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
3c6b733d
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
3c6b733d
编写于
3月 04, 2019
作者:
Q
Qiao Longfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
remove exe context
上级
9573d610
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
24 addition
and
27 deletion
+24
-27
paddle/fluid/operators/distributed/parameter_recv.cc
paddle/fluid/operators/distributed/parameter_recv.cc
+4
-5
paddle/fluid/operators/distributed/parameter_recv.h
paddle/fluid/operators/distributed/parameter_recv.h
+1
-3
paddle/fluid/operators/distributed/parameter_send.cc
paddle/fluid/operators/distributed/parameter_send.cc
+15
-14
paddle/fluid/operators/distributed/parameter_send.h
paddle/fluid/operators/distributed/parameter_send.h
+2
-3
paddle/fluid/operators/distributed_ops/recv_op.cc
paddle/fluid/operators/distributed_ops/recv_op.cc
+1
-1
paddle/fluid/operators/distributed_ops/send_op.cc
paddle/fluid/operators/distributed_ops/send_op.cc
+1
-1
未找到文件。
paddle/fluid/operators/distributed/parameter_recv.cc
浏览文件 @
3c6b733d
...
...
@@ -40,7 +40,6 @@ using DDim = framework::DDim;
template
<
typename
T
>
void
ParameterRecv
<
T
>::
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
ExecutionContext
&
ctx
,
const
framework
::
Scope
&
scope
)
{
framework
::
Scope
*
local_scope
=
scope
.
NewTmpScope
();
...
...
@@ -48,8 +47,7 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
auto
&
cpu_ctx
=
*
pool
.
Get
(
platform
::
CPUPlace
());
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
(
ctx
.
Attr
<
int
>
(
"trainer_id"
));
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
(
0
);
auto
*
recv_var
=
scope
.
FindVar
(
rpc_ctx
.
var_name
);
...
...
@@ -80,12 +78,13 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
size_t
output_offset
=
0
;
framework
::
Tensor
*
recv_tensor
=
recv_var
->
GetMutable
<
framework
::
LoDTensor
>
();
auto
dev_ctx
=
paddle
::
platform
::
CPUDeviceContext
();
for
(
auto
*
in
:
recved_tensors
)
{
auto
in_stride
=
framework
::
stride_numel
(
in
->
dims
());
auto
out_stride
=
framework
::
stride_numel
(
recv_tensor
->
dims
());
StridedNumelCopyWithAxis
<
T
>
(
ctx
.
device_context
(),
0
,
recv_tensor
->
data
<
T
>
()
+
output_offset
,
out_stride
,
in
->
data
<
T
>
(),
in_stride
,
in_stride
[
0
]);
dev_ctx
,
0
,
recv_tensor
->
data
<
T
>
()
+
output_offset
,
out_stride
,
in
->
data
<
T
>
(),
in_stride
,
in_stride
[
0
]);
output_offset
+=
in_stride
[
0
];
}
}
...
...
paddle/fluid/operators/distributed/parameter_recv.h
浏览文件 @
3c6b733d
...
...
@@ -26,9 +26,7 @@ namespace distributed {
template
<
typename
T
>
struct
ParameterRecv
{
void
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
ExecutionContext
&
context
,
const
framework
::
Scope
&
scope
);
void
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
Scope
&
scope
);
};
};
// namespace distributed
...
...
paddle/fluid/operators/distributed/parameter_send.cc
浏览文件 @
3c6b733d
...
...
@@ -39,7 +39,6 @@ using DDim = framework::DDim;
template
<
typename
T
>
void
ParameterSend
<
T
>::
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
ExecutionContext
&
ctx
,
const
framework
::
Scope
&
scope
,
bool
sync
)
{
framework
::
Scope
*
local_scope
=
scope
.
NewTmpScope
();
...
...
@@ -47,8 +46,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
auto
&
cpu_ctx
=
*
pool
.
Get
(
platform
::
CPUPlace
());
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
(
ctx
.
Attr
<
int
>
(
"trainer_id"
));
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
(
0
);
auto
*
send_var
=
scope
.
FindVar
(
rpc_ctx
.
var_name
);
size_t
out_num
=
rpc_ctx
.
splited_var_names
.
size
();
...
...
@@ -105,7 +103,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
outs_rows_idx
[
out_idx
].
push_back
(
send_rows
[
i
]);
outs_dense_idx
[
out_idx
].
push_back
(
i
);
}
auto
place
=
ctx
.
Get
Place
();
auto
place
=
platform
::
CPU
Place
();
for
(
size_t
i
=
0
;
i
<
outs_rows_idx
.
size
();
++
i
)
{
auto
rows_idx
=
outs_rows_idx
[
i
];
...
...
@@ -118,22 +116,25 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
for
(
auto
idx
:
rows_idx
)
{
outs
[
i
]
->
mutable_rows
()
->
push_back
(
idx
-
abs_sections
[
i
]);
}
auto
dst
=
outs
[
i
]
->
mutable_value
()
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()
);
auto
dst
=
outs
[
i
]
->
mutable_value
()
->
mutable_data
<
T
>
(
place
);
for
(
size_t
j
=
0
;
j
<
rows_idx
.
size
();
j
++
)
{
if
(
platform
::
is_cpu_place
(
place
))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
dst
+
j
*
row_numel
,
platform
::
CPUPlace
(),
src
+
outs_dense_idx
[
i
][
j
]
*
row_numel
,
sizeof
(
T
)
*
row_numel
);
}
else
{
#ifdef PADDLE_WITH_CUDA
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
memory
::
Copy
(
platform
::
CUDAPlace
(),
dst
+
j
*
row_numel
,
platform
::
CUDAPlace
(),
src
+
outs_dense_idx
[
i
][
j
]
*
row_numel
,
sizeof
(
T
)
*
row_numel
,
stream
);
#else
PADDLE_THROW
(
"Paddle is not compiled with GPU"
);
#endif
PADDLE_THROW
(
"do not support GPU now"
);
/*
#ifdef PADDLE_WITH_CUDA
auto stream = ctx.cuda_device_context().stream();
memory::Copy(platform::CUDAPlace(), dst + j * row_numel,
platform::CUDAPlace(),
src + outs_dense_idx[i][j] * row_numel,
sizeof(T) * row_numel, stream);
#else
PADDLE_THROW("Paddle is not compiled with GPU");
#endif
*/
}
}
}
...
...
paddle/fluid/operators/distributed/parameter_send.h
浏览文件 @
3c6b733d
...
...
@@ -26,9 +26,8 @@ namespace distributed {
template
<
typename
T
>
struct
ParameterSend
{
void
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
ExecutionContext
&
context
,
const
framework
::
Scope
&
scope
,
bool
sync
);
void
operator
()(
const
RpcContext
&
rpc_ctx
,
const
framework
::
Scope
&
scope
,
bool
sync
);
};
};
// namespace distributed
...
...
paddle/fluid/operators/distributed_ops/recv_op.cc
浏览文件 @
3c6b733d
...
...
@@ -62,7 +62,7 @@ class RecvOp : public framework::OperatorBase {
framework
::
ExecutionContext
(
*
this
,
scope
,
*
dev_ctx
,
ctx
,
nullptr
);
auto
recv_functor
=
distributed
::
ParameterRecv
<
float
>
();
auto
rpc_ctx
=
distributed
::
RpcContext
(
outs
[
0
],
recv_varnames
,
epmap
,
{});
recv_functor
(
rpc_ctx
,
exe_ctx
,
scope
);
recv_functor
(
rpc_ctx
,
scope
);
}
else
{
if
(
with_barrier
)
{
std
::
vector
<
distributed
::
VarHandlePtr
>
rets
;
...
...
paddle/fluid/operators/distributed_ops/send_op.cc
浏览文件 @
3c6b733d
...
...
@@ -56,7 +56,7 @@ class SendOp : public framework::OperatorBase {
auto
send_functor
=
distributed
::
ParameterSend
<
float
>
();
auto
rpc_ctx
=
distributed
::
RpcContext
(
ins
[
0
],
send_varnames
,
epmap
,
height_sections
);
send_functor
(
rpc_ctx
,
exe_ctx
,
scope
,
static_cast
<
bool
>
(
sync_send
));
send_functor
(
rpc_ctx
,
scope
,
static_cast
<
bool
>
(
sync_send
));
}
else
{
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
...
...
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