize / numberThread;
MNN_CONCURRENCY_BEGIN(tId, numberThread) {
int number = sizeDivide;
if (tId == numberThread - 1) {
number = mSize - tId * sizeDivide;
}
auto src = iptr + tId * sizeDivide * mStride;
for (int i = 0; i < number; i++) {
hist_ptr[src[i * mStride]]++;
}
}
MNN_CONCURRENCY_END();
memset(optr, 0, mBinNum * sizeof(float));
for (int i = std::max(mMin, 0); i <= std::min(mMax, 255); i++) {
int bin = std::min((int)(i * mAlpha - mBeta), mBinNum -1);
optr[bin] = hist_map[i];
}
return NO_ERROR;
}
ErrorCode CPUHistogram::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
auto input = inputs[0], output = outputs[0];
/*
1. mAlpha, mBeta
binIdx = (val - mMin) / (mMax - mMin) * mBinNum
mAlpha = mBinNum / (mMax - mMin)
mBeta = mAlpha * mMin
-> binIdx = val * mAlpha - mBeta
2. mChannel, mSize, mStride
mChannel < 0 : compute all element of input
mChannel >= 0 : last dim as channel, and other dim is plane; compute the mChannel plane of input;
*/
if (mChannel < 0) {
mSize = input->elementSize();
mStride = 1;
mChannel = 0;
} else {
mSize = 1;
int lastDim = input->dimensions() - 1;
for (int i = 0; i < lastDim; i++) {
mSize *= input->length(i);
}
mStride = input->length(lastDim);
MNN_ASSERT(mChannel <= mStride);
mChannel = std::min(mChannel, mStride);
}
if (input->getType() == halide_type_of<float>()) {
return histogram<float>(input, output);
}
if (input->getType() == halide_type_of<int>()) {
return histogram<int>(input, output);
}
if (input->getType() == halide_type_of<uint8_t>()) {
return histogram<uint8_t>(input, output);
}
return NOT_SUPPORT;
}
class CPUHistogramCreator : 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 CPUHistogram(backend, op);
}
};
REGISTER_CPU_OP_CREATOR(CPUHistogramCreator, OpType_Histogram);
} // namespace MNN