提交 b45d020f 编写于 作者: H hedaoyuan 提交者: GitHub

Merge pull request #3718 from hedaoyuan/convolution

Depthwise Convolution Optimization
......@@ -21,6 +21,8 @@ if(USE_NNPACK)
endif()
endif()
list(APPEND cpp_files neon/NeonDepthwiseConv.cpp)
add_library(paddle_function STATIC ${cpp_files} ${cu_objs})
add_dependencies(paddle_function ${external_project_dependencies})
add_dependencies(paddle_function paddle_proto)
......@@ -42,11 +44,11 @@ if(WITH_GPU)
add_simple_unittest(RowConvOpTest)
add_simple_unittest(BlockExpandOpTest)
add_simple_unittest(CropOpTest)
add_simple_unittest(DepthwiseConvOpTest)
endif()
add_simple_unittest(Im2ColTest)
add_simple_unittest(GemmConvOpTest)
add_simple_unittest(DepthwiseConvOpTest)
endif()
add_style_check_target(paddle_function ${h_files})
......
......@@ -34,4 +34,13 @@ TEST(DepthwiseConv, BackwardFilter) {
}
#endif
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
TEST(DepthwiseConv, Forward) {
DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
"GemmConv-CPU", "NeonDepthwiseConv-CPU", forward);
}
#endif
} // namespace paddle
......@@ -16,6 +16,7 @@ limitations under the License. */
#include "TensorShape.h"
#include "TensorType.h"
#include "neon/neon_util.h"
namespace paddle {
......@@ -93,4 +94,95 @@ 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
此差异已折叠。
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include <arm_neon.h>
namespace paddle {
namespace neon {
inline float32x4_t vld1q_f32_aligned(const float* p) {
return vld1q_f32(
(const float*)__builtin_assume_aligned(p, sizeof(float32x4_t)));
}
#ifndef __aarch64__
inline float32_t vaddvq_f32(float32x4_t a) {
float32x2_t v = vadd_f32(vget_high_f32(a), vget_low_f32(a));
return vget_lane_f32(vpadd_f32(v, v), 0);
}
inline float32x4_t vmlaq_laneq_f32(float32x4_t a,
float32x4_t b,
float32x4_t v,
const int lane) {
return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
}
#endif
} // namespace neon
} // namespace paddle
#endif
......@@ -29,6 +29,10 @@ namespace paddle {
REGISTER_LAYER(exconv, ExpandConvLayer);
REGISTER_LAYER(exconvt, ExpandConvLayer);
inline bool isDepthwiseConv(int channels, int groups) {
return channels == groups;
}
bool ExpandConvLayer::init(const LayerMap &layerMap,
const ParameterMap &parameterMap) {
/* Initialize the basic convolutional parent class */
......@@ -47,14 +51,27 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]};
std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]};
if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) {
// Convolution Layer uses the GemmConv function by default.
convType = "GemmConv";
convGradInputType = "GemmConvGradInput";
convGradFilterType = "GemmConvGradFilter";
// If depth wise convolution and useGpu == true
if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
convType = "DepthwiseConv";
convGradInputType = "DepthwiseConvGradInput";
convGradFilterType = "DepthwiseConvGradFilter";
} else {
convType = "GemmConv";
convGradInputType = "GemmConvGradInput";
convGradFilterType = "GemmConvGradFilter";
}
// If depth wise convolution and useGpu == false and ARM-NEON
if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
if ((filterSize_[i] == filterSizeY_[i]) &&
(filterSize_[i] == 3 || filterSize_[i] == 4) &&
(stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) {
convType = "NeonDepthwiseConv";
}
#endif
}
if (FLAGS_use_nnpack && !isDeconv_) {
......
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