提交 40d47fae 编写于 作者: H hedaoyuan

[Refine code]Move class Padding into the NeonDepthwiseConv.h.

上级 4b6b7251
......@@ -94,95 +94,4 @@ public:
int paddingWidth);
};
template <class T>
struct Padding {
static void run(const T* src,
T* dest,
int channels,
int inputHeight,
int inputWidth,
int paddingHeight,
int paddingWidth) {
const int destWidth = inputWidth + 2 * paddingWidth;
for (int c = 0; c < channels; c++) {
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(T));
dest += destWidth * paddingHeight;
}
for (int i = 0; i < inputHeight; i++) {
// padding head
for (int j = 0; j < paddingWidth; j++) {
*dest++ = T(0);
}
memcpy(dest, src, inputWidth * sizeof(T));
dest += inputWidth;
src += inputWidth;
// padding tail
for (int j = 0; j < paddingWidth; j++) {
*dest++ = T(0);
}
}
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(T));
dest += destWidth * paddingHeight;
}
}
}
};
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
template <>
struct Padding<float> {
static void run(const float* src,
float* dest,
int channels,
int inputHeight,
int inputWidth,
int paddingHeight,
int paddingWidth) {
const int destWidth = inputWidth + 2 * paddingWidth;
for (int c = 0; c < channels; c++) {
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(float));
dest += destWidth * paddingHeight;
}
for (int i = 0; i < inputHeight; i++) {
// padding head
for (int j = 0; j < paddingWidth; j++) {
*dest++ = float(0);
}
int step = inputWidth >> 2;
int remain = inputWidth & 3;
for (int s = 0; s < step; s++) {
float32x4_t s0 = vld1q_f32(src);
vst1q_f32(dest, s0);
src += 4;
dest += 4;
}
for (int r = 0; r < remain; r++) {
*dest++ = *src++;
}
// padding tail
for (int j = 0; j < paddingWidth; j++) {
*dest++ = float(0);
}
}
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(float));
dest += destWidth * paddingHeight;
}
}
}
};
#endif
} // namespace paddle
......@@ -14,7 +14,6 @@ limitations under the License. */
#include "NeonDepthwiseConv.h"
#include "paddle/function/ConvOp.h"
#include "paddle/function/Im2Col.h"
namespace paddle {
......@@ -70,13 +69,13 @@ public:
(inputWidth + 2 * paddingW());
resizeBuffer<Device>(newSize);
inputPadding = reinterpret_cast<float*>(memory_->getBuf());
Padding<float>::run(inputData,
inputPadding,
batchSize * inputChannels,
inputHeight,
inputWidth,
paddingH(),
paddingW());
neon::Padding<float>::run(inputData,
inputPadding,
batchSize * inputChannels,
inputHeight,
inputWidth,
paddingH(),
paddingW());
// height and width of padding data
inputHeight += 2 * paddingH();
......
......@@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <string.h>
#include "neon_util.h"
namespace paddle {
......@@ -474,6 +475,97 @@ struct DepthwiseConvKernel<4, 2> {
}
};
template <class T>
struct Padding {
static void run(const T* src,
T* dest,
int channels,
int inputHeight,
int inputWidth,
int paddingHeight,
int paddingWidth) {
const int destWidth = inputWidth + 2 * paddingWidth;
for (int c = 0; c < channels; c++) {
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(T));
dest += destWidth * paddingHeight;
}
for (int i = 0; i < inputHeight; i++) {
// padding head
for (int j = 0; j < paddingWidth; j++) {
*dest++ = T(0);
}
memcpy(dest, src, inputWidth * sizeof(T));
dest += inputWidth;
src += inputWidth;
// padding tail
for (int j = 0; j < paddingWidth; j++) {
*dest++ = T(0);
}
}
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(T));
dest += destWidth * paddingHeight;
}
}
}
};
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
template <>
struct Padding<float> {
static void run(const float* src,
float* dest,
int channels,
int inputHeight,
int inputWidth,
int paddingHeight,
int paddingWidth) {
const int destWidth = inputWidth + 2 * paddingWidth;
for (int c = 0; c < channels; c++) {
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(float));
dest += destWidth * paddingHeight;
}
for (int i = 0; i < inputHeight; i++) {
// padding head
for (int j = 0; j < paddingWidth; j++) {
*dest++ = float(0);
}
int step = inputWidth >> 2;
int remain = inputWidth & 3;
for (int s = 0; s < step; s++) {
float32x4_t s0 = vld1q_f32(src);
vst1q_f32(dest, s0);
src += 4;
dest += 4;
}
for (int r = 0; r < remain; r++) {
*dest++ = *src++;
}
// padding tail
for (int j = 0; j < paddingWidth; j++) {
*dest++ = float(0);
}
}
if (paddingHeight > 0) {
memset(dest, 0, destWidth * paddingHeight * sizeof(float));
dest += destWidth * paddingHeight;
}
}
}
};
#endif
#endif
} // namespace neon
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册