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-Lite
提交
0d77fa32
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
331
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
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
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
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
0d77fa32
编写于
7月 30, 2020
作者:
Q
Qi Li
提交者:
GitHub
7月 30, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[Ascend] add interpolate_op, test=develop (#4028)
上级
98e69581
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
162 addition
and
1 deletion
+162
-1
lite/core/program.cc
lite/core/program.cc
+4
-1
lite/kernels/huawei_ascend_npu/bridges/CMakeLists.txt
lite/kernels/huawei_ascend_npu/bridges/CMakeLists.txt
+2
-0
lite/kernels/huawei_ascend_npu/bridges/interpolate_op.cc
lite/kernels/huawei_ascend_npu/bridges/interpolate_op.cc
+138
-0
lite/kernels/huawei_ascend_npu/bridges/paddle_use_bridges.h
lite/kernels/huawei_ascend_npu/bridges/paddle_use_bridges.h
+2
-0
lite/kernels/huawei_ascend_npu/bridges/utility.h
lite/kernels/huawei_ascend_npu/bridges/utility.h
+7
-0
lite/tests/kernels/interp_compute_test.cc
lite/tests/kernels/interp_compute_test.cc
+9
-0
未找到文件。
lite/core/program.cc
浏览文件 @
0d77fa32
...
...
@@ -244,7 +244,10 @@ RuntimeProgram::RuntimeProgram(
ContextScheduler
::
Global
().
NewContext
(
kernel
->
target
()));
}
#else
kernel
->
SetContext
(
ContextScheduler
::
Global
().
NewContext
(
kernel
->
target
()));
if
(
kernel
!=
nullptr
)
{
kernel
->
SetContext
(
ContextScheduler
::
Global
().
NewContext
(
kernel
->
target
()));
}
#endif
instructions_
[
kRootBlockIdx
].
emplace_back
(
std
::
move
(
op
),
std
::
move
(
kernel
));
}
...
...
lite/kernels/huawei_ascend_npu/bridges/CMakeLists.txt
浏览文件 @
0d77fa32
...
...
@@ -9,6 +9,7 @@ set(huawei_ascend_npu_subgraph_bridge_deps subgraph_bridge_registry subgraph_bri
lite_cc_library
(
subgraph_bridge_act_op_huawei_ascend_npu SRCS act_op.cc DEPS
${
huawei_ascend_npu_subgraph_bridge_deps
}
)
lite_cc_library
(
subgraph_bridge_conv_op_huawei_ascend_npu SRCS conv_op.cc DEPS
${
huawei_ascend_npu_subgraph_bridge_deps
}
)
lite_cc_library
(
subgraph_bridge_interpolate_op_huawei_ascend_npu SRCS interpolate_op.cc DEPS
${
huawei_ascend_npu_subgraph_bridge_deps
}
)
set
(
huawei_ascend_npu_subgraph_bridges
subgraph_bridge_registry
...
...
@@ -16,4 +17,5 @@ set(huawei_ascend_npu_subgraph_bridges
subgraph_bridge_graph_huawei_ascend_npu
subgraph_bridge_act_op_huawei_ascend_npu
subgraph_bridge_conv_op_huawei_ascend_npu
subgraph_bridge_interpolate_op_huawei_ascend_npu
CACHE INTERNAL
"huawei_ascend_npu_subgraph_bridges"
)
lite/kernels/huawei_ascend_npu/bridges/interpolate_op.cc
0 → 100644
浏览文件 @
0d77fa32
// Copyright (c) 2019 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.
// #include "lite/operators/interpolate_op.h"
#include "lite/kernels/huawei_ascend_npu/bridges/graph.h"
#include "lite/kernels/huawei_ascend_npu/bridges/utility.h"
#include "lite/kernels/npu/bridges/registry.h"
namespace
paddle
{
namespace
lite
{
namespace
subgraph
{
namespace
huawei_ascend_npu
{
int
InterpolateConverter
(
void
*
ctx
,
OpLite
*
op
,
KernelBase
*
kernel
)
{
CHECK
(
ctx
!=
nullptr
);
CHECK
(
op
!=
nullptr
);
auto
graph
=
static_cast
<
Graph
*>
(
ctx
);
auto
op_info
=
op
->
op_info
();
auto
op_type
=
op_info
->
Type
();
auto
scope
=
op
->
scope
();
VLOG
(
3
)
<<
"[HUAWEI_ASCEND_NPU] Converting "
+
op_type
+
"..."
;
// Get input and output vars and op attributes
auto
x_name
=
op_info
->
Input
(
"X"
).
front
();
auto
x
=
scope
->
FindMutableTensor
(
x_name
);
auto
x_dims
=
x
->
dims
();
auto
x_h
=
x_dims
[
2
];
auto
x_w
=
x_dims
[
3
];
CHECK_EQ
(
x_dims
.
size
(),
4
);
auto
out_name
=
op_info
->
Output
(
"Out"
).
front
();
auto
scale
=
op_info
->
GetAttr
<
float
>
(
"scale"
);
auto
out_w
=
op_info
->
GetAttr
<
int
>
(
"out_w"
);
auto
out_h
=
op_info
->
GetAttr
<
int
>
(
"out_h"
);
auto
align_corners
=
op_info
->
GetAttr
<
bool
>
(
"align_corners"
);
int
align_mode
=
op_info
->
HasAttr
(
"align_mode"
)
?
op_info
->
GetAttr
<
int
>
(
"align_mode"
)
:
1
;
auto
interp_method
=
op_info
->
GetAttr
<
std
::
string
>
(
"interp_method"
);
if
(
align_mode
==
0
&&
!
align_corners
)
{
LOG
(
WARNING
)
<<
"[HUAWEI_ASCEND_NPU] align_mode = 0 && "
"align_corners = false isn't "
"supported in Huawei Ascend NPU DDK"
;
return
FAILED
;
}
// get ge date type
ge
::
DataType
ge_data_type
=
CvtPrecisionType
(
x
->
precision
());
// X node
std
::
shared_ptr
<
Node
>
x_node
=
nullptr
;
if
(
graph
->
Has
(
x_name
))
{
x_node
=
graph
->
Get
(
x_name
);
}
else
{
x_node
=
graph
->
Add
(
x_name
,
*
x
);
}
// Priority: OutSize > scale > out_h/out_w
if
(
scale
>
0
)
{
out_h
=
static_cast
<
int
>
(
x_h
*
scale
);
out_w
=
static_cast
<
int
>
(
x_w
*
scale
);
out_h
=
out_h
>
0
?
out_h
:
-
1
;
out_w
=
out_w
>
0
?
out_w
:
-
1
;
}
// Update out_h and out_w and create out_size node if has OutSize
std
::
shared_ptr
<
Node
>
out_size_node
=
nullptr
;
if
(
HasInputArg
(
op_info
,
scope
,
"OutSize"
))
{
auto
out_size_name
=
op_info
->
Input
(
"OutSize"
).
front
();
if
(
graph
->
Has
(
out_size_name
))
{
out_size_node
=
graph
->
Get
(
out_size_name
);
}
else
{
auto
out_size
=
scope
->
FindMutableTensor
(
out_size_name
);
CHECK_EQ
(
out_size
->
numel
(),
2
);
CHECK
(
out_size
->
persistable
());
auto
out_size_data
=
out_size
->
mutable_data
<
int
>
();
// Update out_h and out_w if has OutSize
out_h
=
out_size_data
[
0
];
out_w
=
out_size_data
[
1
];
}
}
if
(
out_size_node
==
nullptr
)
{
out_size_node
=
graph
->
Add
(
out_name
+
"/out_size"
,
std
::
vector
<
int
>
({
out_h
,
out_w
}));
}
if
(
interp_method
==
"bilinear"
)
{
auto
bilinear_interp_node
=
graph
->
Add
<
ge
::
op
::
ResizeBilinearV2
>
(
out_name
);
auto
bilinear_interp_op
=
bilinear_interp_node
->
data
<
ge
::
op
::
ResizeBilinearV2
>
();
bilinear_interp_op
->
set_input_x
(
*
x_node
->
data
());
bilinear_interp_op
->
set_input_size
(
*
out_size_node
->
data
());
bilinear_interp_op
->
set_attr_align_corners
(
align_corners
);
TENSOR_UPDATE_INPUT
(
bilinear_interp_op
,
x
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
TENSOR_UPDATE_INPUT
(
bilinear_interp_op
,
size
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
TENSOR_UPDATE_OUTPUT
(
bilinear_interp_op
,
y
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
}
else
if
(
interp_method
==
"nearest"
)
{
auto
nearest_interp_node
=
graph
->
Add
<
ge
::
op
::
ResizeNearestNeighborV2
>
(
out_name
);
auto
nearest_interp_op
=
nearest_interp_node
->
data
<
ge
::
op
::
ResizeNearestNeighborV2
>
();
nearest_interp_op
->
set_input_x
(
*
x_node
->
data
());
nearest_interp_op
->
set_input_size
(
*
out_size_node
->
data
());
nearest_interp_op
->
set_attr_align_corners
(
align_corners
);
TENSOR_UPDATE_INPUT
(
nearest_interp_op
,
x
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
TENSOR_UPDATE_INPUT
(
nearest_interp_op
,
size
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
TENSOR_UPDATE_OUTPUT
(
nearest_interp_op
,
y
,
ge
::
FORMAT_NCHW
,
ge_data_type
);
}
else
{
LOG
(
WARNING
)
<<
"[HUAWEI_ASCEND_NPU] Unsupported interpolate method: "
<<
interp_method
;
return
FAILED
;
}
return
REBUILD_WHEN_SHAPE_CHANGED
;
}
}
// namespace huawei_ascend_npu
}
// namespace subgraph
}
// namespace lite
}
// namespace paddle
REGISTER_SUBGRAPH_BRIDGE
(
bilinear_interp
,
kHuaweiAscendNPU
,
paddle
::
lite
::
subgraph
::
huawei_ascend_npu
::
InterpolateConverter
);
REGISTER_SUBGRAPH_BRIDGE
(
nearest_interp
,
kHuaweiAscendNPU
,
paddle
::
lite
::
subgraph
::
huawei_ascend_npu
::
InterpolateConverter
);
lite/kernels/huawei_ascend_npu/bridges/paddle_use_bridges.h
浏览文件 @
0d77fa32
...
...
@@ -25,3 +25,5 @@ USE_SUBGRAPH_BRIDGE(softplus, kHuaweiAscendNPU);
// conv
USE_SUBGRAPH_BRIDGE
(
conv2d
,
kHuaweiAscendNPU
);
USE_SUBGRAPH_BRIDGE
(
depthwise_conv2d
,
kHuaweiAscendNPU
);
USE_SUBGRAPH_BRIDGE
(
bilinear_interp
,
kHuaweiAscendNPU
);
USE_SUBGRAPH_BRIDGE
(
nearest_interp
,
kHuaweiAscendNPU
);
lite/kernels/huawei_ascend_npu/bridges/utility.h
浏览文件 @
0d77fa32
...
...
@@ -30,6 +30,13 @@ namespace lite {
namespace
subgraph
{
namespace
huawei_ascend_npu
{
#define TENSOR_UPDATE_INPUT(op, attr, format, dtype) \
ge::TensorDesc _##op##_input_desc_##attr(ge::Shape(), format, dtype); \
op->update_input_desc_##attr(_##op##_input_desc_##attr);
#define TENSOR_UPDATE_OUTPUT(op, attr, format, dtype) \
ge::TensorDesc _##op##_output_desc_##attr(ge::Shape(), format, dtype); \
op->update_output_desc_##attr(_##op##_output_desc_##attr);
// Type/tensor converters for converting Paddle type/tensor to HiAI type/tensor
bool
HasInputArg
(
const
OpInfo
*
op_info
,
const
Scope
*
scope
,
...
...
lite/tests/kernels/interp_compute_test.cc
浏览文件 @
0d77fa32
...
...
@@ -420,6 +420,12 @@ void TestInterpAlignMode(Place place, float abs_error = 2e-5) {
if
(
place
==
TARGET
(
kARM
)
&&
align_mode
==
1
&&
!
align_corners
)
{
continue
;
}
// align_mode = 0 && align_corners = false NOT supported in Huawei
// Ascend NPU DDK
if
(
place
==
TARGET
(
kHuaweiAscendNPU
)
&&
align_mode
==
0
&&
!
align_corners
)
{
continue
;
}
std
::
unique_ptr
<
arena
::
TestCase
>
tester
(
new
NearestInterpComputeTester
(
place
,
"def"
,
...
...
@@ -443,6 +449,9 @@ TEST(Interp, precision) {
#if defined(LITE_WITH_NPU)
place
=
TARGET
(
kNPU
);
abs_error
=
1e-2
;
// use fp16 in npu
#elif defined(LITE_WITH_HUAWEI_ASCEND_NPU)
place
=
TARGET
(
kHuaweiAscendNPU
);
abs_error
=
1e-2
;
// use fp16 in npu
#elif defined(LITE_WITH_ARM)
place
=
TARGET
(
kARM
);
#else
...
...
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