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
SummerGao.
Paddle
提交
f6b4ed22
P
Paddle
项目概览
SummerGao.
/
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
f6b4ed22
编写于
10月 29, 2021
作者:
B
baoachun
提交者:
GitHub
10月 29, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix matmul error when input's dim is 3 (#36849)
上级
89a8989f
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
257 addition
and
3 deletion
+257
-3
paddle/fluid/inference/tensorrt/convert/matmul_op.cc
paddle/fluid/inference/tensorrt/convert/matmul_op.cc
+43
-2
paddle/fluid/inference/tensorrt/op_teller.cc
paddle/fluid/inference/tensorrt/op_teller.cc
+1
-1
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_matmul.py
...d/tests/unittests/ir/inference/test_trt_convert_matmul.py
+213
-0
未找到文件。
paddle/fluid/inference/tensorrt/convert/matmul_op.cc
浏览文件 @
f6b4ed22
...
...
@@ -61,6 +61,38 @@ class MatMulOpConverter : public OpConverter {
if
(
fabs
(
alpha
-
1.0
)
<
std
::
numeric_limits
<
float
>::
epsilon
())
{
engine_
->
SetITensor
(
output_name
,
layer
->
getOutput
(
0
));
}
else
{
// IScaleLayer requires the input must have at least
// three dimensions in static shape mode and at least
// four dimensions in dynamic shape mode.
auto
*
matmul_out
=
layer
->
getOutput
(
0
);
nvinfer1
::
Dims
out_shape
=
matmul_out
->
getDimensions
();
const
int
out_dims
=
out_shape
.
nbDims
;
bool
need_change_dim
=
false
;
if
(
engine_
->
with_dynamic_shape
())
{
if
(
out_dims
==
3
)
{
need_change_dim
=
true
;
}
}
else
{
if
(
out_dims
==
2
)
{
need_change_dim
=
true
;
}
}
if
(
need_change_dim
)
{
nvinfer1
::
Dims
reshape_dim
;
reshape_dim
.
nbDims
=
out_dims
+
1
;
reshape_dim
.
d
[
out_dims
]
=
1
;
for
(
int
i
=
0
;
i
<
out_dims
;
i
++
)
{
reshape_dim
.
d
[
i
]
=
out_shape
.
d
[
i
];
}
auto
*
reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
matmul_out
);
reshape_layer
->
setReshapeDimensions
(
reshape_dim
);
matmul_out
=
reshape_layer
->
getOutput
(
0
);
}
auto
create_weights
=
[
&
](
float
data
,
const
std
::
string
&
type
)
->
float
*
{
std
::
unique_ptr
<
framework
::
Tensor
>
tmp_tensor
(
new
framework
::
Tensor
());
tmp_tensor
->
Resize
({
1
});
...
...
@@ -80,9 +112,18 @@ class MatMulOpConverter : public OpConverter {
TensorRTEngine
::
Weight
nv_power
{
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
power_data
),
1
};
auto
*
scale_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Scale
,
*
layer
->
getOutput
(
0
)
,
nvinfer1
::
ScaleMode
::
kUNIFORM
,
engine_
,
Scale
,
*
matmul_out
,
nvinfer1
::
ScaleMode
::
kUNIFORM
,
nv_shift
.
get
(),
nv_alpha
.
get
(),
nv_power
.
get
());
engine_
->
SetITensor
(
output_name
,
scale_layer
->
getOutput
(
0
));
auto
*
scale_out
=
scale_layer
->
getOutput
(
0
);
if
(
need_change_dim
)
{
auto
*
reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
scale_out
);
reshape_layer
->
setReshapeDimensions
(
out_shape
);
scale_out
=
reshape_layer
->
getOutput
(
0
);
}
engine_
->
SetITensor
(
output_name
,
scale_out
);
}
if
(
test_mode
)
{
// the test framework can not determine which is the
// output, so place the declaration inside.
...
...
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @
f6b4ed22
...
...
@@ -1550,7 +1550,7 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
!
BOOST_GET_CONST
(
bool
,
desc
.
GetAttr
(
"keep_dim"
)))
return
false
;
}
if
(
desc
.
HasAttr
(
"
reduce_all
"
))
{
if
(
desc
.
HasAttr
(
"
out_dtype
"
))
{
int
out_dtype
=
BOOST_GET_CONST
(
int32_t
,
desc
.
GetAttr
(
"out_dtype"
));
if
(
out_dtype
!=
-
1
)
{
return
false
;
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_matmul.py
0 → 100644
浏览文件 @
f6b4ed22
# Copyright (c) 2021 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
trt_layer_auto_scan_test
import
TrtLayerAutoScanTest
,
SkipReasons
from
program_config
import
TensorConfig
,
ProgramConfig
import
numpy
as
np
import
paddle.inference
as
paddle_infer
from
functools
import
partial
from
typing
import
Optional
,
List
,
Callable
,
Dict
,
Any
,
Set
import
unittest
class
TrtConvertMatmulTest_static
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
return
True
def
sample_program_configs
(
self
):
def
generate_input
(
shape
):
return
np
.
random
.
random
(
shape
).
astype
(
np
.
float32
)
for
batch
in
[
1
,
4
]:
for
trans_x
in
[
True
,
False
]:
for
trans_y
in
[
True
,
False
]:
if
trans_x
and
trans_y
:
input1_shape
=
[
batch
,
6
,
11
]
input2_shape
=
[
batch
,
32
,
6
]
if
trans_x
and
not
trans_y
:
input1_shape
=
[
batch
,
6
,
11
]
input2_shape
=
[
batch
,
6
,
32
]
if
not
trans_x
and
trans_y
:
input1_shape
=
[
batch
,
32
,
6
]
input2_shape
=
[
batch
,
11
,
6
]
if
not
trans_x
and
not
trans_y
:
input1_shape
=
[
batch
,
32
,
6
]
input2_shape
=
[
batch
,
6
,
11
]
for
alpha
in
[
0.3
,
1.0
]:
dics
=
[{
"transpose_X"
:
trans_x
,
"transpose_Y"
:
trans_y
,
"alpha"
:
alpha
,
"fused_reshape_X"
:
[],
"fused_reshape_Y"
:
[],
"fused_transpose_X"
:
[],
"fused_transpose_Y"
:
[],
"fused_reshape_Out"
:
[],
"fused_transpose_Out"
:
[]
}]
ops_config
=
[{
"op_type"
:
"matmul"
,
"op_inputs"
:
{
"X"
:
[
"input1_data"
],
"Y"
:
[
"input2_data"
]
},
"op_outputs"
:
{
"Out"
:
[
"output_data"
]
},
"op_attrs"
:
dics
[
0
]
}]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input1_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input
,
input1_shape
)),
"input2_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input
,
input2_shape
))
},
outputs
=
[
"output_data"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
(
attrs
):
pass
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
opt_input_shape
=
{}
# for static_shape
clear_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
3
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
3
),
1e-5
def
test
(
self
):
self
.
run_test
()
class
TrtConvertMatmulTest_dynamic
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
return
True
def
sample_program_configs
(
self
):
def
generate_input
(
shape
):
return
np
.
random
.
random
(
shape
).
astype
(
np
.
float32
)
for
trans_x
in
[
True
]:
for
trans_y
in
[
True
]:
if
trans_x
and
trans_y
:
input1_shape
=
[
4
,
4
,
4
]
input2_shape
=
[
4
,
4
,
4
]
# if trans_x and not trans_y:
# input1_shape = [4, 4, 4]
# input2_shape = [4, 4, 4]
# if not trans_x and trans_y:
# input1_shape = [batch, 32, 6]
# input2_shape = [batch, 11, 6]
# if not trans_x and not trans_y:
# input1_shape = [batch, 32, 6]
# input2_shape = [batch, 6, 11]
for
alpha
in
[
0.3
,
1.0
]:
dics
=
[{
"transpose_X"
:
trans_x
,
"transpose_Y"
:
trans_y
,
"alpha"
:
alpha
,
"fused_reshape_X"
:
[],
"fused_reshape_Y"
:
[],
"fused_transpose_X"
:
[],
"fused_transpose_Y"
:
[],
"fused_reshape_Out"
:
[],
"fused_transpose_Out"
:
[]
}]
ops_config
=
[{
"op_type"
:
"matmul"
,
"op_inputs"
:
{
"X"
:
[
"input1_data"
],
"Y"
:
[
"input2_data"
]
},
"op_outputs"
:
{
"Out"
:
[
"output_data"
]
},
"op_attrs"
:
dics
[
0
]
}]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input1_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input
,
input1_shape
)),
"input2_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input
,
input2_shape
))
},
outputs
=
[
"output_data"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
(
attrs
):
self
.
dynamic_shape
.
min_input_shape
=
{
"input1_data"
:
[
1
,
4
,
4
],
"input2_data"
:
[
1
,
4
,
4
]
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input1_data"
:
[
16
,
4
,
4
],
"input2_data"
:
[
16
,
4
,
128
]
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input1_data"
:
[
8
,
4
,
4
],
"input2_data"
:
[
8
,
4
,
16
]
}
attrs
=
[
program_config
.
ops
[
i
].
attrs
for
i
in
range
(
len
(
program_config
.
ops
))
]
# for dynamic_shape
generate_dynamic_shape
(
attrs
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
3
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
3
),
1e-5
def
add_skip_trt_case
(
self
):
def
teller1
(
program_config
,
predictor_config
):
if
len
(
self
.
dynamic_shape
.
min_input_shape
)
!=
0
and
self
.
trt_param
.
precision
==
paddle_infer
.
PrecisionType
.
Half
:
return
True
return
False
self
.
add_skip_case
(
teller1
,
SkipReasons
.
TRT_NOT_IMPLEMENTED
,
"Tensorrt MatrixMultiply layer will get error when dynamic shape fp16 mode."
)
def
test
(
self
):
self
.
add_skip_trt_case
()
self
.
run_test
()
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