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
BaiXuePrincess
Paddle
提交
23cb8259
P
Paddle
项目概览
BaiXuePrincess
/
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
0
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
0
Issue
0
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
23cb8259
编写于
7年前
作者:
D
Dong Zhihong
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
"add python test case"
上级
73883bde
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
74 addition
and
63 deletion
+74
-63
paddle/operators/nccl/nccl_gpu_common.cc
paddle/operators/nccl/nccl_gpu_common.cc
+1
-1
paddle/operators/nccl/nccl_gpu_common.h
paddle/operators/nccl/nccl_gpu_common.h
+1
-11
paddle/operators/nccl/nccl_ops.cc
paddle/operators/nccl/nccl_ops.cc
+35
-43
paddle/operators/nccl/nccl_ops.cu
paddle/operators/nccl/nccl_ops.cu
+16
-0
paddle/operators/nccl/nccl_ops.h
paddle/operators/nccl/nccl_ops.h
+21
-8
未找到文件。
paddle/operators/nccl/nccl_gpu_common.cc
浏览文件 @
23cb8259
...
@@ -18,7 +18,7 @@ NCCLManager::~NCCLManager() {
...
@@ -18,7 +18,7 @@ NCCLManager::~NCCLManager() {
int
idx
=
gid
%
gpus_
.
size
();
int
idx
=
gid
%
gpus_
.
size
();
// wait finish
// wait finish
PADDLE_ENFORCE
(
PADDLE_ENFORCE
(
cudaStreamWaitEvent
(
*
comm
->
streams_
[
idx
],
comm
->
events_
[
idx
],
0
));
cudaStreamWaitEvent
(
comm
->
streams_
[
idx
],
comm
->
events_
[
idx
],
0
));
PADDLE_ENFORCE
(
cudaEventDestroy
(
comm
->
events_
[
idx
]));
PADDLE_ENFORCE
(
cudaEventDestroy
(
comm
->
events_
[
idx
]));
...
...
This diff is collapsed.
Click to expand it.
paddle/operators/nccl/nccl_gpu_common.h
浏览文件 @
23cb8259
...
@@ -65,20 +65,10 @@ class WaitGroup {
...
@@ -65,20 +65,10 @@ class WaitGroup {
std
::
condition_variable
cv_
;
std
::
condition_variable
cv_
;
};
};
// class NCCLContext : public DeviceContext {
// public:
// explicit NCCLContext(GPUPlace place);
// virtual ~NCCLContext();
// private:
// std::vector<int> gpu_ids_;
// std::vector<cudaStream_t> streams_;
// };
// TODO(dzh) : make resources managed unified with framework
// TODO(dzh) : make resources managed unified with framework
struct
Communicator
{
struct
Communicator
{
std
::
vector
<
ncclComm_t
>
comms_
;
std
::
vector
<
ncclComm_t
>
comms_
;
std
::
vector
<
cudaStream_t
*
>
streams_
;
std
::
vector
<
cudaStream_t
>
streams_
;
std
::
vector
<
cudaEvent_t
>
events_
;
std
::
vector
<
cudaEvent_t
>
events_
;
std
::
vector
<
int
>
gpus_
;
std
::
vector
<
int
>
gpus_
;
WaitGroup
wg_
;
WaitGroup
wg_
;
...
...
This diff is collapsed.
Click to expand it.
paddle/operators/nccl/nccl_ops.cc
浏览文件 @
23cb8259
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/nccl/nccl_ops.h"
#include "paddle/operators/nccl/nccl_ops.h"
namespace
paddle
{
namespace
paddle
{
...
@@ -9,54 +20,27 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel {
...
@@ -9,54 +20,27 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel {
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
protected:
// allreduce do nothing in infershape
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
PADDLE_ENFORCE_NOT_NULL
(
" Input(X) of AllReduce op input should not be NULL"
);
ctx
.
InputVar
(
"X"
),
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
" Input(X) of AllReduce op input should not be NULL"
);
" Input(X) of AllReduce op input should not be NULL"
);
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
outs
=
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
"Out"
);
PADDLE_ENFORCE
(
ins
.
size
()
==
outs
.
size
(),
"Input(X) and Output(Out) must have same size"
);
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
outs
[
i
]
->
Resize
(
ins
[
i
]
->
dims
());
}
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
PADDLE_ENFORCE
((
reduction
==
"ncclSum"
||
reduction
==
"ncclProd"
||
reduction
==
"ncclMin"
||
reduction
==
"ncclMax"
),
"invalid reduction!"
);
}
};
// BcastSendOp
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
template
<
typename
T
>
class
NCCLBcastSendOp
final
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"X"
),
" Input(X) of BcastSend op input should not be NULL"
);
}
};
// BcastRecvOp
std
::
string
reduction
=
ctx
->
Attrs
().
Get
<
std
::
string
>
(
"reduction"
);
template
<
typename
T
>
PADDLE_ENFORCE
((
reduction
==
"ncclSum"
||
reduction
==
"ncclProd"
||
class
NCCLBcastRecvOp
final
:
public
framework
::
OperatorWithKernel
{
reduction
==
"ncclMin"
||
reduction
==
"ncclMax"
),
public:
"invalid reduction."
);
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
ctx
->
SetOutputsDim
(
"Out"
,
x_dims
);
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
OutputVar
(
"Out"
),
" Input(X) of BcastRecv op input should not be NULL"
);
}
}
};
};
// AllreduceOp
class
NCCLAllReduceOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
class
NCCLAllReduceOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
...
@@ -71,7 +55,9 @@ class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -71,7 +55,9 @@ class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
}
}
};
};
// BcastSendOp
class
NCCLBcastSendOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
class
NCCLBcastSendOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
...
@@ -82,7 +68,9 @@ class NCCLBcastSendOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -82,7 +68,9 @@ class NCCLBcastSendOpMaker : public framework::OpProtoAndCheckerMaker {
}
}
};
};
// BcastRecvOp
class
NCCLBcastRecvOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
class
NCCLBcastRecvOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
NCCLAllReduceOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
...
@@ -93,5 +81,9 @@ class NCCLBcastRecvOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -93,5 +81,9 @@ class NCCLBcastRecvOpMaker : public framework::OpProtoAndCheckerMaker {
}
}
};
};
}
// operators
}
// namespace operators
}
// paddle
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_WITHOUT_GRADIENT
(
ncclAllReduce
,
ops
::
NCCLAllReduceOp
,
ops
::
NCCLAllReduceOpMaker
);
This diff is collapsed.
Click to expand it.
paddle/operators/nccl/nccl_ops.cu
0 → 100644
浏览文件 @
23cb8259
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/operators/nccl/nccl_ops.h"
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_GPU_KERNEL
(
ncclAllReduce
,
ops
::
NCCLAllReduceKernel
<
float
>
);
\ No newline at end of file
This diff is collapsed.
Click to expand it.
paddle/operators/nccl/nccl_ops.h
浏览文件 @
23cb8259
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#pragma once
#include "paddle/framework/op_registry.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/nccl/nccl_gpu_common.h"
#include "paddle/operators/nccl/nccl_gpu_common.h"
...
@@ -14,11 +25,13 @@ class NCCLTypeWrapper;
...
@@ -14,11 +25,13 @@ class NCCLTypeWrapper;
template
<
>
template
<
>
class
NCCLTypeWrapper
<
float
>
{
class
NCCLTypeWrapper
<
float
>
{
public:
static
const
ncclDataType_t
type
=
ncclFloat
;
static
const
ncclDataType_t
type
=
ncclFloat
;
};
};
template
<
>
template
<
>
class
NCCLTypeWrapper
<
double
>
{
class
NCCLTypeWrapper
<
double
>
{
public:
static
const
ncclDataType_t
type
=
ncclDouble
;
static
const
ncclDataType_t
type
=
ncclDouble
;
};
};
...
@@ -49,10 +62,10 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
...
@@ -49,10 +62,10 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
auto
*
comm
=
m
->
GetCommunicator
(
gpus
);
auto
*
comm
=
m
->
GetCommunicator
(
gpus
);
comm
->
wg_
.
Add
(
1
);
comm
->
wg_
.
Add
(
1
);
auto
*
stream
=
&
dev_ctx
.
stream
();
auto
stream
=
dev_ctx
.
stream
();
// device id
// device id
int
gid
=
ctx
.
GetPlace
(
).
GetDeviceId
();
int
gid
=
static_cast
<
platform
::
GPUPlace
>
(
ctx
.
GetPlace
()
).
GetDeviceId
();
int
idx
=
gid
%
gpus
.
size
();
int
idx
=
gid
%
gpus
.
size
();
comm
->
streams_
[
idx
]
=
stream
;
comm
->
streams_
[
idx
]
=
stream
;
...
@@ -60,9 +73,8 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
...
@@ -60,9 +73,8 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE
(
PADDLE_ENFORCE
(
ncclAllReduce
(
ins
[
i
]
->
data
<
T
>
(),
outs
[
i
]
->
mutable_data
<
T
>
(),
ncclAllReduce
(
ins
[
i
]
->
data
<
T
>
(),
outs
[
i
]
->
mutable_data
<
T
>
(),
outs
[
i
]
->
numel
()
*
sizeof
(
T
),
NCCLTypeWrapper
<
T
>::
type
,
outs
[
i
]
->
numel
()
*
sizeof
(
T
),
NCCLTypeWrapper
<
T
>::
type
,
op_type
,
&
comm
->
comms_
[
idx
],
comm
->
streams_
[
idx
]));
op_type
,
comm
->
comms_
[
idx
],
comm
->
streams_
[
idx
]));
PADDLE_ENFORCE
(
PADDLE_ENFORCE
(
cudaEventRecord
(
comm
->
events_
[
idx
],
comm
->
streams_
[
idx
]));
cudaEventRecord
(
comm
->
events_
[
idx
],
*
comms_
->
streams_
[
idx
]));
// wait finish
// wait finish
PADDLE_ENFORCE
(
PADDLE_ENFORCE
(
...
@@ -71,8 +83,9 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
...
@@ -71,8 +83,9 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
comm
->
wg_
.
Done
();
comm
->
wg_
.
Done
();
wg
.
Wait
();
comm
->
wg_
.
Wait
();
}
}
};
};
}
}
}
// namespace operators
}
// namespace paddle
This diff is collapsed.
Click to expand it.
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
新手
引导
客服
返回
顶部