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
野马c
Mace
提交
0d5bf6ce
Mace
项目概览
野马c
/
Mace
与 Fork 源项目一致
Fork自
Xiaomi / Mace
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
DevOps
流水线
流水线任务
计划
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
Mace
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
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
DevOps
DevOps
流水线
流水线任务
计划
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
流水线任务
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
0d5bf6ce
编写于
11月 03, 2017
作者:
L
liuqi
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Change the data type of conv kernel params from uint32_t to int32_t.
上级
b469a945
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
69 addition
and
70 deletion
+69
-70
mace/kernels/opencl/cl/batch_norm.cl
mace/kernels/opencl/cl/batch_norm.cl
+2
-2
mace/kernels/opencl/cl/conv_2d_3x3.cl
mace/kernels/opencl/cl/conv_2d_3x3.cl
+26
-26
mace/kernels/opencl/cl/depthwise_conv_3x3.cl
mace/kernels/opencl/cl/depthwise_conv_3x3.cl
+26
-26
mace/kernels/opencl/conv_2d_opencl_3x3.cc
mace/kernels/opencl/conv_2d_opencl_3x3.cc
+8
-8
mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
+6
-6
mace/ops/conv_2d_benchmark.cc
mace/ops/conv_2d_benchmark.cc
+0
-1
mace/ops/conv_2d_test.cc
mace/ops/conv_2d_test.cc
+1
-1
未找到文件。
mace/kernels/opencl/cl/batch_norm.cl
浏览文件 @
0d5bf6ce
...
...
@@ -4,7 +4,7 @@ void kernel batch_norm(global const float *input,
global
const
float
*mean,
global
const
float
*var,
global
const
float
*epsilon,
private
const
u
int
pixels,
private
const
int
pixels,
global
float
*output,
__local
float4
*new_scale,
__local
float4
*new_offset
)
{
...
...
@@ -12,7 +12,7 @@ void kernel batch_norm(global const float *input,
const
int
channel
=
get_global_id
(
1
)
;
const
int
channels
=
get_global_size
(
1
)
;
const
int
pixel_offset
=
get_global_id
(
2
)
;
const
unsigned
int
local_channel
=
get_local_id
(
1
)
;
const
int
local_channel
=
get_local_id
(
1
)
;
const
int
local_pixel_idx
=
get_local_id
(
2
)
;
if
(
local_pixel_idx
==
0
)
{
...
...
mace/kernels/opencl/cl/conv_2d_3x3.cl
浏览文件 @
0d5bf6ce
...
...
@@ -3,44 +3,44 @@ void kernel conv_2d_3x3(global const float *input,
global
const
float
*filter,
global
const
float
*bias,
global
float
*output,
private
const
u
int
in_chan_num,
private
const
u
int
out_chan_num,
private
const
u
int
in_height,
private
const
u
int
in_width,
private
const
u
int
out_height,
private
const
u
int
out_width,
private
const
u
int
stride_h,
private
const
u
int
stride_w
)
{
const
int
batch
=
get_global_id
(
0
)
;
const
int
out_chan_blk
=
get_global_id
(
1
)
;
const
int
out_pixel_blk
=
get_global_id
(
2
)
;
private
const
int
in_chan_num,
private
const
int
out_chan_num,
private
const
int
in_height,
private
const
int
in_width,
private
const
int
out_height,
private
const
int
out_width,
private
const
int
stride_h,
private
const
int
stride_w
)
{
int
batch
=
get_global_id
(
0
)
;
int
out_chan_blk
=
get_global_id
(
1
)
;
int
out_pixel_blk
=
get_global_id
(
2
)
;
const
u
int
in_pixel
=
in_height
*
in_width
;
const
u
int
out_pixel
=
out_height
*
out_width
;
const
int
in_pixel
=
in_height
*
in_width
;
const
int
out_pixel
=
out_height
*
out_width
;
const
u
int
round_out_width
=
(
out_width
+
3
)
/
4
;
const
u
int
out_pixel_height
=
out_pixel_blk
/
round_out_width
;
const
u
int
out_pixel_width
=
out_pixel_blk
%
round_out_width
;
const
int
round_out_width
=
(
out_width
+
3
)
/
4
;
const
int
out_pixel_height
=
out_pixel_blk
/
round_out_width
;
const
int
out_pixel_width
=
out_pixel_blk
%
round_out_width
;
const
u
int
out_chan_begin
=
out_chan_blk
*
4
;
const
u
int
out_chan_end
=
min
(
out_chan_begin
+
4
,
out_chan_num
)
;
const
u
int
out_pixel_begin
=
out_pixel_height
*
out_width
+
out_pixel_width
*
4
;
const
u
int
out_pixel_end
=
min
(
out_pixel_begin
+
4
,
(
out_pixel_height
+
1
)
*
out_width
)
;
const
u
int
in_pixel_begin
=
out_pixel_height
*
stride_h
*
in_width
+
out_pixel_width
*
stride_w
*
4
;
const
int
out_chan_begin
=
out_chan_blk
*
4
;
const
int
out_chan_end
=
min
(
out_chan_begin
+
4
,
out_chan_num
)
;
const
int
out_pixel_begin
=
out_pixel_height
*
out_width
+
out_pixel_width
*
4
;
const
int
out_pixel_end
=
min
(
out_pixel_begin
+
4
,
(
out_pixel_height
+
1
)
*
out_width
)
;
const
int
in_pixel_begin
=
out_pixel_height
*
stride_h
*
in_width
+
out_pixel_width
*
stride_w
*
4
;
const
u
int
in_offset
=
batch
*
in_chan_num
*
in_pixel
;
const
u
int
out_offset
=
batch
*
out_chan_num
*
out_pixel
;
const
int
in_offset
=
batch
*
in_chan_num
*
in_pixel
;
const
int
out_offset
=
batch
*
out_chan_num
*
out_pixel
;
const
float
*input_base
=
input
+
in_offset
+
in_pixel_begin
;
float
*output_base
=
output
+
out_offset
+
out_pixel_begin
;
uint
pixels
=
out_pixel_end
-
out_pixel_begin
;
for
(
u
int
i
=
out_chan_begin
; i < out_chan_end; ++i) {
for
(
int
i
=
out_chan_begin
; i < out_chan_end; ++i) {
float
*output_ptr
=
output_base
+
i
*
out_pixel
;
const
float
*filter_base
=
filter
+
i
*
in_chan_num
*
9
;
if
(
pixels
==
4
)
{
float4
res
=
(
float4
)
bias[i]
;
for
(
u
int
in_chan_idx
=
0
; in_chan_idx < in_chan_num; ++in_chan_idx) {
for
(
int
in_chan_idx
=
0
; in_chan_idx < in_chan_num; ++in_chan_idx) {
const
float*
input_ptr
=
input_base
+
in_chan_idx
*
in_pixel
;
const
float*
filter_ptr
=
filter_base
+
in_chan_idx
*
9
;
if
(
stride_w
==
1
)
{
...
...
@@ -55,7 +55,7 @@ void kernel conv_2d_3x3(global const float *input,
}
vstore4
(
res,
0
,
output_ptr
)
;
}
else
{
for
(
u
int
p
=
0
; p < pixels; ++p) {
for
(
int
p
=
0
; p < pixels; ++p) {
float
res
=
bias[i]
;
for
(
uint
in_chan_idx
=
0
; in_chan_idx < in_chan_num; ++in_chan_idx) {
const
float*
input_ptr
=
input_base
+
in_chan_idx
*
in_pixel
+
p
*
stride_w
;
...
...
mace/kernels/opencl/cl/depthwise_conv_3x3.cl
浏览文件 @
0d5bf6ce
...
...
@@ -4,40 +4,40 @@ void kernel depthwise_conv_3x3(global const float *input, /* n, c, h, w */
global
const
float
*filter,
/*
m,
i,
kh,
kw
*/
global
const
float
*bias,
/*
o
*/
global
float
*output,
/*
n,
c,
h,
w
*/
private
const
u
int
in_chan_num,
private
const
u
int
out_chan_num,
private
const
u
int
in_height,
private
const
u
int
in_width,
private
const
u
int
out_height,
private
const
u
int
out_width,
private
const
u
int
stride_h,
private
const
u
int
stride_w
)
{
const
int
batch
=
get_global_id
(
0
)
;
const
int
out_chan_blk
=
get_global_id
(
1
)
;
const
int
out_pixel_blk
=
get_global_id
(
2
)
;
private
const
int
in_chan_num,
private
const
int
out_chan_num,
private
const
int
in_height,
private
const
int
in_width,
private
const
int
out_height,
private
const
int
out_width,
private
const
int
stride_h,
private
const
int
stride_w
)
{
int
batch
=
get_global_id
(
0
)
;
int
out_chan_blk
=
get_global_id
(
1
)
;
int
out_pixel_blk
=
get_global_id
(
2
)
;
const
u
int
in_pixel
=
in_height
*
in_width
;
const
u
int
out_pixel
=
out_height
*
out_width
;
const
u
int
multiplier
=
out_chan_num
/
in_chan_num
;
const
int
in_pixel
=
in_height
*
in_width
;
const
int
out_pixel
=
out_height
*
out_width
;
const
int
multiplier
=
out_chan_num
/
in_chan_num
;
const
u
int
round_out_width
=
(
out_width
+
3
)
/
4
;
const
u
int
out_pixel_height
=
out_pixel_blk
/
round_out_width
;
const
u
int
out_pixel_width
=
out_pixel_blk
%
round_out_width
;
const
int
round_out_width
=
(
out_width
+
3
)
/
4
;
const
int
out_pixel_height
=
out_pixel_blk
/
round_out_width
;
const
int
out_pixel_width
=
out_pixel_blk
%
round_out_width
;
const
u
int
out_chan_begin
=
out_chan_blk
*
4
;
const
u
int
out_chan_end
=
min
(
out_chan_begin
+
4
,
out_chan_num
)
;
const
u
int
out_pixel_begin
=
out_pixel_height
*
out_width
+
out_pixel_width
*
4
;
const
u
int
out_pixel_end
=
min
(
out_pixel_begin
+
4
,
(
out_pixel_height
+
1
)
*
out_width
)
;
const
u
int
in_pixel_begin
=
out_pixel_height
*
stride_h
*
in_width
+
out_pixel_width
*
stride_w
*
4
;
const
int
out_chan_begin
=
out_chan_blk
*
4
;
const
int
out_chan_end
=
min
(
out_chan_begin
+
4
,
out_chan_num
)
;
const
int
out_pixel_begin
=
out_pixel_height
*
out_width
+
out_pixel_width
*
4
;
const
int
out_pixel_end
=
min
(
out_pixel_begin
+
4
,
(
out_pixel_height
+
1
)
*
out_width
)
;
const
int
in_pixel_begin
=
out_pixel_height
*
stride_h
*
in_width
+
out_pixel_width
*
stride_w
*
4
;
const
u
int
in_offset
=
batch
*
in_chan_num
*
in_pixel
;
const
u
int
out_offset
=
batch
*
out_chan_num
*
out_pixel
;
const
int
in_offset
=
batch
*
in_chan_num
*
in_pixel
;
const
int
out_offset
=
batch
*
out_chan_num
*
out_pixel
;
const
float
*input_base
=
input
+
in_offset
+
in_pixel_begin
;
float
*output_base
=
output
+
out_offset
+
out_pixel_begin
;
uint
pixels
=
out_pixel_end
-
out_pixel_begin
;
for
(
u
int
i
=
out_chan_begin
; i < out_chan_end; ++i) {
for
(
int
i
=
out_chan_begin
; i < out_chan_end; ++i) {
float
bias_value
=
bias[i]
;
const
float
*input_ptr
=
input_base
+
(
i
/
multiplier
)
*
in_pixel
;
const
float
*filter_ptr
=
filter
+
i
*
9
;
...
...
@@ -55,7 +55,7 @@ void kernel depthwise_conv_3x3(global const float *input, /* n, c, h, w */
}
vstore4
(
res,
0
,
output_ptr
)
;
}
else
{
for
(
u
int
p
=
0
; p < pixels; ++p) {
for
(
int
p
=
0
; p < pixels; ++p) {
float
res
=
bias[i]
;
res
+=
conv3x3
(
input_ptr,
filter_ptr,
in_width
)
;
output_ptr[p]
=
res
;
...
...
mace/kernels/opencl/conv_2d_opencl_3x3.cc
浏览文件 @
0d5bf6ce
...
...
@@ -29,20 +29,20 @@ static void InnerConv2dK3x3S12(const Tensor *input, const Tensor *filter,
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
const
cl
::
Buffer
*>
(
filter
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
const
cl
::
Buffer
*>
(
bias
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
cl
::
Buffer
*>
(
output
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
1
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
channels
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
2
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
3
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
height
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
width
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
1
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
channels
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
2
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
3
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
height
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
width
));
conv_kernel
.
setArg
(
idx
++
,
stride
);
conv_kernel
.
setArg
(
idx
++
,
stride
);
const
uint32_t
gws
[
3
]
=
{
static_cast
<
uint32_t
>
(
output
->
dim
(
0
)),
static_cast
<
uint32_t
>
(
channel_blocks
),
static_cast
<
uint32_t
>
(
pixel_blocks
)};
const
uint32_t
lws
[
3
]
=
{
static_cast
<
uint32_t
>
(
1
),
static_cast
<
uint32_t
>
(
1
),
static_cast
<
uint32_t
>
(
256
)};
static_cast
<
uint32_t
>
(
8
),
static_cast
<
uint32_t
>
(
128
)};
cl_int
error
=
runtime
->
command_queue
().
enqueueNDRangeKernel
(
conv_kernel
,
cl
::
NullRange
,
cl
::
NDRange
(
gws
[
0
],
gws
[
1
],
gws
[
2
]),
...
...
mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
浏览文件 @
0d5bf6ce
...
...
@@ -38,12 +38,12 @@ static void InnerDepthwiseConvOpenclK3x3S12(const Tensor *input,
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
const
cl
::
Buffer
*>
(
filter
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
const
cl
::
Buffer
*>
(
bias
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
*
(
static_cast
<
cl
::
Buffer
*>
(
output
->
buffer
())));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
1
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
channels
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
2
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
input
->
dim
(
3
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
height
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
u
int32_t
>
(
width
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
1
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
channels
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
2
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
input
->
dim
(
3
)));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
height
));
conv_kernel
.
setArg
(
idx
++
,
static_cast
<
int32_t
>
(
width
));
conv_kernel
.
setArg
(
idx
++
,
stride
);
conv_kernel
.
setArg
(
idx
++
,
stride
);
...
...
mace/ops/conv_2d_benchmark.cc
浏览文件 @
0d5bf6ce
...
...
@@ -3,7 +3,6 @@
//
#include <algorithm>
#include <sstream>
#include "mace/core/operator.h"
#include "mace/core/testing/test_benchmark.h"
...
...
mace/ops/conv_2d_test.cc
浏览文件 @
0d5bf6ce
...
...
@@ -296,7 +296,7 @@ static void TestUnalignedConvNxNS12() {
ExpectTensorNear
<
float
>
(
expected
,
*
net
.
GetOutput
(
"Output"
),
0.001
);
};
for
(
int
kernel_size
:
{
1
,
3
,
5
})
{
for
(
int
kernel_size
:
{
3
})
{
for
(
int
stride
:
{
1
,
2
})
{
func
(
kernel_size
,
kernel_size
,
stride
,
stride
,
VALID
);
func
(
kernel_size
,
kernel_size
,
stride
,
stride
,
SAME
);
...
...
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