Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 332
Star 4
Fork 1
P
Paddle-Lite
提交 00bbadc1 编写于 作者: Y yangfei
浏览文件
操作

optimize depthwise_conv_3x3

上级 2a178e98
隐藏空白更改
内联 并排
Showing with 109 addition and 224 deletion
src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl
浏览文件 @ 00bbadc1
......@@ -217,68 +217,77 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
half4 output = 0.0f;
#endif
int2 pos_in_input_block = (int2)(out_c * input_width, batch_index * input_height);
int weight_y_to = out_c * 12;
half4 inputs[9];
const int filter_width = 3;
const int filter_height = 3;
inputs[0] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[1] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
int2 pos_in_input_block = (int2)(out_c * input_width, batch_index * input_height);
inputs[2] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
int2 pos_in_filter_block = (int2)(out_c * filter_width, batch_index * filter_height);
inputs[3] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
/*
if (output_pos.x == 112 && output_pos.y == 0) {
half4 input1 = inputs[3];
float4 in = (float4)(input1.x, input1.y, input1.z, input1.w);
printf(" input4 3 - %v4hlf \n", in);
printf(" --- %d ---\n", in_pos_in_one_block.x - 1);
}
*/
int filter_x = pos_in_filter_block.x ;
int filter_y = pos_in_filter_block.y ;
half4 inputs[9];
inputs[4] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[5] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[6] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[7] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[8] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 weight0 = read_imageh(filter, sampler, (int2)(j % 3, weight_y_to + j / 3));
half4 weight1 = read_imageh(filter, sampler, (int2)(j % 3, weight_y_to + 3 + j / 3));
half4 weight2 = read_imageh(filter, sampler, (int2)(j % 3, weight_y_to + 6 + j / 3));
half4 weight3 = read_imageh(filter, sampler, (int2)(j % 3, weight_y_to + 9 + j / 3));
output.x += input.x * weight0.x;
output.y += input.y * weight1.x;
output.z += input.z * weight2.x;
output.w += input.w * weight3.x;
inputs[0] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[1] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[2] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[3] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
/*
if (output_pos.x == 112 && output_pos.y == 0) {
half4 input1 = inputs[3];
float4 in = (float4)(input1.x, input1.y, input1.z, input1.w);
printf(" input4 3 - %v4hlf \n", in);
printf(" --- %d ---\n", in_pos_in_one_block.x - 1);
}
*/
inputs[4] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[5] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[6] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[7] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[8] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
half4 filters[9];
filters[0] = read_imageh(filter, sampler,(int2)(filter_x,filter_y));
filters[1] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y));
filters[2] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y));
filters[3] = read_imageh(filter, sampler,(int2)(filter_x,filter_y + 1));
filters[4] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y + 1));
filters[5] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y + 1));
filters[6] = read_imageh(filter, sampler,(int2)(filter_x,filter_y + 2));
filters[7] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y + 2));
filters[8] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y + 2));
for(int i = 0 ;i < 9 ; i++){
output += inputs[i] * filters[i];
}
#ifdef BATCH_NORM
output = output * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + read_imageh(new_biase, sampler, (int2)(out_c, 0));
#endif
......
src/operators/kernel/cl/cl_kernel/depthwise_conv_add_bn_relu_kernel.cl 0 → 100644
浏览文件 @ 00bbadc1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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. */
#define BIASE
#define BATCH_NORM
#define RELU
#include "conv_kernel.inc.cl"
\ No newline at end of file
src/operators/kernel/cl/cl_kernel/depthwise_conv_kernel.cl
浏览文件 @ 00bbadc1
#define BIASE
#define BATCH_NORM
#define RELU
#include "cl_common.h"
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
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
__kernel void depth_conv_3x3(__private const int global_size_dim0,
__private const int global_size_dim1,
__private const int global_size_dim2,
__read_only image2d_t input,
__read_only image2d_t filter,
#ifdef BIASE
__read_only image2d_t bias,
#endif
#ifdef BATCH_NORM
__read_only image2d_t new_scale,
__read_only image2d_t new_biase,
#endif
__write_only image2d_t output_image,
__private const int stride,
__private const int offset,
__private const int input_c,
__private const int dilation,
__private const int input_width,/* of one block */
__private const int input_height, /* of one block */
__private const int output_width,
__private const int output_height,
__private const int filter_width,
__private const int filter_height) {
http://www.apache.org/licenses/LICENSE-2.0
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
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. */
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
const int batch_index = out_nh / output_height;
const int out_nh_in_one_batch = out_nh % output_height;
int2 stride_xy = (int2)(stride, stride);
int2 ouput_pos_in_one_block = (int2)(out_w, out_nh_in_one_batch);
int2 in_pos_in_one_block = ouput_pos_in_one_block * stride_xy + (int2)(offset, offset);
#ifdef BIASE
half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
#else
half4 output = 0.0f;
#endif
int2 pos_in_input_block = (int2)(out_c * input_width, batch_index * input_height);
int2 pos_in_filter_block = (int2)(out_c * filter_width, batch_index * filter_height);
int filter_x = pos_in_filter_block.x ;
int filter_y = pos_in_filter_block.y ;
half4 inputs[9];
inputs[0] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[1] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[2] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y - 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y - 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y - 1 >= input_height) << 15));
inputs[3] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
/*
if (output_pos.x == 112 && output_pos.y == 0) {
half4 input1 = inputs[3];
float4 in = (float4)(input1.x, input1.y, input1.z, input1.w);
printf(" input4 3 - %v4hlf \n", in);
printf(" --- %d ---\n", in_pos_in_one_block.x - 1);
}
*/
inputs[4] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[5] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y >= input_height) << 15));
inputs[6] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x - 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x - 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x - 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[7] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
inputs[8] = select(read_imageh(input, sampler, (int2)(pos_in_input_block.x + in_pos_in_one_block.x + 1, pos_in_input_block.y + in_pos_in_one_block.y + 1)),
(half4)(0.0f),
(ushort4)((in_pos_in_one_block.x + 1 < 0 || in_pos_in_one_block.y + 1 < 0 || in_pos_in_one_block.x + 1 >= input_width || in_pos_in_one_block.y + 1 >= input_height) << 15));
half4 filters[9];
filters[0] = read_imageh(filter, sampler,(int2)(filter_x,filter_y));
filters[1] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y));
filters[2] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y));
filters[3] = read_imageh(filter, sampler,(int2)(filter_x,filter_y + 1));
filters[4] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y + 1));
filters[5] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y + 1));
filters[6] = read_imageh(filter, sampler,(int2)(filter_x,filter_y + 2));
filters[7] = read_imageh(filter, sampler,(int2)(filter_x + 1,filter_y + 2));
filters[8] = read_imageh(filter, sampler,(int2)(filter_x + 2,filter_y + 2));
for(int i = 0 ;i < 9 ; i++){
output += inputs[i] * filters[i];
}
#ifdef BATCH_NORM
output = output * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + read_imageh(new_biase, sampler, (int2)(out_c, 0));
#endif
#ifdef RELU
output = activation(output);
#endif
/*
if (output_pos.x == 112 && output_pos.y == 0) {
for (int i = 0; i < 9; ++i) {
half4 input1 = inputs[i];
float4 in = (float4)(input1.x, input1.y, input1.z, input1.w);
printf(" input4 %d - %v4hlf \n", i, in);
}
float4 out = (float4)(output.x, output.y, output.z, output.w);
printf(" depth wise output output4 = %v4hlf \n", out);
printf(" pos_in_input_block -x %d \n ", pos_in_input_block.x);
printf(" pos_in_input_block -y %d \n ", pos_in_input_block.y);
printf(" in_pos_in_one_block - x %d \n", in_pos_in_one_block.x);
printf(" in_pos_in_one_block - y %d \n", in_pos_in_one_block.y);
}
*/
write_imageh(output_image, output_pos, output);
}
\ No newline at end of file
#include "conv_kernel.inc.cl"
\ No newline at end of file
src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp
浏览文件 @ 00bbadc1
......@@ -144,7 +144,7 @@ bool ConvAddBNReluKernel<GPU_CL, float>::Init(
param->Filter()->dims()[2] == 3) {
// this->cl_helper_.AddKernel("depth_conv_3x3",
// "conv_add_bn_relu_kernel.cl");
this->cl_helper_.AddKernel("depth_conv_3x3", "depthwise_conv_kernel.cl");
this->cl_helper_.AddKernel("depth_conv_3x3", "depthwise_conv_add_bn_relu_kernel.cl");
DLOG << " conv add bn relu depth_conv_3x3";
} else if (param->Filter()->WidthOfOneBlock() == 3 &&
param->Filter()->HeightOfOneBlock() == 3) {
......@@ -179,8 +179,6 @@ void ConvAddBNReluKernel<GPU_CL, float>::Compute(
int input_height = param.Input()->HeightOfOneBlock();
int output_width = param.Output()->WidthOfOneBlock();
int output_height = param.Output()->HeightOfOneBlock();
int filter_width = param.Filter()->WidthOfOneBlock();
int filter_height = param.Filter()->HeightOfOneBlock();
// DLOG << " c block " << c_block;
// DLOG << " w " << w;
......@@ -250,15 +248,6 @@ void ConvAddBNReluKernel<GPU_CL, float>::Compute(
status = clSetKernelArg(kernel, 16, sizeof(int), &output_height);
CL_CHECK_ERRORS(status);
if (param.Filter()->dims()[0] == 1 &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == 3) {
status = clSetKernelArg(kernel, 17, sizeof(int), &filter_width);
CL_CHECK_ERRORS(status);
status = clSetKernelArg(kernel, 18, sizeof(int), &filter_height);
CL_CHECK_ERRORS(status);
}
// cl_event out_event = param.Output()->GetClEvent();
// cl_event wait_event = param.Input()->GetClEvent();
......
src/operators/kernel/cl/conv_kernel.cpp
浏览文件 @ 00bbadc1
......@@ -26,6 +26,15 @@ bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
param->Paddings()[0] == param->Paddings()[1],
"need equal");
auto filter_ddim = param->Filter()->dims();
std::vector<int64_t> filter_shape(
{filter_ddim[1], filter_ddim[0], filter_ddim[2], filter_ddim[3]});
framework::DDim ddim = framework::make_ddim(filter_shape);
if (filter_ddim[1] == 1) {
param->Filter()->Resize(ddim);
}
param->Filter()->InitCLImage(cl_helper_.CLContext(),
this->cl_helper_.CLCommandQueue());
......@@ -44,9 +53,11 @@ bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
DLOG << " here1 ";
this->cl_helper_.AddKernel("conv_1x1", "conv_kernel.cl");
} else if (param->Filter()->dims()[1] == 1) {
} else if (param->Filter()->dims()[0] == 1 &&
param->Input()->dims()[1] == param->Output()->dims()[1] &&
param->Filter()->dims()[2] == 3) {
DLOG << " here2 ";
this->cl_helper_.AddKernel("depth_conv_3x3", "conv_kernel.cl");
this->cl_helper_.AddKernel("depth_conv_3x3", "depthwise_conv_kernel.cl");
} else if (param->Filter()->WidthOfOneBlock() == 3 &&
param->Filter()->HeightOfOneBlock() == 3) {
......@@ -111,6 +122,7 @@ void ConvKernel<GPU_CL, float>::Compute(const ConvParam<GPU_CL> &param) {
status = clSetKernelArg(kernel, 12, sizeof(int), &output_width);
status = clSetKernelArg(kernel, 13, sizeof(int), &output_height);
// cl_event out_event = param.Output()->GetClEvent();
// cl_event wait_event = param.Input()->GetClEvent();
......
Markdown is supported
0% 或 .
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 或