Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle-Lite
提交
96b2fcb9
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
331
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
提交
96b2fcb9
编写于
10月 17, 2018
作者:
L
liuruilong
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
update conv kernel cl
上级
a6561a80
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
195 addition
and
1110 deletion
+195
-1110
src/operators/kernel/cl/cl_kernel/conv_add_bn_relu_kernel.cl
src/operators/kernel/cl/cl_kernel/conv_add_bn_relu_kernel.cl
+1
-320
src/operators/kernel/cl/cl_kernel/conv_add_kernel.cl
src/operators/kernel/cl/cl_kernel/conv_add_kernel.cl
+1
-318
src/operators/kernel/cl/cl_kernel/conv_kernel.cl
src/operators/kernel/cl/cl_kernel/conv_kernel.cl
+1
-317
src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl
src/operators/kernel/cl/cl_kernel/conv_kernel.inc.cl
+192
-155
未找到文件。
src/operators/kernel/cl/cl_kernel/conv_add_bn_relu_kernel.cl
浏览文件 @
96b2fcb9
...
@@ -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 @
96b2fcb9
...
@@ -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 @
96b2fcb9
...
@@ -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 @
96b2fcb9
...
@@ -23,84 +23,14 @@ conv_add_bn_relu
...
@@ -23,84 +23,14 @@ conv_add_bn_relu
*/
*/
#
include
"cl_common.h"
#
include
"cl_kernel/cl_common.h"
__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,
__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
)
;
int
input_c
;
#
ifdef
BIASE
half4
output
=
read_imageh
(
bias,
sampler,
int2
(
out_c,
0
))
;
#
else
half4
output
=
0.0
;
#
endif
for
(
int
i
=
0
; i < input_c;h ++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
)
{
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
))
;
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
(
out_c
*
global_size_dim1
+
out_w,
out_nh
)
;
write_imageh
(
output_image,
output_pos,
output
)
;
}
__kernel
void
conv_3x3
(
__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
...
@@ -109,6 +39,7 @@ __kernel void conv_3x3(__private const int global_size_dim0,
...
@@ -109,6 +39,7 @@ __kernel void conv_3x3(__private const int global_size_dim0,
__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,
...
@@ -118,89 +49,121 @@ __kernel void conv_3x3(__private const int global_size_dim0,
...
@@ -118,89 +49,121 @@ __kernel void conv_3x3(__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
)
{
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
))
;
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
#
else
half4
output
=
0.0
;
half4
output
=
0.0
f
;
#
endif
#
endif
half4
input[9]
;
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,
sampler,
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,
for
(
int
i
=
0
; i < input_c; ++i) {
int2(pos_in.x, pos_in.y - dilation)),
int2
pos_in
=
(
int2
)(
i
*
input_width
+
in_pos_in_one_block.x,
in_pos_in_one_block.y
)
;
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[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[2] = select(read_imageh(input, sampler,
input[1] = select(read_imageh(input_image, sampler,
int2(pos_in.x + dilation, pos_in.y - dilation)),
(int2)(pos_in.x, 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);
(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[3] = select(read_imageh(input, sampler,
input[2] = select(read_imageh(input_image, sampler,
int2(pos_in.x - dilation, pos_in.y)),
(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 < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y >= input_height);
(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[4] = select(read_imageh(input, sampler,
input[3] = select(read_imageh(input_image, sampler,
int2(pos_in.x, pos_in.y)),
(int2)(pos_in.x - dilation, pos_in.y)),
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);
(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[5] = select(read_imageh(input, sampler,
input[4] = select(read_imageh(input_image, sampler,
int2(pos_in.x + dilation, pos_in.y)),
(int2)(pos_in.x, 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);
(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[6] = select(read_imageh(input, sampler,
input[5] = select(read_imageh(input_image, sampler,
int2(pos_in.x - dilation, pos_in.y + dilation)),
(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 + dilation < 0 || in_pos_in_one_block.x - dilation >= input_width || in_pos_in_one_block.y + dilation >= input_height);
(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[7] = select(read_imageh(input, sampler,
input[6] = select(read_imageh(input_image, sampler,
int2(pos_in.x, pos_in.y + dilation)),
(int2)(pos_in.x - dilation, 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);
(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[8] = select(read_imageh(input, sampler,
input[7] = select(read_imageh(input_image, sampler,
int2(pos_in.x + dilation, pos_in.y + dilation)),
(int2)(pos_in.x, pos_in.y + dilation)),
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);
(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) {
for (int j = 0; j < 9; ++j) {
int2 fuck;
half4 weight_x = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 0 * out_c * 3 + j / 3));
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);
output.x += dot(input[j], weight_x);
half4 weight_y = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 1 * out_c * 3 + j / 3));
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);
output.y += dot(input[j], weight_y);
half4 weight_z = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 2 * out_c * 3 + j / 3));
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);
output.z += dot(input[j], weight_z);
half4 weight_w = read_imageh(filter, sampler, int2(i * 3 + j % 3, out_c * 4 * 3 + 3 * out_c * 3 + j / 3));
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);
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 depth_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,
...
@@ -237,60 +200,61 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
...
@@ -237,60 +200,61 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
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_in_one_batch);
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 = 0.0
f
;
#endif
#endif
int2 pos_in_input_block = int2(out_c * input_width, batch_index * input_height);
int2 pos_in_input_block = (int2)(out_c * input_width, batch_index * input_height);
int weight_x_to = out_c * 3;
int weight_x_to = out_c * 3;
half4 inputs[9];
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)),
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,
(half4)(0.0f),
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);
(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[
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)),
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
,
(half4)(0.0f)
,
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
);
(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[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)),
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,
(half4)(0.0f),
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);
(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[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)),
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
,
(half4)(0.0f)
,
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
);
(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[
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
)),
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
)),
0.0
,
(half4)(0.0f)
,
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
);
(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[
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)),
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)),
0.0
,
(half4)(0.0f)
,
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
);
(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[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)),
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)),
0.0,
(half4)(0.0f),
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
)
;
(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) {
for (int j = 0; j < 9; ++j) {
half4 input = inputs[j];
half4 input = inputs[j];
half4
weight
=
read_imageh
(
filter,
sampler,
int2
(
weight_x_to
+
j
%
3
,
j
/
3
))
;
half4 weight = read_imageh(filter, sampler,
(int2)
(weight_x_to + j % 3, j / 3));
output.x += input.x * weight.x;
output.x += input.x * weight.x;
output.y += input.y * weight.y;
output.y += input.y * weight.y;
output.z += input.z * weight.z;
output.z += input.z * weight.z;
...
@@ -298,13 +262,86 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
...
@@ -298,13 +262,86 @@ __kernel void depth_conv_3x3(__private const int global_size_dim0,
}
}
#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 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
)
;
}
}
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录