mirror of https://github.com/alibaba/MNN.git
209 lines
8.3 KiB
C++
209 lines
8.3 KiB
C++
|
//
|
||
|
|
// CPUPoolInt8.cpp
|
||
|
|
// MNN
|
||
|
|
//
|
||
|
|
// Created by MNN on 2019/06/10.
|
||
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
||
|
|
//
|
||
|
|
|
||
|
|
#include "CPUPoolInt8.hpp"
|
||
|
|
#include "Macro.h"
|
||
|
|
|
||
|
|
#ifdef MNN_USE_NEON
|
||
|
|
#include <arm_neon.h>
|
||
|
|
#endif
|
||
|
|
#include "Concurrency.h"
|
||
|
|
|
||
|
|
#define UNIT 4
|
||
|
|
|
||
|
|
namespace MNN {
|
||
|
|
|
||
|
|
static void poolingInt8Max(int8_t *dst, int outputWidth, int outputHeight, const int8_t *src, int inputWidth,
|
||
|
|
int inputHeight, int kernelWidth, int kernelHeight, int strideWidth, int strideHeight,
|
||
|
|
int padWidth, int padHeight) {
|
||
|
|
int8_t maxValue[2 * UNIT];
|
||
|
|
for (int oy = 0; oy < outputHeight; ++oy) {
|
||
|
|
for (int ox = 0; ox < outputWidth; ++ox) {
|
||
|
|
memset(maxValue, INT8_MIN, 2 * UNIT * sizeof(int8_t));
|
||
|
|
const int srcOriginX = ox * strideWidth - padWidth;
|
||
|
|
const int srcOriginY = oy * strideHeight - padHeight;
|
||
|
|
const int kxs = std::max(0, -srcOriginX);
|
||
|
|
const int kxe = std::min(kernelWidth, inputWidth - srcOriginX);
|
||
|
|
const int kys = std::max(0, -srcOriginY);
|
||
|
|
const int kye = std::min(kernelHeight, inputHeight - srcOriginY);
|
||
|
|
|
||
|
|
const auto srcPtr = src + (srcOriginY * inputWidth + srcOriginX) * UNIT;
|
||
|
|
auto dstPtr = dst + (oy * outputWidth + ox) * UNIT;
|
||
|
|
// find kernel_w * kernel_h max value
|
||
|
|
for (int ky = kys; ky < kye; ++ky) {
|
||
|
|
const auto srcPtrRow = srcPtr + ky * inputWidth + kxs;
|
||
|
|
int kx = kxs;
|
||
|
|
#ifdef MNN_USE_NEON
|
||
|
|
// process two data together
|
||
|
|
int8x8_t max_reg = vld1_s8(maxValue);
|
||
|
|
for (; kx < kxe - 2; kx += 2) {
|
||
|
|
const auto srcPtrStart = srcPtrRow + kx * UNIT;
|
||
|
|
int8x8_t input_reg = vld1_s8(srcPtrStart);
|
||
|
|
max_reg = vmax_s8(max_reg, input_reg);
|
||
|
|
vst1_s8(maxValue, max_reg);
|
||
|
|
}
|
||
|
|
for (int j = 0; j < UNIT; ++j) {
|
||
|
|
maxValue[j] = std::max(maxValue[j], maxValue[j + UNIT]);
|
||
|
|
}
|
||
|
|
#else
|
||
|
|
for (; kx < kxe; ++kx) {
|
||
|
|
const auto srcPtrStart = srcPtrRow + kx * UNIT;
|
||
|
|
for (int j = 0; j < UNIT; ++j) {
|
||
|
|
maxValue[j] = std::max(maxValue[j], srcPtrStart[j]);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
#endif
|
||
|
|
}
|
||
|
|
// output
|
||
|
|
memcpy(dstPtr, maxValue, UNIT * sizeof(int8_t));
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
static void poolingInt8Avg(int8_t *dst, int outputWidth, int outputHeight, const int8_t *src, int inputWidth,
|
||
|
|
int inputHeight, int kernelWidth, int kernelHeight, int strideWidth, int strideHeight,
|
||
|
|
int padWidth, int padHeight) {
|
||
|
|
int16_t sum[2 * UNIT];
|
||
|
|
for (int oy = 0; oy < outputHeight; ++oy) {
|
||
|
|
for (int ox = 0; ox < outputWidth; ++ox) {
|
||
|
|
memset(sum, 0, 2 * UNIT * sizeof(int16_t));
|
||
|
|
const int srcOriginX = ox * strideWidth - padWidth;
|
||
|
|
const int srcOriginY = oy * strideHeight - padHeight;
|
||
|
|
const int kxs = std::max(0, -srcOriginX);
|
||
|
|
const int kxe = std::min(kernelWidth, inputWidth - srcOriginX);
|
||
|
|
const int kys = std::max(0, -srcOriginY);
|
||
|
|
const int kye = std::min(kernelHeight, inputHeight - srcOriginY);
|
||
|
|
const int kernelCount = (kxe - kxs) * (kye - kys);
|
||
|
|
|
||
|
|
const auto srcPtr = src + (srcOriginY * inputWidth + srcOriginX) * UNIT;
|
||
|
|
auto dstPtr = dst + (oy * outputWidth + ox) * UNIT;
|
||
|
|
// compute kernel_w * kernel_h sum
|
||
|
|
for (int ky = kys; ky < kye; ++ky) {
|
||
|
|
const auto srcPtrRow = srcPtr + ky * inputWidth + kxs;
|
||
|
|
int kx = kxs;
|
||
|
|
#ifdef MNN_USE_NEON
|
||
|
|
// process two data together
|
||
|
|
int16x8_t sum_reg = vld1q_s16(sum);
|
||
|
|
for (; kx < kxe - 2; kx += 2) {
|
||
|
|
const auto srcPtrStart = srcPtrRow + kx * UNIT;
|
||
|
|
int8x8_t input_reg = vld1_s8(srcPtrStart);
|
||
|
|
sum_reg = vaddw_s8(sum_reg, input_reg);
|
||
|
|
vst1_s16(sum, vadd_s16(vget_high_s16(sum_reg), vget_low_s16(sum_reg)));
|
||
|
|
}
|
||
|
|
#else
|
||
|
|
for (; kx < kxe; ++kx) {
|
||
|
|
const auto srcPtrStart = srcPtrRow + kx * UNIT;
|
||
|
|
for (int j = 0; j < UNIT; ++j) {
|
||
|
|
sum[j] += srcPtrStart[j];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
#endif
|
||
|
|
}
|
||
|
|
// avg
|
||
|
|
for (int j = 0; j < UNIT; ++j) {
|
||
|
|
sum[j + UNIT] =
|
||
|
|
sum[j] > 0 ? (sum[j] + kernelCount / 2) / kernelCount : (sum[j] - kernelCount / 2) / kernelCount;
|
||
|
|
dstPtr[j] = static_cast<int8_t>(sum[j + UNIT]);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
CPUPoolInt8::CPUPoolInt8(Backend *b, const Pool *parameter) : Execution(b), mParameter(parameter) {
|
||
|
|
}
|
||
|
|
|
||
|
|
ErrorCode CPUPoolInt8::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
||
|
|
const auto input = inputs[0];
|
||
|
|
auto output = outputs[0];
|
||
|
|
|
||
|
|
int strideWidth = mParameter->strideX();
|
||
|
|
int strideHeight = mParameter->strideY();
|
||
|
|
int padWidth = mParameter->padX();
|
||
|
|
int padHeight = mParameter->padY();
|
||
|
|
int kernelWidth = mParameter->kernelX();
|
||
|
|
int kernelHeight = mParameter->kernelY();
|
||
|
|
|
||
|
|
const int inputWidth = input->width();
|
||
|
|
const int inputHeight = input->height();
|
||
|
|
const int outputWidth = output->width();
|
||
|
|
const int outputHeight = output->height();
|
||
|
|
|
||
|
|
kernelWidth = std::min(kernelWidth, inputWidth);
|
||
|
|
kernelHeight = std::min(kernelHeight, inputHeight);
|
||
|
|
if (mParameter->isGlobal()) {
|
||
|
|
kernelWidth = inputWidth;
|
||
|
|
kernelHeight = inputHeight;
|
||
|
|
strideWidth = inputWidth;
|
||
|
|
strideHeight = inputHeight;
|
||
|
|
padWidth = 0;
|
||
|
|
padHeight = 0;
|
||
|
|
}
|
||
|
|
if (mParameter->padType() == PoolPadType_SAME) {
|
||
|
|
int padNeededWidth = (outputWidth - 1) * strideWidth + kernelWidth - inputWidth;
|
||
|
|
int padNeededHeight = (outputHeight - 1) * strideHeight + kernelHeight - inputHeight;
|
||
|
|
padWidth = padNeededWidth > 0 ? padNeededWidth / 2 : 0;
|
||
|
|
padHeight = padNeededHeight > 0 ? padNeededHeight / 2 : 0;
|
||
|
|
}
|
||
|
|
|
||
|
|
const int inputPlaneStride = inputWidth * inputHeight * 4;
|
||
|
|
const int outputPlaneStride = outputWidth * outputHeight * 4;
|
||
|
|
|
||
|
|
auto planeFunc = poolingInt8Max;
|
||
|
|
if (mParameter->type() == MNN::PoolType_AVEPOOL) {
|
||
|
|
planeFunc = poolingInt8Avg;
|
||
|
|
}
|
||
|
|
|
||
|
|
const int channelDiv4 = UP_DIV(input->channel(), 4);
|
||
|
|
const int threadNumer = static_cast<CPUBackend *>(backend())->threadNumber();
|
||
|
|
mThreadNumber = std::min(threadNumer, channelDiv4);
|
||
|
|
|
||
|
|
mThreadFunction = [=](int tId, const int8_t *src, int8_t *dst) {
|
||
|
|
for (int depth = tId; depth < channelDiv4; depth += mThreadNumber) {
|
||
|
|
planeFunc(dst + depth * outputPlaneStride, outputWidth, outputHeight, src + depth * inputPlaneStride,
|
||
|
|
inputWidth, inputHeight, kernelWidth, kernelHeight, strideWidth, strideHeight, padWidth,
|
||
|
|
padHeight);
|
||
|
|
}
|
||
|
|
};
|
||
|
|
|
||
|
|
return NO_ERROR;
|
||
|
|
}
|
||
|
|
|
||
|
|
ErrorCode CPUPoolInt8::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
||
|
|
auto input = inputs[0];
|
||
|
|
auto output = outputs[0];
|
||
|
|
const int batch = input->batch();
|
||
|
|
const int src_b_step = input->stride(0);
|
||
|
|
const int dst_b_step = output->stride(0);
|
||
|
|
|
||
|
|
const auto inputPtr = input->host<int8_t>();
|
||
|
|
auto outputPtr = output->host<int8_t>();
|
||
|
|
|
||
|
|
for (int bIndex = 0; bIndex < batch; ++bIndex) {
|
||
|
|
const auto srcOrigin = inputPtr + bIndex * src_b_step;
|
||
|
|
auto dstOrigin = outputPtr + bIndex * dst_b_step;
|
||
|
|
|
||
|
|
MNN_CONCURRENCY_BEGIN(tId, mThreadNumber) {
|
||
|
|
mThreadFunction((int)tId, srcOrigin, dstOrigin);
|
||
|
|
}
|
||
|
|
MNN_CONCURRENCY_END();
|
||
|
|
}
|
||
|
|
return NO_ERROR;
|
||
|
|
}
|
||
|
|
|
||
|
|
class CPUPoolInt8Creator : 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 CPUPoolInt8(backend, op->main_as_Pool());
|
||
|
|
}
|
||
|
|
};
|
||
|
|
|
||
|
|
REGISTER_CPU_OP_CREATOR(CPUPoolInt8Creator, OpType_PoolInt8);
|
||
|
|
|
||
|
|
} // namespace MNN
|