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
3067114f
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
体验新版 GitCode,发现更多精彩内容 >>
提交
3067114f
编写于
2月 11, 2018
作者:
Y
Yang Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
clean up
上级
cd9e660d
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
3 addition
and
39 deletion
+3
-39
paddle/fluid/operators/nccl_op.cu.cc
paddle/fluid/operators/nccl_op.cu.cc
+0
-8
paddle/fluid/operators/parallel_do_op.cc
paddle/fluid/operators/parallel_do_op.cc
+0
-5
python/paddle/v2/fluid/backward.py
python/paddle/v2/fluid/backward.py
+0
-25
python/paddle/v2/fluid/tests/test_parallel_op.py
python/paddle/v2/fluid/tests/test_parallel_op.py
+3
-1
未找到文件。
paddle/fluid/operators/nccl_op.cu.cc
浏览文件 @
3067114f
...
...
@@ -47,11 +47,8 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
auto
ins
=
ctx
.
MultiInput
<
LoDTensor
>
(
"X"
);
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out"
);
LOG
(
INFO
)
<<
"------------------"
;
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
LOG
(
INFO
)
<<
"------------------"
;
ncclRedOp_t
reduction_op_
=
ncclSum
;
LOG
(
INFO
)
<<
"------------------"
;
if
(
reduction
==
"ncclMin"
)
{
reduction_op_
=
ncclMin
;
...
...
@@ -65,19 +62,14 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
PADDLE_THROW
(
"Invalid reduction. default ncclSum."
);
}
LOG
(
INFO
)
<<
"------------------"
;
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
LOG
(
INFO
)
<<
"------------------"
;
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
LOG
(
INFO
)
<<
"------------------"
;
// device id
int
gpu_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
ctx
.
GetPlace
()).
GetDeviceId
();
LOG
(
INFO
)
<<
"------------------"
;
int
idx
=
comm
->
GetCommId
(
gpu_id
);
LOG
(
INFO
)
<<
"------------------"
;
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
VLOG
(
1
)
<<
"gpu : "
<<
" invoke allreduce. send "
<<
ins
[
i
]
->
numel
()
<<
" recv "
...
...
paddle/fluid/operators/parallel_do_op.cc
浏览文件 @
3067114f
...
...
@@ -151,7 +151,6 @@ class ParallelDoOp : public framework::OperatorBase {
}
WaitOnPlaces
(
places
);
// PADDLE_ENFORCE_EQ(places.size(), sub_scopes.size());
std
::
vector
<
std
::
future
<
void
>>
workers
;
workers
.
reserve
(
places
.
size
());
for
(
size_t
place_idx
=
0
;
place_idx
<
sub_scopes
.
size
();
++
place_idx
)
{
...
...
@@ -219,21 +218,18 @@ class ParallelDoGradOp : public framework::OperatorBase {
auto
&
sub_scopes
=
scope
.
FindVar
(
Input
(
kParallelScopes
))
->
Get
<
std
::
vector
<
framework
::
Scope
*>>
();
auto
&
places
=
scope
.
FindVar
(
Input
(
kPlaces
))
->
Get
<
platform
::
PlaceList
>
();
// PADDLE_ENFORCE_EQ(places.size(), sub_scopes.size());
// feed output@grad
SplitTensorAndMoveTensorToScopes
(
scope
,
const_cast
<
std
::
vector
<
framework
::
Scope
*>
*>
(
&
sub_scopes
),
places
,
Inputs
(
framework
::
GradVarName
(
kOutputs
)));
WaitOnPlaces
(
places
);
LOG
(
INFO
)
<<
"places "
<<
places
.
size
();
// exe run
std
::
vector
<
std
::
future
<
void
>>
workers
;
for
(
size_t
i
=
0
;
i
<
sub_scopes
.
size
();
++
i
)
{
auto
&
place
=
places
[
i
];
auto
*
cur_scope
=
sub_scopes
[
i
];
LOG
(
INFO
)
<<
place
;
// execute
workers
.
emplace_back
(
framework
::
Async
([
program
,
cur_scope
,
place
,
block
]
{
...
...
@@ -242,7 +238,6 @@ class ParallelDoGradOp : public framework::OperatorBase {
false
/*create_local_scope*/
);
}));
}
LOG
(
INFO
)
<<
"places "
<<
places
.
size
();
for
(
auto
&
worker
:
workers
)
{
worker
.
wait
();
}
...
...
python/paddle/v2/fluid/backward.py
浏览文件 @
3067114f
...
...
@@ -230,44 +230,19 @@ def _callback_lookup_(op):
def
__call__
(
self
,
block
,
context
):
if
not
self
.
has_inserted_nccl_init
:
# global_block = block.program.global_block()
# op_desc = global_block.desc.append_op()
# var_desc = global_block.desc.var('nccl_com__do_not_change_')
# var_desc.set_type(core.VarDesc.VarType.NCCL_COM)
# self.nccl_com = global_block.create_var(
# name='nccl_com', type=core.VarDesc.VarType.NCCL_COM)
# framework.Operator(
# global_block,
# type='ncclInit',
# desc=op_desc,
# inputs={},
# outputs={'Communicator': [self.nccl_com]})
op_desc
=
_create_op_desc_
(
"ncclInit"
,
{
"parallel_scopes"
:
self
.
parallel_scopes_name
},
{
"Communicator"
:
[
'nccl_com__do_not_change_'
]},
{})
# block.desc.append_op().copy_from(op_desc)
print
(
serialize_op_decs
(
op_desc
))
block
.
program
.
global_block
().
desc
.
append_op
().
copy_from
(
op_desc
)
self
.
has_inserted_nccl_init
=
True
current_op_desc
=
context
[
"__current_op_desc__"
]
# print(serialize_op_decs(context))
for
o_param
in
current_op_desc
.
output_names
():
for
o_argu
in
current_op_desc
.
output
(
o_param
):
if
o_argu
in
self
.
param_grad_names
:
# # print("reduce", o_argu)
# op_desc = block.desc.append_op()
# op_desc.set_type("ncclAllReduce")
# op_desc.set_input("X", [o_argu])
#
# # FIXME(tonyyang-svail):
# # Looks like nccl_com has been changed to nccl_com_0
# op_desc.set_input("Communicator", ['nccl_com_0'])
# out_var = block.create_var()
# op_desc.set_output("Out", [out_var.name])
# op_desc.set_attr("reduction", "ncclSum")
allreduce_out_name
=
o_argu
+
"__nccl_all_reduce__"
op_desc
=
_create_op_desc_
(
"ncclAllReduce"
,
{
...
...
python/paddle/v2/fluid/tests/test_parallel_op.py
浏览文件 @
3067114f
...
...
@@ -175,7 +175,9 @@ class ParallelOpTest(BaseParallelForTest):
def
test_simple_fc
(
self
):
self
.
run_test
(
callback
=
self
.
__network__
,
feed
=
{
'img'
:
numpy
.
random
.
random
(
size
=
(
8
,
784
)).
astype
(
'float32'
)},
feed
=
{
'img'
:
numpy
.
random
.
random
(
size
=
(
51
,
784
)).
astype
(
'float32'
)
},
fetch
=
[
'fc1.w@GRAD'
])
def
test_fc_with_tiny_data
(
self
):
...
...
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