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
c8de7284
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
c8de7284
编写于
10月 13, 2019
作者:
Z
Zeng Jinle
提交者:
XiaoguangHu
10月 13, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix cuda dev_ctx by event, test=release/1.6 (#20559)
上级
66c084b7
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
29 addition
and
4 deletion
+29
-4
paddle/fluid/framework/details/sparse_all_reduce_op_handle.cc
...le/fluid/framework/details/sparse_all_reduce_op_handle.cc
+27
-4
paddle/fluid/framework/details/sparse_all_reduce_op_handle.h
paddle/fluid/framework/details/sparse_all_reduce_op_handle.h
+2
-0
未找到文件。
paddle/fluid/framework/details/sparse_all_reduce_op_handle.cc
浏览文件 @
c8de7284
...
...
@@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/details/sparse_all_reduce_op_handle.h"
#include <algorithm>
#include <utility>
#include "dgc/dgc.h"
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
...
...
@@ -41,11 +42,30 @@ SparseAllReduceOpHandle::SparseAllReduceOpHandle(
}
}
void
SparseAllReduceOpHandle
::
WaitInputVarGenerated
()
{
#ifdef PADDLE_WITH_CUDA
for
(
auto
&
p
:
dev_ctxes_
)
{
if
(
platform
::
is_gpu_place
(
p
.
first
))
{
int
dev_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
p
.
first
).
device
;
auto
*
compute_dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
GetByPlace
(
platform
::
CUDAPlace
(
dev_id
));
auto
*
dev_ctx
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
p
.
second
);
if
(
compute_dev_ctx
->
stream
()
!=
dev_ctx
->
stream
())
{
auto
&
event
=
events_
.
at
(
dev_id
);
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaEventRecord
(
event
,
compute_dev_ctx
->
stream
()));
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamWaitEvent
(
dev_ctx
->
stream
(),
event
,
0
));
}
}
}
#endif
}
void
SparseAllReduceOpHandle
::
RunImplEncoded
()
{
platform
::
RecordEvent
record_event
(
Name
());
WaitInputVarGenerated
();
auto
in_var_handles
=
DynamicCast
<
VarHandle
>
(
this
->
Inputs
());
auto
out_var_handles
=
DynamicCast
<
VarHandle
>
(
this
->
Outputs
());
PADDLE_ENFORCE_EQ
(
...
...
@@ -87,6 +107,8 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
PADDLE_ENFORCE
(
nranks_
>
1
);
std
::
vector
<
std
::
function
<
void
()
>>
all_reduce_calls
;
std
::
vector
<
memory
::
AllocationPtr
>
allocations
;
for
(
size_t
i
=
0
;
i
<
local_scopes_
.
size
();
++
i
)
{
auto
&
place
=
places_
[
i
];
auto
&
in
=
*
ins
[
i
];
...
...
@@ -104,7 +126,6 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
int
dev_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
place
).
device
;
auto
*
nccl_ctxs
=
nccl_ctxs_
->
GetRunEnvNCCLCtx
(
run_order_
,
false
);
auto
&
nccl_ctx
=
nccl_ctxs
->
at
(
dev_id
);
auto
*
dev_ctx
=
nccl_ctxs
->
DevCtx
(
dev_id
);
auto
stream
=
nccl_ctx
.
stream
();
auto
comm
=
nccl_ctx
.
comm_
;
...
...
@@ -112,8 +133,9 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
// dgc use ncclAllGather to get all the encoded data
// so the buffer need nranks.
int
buf_size
=
nranks_
*
encode_size
;
auto
tmp_ious_data
=
memory
::
Alloc
(
*
dev_ctx
,
buf_size
);
auto
tmp_ious_data
=
memory
::
Alloc
(
place
,
buf_size
);
void
*
gather_buff
=
reinterpret_cast
<
void
*>
(
tmp_ious_data
->
ptr
());
allocations
.
emplace_back
(
std
::
move
(
tmp_ious_data
));
VLOG
(
10
)
<<
"in_numel:"
<<
in_numel
<<
", out_numel:"
<<
out_numel
<<
", nranks:"
<<
nranks_
<<
", gather_buf size:"
<<
buf_size
...
...
@@ -126,6 +148,7 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
});
}
WaitInputVarGenerated
();
NCCLAllReduceFunc
(
all_reduce_calls
);
}
...
...
paddle/fluid/framework/details/sparse_all_reduce_op_handle.h
浏览文件 @
c8de7284
...
...
@@ -36,6 +36,8 @@ class SparseAllReduceOpHandle : public AllReduceOpHandle {
bool
is_encoded
=
false
,
int
nranks
=
-
1
);
std
::
string
Name
()
const
override
;
void
WaitInputVarGenerated
()
override
;
protected:
void
RunImpl
()
override
;
int
GetKValue
(
const
std
::
string
&
grad_name
);
...
...
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