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
173660be
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 2 年 前同步成功
通知
2325
Star
20933
Fork
5424
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
未验证
提交
173660be
编写于
1月 25, 2021
作者:
J
Jacek Czaja
提交者:
GitHub
1月 25, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[oneDNN] Cache oneDNN stream not to recreate in each oneDNN op (#30358)
上级
ae0f88a9
变更
26
隐藏空白更改
内联
并排
Showing
26 changed file
with
81 addition
and
55 deletion
+81
-55
paddle/fluid/framework/data_layout_transform.cc
paddle/fluid/framework/data_layout_transform.cc
+1
-1
paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
...operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
+1
-1
paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
...luid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
+1
-1
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
+4
-4
paddle/fluid/operators/fused/mkldnn/multi_gru_mkldnn_op.cc
paddle/fluid/operators/fused/mkldnn/multi_gru_mkldnn_op.cc
+5
-5
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
+2
-2
paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
+2
-2
paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
+3
-3
paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc
paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
+3
-3
paddle/fluid/operators/mkldnn/interpolate_mkldnn_op.cc
paddle/fluid/operators/mkldnn/interpolate_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc
paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc
+2
-2
paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
+4
-5
paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
+2
-2
paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc
paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc
paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
+2
-2
paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc
paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc
+1
-1
paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
+2
-2
paddle/fluid/platform/device_context.cc
paddle/fluid/platform/device_context.cc
+26
-4
paddle/fluid/platform/device_context.h
paddle/fluid/platform/device_context.h
+7
-2
paddle/fluid/platform/mkldnn_helper.h
paddle/fluid/platform/mkldnn_helper.h
+1
-1
paddle/fluid/platform/mkldnn_reuse.h
paddle/fluid/platform/mkldnn_reuse.h
+5
-5
未找到文件。
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @
173660be
...
@@ -193,7 +193,7 @@ void innerTransDataLayoutFromMKLDNN(DataLayout in_layout, DataLayout out_layout,
...
@@ -193,7 +193,7 @@ void innerTransDataLayoutFromMKLDNN(DataLayout in_layout, DataLayout out_layout,
auto
reorder_p
=
auto
reorder_p
=
handler
.
AcquireReorder
(
reorder_dst_memory_p
,
reorder_src_memory_p
);
handler
.
AcquireReorder
(
reorder_dst_memory_p
,
reorder_src_memory_p
);
mkldnn
::
stream
astream
(
cpu_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
platform
::
RecordEvent
record_reorder
(
"ext_reorder"
,
platform
::
RecordEvent
record_reorder
(
"ext_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
*
reorder_dst_memory_p
);
reorder_p
->
execute
(
astream
,
*
reorder_src_memory_p
,
*
reorder_dst_memory_p
);
...
...
paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -48,7 +48,7 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
...
@@ -48,7 +48,7 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
platform
::
ReorderMKLDNNHandler
handler
(
tz
,
dout
->
type
(),
dout_type
,
dev_ctx
,
platform
::
ReorderMKLDNNHandler
handler
(
tz
,
dout
->
type
(),
dout_type
,
dev_ctx
,
onednn_engine
,
key
);
onednn_engine
,
key
);
mkldnn
::
stream
astream
(
onednn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
auto
reorder_src_memory_p
=
handler
.
AcquireSrcMemory
(
auto
reorder_src_memory_p
=
handler
.
AcquireSrcMemory
(
dout
->
format
(),
platform
::
to_void_cast
(
dout
->
data
<
T
>
()));
dout
->
format
(),
platform
::
to_void_cast
(
dout
->
data
<
T
>
()));
...
...
paddle/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h
浏览文件 @
173660be
...
@@ -68,7 +68,7 @@ class EltwiseMKLDNNKernel : public framework::OpKernel<T> {
...
@@ -68,7 +68,7 @@ class EltwiseMKLDNNKernel : public framework::OpKernel<T> {
const
auto
binary_prim
=
handler
.
AcquireForwardPrimitive
();
const
auto
binary_prim
=
handler
.
AcquireForwardPrimitive
();
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
{
DNNL_ARG_SRC_0
,
*
src_x_memory
},
{
DNNL_ARG_SRC_0
,
*
src_x_memory
},
...
...
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -246,7 +246,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
...
@@ -246,7 +246,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
this
->
fwd_pd_
->
src_iter_desc
(),
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
this
->
fwd_pd_
->
src_iter_desc
(),
this
->
engine_
);
this
->
engine_
);
dnnl
::
stream
astream
(
this
->
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_h0_memory
,
*
memory_p
,
attr_
)
dnnl
::
reorder
(
user_h0_memory
,
*
memory_p
,
attr_
)
.
execute
(
astream
,
user_h0_memory
,
*
memory_p
);
.
execute
(
astream
,
user_h0_memory
,
*
memory_p
);
...
@@ -284,7 +284,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
...
@@ -284,7 +284,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
this
->
fwd_pd_
->
weights_layer_desc
(),
this
->
engine_
);
this
->
fwd_pd_
->
weights_layer_desc
(),
this
->
engine_
);
dnnl
::
stream
astream
(
this
->
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attr_
)
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attr_
)
.
execute
(
astream
,
user_memory
,
*
memory_p
);
.
execute
(
astream
,
user_memory
,
*
memory_p
);
...
@@ -337,7 +337,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
...
@@ -337,7 +337,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
this
->
fwd_pd_
->
weights_iter_desc
(),
this
->
engine_
);
this
->
fwd_pd_
->
weights_iter_desc
(),
this
->
engine_
);
dnnl
::
stream
astream
(
this
->
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attr_
)
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attr_
)
.
execute
(
astream
,
user_memory
,
*
memory_p
);
.
execute
(
astream
,
user_memory
,
*
memory_p
);
...
@@ -469,7 +469,7 @@ class FusionGRUMKLDNNKernel : public framework::OpKernel<T> {
...
@@ -469,7 +469,7 @@ class FusionGRUMKLDNNKernel : public framework::OpKernel<T> {
auto
gru_forward_p
=
handler
.
AcquireForwardPrimitive
();
auto
gru_forward_p
=
handler
.
AcquireForwardPrimitive
();
dnnl
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
gru_forward_p
->
execute
(
astream
,
gru_args
);
gru_forward_p
->
execute
(
astream
,
gru_args
);
astream
.
wait
();
astream
.
wait
();
...
...
paddle/fluid/operators/fused/mkldnn/multi_gru_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -292,7 +292,7 @@ class MultiGRUHandler {
...
@@ -292,7 +292,7 @@ class MultiGRUHandler {
auto
gru_forward_p0
=
AcquireGruPrimitive
(
layer
,
dir
);
auto
gru_forward_p0
=
AcquireGruPrimitive
(
layer
,
dir
);
dnnl
::
stream
astream
(
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
gru_forward_p0
->
execute
(
astream
,
gru_args
);
gru_forward_p0
->
execute
(
astream
,
gru_args
);
astream
.
wait
();
astream
.
wait
();
return
out_mem
;
return
out_mem
;
...
@@ -315,7 +315,7 @@ class MultiGRUHandler {
...
@@ -315,7 +315,7 @@ class MultiGRUHandler {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
gru_pds_
[{
layer
,
dir
}]
->
src_iter_desc
(),
engine_
);
gru_pds_
[{
layer
,
dir
}]
->
src_iter_desc
(),
engine_
);
dnnl
::
stream
astream
(
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_h0_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
dnnl
::
reorder
(
user_h0_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
.
execute
(
astream
,
user_h0_memory
,
*
memory_p
);
.
execute
(
astream
,
user_h0_memory
,
*
memory_p
);
...
@@ -354,7 +354,7 @@ class MultiGRUHandler {
...
@@ -354,7 +354,7 @@ class MultiGRUHandler {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
gru_pds_
[{
layer
,
dir
}]
->
weights_layer_desc
(),
engine_
);
gru_pds_
[{
layer
,
dir
}]
->
weights_layer_desc
(),
engine_
);
dnnl
::
stream
astream
(
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
.
execute
(
astream
,
user_memory
,
*
memory_p
);
.
execute
(
astream
,
user_memory
,
*
memory_p
);
...
@@ -410,7 +410,7 @@ class MultiGRUHandler {
...
@@ -410,7 +410,7 @@ class MultiGRUHandler {
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
memory_p
=
std
::
make_shared
<
dnnl
::
memory
>
(
gru_pds_
[{
layer
,
dir
}]
->
weights_iter_desc
(),
engine_
);
gru_pds_
[{
layer
,
dir
}]
->
weights_iter_desc
(),
engine_
);
dnnl
::
stream
astream
(
engine_
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
dnnl
::
reorder
(
user_memory
,
*
memory_p
,
attrs_
[
2
*
layer
+
(
dir
==
R2L
)])
.
execute
(
astream
,
user_memory
,
*
memory_p
);
.
execute
(
astream
,
user_memory
,
*
memory_p
);
...
@@ -516,7 +516,7 @@ class MultiGRUHandler {
...
@@ -516,7 +516,7 @@ class MultiGRUHandler {
auto
concat_p
=
AcquireConcatPrimitive
(
layer
);
auto
concat_p
=
AcquireConcatPrimitive
(
layer
);
dnnl
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
concat_p
->
execute
(
astream
,
concat_args
);
concat_p
->
execute
(
astream
,
concat_args
);
astream
.
wait
();
astream
.
wait
();
return
out_mem
;
return
out_mem
;
...
...
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -112,7 +112,7 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
...
@@ -112,7 +112,7 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
auto
dst_memory_p
=
is_inplaced
?
src_memory_p
:
handler
.
AcquireDstMemory
(
y
);
auto
dst_memory_p
=
is_inplaced
?
src_memory_p
:
handler
.
AcquireDstMemory
(
y
);
auto
activation_p
=
handler
.
AcquireForwardPrimitive
();
auto
activation_p
=
handler
.
AcquireForwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
activation_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
src_memory_p
},
activation_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
src_memory_p
},
{
MKLDNN_ARG_TO
,
*
dst_memory_p
}});
{
MKLDNN_ARG_TO
,
*
dst_memory_p
}});
astream
.
wait
();
astream
.
wait
();
...
@@ -158,7 +158,7 @@ void eltwise_grad(const framework::ExecutionContext &ctx,
...
@@ -158,7 +158,7 @@ void eltwise_grad(const framework::ExecutionContext &ctx,
auto
diff_src_memory_p
=
handler
.
AcquireDiffSrcMemory
(
diff_x
);
auto
diff_src_memory_p
=
handler
.
AcquireDiffSrcMemory
(
diff_x
);
auto
activation_backward_p
=
handler
.
AcquireBackwardPrimitive
();
auto
activation_backward_p
=
handler
.
AcquireBackwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
activation_backward_p
->
execute
(
astream
,
activation_backward_p
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory_p
},
{{
MKLDNN_ARG_SRC
,
*
src_memory_p
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory_p
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory_p
},
...
...
paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -220,7 +220,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -220,7 +220,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
y
->
set_layout
(
DataLayout
::
kMKLDNN
);
y
->
set_layout
(
DataLayout
::
kMKLDNN
);
y
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
dst_memory
));
y
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
dst_memory
));
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
batch_norm_p
->
execute
(
astream
,
batch_norm_p
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
{
MKLDNN_ARG_SCALE_SHIFT
,
*
scaleshift_memory
},
{
MKLDNN_ARG_SCALE_SHIFT
,
*
scaleshift_memory
},
...
@@ -321,7 +321,7 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -321,7 +321,7 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
// finally create batch_norm backward primitive
// finally create batch_norm backward primitive
auto
batch_norm_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
auto
batch_norm_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
batch_norm_bwd_p
->
execute
(
batch_norm_bwd_p
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
{
MKLDNN_ARG_MEAN
,
*
mean_memory
},
{
MKLDNN_ARG_MEAN
,
*
mean_memory
},
...
...
paddle/fluid/operators/mkldnn/concat_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -202,7 +202,7 @@ class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -202,7 +202,7 @@ class ConcatMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
output
->
mutable_data
<
T
>
(
place
,
concat_pd
->
dst_desc
().
get_size
()));
output
->
mutable_data
<
T
>
(
place
,
concat_pd
->
dst_desc
().
get_size
()));
}
}
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
std
::
unordered_map
<
int
,
memory
>
args
;
std
::
unordered_map
<
int
,
memory
>
args
;
for
(
size_t
i
=
0
;
i
<
multi_input
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
multi_input
.
size
();
++
i
)
{
args
.
insert
({
MKLDNN_ARG_MULTIPLE_SRC
+
i
,
(
*
srcs
).
at
(
i
)});
args
.
insert
({
MKLDNN_ARG_MULTIPLE_SRC
+
i
,
(
*
srcs
).
at
(
i
)});
...
...
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -471,7 +471,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -471,7 +471,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
args
.
insert
({
MKLDNN_ARG_BIAS
,
*
bias_memory_p
});
args
.
insert
({
MKLDNN_ARG_BIAS
,
*
bias_memory_p
});
}
}
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
conv_p
->
execute
(
astream
,
args
);
conv_p
->
execute
(
astream
,
args
);
astream
.
wait
();
astream
.
wait
();
...
@@ -553,7 +553,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -553,7 +553,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
conv_p
=
std
::
static_pointer_cast
<
mkldnn
::
convolution_forward
>
(
conv_p
=
std
::
static_pointer_cast
<
mkldnn
::
convolution_forward
>
(
dev_ctx
.
GetBlob
(
prim_key
));
dev_ctx
.
GetBlob
(
prim_key
));
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
conv_p
==
nullptr
||
!
is_test
)
{
if
(
conv_p
==
nullptr
||
!
is_test
)
{
float
fuse_alpha
=
ctx
.
Attr
<
float
>
(
"fuse_alpha"
);
float
fuse_alpha
=
ctx
.
Attr
<
float
>
(
"fuse_alpha"
);
...
@@ -1045,7 +1045,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -1045,7 +1045,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
user_weights_md
,
to_void_cast
<
T
>
(
filter_data
));
user_weights_md
,
to_void_cast
<
T
>
(
filter_data
));
auto
user_diff_dst_memory_p
=
handler
.
AcquireDiffDstMemory
(
auto
user_diff_dst_memory_p
=
handler
.
AcquireDiffDstMemory
(
user_diff_dst_md
,
to_void_cast
<
T
>
(
output_grad_data
));
user_diff_dst_md
,
to_void_cast
<
T
>
(
output_grad_data
));
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
filter_grad
)
{
if
(
filter_grad
)
{
auto
src_memory_p
=
handler
.
AcquireSrcMemoryFromWeightsPrimitive
(
auto
src_memory_p
=
handler
.
AcquireSrcMemoryFromWeightsPrimitive
(
user_src_memory_p
,
pipeline
);
user_src_memory_p
,
pipeline
);
...
...
paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -242,7 +242,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -242,7 +242,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
conv_p
=
handler
.
AcquireConvolution
();
auto
conv_p
=
handler
.
AcquireConvolution
();
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
bias
)
{
if
(
bias
)
{
const
T
*
bias_data
=
bias
->
data
<
T
>
();
const
T
*
bias_data
=
bias
->
data
<
T
>
();
auto
user_bias_md
=
platform
::
MKLDNNMemDesc
(
auto
user_bias_md
=
platform
::
MKLDNNMemDesc
(
...
...
paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -124,7 +124,7 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
...
@@ -124,7 +124,7 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
dst_memory
->
set_data_handle
(
output
->
mutable_data
<
float
>
(
ctx
.
GetPlace
()));
dst_memory
->
set_data_handle
(
output
->
mutable_data
<
float
>
(
ctx
.
GetPlace
()));
}
}
mkldnn
::
stream
astream
(
engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
reorder_p
->
execute
(
astream
,
*
src_memory
,
*
dst_memory
);
reorder_p
->
execute
(
astream
,
*
src_memory
,
*
dst_memory
);
astream
.
wait
();
astream
.
wait
();
...
...
paddle/fluid/operators/mkldnn/fc_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -137,7 +137,7 @@ class FCPrimitiveFactory {
...
@@ -137,7 +137,7 @@ class FCPrimitiveFactory {
}
}
void
Execute
()
{
void
Execute
()
{
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
bias_
)
{
if
(
bias_
)
{
fc_
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
input_
},
fc_
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
input_
},
{
MKLDNN_ARG_WEIGHTS
,
*
weights_
},
{
MKLDNN_ARG_WEIGHTS
,
*
weights_
},
...
@@ -280,7 +280,7 @@ class FCPrimitiveFactory {
...
@@ -280,7 +280,7 @@ class FCPrimitiveFactory {
auto
dst_mem
=
std
::
make_shared
<
memory
>
(
dst_desc
,
engine_
);
auto
dst_mem
=
std
::
make_shared
<
memory
>
(
dst_desc
,
engine_
);
auto
reorder
=
mkldnn
::
reorder
(
src_mem
,
*
dst_mem
);
auto
reorder
=
mkldnn
::
reorder
(
src_mem
,
*
dst_mem
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
...
@@ -309,7 +309,7 @@ class FCPrimitiveFactory {
...
@@ -309,7 +309,7 @@ class FCPrimitiveFactory {
attributes
.
set_output_scales
(
mask
,
scale_data
);
attributes
.
set_output_scales
(
mask
,
scale_data
);
auto
reorder
=
mkldnn
::
reorder
(
*
src_mem
,
*
dst_mem
,
attributes
);
auto
reorder
=
mkldnn
::
reorder
(
*
src_mem
,
*
dst_mem
,
attributes
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
...
paddle/fluid/operators/mkldnn/interpolate_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -154,7 +154,7 @@ class InterpolateMKLDNNKernel : public framework::OpKernel<T> {
...
@@ -154,7 +154,7 @@ class InterpolateMKLDNNKernel : public framework::OpKernel<T> {
auto
resampling_prim
=
handler
.
AcquireForwardPrimitive
();
auto
resampling_prim
=
handler
.
AcquireForwardPrimitive
();
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
const
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
=
{
{
DNNL_ARG_SRC
,
*
src_memory_p
},
{
DNNL_ARG_DST
,
*
dst_memory_p
}};
{
DNNL_ARG_SRC
,
*
src_memory_p
},
{
DNNL_ARG_DST
,
*
dst_memory_p
}};
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
resampling_prim
->
execute
(
astream
,
args
);
resampling_prim
->
execute
(
astream
,
args
);
astream
.
wait
();
astream
.
wait
();
...
...
paddle/fluid/operators/mkldnn/layer_norm_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -120,7 +120,7 @@ class LayerNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -120,7 +120,7 @@ class LayerNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
layer_norm_p
=
handler
.
AcquireForwardPrimitive
();
auto
layer_norm_p
=
handler
.
AcquireForwardPrimitive
();
dnnl
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
;
std
::
unordered_map
<
int
,
dnnl
::
memory
>
args
;
args
.
insert
({
DNNL_ARG_SRC
,
*
src_memory
});
args
.
insert
({
DNNL_ARG_SRC
,
*
src_memory
});
...
...
paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -59,7 +59,7 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -59,7 +59,7 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
workspace_memory
=
handler
.
AcquireWorkspaceMemory
(
mid
);
auto
workspace_memory
=
handler
.
AcquireWorkspaceMemory
(
mid
);
mid
->
set_layout
(
framework
::
DataLayout
::
kMKLDNN
);
mid
->
set_layout
(
framework
::
DataLayout
::
kMKLDNN
);
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
!
workspace_memory
->
get_desc
().
is_zero
())
{
if
(
!
workspace_memory
->
get_desc
().
is_zero
())
{
mid
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
workspace_memory
));
mid
->
set_format
(
platform
::
GetMKLDNNFormat
(
*
workspace_memory
));
lrn_p
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
lrn_p
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
...
@@ -118,7 +118,7 @@ class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -118,7 +118,7 @@ class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
auto
lrn_bwd
=
handler
.
AcquireBackwardPrimitive
();
auto
lrn_bwd
=
handler
.
AcquireBackwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
lrn_bwd
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
lrn_bwd
->
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
src_memory
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory
},
{
MKLDNN_ARG_DIFF_SRC
,
*
diff_src_memory
},
{
MKLDNN_ARG_DIFF_SRC
,
*
diff_src_memory
},
...
...
paddle/fluid/operators/mkldnn/mul_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -109,7 +109,7 @@ class MulPrimitiveFactory {
...
@@ -109,7 +109,7 @@ class MulPrimitiveFactory {
auto
reorder
=
mkldnn
::
reorder
(
reorder_pd
);
auto
reorder
=
mkldnn
::
reorder
(
reorder_pd
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
@@ -184,7 +184,7 @@ class MulPrimitiveFactory {
...
@@ -184,7 +184,7 @@ class MulPrimitiveFactory {
}
}
void
Execute
()
{
void
Execute
()
{
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
(
*
mul_
).
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
x_input_
},
(
*
mul_
).
execute
(
astream
,
{{
MKLDNN_ARG_SRC
,
*
x_input_
},
{
MKLDNN_ARG_WEIGHTS
,
*
y_input_
},
{
MKLDNN_ARG_WEIGHTS
,
*
y_input_
},
{
MKLDNN_ARG_DST
,
*
output_
}});
{
MKLDNN_ARG_DST
,
*
output_
}});
...
@@ -270,8 +270,7 @@ class MulPrimitiveFactory {
...
@@ -270,8 +270,7 @@ class MulPrimitiveFactory {
auto
reorder
=
mkldnn
::
reorder
(
src_mem
,
dst_mem
);
auto
reorder
=
mkldnn
::
reorder
(
src_mem
,
dst_mem
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
();
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
@@ -355,7 +354,7 @@ class MulMKLDNNKernel : public framework::OpKernel<XT> {
...
@@ -355,7 +354,7 @@ class MulMKLDNNKernel : public framework::OpKernel<XT> {
"Operator DNNL Mul must use CPUPlace"
));
"Operator DNNL Mul must use CPUPlace"
));
platform
::
MKLDNNDeviceContext
::
tls
().
log_lib_version
();
platform
::
MKLDNNDeviceContext
::
tls
().
log_lib_version
();
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
const
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
const
Tensor
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
const
Tensor
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
const
Tensor
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
const
Tensor
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
...
...
paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -51,7 +51,7 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -51,7 +51,7 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
pool_p
=
handler
.
AcquireForwardPrimitive
();
auto
pool_p
=
handler
.
AcquireForwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
((
ctx
.
Attr
<
bool
>
(
"is_test"
)
==
false
)
&&
if
((
ctx
.
Attr
<
bool
>
(
"is_test"
)
==
false
)
&&
(
ctx
.
Attr
<
std
::
string
>
(
"pooling_type"
)
==
"max"
))
{
(
ctx
.
Attr
<
std
::
string
>
(
"pooling_type"
)
==
"max"
))
{
// Training
// Training
...
@@ -154,7 +154,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -154,7 +154,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
auto
pool_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
auto
pool_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
pooling_type
==
"max"
)
{
if
(
pooling_type
==
"max"
)
{
// Max - pooling needs Workspace
// Max - pooling needs Workspace
auto
workspace_memory
=
handler
.
AcquireWorkspaceMemory
();
auto
workspace_memory
=
handler
.
AcquireWorkspaceMemory
();
...
...
paddle/fluid/operators/mkldnn/quantize_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -140,7 +140,7 @@ class QuantOpKernel : public framework::OpKernel<T> {
...
@@ -140,7 +140,7 @@ class QuantOpKernel : public framework::OpKernel<T> {
}
}
}
}
mkldnn
::
stream
astream
(
engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
...
paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -137,7 +137,7 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
...
@@ -137,7 +137,7 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
}
}
}
}
dnnl
::
stream
astream
(
engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
{
{
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
...
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -117,7 +117,7 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
...
@@ -117,7 +117,7 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
auto
softmax_p
=
handler
.
AcquireForwardPrimitive
();
auto
softmax_p
=
handler
.
AcquireForwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
paddle
::
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
softmax_p
->
execute
(
astream
,
{{
DNNL_ARG_SRC
,
*
softmax_src_memory_p
},
softmax_p
->
execute
(
astream
,
{{
DNNL_ARG_SRC
,
*
softmax_src_memory_p
},
{
DNNL_ARG_DST
,
*
softmax_dst_memory_p
}});
{
DNNL_ARG_DST
,
*
softmax_dst_memory_p
}});
astream
.
wait
();
astream
.
wait
();
...
@@ -169,7 +169,7 @@ class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel<T> {
...
@@ -169,7 +169,7 @@ class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel<T> {
auto
softmax_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
auto
softmax_bwd_p
=
handler
.
AcquireBackwardPrimitive
();
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
softmax_bwd_p
->
execute
(
astream
,
softmax_bwd_p
->
execute
(
astream
,
{{
MKLDNN_ARG_DST
,
*
dst_memory_p
},
{{
MKLDNN_ARG_DST
,
*
dst_memory_p
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory_p
},
{
MKLDNN_ARG_DIFF_DST
,
*
diff_dst_memory_p
},
...
...
paddle/fluid/operators/mkldnn/sum_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -178,7 +178,7 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -178,7 +178,7 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
}
}
args
.
insert
({
MKLDNN_ARG_DST
,
*
dst_mem
});
args
.
insert
({
MKLDNN_ARG_DST
,
*
dst_mem
});
mkldnn
::
stream
astream
(
dev_ctx
.
GetEngine
()
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
sum_p
->
execute
(
astream
,
args
);
sum_p
->
execute
(
astream
,
args
);
astream
.
wait
();
astream
.
wait
();
...
...
paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc
浏览文件 @
173660be
...
@@ -61,7 +61,7 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -61,7 +61,7 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto
transpose_p
=
handler
.
AcquireTranspose
(
transpose_dst_memory_p
,
auto
transpose_p
=
handler
.
AcquireTranspose
(
transpose_dst_memory_p
,
transpose_src_memory_p
);
transpose_src_memory_p
);
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
transpose_p
->
execute
(
astream
,
*
transpose_src_memory_p
,
transpose_p
->
execute
(
astream
,
*
transpose_src_memory_p
,
*
transpose_dst_memory_p
);
*
transpose_dst_memory_p
);
astream
.
wait
();
astream
.
wait
();
...
@@ -116,7 +116,7 @@ class TransposeMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
...
@@ -116,7 +116,7 @@ class TransposeMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
auto
transpose_p
=
handler
.
AcquireTranspose
(
transpose_dst_memory_p
,
auto
transpose_p
=
handler
.
AcquireTranspose
(
transpose_dst_memory_p
,
transpose_src_memory_p
);
transpose_src_memory_p
);
mkldnn
::
stream
astream
(
mkldnn_engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
transpose_p
->
execute
(
astream
,
*
transpose_src_memory_p
,
transpose_p
->
execute
(
astream
,
*
transpose_src_memory_p
,
*
transpose_dst_memory_p
);
*
transpose_dst_memory_p
);
astream
.
wait
();
astream
.
wait
();
...
...
paddle/fluid/platform/device_context.cc
浏览文件 @
173660be
...
@@ -458,20 +458,34 @@ Place CUDAPinnedDeviceContext::GetPlace() const { return place_; }
...
@@ -458,20 +458,34 @@ Place CUDAPinnedDeviceContext::GetPlace() const { return place_; }
#ifdef PADDLE_WITH_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
MKLDNNDeviceContext
::
MKLDNNDeviceContext
(
CPUPlace
place
)
MKLDNNDeviceContext
::
MKLDNNDeviceContext
(
CPUPlace
place
)
:
CPUDeviceContext
(
place
),
:
CPUDeviceContext
(
place
),
p_blobmap_
()
{
engine_
(
mkldnn
::
engine
::
kind
::
cpu
,
0
),
p_blobmap_
()
{
p_blobmap_
.
reset
(
new
BlobMap
());
p_blobmap_
.
reset
(
new
BlobMap
());
p_mutex_
.
reset
(
new
std
::
mutex
());
p_mutex_
.
reset
(
new
std
::
mutex
());
}
}
MKLDNNDeviceContextThreadLocals
::
Body
::
Body
()
{
MKLDNNDeviceContextThreadLocals
::
Body
::
Body
()
:
cur_engine
(
mkldnn
::
engine
::
kind
::
cpu
,
0
),
cur_stream
(
cur_engine
)
{
cur_mkldnn_session_id
=
kMKLDNNSessionID_Default
;
cur_mkldnn_session_id
=
kMKLDNNSessionID_Default
;
cur_input_shape_str
=
""
;
cur_input_shape_str
=
""
;
cur_input_shape_cache_capacity
=
1
;
cur_input_shape_cache_capacity
=
1
;
cur_paddle_data_layout
=
paddle
::
framework
::
DataLayout
::
kNCHW
;
cur_paddle_data_layout
=
paddle
::
framework
::
DataLayout
::
kNCHW
;
}
}
// When Thread finish we clear oneDNN cache
// This is needed when we have one executor used by many threads
// e.g. test_analyzer_detect. Thread ID is not part of caching key
// (for naive executor) so we need to clear cache when one thread finish
// and other is to start inference
// TODO(jczaja): Ideally it would be good to clear only part of cache
// related to thread that is to be terminated
MKLDNNDeviceContextThreadLocals
::
Body
::~
Body
()
{
auto
cpu_place
=
paddle
::
platform
::
CPUPlace
();
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
platform
::
MKLDNNDeviceContext
*
dev_ctx
=
(
platform
::
MKLDNNDeviceContext
*
)
pool
.
Get
(
cpu_place
);
dev_ctx
->
ResetBlobMap
();
}
void
MKLDNNDeviceContextThreadLocals
::
Body
::
set_cur_mkldnn_session_id
(
void
MKLDNNDeviceContextThreadLocals
::
Body
::
set_cur_mkldnn_session_id
(
size_t
sid
)
{
size_t
sid
)
{
cur_mkldnn_session_id
=
sid
;
cur_mkldnn_session_id
=
sid
;
...
@@ -508,6 +522,14 @@ void MKLDNNDeviceContextThreadLocals::Body::log_lib_version(void) {
...
@@ -508,6 +522,14 @@ void MKLDNNDeviceContextThreadLocals::Body::log_lib_version(void) {
}
}
}
}
const
mkldnn
::
engine
&
MKLDNNDeviceContextThreadLocals
::
Body
::
get_engine
(
void
)
{
return
cur_engine
;
}
mkldnn
::
stream
&
MKLDNNDeviceContextThreadLocals
::
Body
::
get_stream
(
void
)
{
return
cur_stream
;
}
void
MKLDNNDeviceContext
::
ResetBlobMap
()
{
void
MKLDNNDeviceContext
::
ResetBlobMap
()
{
std
::
lock_guard
<
decltype
(
*
p_mutex_
)
>
lock
(
*
p_mutex_
);
std
::
lock_guard
<
decltype
(
*
p_mutex_
)
>
lock
(
*
p_mutex_
);
if
(
!
block_next_cache_clearing_
)
{
if
(
!
block_next_cache_clearing_
)
{
...
...
paddle/fluid/platform/device_context.h
浏览文件 @
173660be
...
@@ -525,8 +525,12 @@ class MKLDNNDeviceContextThreadLocals {
...
@@ -525,8 +525,12 @@ class MKLDNNDeviceContextThreadLocals {
// Recently registered data_format. This is needed to
// Recently registered data_format. This is needed to
// know for converting MKL-DNN Tensor to non MKL-DNN
// know for converting MKL-DNN Tensor to non MKL-DNN
paddle
::
framework
::
DataLayout
cur_paddle_data_layout
;
paddle
::
framework
::
DataLayout
cur_paddle_data_layout
;
// MKL-DNN stream used for execution of primitives (per-thread)
mkldnn
::
engine
cur_engine
;
mkldnn
::
stream
cur_stream
;
Body
();
Body
();
~
Body
();
void
set_cur_mkldnn_session_id
(
size_t
sid
);
void
set_cur_mkldnn_session_id
(
size_t
sid
);
size_t
get_cur_mkldnn_session_id
(
void
);
size_t
get_cur_mkldnn_session_id
(
void
);
void
set_cur_input_shape_str
(
std
::
string
input_shape_str
);
void
set_cur_input_shape_str
(
std
::
string
input_shape_str
);
...
@@ -534,6 +538,8 @@ class MKLDNNDeviceContextThreadLocals {
...
@@ -534,6 +538,8 @@ class MKLDNNDeviceContextThreadLocals {
void
set_cur_paddle_data_layout
(
framework
::
DataLayout
dl
);
void
set_cur_paddle_data_layout
(
framework
::
DataLayout
dl
);
framework
::
DataLayout
get_cur_paddle_data_layout
(
void
);
framework
::
DataLayout
get_cur_paddle_data_layout
(
void
);
void
log_lib_version
(
void
);
void
log_lib_version
(
void
);
const
mkldnn
::
engine
&
get_engine
(
void
);
mkldnn
::
stream
&
get_stream
(
void
);
};
};
MKLDNNDeviceContextThreadLocals
()
=
default
;
MKLDNNDeviceContextThreadLocals
()
=
default
;
MKLDNNDeviceContextThreadLocals
(
const
MKLDNNDeviceContextThreadLocals
&
c
)
=
MKLDNNDeviceContextThreadLocals
(
const
MKLDNNDeviceContextThreadLocals
&
c
)
=
...
@@ -572,7 +578,7 @@ class MKLDNNDeviceContext : public CPUDeviceContext {
...
@@ -572,7 +578,7 @@ class MKLDNNDeviceContext : public CPUDeviceContext {
explicit
MKLDNNDeviceContext
(
CPUPlace
place
);
explicit
MKLDNNDeviceContext
(
CPUPlace
place
);
/* \brief Get the active engine */
/* \brief Get the active engine */
const
mkldnn
::
engine
&
GetEngine
()
const
{
return
engine_
;
}
const
mkldnn
::
engine
&
GetEngine
()
const
{
return
tls
().
get_engine
()
;
}
// Remove all entries from the blob map
// Remove all entries from the blob map
void
ResetBlobMap
();
void
ResetBlobMap
();
...
@@ -605,7 +611,6 @@ class MKLDNNDeviceContext : public CPUDeviceContext {
...
@@ -605,7 +611,6 @@ class MKLDNNDeviceContext : public CPUDeviceContext {
}
}
private:
private:
mkldnn
::
engine
engine_
;
std
::
shared_ptr
<
BlobMap
>
p_blobmap_
;
std
::
shared_ptr
<
BlobMap
>
p_blobmap_
;
std
::
shared_ptr
<
std
::
mutex
>
p_mutex_
;
std
::
shared_ptr
<
std
::
mutex
>
p_mutex_
;
bool
block_next_cache_clearing_
=
false
;
bool
block_next_cache_clearing_
=
false
;
...
...
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @
173660be
...
@@ -188,7 +188,7 @@ MKLDNNGetDataType<paddle::platform::bfloat16>() {
...
@@ -188,7 +188,7 @@ MKLDNNGetDataType<paddle::platform::bfloat16>() {
inline
void
Reorder
(
mkldnn
::
memory
src
,
mkldnn
::
memory
dst
,
inline
void
Reorder
(
mkldnn
::
memory
src
,
mkldnn
::
memory
dst
,
const
mkldnn
::
engine
&
engine
)
{
const
mkldnn
::
engine
&
engine
)
{
auto
reorder_prim
=
mkldnn
::
reorder
(
src
,
dst
);
auto
reorder_prim
=
mkldnn
::
reorder
(
src
,
dst
);
mkldnn
::
stream
astream
(
engine
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
reorder_prim
.
execute
(
astream
,
src
,
dst
);
reorder_prim
.
execute
(
astream
,
src
,
dst
);
...
...
paddle/fluid/platform/mkldnn_reuse.h
浏览文件 @
173660be
...
@@ -232,7 +232,7 @@ class MKLDNNHandlerT {
...
@@ -232,7 +232,7 @@ class MKLDNNHandlerT {
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
}
}
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
...
@@ -261,7 +261,7 @@ class MKLDNNHandlerT {
...
@@ -261,7 +261,7 @@ class MKLDNNHandlerT {
std
::
make_shared
<
dnnl
::
reorder
>
(
*
user_memory_p
,
*
target_memory_p
);
std
::
make_shared
<
dnnl
::
reorder
>
(
*
user_memory_p
,
*
target_memory_p
);
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
reorder_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
user_memory_p
},
reorder_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
user_memory_p
},
...
@@ -273,7 +273,7 @@ class MKLDNNHandlerT {
...
@@ -273,7 +273,7 @@ class MKLDNNHandlerT {
dev_ctx_
.
SetBlob
(
user_key
,
user_memory_p
);
dev_ctx_
.
SetBlob
(
user_key
,
user_memory_p
);
dev_ctx_
.
SetBlob
(
target_key
,
target_memory_p
);
dev_ctx_
.
SetBlob
(
target_key
,
target_memory_p
);
}
else
if
(
!
is_persistent
)
{
}
else
if
(
!
is_persistent
)
{
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
auto
user_memory_p
=
auto
user_memory_p
=
std
::
static_pointer_cast
<
dnnl
::
memory
>
(
dev_ctx_
.
GetBlob
(
user_key
));
std
::
static_pointer_cast
<
dnnl
::
memory
>
(
dev_ctx_
.
GetBlob
(
user_key
));
...
@@ -425,7 +425,7 @@ class MKLDNNHandler {
...
@@ -425,7 +425,7 @@ class MKLDNNHandler {
auto
reorder_p
=
auto
reorder_p
=
std
::
make_shared
<
mkldnn
::
reorder
>
(
*
user_memory_p
,
*
target_memory_p
);
std
::
make_shared
<
mkldnn
::
reorder
>
(
*
user_memory_p
,
*
target_memory_p
);
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
dev_ctx_
.
SetBlob
(
key_reorder_p
,
reorder_p
);
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
RecordEvent
record_reorder
(
"int_reorder"
,
platform
::
EventRole
::
kUniqueOp
);
platform
::
EventRole
::
kUniqueOp
);
reorder_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
user_memory_p
},
reorder_p
->
execute
(
astream
,
{{
MKLDNN_ARG_FROM
,
*
user_memory_p
},
...
@@ -451,7 +451,7 @@ class MKLDNNHandler {
...
@@ -451,7 +451,7 @@ class MKLDNNHandler {
auto
target_memory_p
=
auto
target_memory_p
=
std
::
static_pointer_cast
<
mkldnn
::
memory
>
(
dev_ctx_
.
GetBlob
(
local_key
));
std
::
static_pointer_cast
<
mkldnn
::
memory
>
(
dev_ctx_
.
GetBlob
(
local_key
));
mkldnn
::
stream
astream
(
engine_
);
auto
&
astream
=
platform
::
MKLDNNDeviceContext
::
tls
().
get_stream
(
);
if
(
target_memory_p
==
nullptr
)
{
if
(
target_memory_p
==
nullptr
)
{
target_memory_p
=
user_memory_p
;
target_memory_p
=
user_memory_p
;
...
...
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