Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 e82ebee8 编写于 作者: X xiebaiyuan 提交者： GitHub
浏览文件
操作

optimise conv 1x1 ,test=develop (#2248)

上级 8006f5e6
显示空白变更内容
内联 并排
Showing with 62 addition and 103 deletion
mobile/src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl
浏览文件 @ e82ebee8
...@@ -1093,138 +1093,97 @@ __kernel void conv_1x1_spl( ...@@ -1093,138 +1093,97 @@ __kernel void conv_1x1_spl(
half4 weight2 = read_imageh(filter, sampler, (int2)(out_c, i * 4 + 2)); half4 weight2 = read_imageh(filter, sampler, (int2)(out_c, i * 4 + 2));
half4 weight3 = read_imageh(filter, sampler, (int2)(out_c, i * 4 + 3)); half4 weight3 = read_imageh(filter, sampler, (int2)(out_c, i * 4 + 3));
if ((max_w_bound - pos_in.x-1) < input_width && (max_w_bound - pos_in.x-1)>=0 ){ int bound_gap = max_w_bound - pos_in.x - 1;
if (bound_gap < input_width && bound_gap >= 0){
if (burndary_index==0){ if (burndary_index==0){
output0 = mad(input0.x, weight0, output0); // do nothing
output0 = mad(input0.y, weight1, output0);
output0 = mad(input0.z, weight2, output0);
output0 = mad(input0.w, weight3, output0);
} else if (burndary_index==1){ } else if (burndary_index==1){
output0 = mad(input0.x, weight0, output0); input0.w = 0.0f;
output0 = mad(input0.y, weight1, output0);
output0 = mad(input0.z, weight2, output0);
output0 = mad(0.0f, weight3, output0);
} else if (burndary_index==2){ } else if (burndary_index==2){
output0 = mad(input0.x, weight0, output0); input0.z = 0.0f;
output0 = mad(input0.y, weight1, output0); input0.w = 0.0f;
output0 = mad(0.0f, weight2, output0);
output0 = mad(0.0f, weight3, output0);
} else if (burndary_index==3){ } else if (burndary_index==3){
output0 = mad(input0.x, weight0, output0); input0.y = 0.0f;
output0 = mad(0.0f, weight1, output0); input0.z = 0.0f;
output0 = mad(0.0f, weight2, output0); input0.w = 0.0f;
output0 = mad(0.0f, weight3, output0); }
} }
}else {
output0 = mad(input0.x, weight0, output0); output0 = mad(input0.x, weight0, output0);
output0 = mad(input0.y, weight1, output0); output0 = mad(input0.y, weight1, output0);
output0 = mad(input0.z, weight2, output0); output0 = mad(input0.z, weight2, output0);
output0 = mad(input0.w, weight3, output0); output0 = mad(input0.w, weight3, output0);
}
// -------------1-------------- // -------------1--------------
pos_in = (int2)(i * input_width + in_pos_in_one_block1.x, in_pos_in_one_block1.y); pos_in = (int2)(i * input_width + in_pos_in_one_block1.x, in_pos_in_one_block1.y);
half4 input1 = read_imageh(input_image, sampler, pos_in); half4 input1 = read_imageh(input_image, sampler, pos_in);
if (abs(max_w_bound - pos_in.x) < input_width){ bound_gap = max_w_bound - pos_in.x - 1;
if (bound_gap < input_width && bound_gap >= 0){
if (burndary_index==0){ if (burndary_index==0){
output1 = mad(input1.x, weight0, output1); // do nothing
output1 = mad(input1.y, weight1, output1);
output1 = mad(input1.z, weight2, output1);
output1 = mad(input1.w, weight3, output1);
} else if (burndary_index==1){ } else if (burndary_index==1){
output1 = mad(input1.x, weight0, output1); input1.w = 0.0f;
output1 = mad(input1.y, weight1, output1);
output1 = mad(input1.z, weight2, output1);
output1 = mad(0.0f, weight3, output1);
} else if (burndary_index==2){ } else if (burndary_index==2){
output1 = mad(input1.x, weight0, output1); input1.z = 0.0f;
output1 = mad(input1.y, weight1, output1); input1.w = 0.0f;
output1 = mad(0.0f, weight2, output1);
output1 = mad(0.0f, weight3, output1);
} else if (burndary_index==3){ } else if (burndary_index==3){
output1 = mad(input1.x, weight0, output1); input1.y = 0.0f;
output1 = mad(0.0f, weight1, output1); input1.z = 0.0f;
output1 = mad(0.0f, weight2, output1); input1.w = 0.0f;
output1 = mad(0.0f, weight3, output1); }
} }
}else {
output1 = mad(input1.x, weight0, output1); output1 = mad(input1.x, weight0, output1);
output1 = mad(input1.y, weight1, output1); output1 = mad(input1.y, weight1, output1);
output1 = mad(input1.z, weight2, output1); output1 = mad(input1.z, weight2, output1);
output1 = mad(input1.w, weight3, output1); output1 = mad(input1.w, weight3, output1);
}
// -------------2-------------- // -------------2--------------
pos_in = (int2)(i * input_width + in_pos_in_one_block2.x, in_pos_in_one_block2.y); pos_in = (int2)(i * input_width + in_pos_in_one_block2.x, in_pos_in_one_block2.y);
half4 input2 = read_imageh(input_image, sampler, pos_in); half4 input2 = read_imageh(input_image, sampler, pos_in);
if (abs(max_w_bound - pos_in.x) < input_width){ bound_gap = max_w_bound - pos_in.x - 1;
if (bound_gap < input_width && bound_gap >= 0){
if (burndary_index==0){ if (burndary_index==0){
output2 = mad(input2.x, weight0, output2); // do nothing
output2 = mad(input2.y, weight1, output2);
output2 = mad(input2.z, weight2, output2);
output2 = mad(input2.w, weight3, output2);
} else if (burndary_index==1){ } else if (burndary_index==1){
output2 = mad(input2.x, weight0, output2); input2.w = 0.0f;
output2 = mad(input2.y, weight1, output2);
output2 = mad(input2.z, weight2, output2);
output2 = mad(0.0f, weight3, output2);
} else if (burndary_index==2){ } else if (burndary_index==2){
output2 = mad(input2.x, weight0, output2); input2.z = 0.0f;
output2 = mad(input2.y, weight1, output2); input2.w = 0.0f;
output2 = mad(0.0f, weight2, output2);
output2 = mad(0.0f, weight3, output2);
} else if (burndary_index==3){ } else if (burndary_index==3){
output2 = mad(input2.x, weight0, output2); input2.y = 0.0f;
output2 = mad(0.0f, weight1, output2); input2.z = 0.0f;
output2 = mad(0.0f, weight2, output2); input2.w = 0.0f;
output2 = mad(0.0f, weight3, output2); }
} }
}else {
output2 = mad(input2.x, weight0, output2); output2 = mad(input2.x, weight0, output2);
output2 = mad(input2.y, weight1, output2); output2 = mad(input2.y, weight1, output2);
output2 = mad(input2.z, weight2, output2); output2 = mad(input2.z, weight2, output2);
output2 = mad(input2.w, weight3, output2); output2 = mad(input2.w, weight3, output2);
}
// -------------3-------------- // -------------3--------------
pos_in = (int2)(i * input_width + in_pos_in_one_block3.x, in_pos_in_one_block3.y); pos_in = (int2)(i * input_width + in_pos_in_one_block3.x, in_pos_in_one_block3.y);
half4 input3 = read_imageh(input_image, sampler, pos_in); half4 input3 = read_imageh(input_image, sampler, pos_in);
bound_gap = max_w_bound - pos_in.x - 1;
if (abs(max_w_bound - pos_in.x) < input_width){ if (bound_gap < input_width && bound_gap >= 0){
if (burndary_index==0){ if (burndary_index==0){
output3 = mad(input3.x, weight0, output3); // do nothing
output3 = mad(input3.y, weight1, output3);
output3 = mad(input3.z, weight2, output3);
output3 = mad(input3.w, weight3, output3);
} else if (burndary_index==1){ } else if (burndary_index==1){
output3 = mad(input3.x, weight0, output3); input3.w = 0.0f;
output3 = mad(input3.y, weight1, output3);
output3 = mad(input3.z, weight2, output3);
output3 = mad(0.0f, weight3, output3);
} else if (burndary_index==2){ } else if (burndary_index==2){
output3 = mad(input3.x, weight0, output3); input3.z = 0.0f;
output3 = mad(input3.y, weight1, output3); input3.w = 0.0f;
output3 = mad(0.0f, weight2, output3);
output3 = mad(0.0f, weight3, output3);
} else if (burndary_index==3){ } else if (burndary_index==3){
output3 = mad(input3.x, weight0, output3); input3.y = 0.0f;
output3 = mad(0.0f, weight1, output3); input3.z = 0.0f;
output3 = mad(0.0f, weight2, output3); input3.w = 0.0f;
output3 = mad(0.0f, weight3, output3);
} }
}else { }
output3 = mad(input3.x, weight0, output3); output3 = mad(input3.x, weight0, output3);
output3 = mad(input3.y, weight1, output3); output3 = mad(input3.y, weight1, output3);
output3 = mad(input3.z, weight2, output3); output3 = mad(input3.z, weight2, output3);
output3 = mad(input3.w, weight3, output3); output3 = mad(input3.w, weight3, output3);
} }
}
#ifdef BATCH_NORM #ifdef BATCH_NORM
output0 = output0 * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + output0 = output0 * read_imageh(new_scale, sampler, (int2)(out_c, 0)) +
......
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