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
76156d12
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
未验证
提交
76156d12
编写于
7月 01, 2022
作者:
Z
zhoutianzi666
提交者:
GitHub
7月 01, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[inference TRT]template GetWeightCPUData (#43993)
* template GetWeightCPUData
上级
267d3191
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
33 addition
and
22 deletion
+33
-22
paddle/fluid/inference/tensorrt/engine.cc
paddle/fluid/inference/tensorrt/engine.cc
+21
-18
paddle/fluid/inference/tensorrt/engine.h
paddle/fluid/inference/tensorrt/engine.h
+12
-4
未找到文件。
paddle/fluid/inference/tensorrt/engine.cc
浏览文件 @
76156d12
...
...
@@ -390,33 +390,36 @@ nvinfer1::ITensor *TensorRTEngine::GetITensor(const std::string &name) {
return
itensor_map_
[
name
];
}
std
::
unordered_map
<
std
::
string
,
nvinfer1
::
ITensor
*>
*
TensorRTEngine
::
GetITensorMap
()
{
return
&
itensor_map_
;
}
void
TensorRTEngine
::
SetRuntimeBatch
(
size_t
batch_size
)
{
runtime_batch_
=
batch_size
;
}
float
*
TensorRTEngine
::
GetWeightCPUData
(
const
std
::
string
&
name
,
template
<
typename
T
=
float
>
T
*
TensorRTEngine
::
GetWeightCPUData
(
const
std
::
string
&
name
,
framework
::
Tensor
*
weight_tensor
)
{
static
int
name_suffix_counter
=
0
;
std
::
string
name_suffix
=
std
::
to_string
(
name_suffix_counter
);
std
::
string
splitter
=
"__"
;
std
::
string
name_with_suffix
=
name
+
splitter
+
name_suffix
;
std
::
unique_ptr
<
framework
::
Tensor
>
cpu_weight_tensor
(
new
framework
::
Tensor
());
platform
::
CPUPlace
cpu_place
;
PADDLE_ENFORCE_EQ
(
weight_map
.
count
(
name_with_suffix
),
0
,
platform
::
errors
::
AlreadyExists
(
"The weight named %s is set into the weight map "
"twice in TRT OP converter."
,
name_with_suffix
));
weight_map
[
name_with_suffix
].
reset
(
new
framework
::
Tensor
());
weight_map
[
name_with_suffix
]
->
Resize
(
weight_tensor
->
dims
());
cpu_weight_tensor
->
Resize
(
weight_tensor
->
dims
());
paddle
::
framework
::
TensorCopySync
(
*
weight_tensor
,
cpu_place
,
weight_map
[
name_with_suffix
].
get
());
float
*
weight_data
=
weight_map
[
name_with_suffix
]
->
mutable_data
<
float
>
(
cpu_place
);
name_suffix_counter
+=
1
;
*
weight_tensor
,
cpu_place
,
cpu_weight_tensor
.
get
());
T
*
weight_data
=
cpu_weight_tensor
->
mutable_data
<
T
>
(
cpu_place
);
SetWeights
(
name
,
std
::
move
(
cpu_weight_tensor
));
return
weight_data
;
}
template
float
*
TensorRTEngine
::
GetWeightCPUData
(
const
std
::
string
&
name
,
framework
::
Tensor
*
weight_tensor
);
template
int32_t
*
TensorRTEngine
::
GetWeightCPUData
(
const
std
::
string
&
name
,
framework
::
Tensor
*
weight_tensor
);
template
int64_t
*
TensorRTEngine
::
GetWeightCPUData
(
const
std
::
string
&
name
,
framework
::
Tensor
*
weight_tensor
);
int
TensorRTEngine
::
GetRuntimeBatch
()
{
return
runtime_batch_
;
}
nvinfer1
::
IPluginV2Layer
*
TensorRTEngine
::
AddPlugin
(
...
...
paddle/fluid/inference/tensorrt/engine.h
浏览文件 @
76156d12
...
...
@@ -268,6 +268,7 @@ class TensorRTEngine {
void
SetITensor
(
const
std
::
string
&
name
,
nvinfer1
::
ITensor
*
tensor
);
// Get an ITensor called name.
nvinfer1
::
ITensor
*
GetITensor
(
const
std
::
string
&
name
);
std
::
unordered_map
<
std
::
string
,
nvinfer1
::
ITensor
*>*
GetITensorMap
();
nvinfer1
::
ICudaEngine
*
engine
()
{
return
infer_engine_
.
get
();
}
nvinfer1
::
IExecutionContext
*
context
()
{
...
...
@@ -405,8 +406,8 @@ class TensorRTEngine {
void
SetTensorDynamicRange
(
nvinfer1
::
ITensor
*
tensor
,
float
range
)
{
quant_dynamic_range_
[
tensor
]
=
range
;
}
float
*
GetWeightCPUData
(
const
std
::
string
&
name
,
template
<
typename
T
=
float
>
T
*
GetWeightCPUData
(
const
std
::
string
&
name
,
framework
::
Tensor
*
weight_tensor
);
// A pointer to CPU memory is needed of the TRT weight.
...
...
@@ -424,7 +425,14 @@ class TensorRTEngine {
static
int
suffix_counter
=
0
;
std
::
string
suffix
=
std
::
to_string
(
suffix_counter
);
std
::
string
splitter
=
"__"
;
weight_map
[
w_name
+
splitter
+
suffix
]
=
std
::
move
(
w_tensor
);
std
::
string
name_with_suffix
=
w_name
+
splitter
+
suffix
;
PADDLE_ENFORCE_EQ
(
weight_map
.
count
(
name_with_suffix
),
0
,
platform
::
errors
::
AlreadyExists
(
"The weight named %s is set into the weight map "
"twice in TRT OP converter."
,
name_with_suffix
));
weight_map
[
name_with_suffix
]
=
std
::
move
(
w_tensor
);
suffix_counter
+=
1
;
}
...
...
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