Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 e9c54cbb 编写于 作者: R Ray Liu 提交者： GitHub
浏览文件
操作

Merge pull request #1124 from codeWorm2015/opencl 

 update conv kernel cl
上级 a6561a80 96b2fcb9
显示空白变更内容
内联 并排
Showing with 195 addition and 1110 deletion
src/operators/kernel/cl/cl_kernel/conv_add_bn_relu_kernel.cl
浏览文件 @ e9c54cbb
...@@ -17,323 +17,4 @@ limitations under the License. */ ...@@ -17,323 +17,4 @@ limitations under the License. */
#define BATCH_NORM #define BATCH_NORM
#define RELU #define RELU
#include "cl_kernel/cl_common.h" #include "cl_kernel/conv_kernel.inc.cl"
__kernel void 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_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
int2 stride_xy;
stride_xy.x = stride;
stride_xy.y = stride;
int2 ouput_pos_in_one_block;
ouput_pos_in_one_block.x = out_w;
ouput_pos_in_one_block.y = out_nh;
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
int2 in_pos_in_one_block;
in_pos_in_one_block.x = ouput_pos_in_one_block.x * stride + offset;
in_pos_in_one_block.y = ouput_pos_in_one_block.y * stride + offset;
#ifdef BIASE
half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
#else
half4 output = 0.0f;
#endif
half4 input[9];
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
input[0] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[1] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[2] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[3] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[4] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.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));
input[5] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[6] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[7] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[8] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(pos_in.x + dilation < 0 || in_pos_in_one_block.y + dilation < 0 || pos_in.x + dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
for (int j = 0; j < 9; ++j) {
int2 fuck;
fuck.x = i * 3 + j % 3;
fuck.y = out_c * 4 * 3 + 0 * out_c * 3 + j / 3;
half4 weight_x = read_imageh(filter, sampler, fuck);
output.x += dot(input[j], weight_x);
fuck.y = out_c * 4 * 3 + 1 * out_c * 3 + j / 3;
half4 weight_y = read_imageh(filter, sampler, fuck);
output.y += dot(input[j], weight_y);
fuck.y = out_c * 4 * 3 + 2 * out_c * 3 + j / 3;
half4 weight_z = read_imageh(filter, sampler, fuck);
output.z += dot(input[j], weight_z);
fuck.y = out_c * 4 * 3 + 3 * out_c * 3 + j / 3;
half4 weight_w = read_imageh(filter, sampler, fuck);
output.w += dot(input[j], weight_w);
}
}
#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
write_imageh(output_image, (int2)(out_c * global_size_dim1 + out_w, out_nh), output);
}
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
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;
const uint kernelHXW = 1;
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);
int weight_x_to = out_c * 3;
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));
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));
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));
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));
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));
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));
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));
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));
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));
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 weight = read_imageh(filter, sampler, (int2)(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.w += input.w * weight.w;
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
__kernel void conv_1x1(__private const int global_size_dim0,
__private const int global_size_dim1,
__private const int global_size_dim2,
__read_only image2d_t input_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
const uint kernelHXW = 1;
int2 stride_xy = (int2)(stride, stride);
int2 ouput_pos_in_one_block = (int2)(out_w, out_nh);
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
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
if (pos_in.x >=0 && pos_in.y >= 0 && pos_in.x < input_width && pos_in.y < input_height) {
half4 input = read_imageh(input_image, sampler, pos_in);
half4 weight_x = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 0));
output.x += dot(input, weight_x);
half4 weight_y = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 1));
output.y += dot(input, weight_y);
half4 weight_z = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 2));
output.z += dot(input, weight_z);
half4 weight_w = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 3));
output.w += dot(input, weight_w);
}
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
src/operators/kernel/cl/cl_kernel/conv_add_kernel.cl
浏览文件 @ e9c54cbb
...@@ -12,324 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,324 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#define BIASE #define BIASE
__kernel void conv_3x3(__private const int global_size_dim0, #include "cl_kernel/conv_kernel.inc.cl"
__private const int global_size_dim1,
__private const int global_size_dim2,
__read_only image2d_t input_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
int2 stride_xy;
stride_xy.x = stride;
stride_xy.y = stride;
int2 ouput_pos_in_one_block;
ouput_pos_in_one_block.x = out_w;
ouput_pos_in_one_block.y = out_nh;
int2 in_pos_in_one_block;
in_pos_in_one_block.x = ouput_pos_in_one_block.x * stride + offset;
in_pos_in_one_block.y = ouput_pos_in_one_block.y * stride + offset;
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
#ifdef BIASE
half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
#else
half4 output = 0.0f;
#endif
half4 input[9];
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
input[0] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[1] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[2] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[3] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[4] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.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));
input[5] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[6] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[7] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[8] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(pos_in.x + dilation < 0 || in_pos_in_one_block.y + dilation < 0 || pos_in.x + dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
for (int j = 0; j < 9; ++j) {
int2 fuck;
fuck.x = i * 3 + j % 3;
fuck.y = out_c * 4 * 3 + 0 * out_c * 3 + j / 3;
half4 weight_x = read_imageh(filter, sampler, fuck);
output.x += dot(input[j], weight_x);
fuck.y = out_c * 4 * 3 + 1 * out_c * 3 + j / 3;
half4 weight_y = read_imageh(filter, sampler, fuck);
output.y += dot(input[j], weight_y);
fuck.y = out_c * 4 * 3 + 2 * out_c * 3 + j / 3;
half4 weight_z = read_imageh(filter, sampler, fuck);
output.z += dot(input[j], weight_z);
fuck.y = out_c * 4 * 3 + 3 * out_c * 3 + j / 3;
half4 weight_w = read_imageh(filter, sampler, fuck);
output.w += dot(input[j], weight_w);
}
}
#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
write_imageh(output_image, (int2)(out_c * global_size_dim1 + out_w, out_nh), output);
}
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
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;
const uint kernelHXW = 1;
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);
int weight_x_to = out_c * 3;
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));
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));
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));
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));
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));
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));
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));
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));
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));
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 weight = read_imageh(filter, sampler, (int2)(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.w += input.w * weight.w;
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
__kernel void conv_1x1(__private const int global_size_dim0,
__private const int global_size_dim1,
__private const int global_size_dim2,
__read_only image2d_t input_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
const uint kernelHXW = 1;
int2 stride_xy = (int2)(stride, stride);
int2 ouput_pos_in_one_block = (int2)(out_w, out_nh);
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
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
if (pos_in.x >=0 && pos_in.y >= 0 && pos_in.x < input_width && pos_in.y < input_height) {
half4 input = read_imageh(input_image, sampler, pos_in);
half4 weight_x = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 0));
output.x += dot(input, weight_x);
half4 weight_y = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 1));
output.y += dot(input, weight_y);
half4 weight_z = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 2));
output.z += dot(input, weight_z);
half4 weight_w = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 3));
output.w += dot(input, weight_w);
}
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
src/operators/kernel/cl/cl_kernel/conv_kernel.cl
浏览文件 @ e9c54cbb
...@@ -12,320 +12,4 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,320 +12,4 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#pragma OPENCL EXTENSION cl_khr_fp16 : enable #include "cl_kernel/conv_kernel.inc.cl"
__kernel void 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_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
int2 stride_xy;
stride_xy.x = stride;
stride_xy.y = stride;
int2 ouput_pos_in_one_block;
ouput_pos_in_one_block.x = out_w;
ouput_pos_in_one_block.y = out_nh;
int2 in_pos_in_one_block;
in_pos_in_one_block.x = ouput_pos_in_one_block.x * stride + offset;
in_pos_in_one_block.y = ouput_pos_in_one_block.y * stride + offset;
#ifdef BIASE
half4 output = read_imageh(bias, sampler, int2(out_c, 0));
#else
half4 output = 0.0f;
#endif
half4 input[9];
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
input[0] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[1] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[2] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[3] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[4] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.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));
input[5] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[6] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[7] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[8] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(pos_in.x + dilation < 0 || in_pos_in_one_block.y + dilation < 0 || pos_in.x + dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
for (int j = 0; j < 9; ++j) {
int2 fuck;
fuck.x = i * 3 + j % 3;
fuck.y = out_c * 4 * 3 + 0 * out_c * 3 + j / 3;
half4 weight_x = read_imageh(filter, sampler, fuck);
output.x += dot(input[j], weight_x);
fuck.y = out_c * 4 * 3 + 1 * out_c * 3 + j / 3;
half4 weight_y = read_imageh(filter, sampler, fuck);
output.y += dot(input[j], weight_y);
fuck.y = out_c * 4 * 3 + 2 * out_c * 3 + j / 3;
half4 weight_z = read_imageh(filter, sampler, fuck);
output.z += dot(input[j], weight_z);
fuck.y = out_c * 4 * 3 + 3 * out_c * 3 + j / 3;
half4 weight_w = read_imageh(filter, sampler, fuck);
output.w += dot(input[j], weight_w);
}
}
#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
write_imageh(output_image, (int2)(out_c * global_size_dim1 + out_w, out_nh), output);
}
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
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;
const uint kernelHXW = 1;
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);
int weight_x_to = out_c * 3;
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));
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));
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));
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));
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));
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));
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));
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));
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));
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 weight = read_imageh(filter, sampler, (int2)(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.w += input.w * weight.w;
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
__kernel void conv_1x1(__private const int global_size_dim0,
__private const int global_size_dim1,
__private const int global_size_dim2,
__read_only image2d_t input_image,
__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) {
const int out_c = get_global_id(0);
const int out_w = get_global_id(1);
const int out_nh = get_global_id(2);
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
const uint kernelHXW = 1;
int2 stride_xy = (int2)(stride, stride);
int2 ouput_pos_in_one_block = (int2)(out_w, out_nh);
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
for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
if (pos_in.x >=0 && pos_in.y >= 0 && pos_in.x < input_width && pos_in.y < input_height) {
half4 input = read_imageh(input_image, sampler, pos_in);
half4 weight_x = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 0));
output.x += dot(input, weight_x);
half4 weight_y = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 1));
output.y += dot(input, weight_y);
half4 weight_z = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 2));
output.z += dot(input, weight_z);
half4 weight_w = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 3));
output.w += dot(input, weight_w);
}
}
#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
int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output);
}
src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl
浏览文件 @ e9c54cbb
...@@ -23,20 +23,23 @@ conv_add_bn_relu ...@@ -23,20 +23,23 @@ conv_add_bn_relu
*/ */
#include "cl_common.h" #include "cl_kernel/cl_common.h"
__kernel void conv_1x1(__private const int global_size_dim0, __kernel void conv_3x3(__private const int global_size_dim0,
__private const int global_size_dim1, __private const int global_size_dim1,
__private const int global_size_dim2, __private const int global_size_dim2,
__read_only image2d_t input, __read_only image2d_t input_image,
__read_only image2d_t filter, __read_only image2d_t filter,
#ifdef BIASE #ifdef BIASE
__read_only image2d_t bias, __read_only image2d_t bias,
#endif #endif
#ifdef BATCH_NORM #ifdef BATCH_NORM
__read_only image2d_t new_scale, __read_only image2d_t new_scale,
__read_only image2d_t new_biase, __read_only image2d_t new_biase,
#endif #endif
__write_only image2d_t output_image, __write_only image2d_t output_image,
__private const int stride, __private const int stride,
__private const int offset, __private const int offset,
...@@ -46,57 +49,122 @@ __kernel void conv_1x1(__private const int global_size_dim0, ...@@ -46,57 +49,122 @@ __kernel void conv_1x1(__private const int global_size_dim0,
__private const int input_height,/* of one block */ __private const int input_height,/* of one block */
__private const int output_width, __private const int output_width,
__private const int output_height) { __private const int output_height) {
const int out_c = get_global_id(0); const int out_c = get_global_id(0);
const int out_w = get_global_id(1); const int out_w = get_global_id(1);
const int out_nh = get_global_id(2); const int out_nh = get_global_id(2);
int2 stride_xy;
stride_xy.x = stride;
stride_xy.y = stride;
int2 ouput_pos_in_one_block;
ouput_pos_in_one_block.x = out_w;
ouput_pos_in_one_block.y = out_nh;
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE | const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP | CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST; CLK_FILTER_NEAREST;
const uint kernelHXW = 1;
int2 stride_xy = int2(stride, stride); int2 in_pos_in_one_block;
int2 ouput_pos_in_one_block = int2(out_w, out_nh); in_pos_in_one_block.x = ouput_pos_in_one_block.x * stride + offset;
int2 in_pos_in_one_block = ouput_pos_in_one_block * stride_xy + int2(offset, offset); in_pos_in_one_block.y = ouput_pos_in_one_block.y * stride + offset;
int input_c;
#ifdef BIASE #ifdef BIASE
half4 output = read_imageh(bias, sampler, int2(out_c, 0)); half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
#else #else
half4 output = 0.0; half4 output = 0.0f;
#endif #endif
for (int i = 0; i < input_c;h ++i) { half4 input[9];
int2 pos_in = int2(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
if (pos_in.x >=0 && pos_in.y >= 0 && pos_in.x < input_width && pos_in.y < input_height) {
hafl4 input = read_imageh(input, sampler, pos_in);
half4 weight_x = read_imageh(filter, sampler, int2(i, out_c * 4 + 0));
output.x += dot(input, weight_x);
half4 weight_y = read_imageh(filter, sampler, int2(i, out_c * 4 + 1));
output.y += dot(input, weight_y);
half4 weight_z = read_imageh(filter, sampler, int2(i, out_c * 4 + 2)); for (int i = 0; i < input_c; ++i) {
output.z += dot(input, weight_z); int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
input[0] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[1] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[2] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y - dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height));
input[3] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[4] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.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));
input[5] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y >= input_height));
input[6] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x - dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[7] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + dilation >= input_height));
input[8] = select(read_imageh(input_image, sampler,
(int2)(pos_in.x + dilation, pos_in.y + dilation)),
(half4)(0.0f),
(ushort4)(pos_in.x + dilation < 0 || in_pos_in_one_block.y + dilation < 0 || pos_in.x + dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height));
half4 weight_w = read_imageh(filter, sampler, int2(i, out_c * 4 + 3)); for (int j = 0; j < 9; ++j) {
output.w += dot(input, weight_w); int2 fuck;
fuck.x = i * 3 + j % 3;
fuck.y = out_c * 4 * 3 + 0 * out_c * 3 + j / 3;
half4 weight_x = read_imageh(filter, sampler, fuck);
output.x += dot(input[j], weight_x);
fuck.y = out_c * 4 * 3 + 1 * out_c * 3 + j / 3;
half4 weight_y = read_imageh(filter, sampler, fuck);
output.y += dot(input[j], weight_y);
fuck.y = out_c * 4 * 3 + 2 * out_c * 3 + j / 3;
half4 weight_z = read_imageh(filter, sampler, fuck);
output.z += dot(input[j], weight_z);
fuck.y = out_c * 4 * 3 + 3 * out_c * 3 + j / 3;
half4 weight_w = read_imageh(filter, sampler, fuck);
output.w += dot(input[j], weight_w);
} }
} }
#ifdef BATCH_NORM #ifdef BATCH_NORM
output = output * read_imageh(new_scale, sampler, int2(out_c, 0)) + read_imageh(new_biase, sampler, int2(out_c, 0)) output = output * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + read_imageh(new_biase, sampler, (int2)(out_c, 0));
#endif #endif
#ifdef RELU #ifdef RELU
output = activation(output); output = activation(output);
#endif #endif
int2 output_pos(out_c * global_size_dim1 + out_w, out_nh); write_imageh(output_image, (int2)(out_c * global_size_dim1 + out_w, out_nh), output);
write_imageh(output_image, output_pos, output);
} }
__kernel void conv_3x3(__private const int global_size_dim0,
__kernel void depth_conv_3x3(__private const int global_size_dim0,
__private const int global_size_dim1, __private const int global_size_dim1,
__private const int global_size_dim2, __private const int global_size_dim2,
__read_only image2d_t input, __read_only image2d_t input,
...@@ -104,7 +172,6 @@ __kernel void conv_3x3(__private const int global_size_dim0, ...@@ -104,7 +172,6 @@ __kernel void conv_3x3(__private const int global_size_dim0,
#ifdef BIASE #ifdef BIASE
__read_only image2d_t bias, __read_only image2d_t bias,
#endif #endif
#ifdef BATCH_NORM #ifdef BATCH_NORM
__read_only image2d_t new_scale, __read_only image2d_t new_scale,
__read_only image2d_t new_biase, __read_only image2d_t new_biase,
...@@ -115,96 +182,102 @@ __kernel void conv_3x3(__private const int global_size_dim0, ...@@ -115,96 +182,102 @@ __kernel void conv_3x3(__private const int global_size_dim0,
__private const int input_c, __private const int input_c,
__private const int dilation, __private const int dilation,
__private const int input_width,/* of one block */ __private const int input_width,/* of one block */
__private const int input_height,/* of one block */ __private const int input_height, /* of one block */
__private const int output_width, __private const int output_width,
__private const int output_height) { __private const int output_height) {
int2 stride_xy = int2(stride, stride);
int2 ouput_pos_in_one_block = int2(out_w, out_nh);
int2 in_pos_in_one_block = ouput_pos_in_one_block * stride_xy + int2(offset, offset);
#ifdef BIASE const int out_c = get_global_id(0);
half4 output = read_imageh(bias, sampler, int2(out_c, 0)); const int out_w = get_global_id(1);
#else const int out_nh = get_global_id(2);
half4 output = 0.0;
#endif
half4 input[9]; const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST;
for (int i = 0; i < input_c; ++i) { const int batch_index = out_nh / output_height;
int2 pos_in = int2(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
input[0] = select(read_imageh(input, sampler, const int out_nh_in_one_batch = out_nh % output_height;
int2(pos_in.x - dilation, pos_in.y - dilation)),
half4(0.0),in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height);
input[1] = select(read_imageh(input, sampler, const uint kernelHXW = 1;
int2(pos_in.x, pos_in.y - dilation)),
half4(0.0),in_pos_in_one_block.x < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y - dilation >= input_height);
input[2] = select(read_imageh(input, sampler, int2 stride_xy = (int2)(stride, stride);
int2(pos_in.x + dilation, pos_in.y - dilation)), int2 ouput_pos_in_one_block = (int2)(out_w, out_nh_in_one_batch);
half4(0.0),in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y - dilation < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y - dilation >= input_height);
input[3] = select(read_imageh(input, sampler, int2 in_pos_in_one_block = ouput_pos_in_one_block * stride_xy + (int2)(offset, offset);
int2(pos_in.x - dilation, pos_in.y)),
half4(0.0), in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height);
input[4] = select(read_imageh(input, sampler, #ifdef BIASE
int2(pos_in.x, pos_in.y)), half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
half4(0.0), 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); #else
half4 output = 0.0f;
#endif
input[5] = select(read_imageh(input, sampler, int2 pos_in_input_block = (int2)(out_c * input_width, batch_index * input_height);
int2(pos_in.x + dilation, pos_in.y)), int weight_x_to = out_c * 3;
half4(0.0), in_pos_in_one_block.x + dilation < 0 || in_pos_in_one_block.y < 0 || in_pos_in_one_block.x + dilation >= input_width || in_pos_in_one_block.y >= input_height);
input[6] = select(read_imageh(input, sampler, half4 inputs[9];
int2(pos_in.x - dilation, pos_in.y + dilation)),
half4(0.0), in_pos_in_one_block.x - dilation < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height);
input[7] = select(read_imageh(input, sampler, 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)),
int2(pos_in.x, pos_in.y + dilation)), (half4)(0.0f),
half4(0.0), in_pos_in_one_block.x < 0 || in_pos_in_one_block.y + dilation < 0 || in_pos_in_one_block.x >= input_width || in_pos_in_one_block.y + dilation >= input_height); (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));
input[8] = select(read_imageh(input, sampler, 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)),
int2(pos_in.x + dilation, pos_in.y + dilation)), (half4)(0.0f),
half4(0.0), pos_in.x + dilation < 0 || in_pos_in_one_block.y + dilation < 0 || pos_in.x + dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height); (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));
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));
for (int j = 0; j < 9; ++j) { 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));
half4 weight_x = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 0 * out_c * 3 + j / 3)); 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)),
output.x += dot(input[j], weight_x); (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));
half4 weight_y = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 1 * out_c * 3 + j / 3)); 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)),
output.y += dot(input[j], weight_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));
half4 weight_z = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 2 * out_c * 3 + j / 3)); 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)),
output.z += dot(input[j], weight_z); (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));
half4 weight_w = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 3 * out_c * 3 + j / 3)); 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)),
output.w += dot(input[j], weight_w); (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));
} 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));
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 weight = read_imageh(filter, sampler, (int2)(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.w += input.w * weight.w;
} }
#ifdef BATCH_NORM #ifdef BATCH_NORM
output = output * read_imageh(new_scale, sampler, int2(out_c, 0)) + read_imageh(new_biase, sampler, int2(out_c, 0)) output = output * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + read_imageh(new_biase, sampler, (int2)(out_c, 0));
#endif #endif
#ifdef RELU #ifdef RELU
output = activation(output); output = activation(output);
#endif #endif
int2 output_pos(out_c * global_size_dim1 + out_w, out_nh); int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output); write_imageh(output_image, output_pos, output);
}
}
__kernel void depth_conv_3x3(__private const int global_size_dim0, __kernel void conv_1x1(__private const int global_size_dim0,
__private const int global_size_dim1, __private const int global_size_dim1,
__private const int global_size_dim2, __private const int global_size_dim2,
__read_only image2d_t input, __read_only image2d_t input_image,
__read_only image2d_t filter, __read_only image2d_t filter,
#ifdef BIASE #ifdef BIASE
__read_only image2d_t bias, __read_only image2d_t bias,
...@@ -219,10 +292,9 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0, ...@@ -219,10 +292,9 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
__private const int input_c, __private const int input_c,
__private const int dilation, __private const int dilation,
__private const int input_width,/* of one block */ __private const int input_width,/* of one block */
__private const int input_height, /* of one block */ __private const int input_height,/* of one block */
__private const int output_width, __private const int output_width,
__private const int output_height) { __private const int output_height) {
const int out_c = get_global_id(0); const int out_c = get_global_id(0);
const int out_w = get_global_id(1); const int out_w = get_global_id(1);
const int out_nh = get_global_id(2); const int out_nh = get_global_id(2);
...@@ -230,81 +302,46 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0, ...@@ -230,81 +302,46 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE | const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
CLK_ADDRESS_CLAMP | CLK_ADDRESS_CLAMP |
CLK_FILTER_NEAREST; CLK_FILTER_NEAREST;
const int batch_index = out_nh / output_height;
const int out_nh_in_one_batch = out_nh % output_height;
const uint kernelHXW = 1; const uint kernelHXW = 1;
int2 stride_xy = (int2)(stride, stride);
int2 stride_xy = int2(stride, stride); int2 ouput_pos_in_one_block = (int2)(out_w, out_nh);
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);
int2 in_pos_in_one_block = ouput_pos_in_one_block * stride_xy + int2(offset, offset);
#ifdef BIASE #ifdef BIASE
half4 output = read_imageh(bias, sampler, int2(out_c, 0)); half4 output = read_imageh(bias, sampler, (int2)(out_c, 0));
#else #else
half4 output = 0.0; half4 output = 0.0f;
#endif #endif
int2 pos_in_input_block = int2(out_c * input_width, batch_index * input_height); for (int i = 0; i < input_c; ++i) {
int2 pos_in = (int2)(i * input_width + in_pos_in_one_block.x, in_pos_in_one_block.y);
int weight_x_to = out_c * 3; if (pos_in.x >=0 && pos_in.y >= 0 && pos_in.x < input_width && pos_in.y < input_height) {
half4 input = read_imageh(input_image, sampler, pos_in);
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)),
0.0,
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);
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)),
0.0,
n_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);
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)),
0.0,
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);
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)),
0.0,
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);
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)),
0.0,
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);
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 weight_x = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 0));
0.0, output.x += dot(input, weight_x);
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);
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 weight_y = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 1));
0.0, output.y += dot(input, weight_y);
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);
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 weight_z = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 2));
0.0, output.z += dot(input, weight_z);
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);
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 weight_w = read_imageh(filter, sampler, (int2)(i, out_c * 4 + 3));
0.0, output.w += dot(input, weight_w);
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);
for (int j = 0; j < 9; ++j) { }
half4 input = inputs[j];
half4 weight = read_imageh(filter, sampler, int2(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.w += input.w * weight.w;
} }
#ifdef BATCH_NORM #ifdef BATCH_NORM
output = output * read_imageh(new_scale, sampler, int2(out_c, 0)) + read_imageh(new_biase, sampler, int2(out_c, 0)) output = output * read_imageh(new_scale, sampler, (int2)(out_c, 0)) + read_imageh(new_biase, sampler, (int2)(out_c, 0));
#endif #endif
#ifdef RELU #ifdef RELU
output = activation(output); output = activation(output);
#endif #endif
int2 output_pos(out_c * global_size_dim1 + out_w, out_nh); int2 output_pos = (int2)(out_c * global_size_dim1 + out_w, out_nh);
write_imageh(output_image, output_pos, output); write_imageh(output_image, output_pos, output);
} }
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