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
TonyTonyFun
Paddle
提交
606611d3
P
Paddle
项目概览
TonyTonyFun
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
606611d3
编写于
10月 09, 2020
作者:
J
Jacek Czaja
提交者:
GitHub
10月 09, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[oneDNN] GRU BF16 kernel (#27731)
上级
6c1acf34
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
123 addition
and
6 deletion
+123
-6
paddle/fluid/framework/ir/graph_pattern_detector.cc
paddle/fluid/framework/ir/graph_pattern_detector.cc
+2
-1
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
+8
-5
python/paddle/fluid/tests/unittests/mkldnn/test_fusion_gru_bf16_mkldnn_op.py
.../tests/unittests/mkldnn/test_fusion_gru_bf16_mkldnn_op.py
+113
-0
未找到文件。
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @
606611d3
...
...
@@ -1894,7 +1894,8 @@ PDNode *patterns::QuantizePlacement::operator()(
PDNode
*
patterns
::
Bfloat16Placement
::
operator
()(
const
std
::
unordered_set
<
std
::
string
>
&
bfloat16_enabled_op_types
)
{
std
::
unordered_set
<
std
::
string
>
supported_op_types
{
"conv2d"
};
std
::
unordered_set
<
std
::
string
>
supported_op_types
=
std
::
unordered_set
<
std
::
string
>
({
"conv2d"
,
"fusion_gru"
});
if
(
!
bfloat16_enabled_op_types
.
empty
())
{
supported_op_types
=
bfloat16_enabled_op_types
;
}
...
...
paddle/fluid/operators/fused/mkldnn/fusion_gru_mkldnn_op.cc
浏览文件 @
606611d3
...
...
@@ -86,7 +86,7 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
// Weights for int8 kernel are of a type s8
const
auto
weights_dt
=
is_INT8
?
dnnl
::
memory
::
data_type
::
s8
:
dnnl
::
memory
::
data_type
::
f32
;
is_INT8
?
dnnl
::
memory
::
data_type
::
s8
:
MKLDNNGetDataType
<
T
>
()
;
// oneDNN RNN dimensions
const
int64_t
D
=
1
;
// Directions
...
...
@@ -226,6 +226,8 @@ class GRUMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::gru_forward> {
}
// TODO(grygielski) H0 is for now persistable
// TODO(jczaja) H0 should be updated each iter and of T type (Fusion pass does
// not support in yet)
std
::
shared_ptr
<
dnnl
::
memory
>
AcquireH0Memory
(
const
Tensor
*
h0
)
{
const
std
::
string
h0_key
=
memory_key_
+
"@h0"
;
auto
memory_p
=
...
...
@@ -397,14 +399,14 @@ template <typename T>
class
FusionGRUMKLDNNKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
bool
is_
INT8
=
std
::
is_same
<
T
,
uint8_t
>::
value
;
const
bool
is_
bf16
=
std
::
is_same
<
T
,
paddle
::
platform
::
bfloat16
>::
value
;
const
bool
force_fp32_output
=
ctx
.
Attr
<
bool
>
(
"force_fp32_output"
);
//
TODO(grygielski) Add option for bfloa
t
if
(
!
is_
INT8
||
force_fp32_output
)
{
//
BF16 does not support force outpu
t
if
(
!
is_
bf16
&&
force_fp32_output
)
{
RunKernel
<
float
>
(
ctx
);
}
else
{
RunKernel
<
uint8_t
>
(
ctx
);
RunKernel
<
T
>
(
ctx
);
}
}
...
...
@@ -495,4 +497,5 @@ class FusionGRUMKLDNNKernel : public framework::OpKernel<T> {
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_KERNEL
(
fusion_gru
,
MKLDNN
,
paddle
::
platform
::
CPUPlace
,
ops
::
FusionGRUMKLDNNKernel
<
float
>
,
ops
::
FusionGRUMKLDNNKernel
<
paddle
::
platform
::
bfloat16
>
,
ops
::
FusionGRUMKLDNNKernel
<
uint8_t
>
);
python/paddle/fluid/tests/unittests/mkldnn/test_fusion_gru_bf16_mkldnn_op.py
0 → 100644
浏览文件 @
606611d3
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
from
__future__
import
print_function
import
unittest
import
numpy
as
np
import
struct
import
paddle.fluid.core
as
core
from
paddle.fluid.tests.unittests.op_test
import
OpTest
,
convert_float_to_uint16
from
paddle.fluid.tests.unittests.op_test
import
OpTest
from
paddle.fluid.tests.unittests.test_fusion_gru_op
import
fusion_gru
from
paddle.fluid.tests.unittests.test_fusion_lstm_op
import
fc
,
ACTIVATION
@
unittest
.
skipIf
(
not
core
.
supports_bfloat16
(),
"place does not support BF16 evaluation"
)
class
TestFusionGRUBF16MKLDNNOp
(
OpTest
):
def
set_confs
(
self
):
self
.
mkldnn_data_type
=
False
def
setUp
(
self
):
self
.
op_type
=
"fusion_gru"
self
.
lod
=
[[
2
,
4
,
3
]]
self
.
M
=
3
self
.
D
=
5
self
.
is_reverse
=
False
self
.
with_h0
=
False
self
.
use_mkldnn
=
True
self
.
_cpu_only
=
True
self
.
with_bias
=
True
self
.
act_state
=
'tanh'
self
.
act_gate
=
'sigmoid'
self
.
origin_mode
=
False
self
.
use_mkldnn
=
True
self
.
force_fp32_output
=
False
self
.
set_confs
()
T
=
sum
(
self
.
lod
[
0
])
N
=
len
(
self
.
lod
[
0
])
# fp32 X input for reference implementation and
# corressponding bf16 data as input to GRU oneDNN bf16 kernel
x_fp32
=
np
.
random
.
rand
(
T
,
self
.
M
).
astype
(
'float32'
)
x_bf16
=
convert_float_to_uint16
(
x_fp32
)
wx_fp32
=
np
.
random
.
rand
(
self
.
M
,
3
*
self
.
D
).
astype
(
'float32'
)
wh_fp32
=
np
.
random
.
rand
(
self
.
D
,
3
*
self
.
D
).
astype
(
'float32'
)
# bias is fp32 despite other inputs being in bf16
bias
=
np
.
random
.
rand
(
1
,
3
*
self
.
D
).
astype
(
'float32'
)
if
self
.
with_bias
else
np
.
zeros
(
(
1
,
3
*
self
.
D
),
dtype
=
'float32'
)
h0_fp32
=
np
.
random
.
rand
(
N
,
self
.
D
).
astype
(
'float32'
)
if
self
.
with_h0
else
np
.
zeros
(
(
N
,
self
.
D
),
dtype
=
'float32'
)
_
,
_
,
_
,
hidden
=
fusion_gru
(
x_fp32
,
self
.
lod
,
h0_fp32
,
wx_fp32
,
wh_fp32
,
bias
,
self
.
is_reverse
,
self
.
origin_mode
,
ACTIVATION
[
self
.
act_state
],
ACTIVATION
[
self
.
act_gate
])
hidden_bf16
=
convert_float_to_uint16
(
hidden
)
self
.
inputs
=
{
'X'
:
(
x_bf16
,
self
.
lod
),
'WeightX'
:
wx_fp32
,
'WeightH'
:
wh_fp32
}
if
self
.
with_bias
:
self
.
inputs
[
'Bias'
]
=
bias
if
self
.
with_h0
:
self
.
inputs
[
'H0'
]
=
h0_bf16
h0_bf16
=
convert_float_to_uint16
(
h0_fp32
)
self
.
outputs
=
{
'Hidden'
:
(
hidden_bf16
,
self
.
lod
)}
self
.
attrs
=
{
'activation'
:
self
.
act_state
,
'gate_activation'
:
self
.
act_gate
,
'is_reverse'
:
self
.
is_reverse
,
'origin_mode'
:
self
.
origin_mode
,
'force_fp32_output'
:
self
.
force_fp32_output
,
'use_mkldnn'
:
self
.
use_mkldnn
}
class
TestFusionGRUINT8MKLDNNOp2
(
TestFusionGRUBF16MKLDNNOp
):
def
set_confs
(
self
):
self
.
origin_mode
=
False
class
TestFusionGRUINT8MKLDNNOp3
(
TestFusionGRUBF16MKLDNNOp
):
def
set_confs
(
self
):
self
.
with_bias
=
False
if
__name__
==
"__main__"
:
unittest
.
main
()
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