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
71636e67
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 >>
提交
71636e67
编写于
12月 26, 2018
作者:
N
nhzlx
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add min_subgraph_size attr to tensorrt config
test=develop
上级
a6aa8ea7
变更
6
显示空白变更内容
内联
并排
Showing
6 changed file
with
26 addition
and
5 deletion
+26
-5
paddle/fluid/inference/analysis/argument.h
paddle/fluid/inference/analysis/argument.h
+1
-0
paddle/fluid/inference/analysis/ir_pass_manager.cc
paddle/fluid/inference/analysis/ir_pass_manager.cc
+2
-0
paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
...id/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
+4
-2
paddle/fluid/inference/api/analysis_config.cc
paddle/fluid/inference/api/analysis_config.cc
+6
-2
paddle/fluid/inference/api/analysis_predictor.cc
paddle/fluid/inference/api/analysis_predictor.cc
+1
-0
paddle/fluid/inference/api/paddle_analysis_config.h
paddle/fluid/inference/api/paddle_analysis_config.h
+12
-1
未找到文件。
paddle/fluid/inference/analysis/argument.h
浏览文件 @
71636e67
...
...
@@ -127,6 +127,7 @@ struct Argument {
std
::
function
<
bool
(
const
framework
::
ir
::
Node
*
)
>
);
DECL_ARGUMENT_FIELD
(
tensorrt_max_batch_size
,
TensorRtMaxBatchSize
,
int
);
DECL_ARGUMENT_FIELD
(
tensorrt_workspace_size
,
TensorRtWorkspaceSize
,
int
);
DECL_ARGUMENT_FIELD
(
tensorrt_min_subgraph_size
,
TensorRtMinSubgraphSize
,
int
);
// The program transformed by IR analysis phase.
DECL_ARGUMENT_UNIQUE_FIELD
(
ir_analyzed_program
,
IrAnalyzedProgram
,
...
...
paddle/fluid/inference/analysis/ir_pass_manager.cc
浏览文件 @
71636e67
...
...
@@ -75,6 +75,8 @@ void IRPassManager::CreatePasses(Argument *argument,
argument
->
tensorrt_node_teller_ptr
());
pass
->
Set
(
"workspace_size"
,
new
int
(
argument
->
tensorrt_workspace_size
()));
pass
->
Set
(
"max_batch_size"
,
new
int
(
argument
->
tensorrt_max_batch_size
()));
pass
->
Set
(
"min_subgraph_size"
,
new
int
(
argument
->
tensorrt_min_subgraph_size
()));
}
// graph_ = pass->Apply(std::move(graph_));
...
...
paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
浏览文件 @
71636e67
...
...
@@ -38,7 +38,8 @@ std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
auto
teller
=
Get
<
SubgraphDetector
::
NodeInsideSubgraphTeller
>
(
"tensorrt_node_teller"
);
SubGraphFuser
fuser
(
graph
.
get
(),
teller
,
2
/*min subgraph size*/
);
SubGraphFuser
fuser
(
graph
.
get
(),
teller
,
Get
<
int
>
(
"min_subgraph_size"
)
/*min subgraph size*/
);
fuser
();
for
(
auto
*
node
:
graph
->
Nodes
())
{
...
...
@@ -233,4 +234,5 @@ REGISTER_PASS(tensorrt_subgraph_pass,
paddle
::
inference
::
analysis
::
TensorRtSubgraphPass
)
.
RequirePassAttr
(
"tensorrt_node_teller"
)
.
RequirePassAttr
(
"max_batch_size"
)
.
RequirePassAttr
(
"workspace_size"
);
.
RequirePassAttr
(
"workspace_size"
)
.
RequirePassAttr
(
"min_subgraph_size"
);
paddle/fluid/inference/api/analysis_config.cc
浏览文件 @
71636e67
...
...
@@ -57,6 +57,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
use_tensorrt_
=
other
.
use_tensorrt_
;
tensorrt_max_batchsize_
=
other
.
tensorrt_max_batchsize_
;
tensorrt_workspace_size_
=
other
.
tensorrt_workspace_size_
;
tensorrt_min_subgraph_size_
=
other
.
tensorrt_min_subgraph_size_
;
model_from_memory_
=
other
.
model_from_memory_
;
if
(
use_gpu
)
{
...
...
@@ -89,6 +90,7 @@ contrib::AnalysisConfig::AnalysisConfig(contrib::AnalysisConfig &&other) {
use_tensorrt_
=
other
.
use_tensorrt_
;
tensorrt_max_batchsize_
=
other
.
tensorrt_max_batchsize_
;
tensorrt_workspace_size_
=
other
.
tensorrt_workspace_size_
;
tensorrt_min_subgraph_size_
=
other
.
tensorrt_min_subgraph_size_
;
model_from_memory_
=
other
.
model_from_memory_
;
pass_builder_
=
std
::
move
(
other
.
pass_builder_
);
...
...
@@ -105,11 +107,13 @@ void contrib::AnalysisConfig::EnableMKLDNN() {
}
void
contrib
::
AnalysisConfig
::
EnableTensorRtEngine
(
int
workspace_size
,
int
max_batch_size
)
{
int
max_batch_size
,
int
min_subgraph_size
)
{
use_tensorrt_
=
true
;
tensorrt_workspace_size_
=
workspace_size
;
tensorrt_max_batchsize_
=
max_batch_size
;
// Append after the infer_clean pass.
tensorrt_min_subgraph_size_
=
min_subgraph_size
;
// Append after the conv+affine_channel fuse pass.
pass_builder
()
->
InsertPass
(
3
,
"tensorrt_subgraph_pass"
);
}
...
...
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @
71636e67
...
...
@@ -328,6 +328,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_
.
SetUseTensorRT
(
true
);
argument_
.
SetTensorRtWorkspaceSize
(
config_
.
tensorrt_workspace_size_
);
argument_
.
SetTensorRtMaxBatchSize
(
config_
.
tensorrt_max_batchsize_
);
argument_
.
SetTensorRtMinSubgraphSize
(
config_
.
tensorrt_min_subgraph_size_
);
}
if
(
config_
.
use_mkldnn_
)
{
...
...
paddle/fluid/inference/api/paddle_analysis_config.h
浏览文件 @
71636e67
...
...
@@ -49,7 +49,7 @@ struct AnalysisConfig : public NativeConfig {
bool
use_feed_fetch_ops
{
true
};
void
EnableTensorRtEngine
(
int
workspace_size
=
1
<<
20
,
int
max_batch_size
=
1
);
int
max_batch_size
=
1
,
int
min_subgraph_size
=
3
);
bool
use_tensorrt
()
const
{
return
use_tensorrt_
;
}
void
EnableMKLDNN
();
...
...
@@ -69,8 +69,19 @@ struct AnalysisConfig : public NativeConfig {
bool
use_tensorrt_
{
false
};
bool
use_mkldnn_
{
false
};
std
::
unordered_set
<
std
::
string
>
mkldnn_enabled_op_types_
;
// For workspace_size, refer it from here:
// https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting
int
tensorrt_workspace_size_
;
// While TensorRT allows an engine optimized for a given max batch size
// to run at any smaller size, the performance for those smaller
// sizes may not be as well-optimized. Therefore, Max batch is best
// equivalent to the runtime batch size.
int
tensorrt_max_batchsize_
;
// We transform the Ops that can be converted into TRT layer in the model,
// and aggregate these Ops into subgraphs for TRT execution.
// We set this variable to control the minimum number of nodes in the
// subgraph, 3 as default value.
int
tensorrt_min_subgraph_size_
{
3
};
std
::
unique_ptr
<
PassStrategy
>
pass_builder_
;
bool
model_from_memory_
{
false
};
};
...
...
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