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
提交
4e91d8d2
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
4e91d8d2
编写于
1月 22, 2019
作者:
W
WangZhen
浏览文件
操作
浏览文件
下载
差异文件
Merge branch 'develop' of
https://github.com/PaddlePaddle/Paddle
into graph_quantization
test=develop
上级
3b668c15
b3fdf708
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
26 addition
and
27 deletion
+26
-27
paddle/fluid/operators/distributed/CMakeLists.txt
paddle/fluid/operators/distributed/CMakeLists.txt
+1
-1
paddle/fluid/operators/group_norm_op.cu
paddle/fluid/operators/group_norm_op.cu
+8
-8
python/paddle/fluid/executor.py
python/paddle/fluid/executor.py
+17
-18
未找到文件。
paddle/fluid/operators/distributed/CMakeLists.txt
浏览文件 @
4e91d8d2
...
...
@@ -37,7 +37,7 @@ else()
variable_response.cc
collective_client.cc collective_server.cc
${
BRPC_SRCS
}
PROTO
${
CMAKE_CURRENT_BINARY_DIR
}
/
send_recv.proto
PROTO send_recv.proto
DEPS lod_tensor selected_rows memory
)
set
(
RPC_DEPS sendrecvop_rpc brpc ssl crypto protobuf leveldb snappystream snappy zlib
)
...
...
paddle/fluid/operators/group_norm_op.cu
浏览文件 @
4e91d8d2
...
...
@@ -21,20 +21,20 @@ namespace operators {
enum
GroupNormKernelFlags
{
kHasScale
=
1
,
kHasBias
=
2
};
#define CHECK_CASE(i, flags, kernel_name,
args...)
\
#define CHECK_CASE(i, flags, kernel_name,
...)
\
if (i == flags) { \
kernel_name<T, i><<<grid, threads, 0, dev_ctx.stream()>>>(
args
); \
kernel_name<T, i><<<grid, threads, 0, dev_ctx.stream()>>>(
__VA_ARGS__
); \
}
// 0 for no scale, no bias
// 1 for has scale, no bias
// 2 for no scale, has bias
// 3 for has scale, has bias
#define UNROLL_ALL_CASES(flags, kernel_name,
args
...) \
CHECK_CASE(0, flags, kernel_name,
args)
\
CHECK_CASE(1, flags, kernel_name,
args)
\
CHECK_CASE(2, flags, kernel_name,
args)
\
CHECK_CASE(3, flags, kernel_name,
args
)
#define UNROLL_ALL_CASES(flags, kernel_name, ...) \
CHECK_CASE(0, flags, kernel_name,
__VA_ARGS__)
\
CHECK_CASE(1, flags, kernel_name,
__VA_ARGS__)
\
CHECK_CASE(2, flags, kernel_name,
__VA_ARGS__)
\
CHECK_CASE(3, flags, kernel_name,
__VA_ARGS__
)
template
<
typename
T
>
__device__
__inline__
void
CudaAtomicAddWithWarp
(
T
*
sum
,
T
value
)
{
...
...
python/paddle/fluid/executor.py
浏览文件 @
4e91d8d2
...
...
@@ -305,7 +305,9 @@ class Executor(object):
def
__init__
(
self
,
place
):
self
.
place
=
place
self
.
program_caches
=
dict
()
self
.
executor
=
None
p
=
core
.
Place
()
p
.
set_place
(
self
.
place
)
self
.
_default_executor
=
core
.
Executor
(
p
)
self
.
_closed
=
False
def
_get_program_cache
(
self
,
program_cache_key
):
...
...
@@ -397,12 +399,13 @@ class Executor(object):
>>> ...
>>> exe.close()
"""
if
not
self
.
_closed
and
self
.
executor
:
self
.
executor
.
close
()
if
not
self
.
_closed
:
self
.
_default_
executor
.
close
()
self
.
_closed
=
True
def
_run_parallel
(
self
,
program
,
scope
,
feed
,
fetch_list
,
fetch_var_name
,
return_numpy
):
exe
=
program
.
_executor
if
isinstance
(
feed
,
dict
):
feed_tensor_dict
=
dict
()
for
feed_name
in
feed
:
...
...
@@ -414,8 +417,7 @@ class Executor(object):
feed_tensor
.
set
(
feed
[
feed_name
],
core
.
CPUPlace
())
feed_tensor_dict
[
feed_name
]
=
feed_tensor
self
.
executor
.
feed_and_split_tensor_into_local_scopes
(
feed_tensor_dict
)
exe
.
feed_and_split_tensor_into_local_scopes
(
feed_tensor_dict
)
elif
isinstance
(
feed
,
list
)
or
isinstance
(
feed
,
tuple
):
if
len
(
feed
)
!=
len
(
program
.
_places
):
raise
ValueError
(
...
...
@@ -436,10 +438,10 @@ class Executor(object):
tensor
=
tmp
res_dict
[
feed_name
]
=
tensor
res
.
append
(
res_dict
)
self
.
executor
.
feed_tensors_into_local_scopes
(
res
)
exe
.
feed_tensors_into_local_scopes
(
res
)
fetch_var_names
=
list
(
map
(
_to_name_str
,
fetch_list
))
self
.
executor
.
run
(
fetch_var_names
,
fetch_var_name
)
exe
.
run
(
fetch_var_names
,
fetch_var_name
)
arr
=
scope
.
find_var
(
fetch_var_name
).
get_lod_tensor_array
()
if
return_numpy
:
...
...
@@ -511,12 +513,9 @@ class Executor(object):
compiled
=
isinstance
(
program
,
compiler
.
CompiledProgram
)
# For backward compatibility, run directly.
if
not
compiled
:
if
not
self
.
executor
:
p
=
core
.
Place
()
p
.
set_place
(
self
.
place
)
self
.
executor
=
core
.
Executor
(
p
)
return
self
.
_run
(
program
,
self
.
_default_executor
,
feed
=
feed
,
fetch_list
=
fetch_list
,
feed_var_name
=
feed_var_name
,
...
...
@@ -526,7 +525,6 @@ class Executor(object):
use_program_cache
=
use_program_cache
)
program
.
_compile
(
scope
,
self
.
place
)
self
.
executor
=
program
.
_executor
if
program
.
_is_data_parallel
:
return
self
.
_run_parallel
(
program
,
...
...
@@ -536,12 +534,13 @@ class Executor(object):
fetch_var_name
=
fetch_var_name
,
return_numpy
=
return_numpy
)
elif
program
.
_is_inference
:
return
self
.
_run_inference
(
program
,
feed
)
return
self
.
_run_inference
(
program
.
_executor
,
feed
)
else
:
# TODO(panyx0718): Can compile program to optimize executor
# performance.
return
self
.
_run
(
program
.
_program
,
self
.
_default_executor
,
feed
=
feed
,
fetch_list
=
fetch_list
,
feed_var_name
=
feed_var_name
,
...
...
@@ -550,8 +549,8 @@ class Executor(object):
return_numpy
=
return_numpy
,
use_program_cache
=
use_program_cache
)
def
_run
(
self
,
program
,
feed
,
fetch_list
,
feed_var_name
,
fetch
_var_name
,
scope
,
return_numpy
,
use_program_cache
):
def
_run
(
self
,
program
,
exe
,
feed
,
fetch_list
,
feed
_var_name
,
fetch_var_name
,
scope
,
return_numpy
,
use_program_cache
):
if
feed
is
None
:
feed
=
{}
...
...
@@ -589,11 +588,11 @@ class Executor(object):
fetch_var_name
=
fetch_var_name
)
self
.
_feed_data
(
program
,
feed
,
feed_var_name
,
scope
)
self
.
executor
.
run
(
program
.
desc
,
scope
,
0
,
True
,
True
)
exe
.
run
(
program
.
desc
,
scope
,
0
,
True
,
True
)
outs
=
self
.
_fetch_data
(
fetch_list
,
fetch_var_name
,
scope
)
if
return_numpy
:
outs
=
as_numpy
(
outs
)
return
outs
def
_run_inference
(
self
,
program
,
feed
):
return
self
.
executor
.
run
(
feed
)
def
_run_inference
(
self
,
exe
,
feed
):
return
exe
.
run
(
feed
)
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