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
机器未来
Paddle
提交
4b4e558a
P
Paddle
项目概览
机器未来
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
4b4e558a
编写于
8月 21, 2020
作者:
S
seiriosPlus
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
geo sparse init from pserver
上级
69c349a0
变更
1
显示空白变更内容
内联
并排
Showing
1 changed file
with
22 addition
and
15 deletion
+22
-15
paddle/fluid/operators/distributed/communicator.cc
paddle/fluid/operators/distributed/communicator.cc
+22
-15
未找到文件。
paddle/fluid/operators/distributed/communicator.cc
浏览文件 @
4b4e558a
...
@@ -676,8 +676,7 @@ void GeoCommunicator::RecvSparse(const std::string &varname) {
...
@@ -676,8 +676,7 @@ void GeoCommunicator::RecvSparse(const std::string &varname) {
v_delta
.
resize
(
numel
);
v_delta
.
resize
(
numel
);
auto
cpu_ctx
=
paddle
::
platform
::
CPUDeviceContext
();
auto
cpu_ctx
=
paddle
::
platform
::
CPUDeviceContext
();
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
float
>
(
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
float
>
(
cpu_ctx
);
paddle
::
platform
::
CPUDeviceContext
());
for
(
auto
j
=
0
;
j
<
static_cast
<
int
>
(
ids
.
size
());
++
j
)
{
for
(
auto
j
=
0
;
j
<
static_cast
<
int
>
(
ids
.
size
());
++
j
)
{
blas
.
VSUB
(
dims1
,
t_psrever
.
data
<
float
>
()
+
j
*
dims1
,
blas
.
VSUB
(
dims1
,
t_psrever
.
data
<
float
>
()
+
j
*
dims1
,
...
@@ -791,29 +790,35 @@ void GeoCommunicator::InitSparse() {
...
@@ -791,29 +790,35 @@ void GeoCommunicator::InitSparse() {
framework
::
Scope
&
local_scope
=
send_scope_
->
NewScope
();
framework
::
Scope
&
local_scope
=
send_scope_
->
NewScope
();
for
(
size_t
i
=
0
;
i
<
metas
.
size
();
i
++
)
{
for
(
auto
&
meta
:
metas
)
{
auto
&
meta
=
metas
[
i
];
auto
&
ctx
=
recv_varname_to_ctx_
.
at
(
meta
.
name
);
auto
&
ctx
=
recv_varname_to_ctx_
.
at
(
meta
.
name
);
auto
pserver_num
=
ctx
.
splited_varnames
.
size
();
auto
pserver_num
=
ctx
.
splited_varnames
.
size
();
for
(
size_t
j
=
0
;
j
<
ctx
.
splited_varnames
.
size
();
j
++
)
{
for
(
size_t
i
=
0
;
i
<
ctx
.
splited_varnames
.
size
();
i
++
)
{
auto
&
recv_var_name
=
ctx
.
splited_varnames
[
i
];
auto
&
recv_var_name
=
ctx
.
splited_varnames
[
i
];
auto
*
var
=
local_scope
.
Var
(
recv_var_name
);
var
->
GetMutable
<
framework
::
LoDTensor
>
();
distributed
::
VarHandlePtr
ret
;
distributed
::
VarHandlePtr
ret
;
ret
=
rpc_client
->
AsyncGetVarNoBarrier
(
ctx
.
epmap
[
i
],
cpu_ctx
,
local_scope
,
ret
=
rpc_client
->
AsyncGetVarNoBarrier
(
ctx
.
epmap
[
i
],
cpu_ctx
,
local_scope
,
recv_var_name
,
recv_var_name
);
recv_var_name
,
recv_var_name
);
width
=
recv_t
.
value
().
dims
()[
1
];
auto
*
recv_var
=
local_scope
.
FindVar
(
recv_var_name
);
auto
&
recv_t
=
recv_var
->
Get
<
framework
::
LoDTensor
>
();
auto
width
=
recv_t
.
dims
()[
1
];
auto
rows
=
recv_t
.
dims
()[
0
];
PADDLE_ENFORCE_EQ
(
PADDLE_ENFORCE_EQ
(
width
,
meta
.
value_dims
[
0
],
width
,
meta
.
value_dims
[
0
],
platform
::
errors
::
InvalidArgument
(
"sparse params do not match"
));
platform
::
errors
::
InvalidArgument
(
"sparse params do not match"
));
auto
*
recv_var
=
local_scope
->
FindVar
(
recv_var_name
);
auto
&
recv_t
=
recv_var
->
Get
<
framework
::
SelectedRows
>
();
std
::
vector
<
int64_t
>
ids
;
std
::
vector
<
int64_t
>
ids
;
std
::
transform
(
recv_t
.
rows
().
begin
(),
recv_t
.
rows
().
end
(),
std
::
back_inserter
(
ids
),
for
(
int
x
=
0
;
x
<
rows
;
x
++
)
{
[
&
](
int64_t
id
)
{
return
id
*
pserver_num
+
i
;
});
ids
.
push_back
(
x
*
pserver_num
+
i
);
}
std
::
vector
<
std
::
vector
<
std
::
vector
<
float
>
*>>
values
;
std
::
vector
<
std
::
vector
<
std
::
vector
<
float
>
*>>
values
;
auto
*
ins
=
distributed
::
LargeScaleKV
::
GetInstance
();
auto
*
ins
=
distributed
::
LargeScaleKV
::
GetInstance
();
...
@@ -824,18 +829,20 @@ void GeoCommunicator::InitSparse() {
...
@@ -824,18 +829,20 @@ void GeoCommunicator::InitSparse() {
PADDLE_ENFORCE_NE
(
ret
->
Wait
(),
0U
,
platform
::
errors
::
ExecutionTimeout
(
PADDLE_ENFORCE_NE
(
ret
->
Wait
(),
0U
,
platform
::
errors
::
ExecutionTimeout
(
"internal error in RPCClient"
));
"internal error in RPCClient"
));
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
float
>
(
cpu_ctx
);
auto
blas
=
math
::
GetBlas
<
platform
::
CPUDeviceContext
,
float
>
(
paddle
::
platform
::
CPUDeviceContext
());
for
(
size_t
k
=
0
;
k
<
ids
.
size
();
++
k
)
{
for
(
size_t
k
=
0
;
k
<
ids
.
size
();
++
k
)
{
blas
.
VCOPY
(
width
,
recv_t
.
value
().
data
<
float
>
()
+
k
*
width
,
blas
.
VCOPY
(
width
,
recv_t
.
data
<
float
>
()
+
k
*
width
,
values
[
k
][
0
]
->
data
());
values
[
k
][
0
]
->
data
());
}
}
local_scope
.
EraseVars
({
recv_var_name
});
}
}
}
}
send_scope_
->
DeleteScope
(
&
local_scope
);
send_scope_
->
DeleteScope
(
&
local_scope
);
VLOG
(
3
)
<<
"GeoCommunicator init sparse "
<<
varname
<<
" done "
;
VLOG
(
3
)
<<
"init sparse variable done"
;
VLOG
(
3
)
<<
"init sparse variable done"
;
}
}
...
...
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