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
PaddleDetection
提交
94a57f1d
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
12 个月 前同步成功
通知
692
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
94a57f1d
编写于
9月 19, 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 trt config to arguments
上级
68fb818a
变更
10
显示空白变更内容
内联
并排
Showing
10 changed file
with
49 addition
and
19 deletion
+49
-19
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc
...fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc
+6
-2
paddle/fluid/inference/analysis/subgraph_splitter.cc
paddle/fluid/inference/analysis/subgraph_splitter.cc
+2
-1
paddle/fluid/inference/analysis/subgraph_splitter.h
paddle/fluid/inference/analysis/subgraph_splitter.h
+7
-2
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc
+1
-1
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h
+5
-1
paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc
paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc
+10
-2
paddle/fluid/inference/api/paddle_inference_api.h
paddle/fluid/inference/api/paddle_inference_api.h
+9
-0
paddle/fluid/inference/tests/api/trt_models_tester.cc
paddle/fluid/inference/tests/api/trt_models_tester.cc
+1
-1
paddle/fluid/operators/tensorrt_engine_op.cc
paddle/fluid/operators/tensorrt_engine_op.cc
+2
-2
paddle/fluid/operators/tensorrt_engine_op.h
paddle/fluid/operators/tensorrt_engine_op.h
+6
-7
未找到文件。
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc
浏览文件 @
94a57f1d
...
...
@@ -97,8 +97,9 @@ void DataFlowGraphToFluidPass::AddFluidOp(Node *node) {
}
}
void
CreateTrtEngineOp
(
Node
*
node
,
const
DataFlowGraph
&
graph
,
void
CreateTrtEngineOp
(
Node
*
node
,
Argument
*
argument
,
framework
::
proto
::
BlockDesc
*
block
)
{
const
DataFlowGraph
&
graph
=
*
(
argument
->
main_dfg
);
static
int
counter
{
0
};
PADDLE_ENFORCE
(
node
->
IsFunctionBlock
());
framework
::
OpDesc
desc
;
...
...
@@ -204,7 +205,10 @@ void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph,
PADDLE_ENFORCE
(
!
block
->
vars
().
empty
(),
"the block has no var-desc"
);
// Set attrs
SetAttr
(
desc
.
Proto
(),
"subgraph"
,
block
->
SerializeAsString
());
SetAttr
(
desc
.
Proto
(),
"max_batch_size"
,
argument
->
Get
<
int
>
(
"max_batch_size"
));
SetAttr
(
desc
.
Proto
(),
"workspace_size"
,
argument
->
Get
<
int
>
(
"workspace_size"
));
SetAttr
(
desc
.
Proto
(),
"engine_uniq_key"
,
"trt-"
+
std
::
to_string
(
counter
++
));
SetAttr
(
desc
.
Proto
(),
"parameters"
,
ExtractParameters
(
graph
.
nodes
.
nodes
()));
SetAttr
(
desc
.
Proto
(),
"output_name_mapping"
,
output_mapping
);
...
...
@@ -248,7 +252,7 @@ void DataFlowGraphToFluidPass::AddEngineOp(Node *node) {
*
block_desc
.
Proto
()
->
mutable_vars
()
=
argument_
->
origin_program_desc
->
blocks
(
0
).
vars
();
PADDLE_ENFORCE
(
!
block_desc
.
Proto
()
->
vars
().
empty
());
CreateTrtEngineOp
(
node
,
*
argument_
->
main_dfg
,
block_desc
.
Proto
());
CreateTrtEngineOp
(
node
,
argument_
,
block_desc
.
Proto
());
auto
*
main_block
=
desc_
->
mutable_blocks
(
framework
::
kRootBlockIndex
);
auto
*
op
=
main_block
->
add_ops
();
PADDLE_ENFORCE
(
!
node
->
pb_msg
().
empty
(),
"failed to set desc for block"
);
...
...
paddle/fluid/inference/analysis/subgraph_splitter.cc
浏览文件 @
94a57f1d
...
...
@@ -309,7 +309,8 @@ void SubGraphFuse::operator()() { ReplaceNodesWithSubGraphs(); }
void
SubGraphFuse
::
ReplaceNodesWithSubGraphs
()
{
auto
subgraphs
=
SubGraphSplitter
(
graph_
,
node_inside_subgraph_teller_
)();
for
(
auto
&
subgraph
:
subgraphs
)
{
if
(
subgraph
.
size
()
<=
3
)
continue
;
if
(
subgraph
.
size
()
<=
argument_
->
Get
<
int
>
(
"minimun_subgraph_size"
))
continue
;
std
::
unordered_set
<
Node
*>
subgraph_uniq
(
subgraph
.
begin
(),
subgraph
.
end
());
// replace this sub-graph with the first node. Two steps: 1. Create a Block
// Node that contains this subgraph 2. Mark the nodes inside the sub-graph
...
...
paddle/fluid/inference/analysis/subgraph_splitter.h
浏览文件 @
94a57f1d
...
...
@@ -20,6 +20,7 @@ limitations under the License. */
#include <vector>
#include "paddle/fluid/inference/analysis/argument.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/node.h"
...
...
@@ -63,8 +64,11 @@ class SubGraphFuse {
public:
using
NodeInsideSubgraphTeller
=
SubGraphSplitter
::
NodeInsideSubgraphTeller
;
SubGraphFuse
(
DataFlowGraph
*
graph
,
const
NodeInsideSubgraphTeller
&
teller
)
:
graph_
(
graph
),
node_inside_subgraph_teller_
(
teller
)
{}
SubGraphFuse
(
DataFlowGraph
*
graph
,
const
NodeInsideSubgraphTeller
&
teller
,
Argument
*
argument
)
:
graph_
(
graph
),
node_inside_subgraph_teller_
(
teller
),
argument_
(
argument
)
{}
// The main method which run all the logic.
void
operator
()();
...
...
@@ -76,6 +80,7 @@ class SubGraphFuse {
private:
DataFlowGraph
*
graph_
;
NodeInsideSubgraphTeller
node_inside_subgraph_teller_
;
Argument
*
argument_
;
};
}
// namespace analysis
...
...
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc
浏览文件 @
94a57f1d
...
...
@@ -24,7 +24,7 @@ TensorRTSubGraphPass::TensorRTSubGraphPass(
:
node_inside_subgraph_teller_
(
teller
)
{}
void
TensorRTSubGraphPass
::
Run
(
DataFlowGraph
*
graph
)
{
SubGraphFuse
(
graph
,
node_inside_subgraph_teller_
)();
SubGraphFuse
(
graph
,
node_inside_subgraph_teller_
,
argument_
)();
VLOG
(
4
)
<<
"debug info "
<<
graph
->
HumanReadableInfo
(
false
/*show_values*/
,
true
/*show_functions*/
);
...
...
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h
浏览文件 @
94a57f1d
...
...
@@ -33,7 +33,10 @@ class TensorRTSubGraphPass : public DataFlowGraphPass {
explicit
TensorRTSubGraphPass
(
const
NodeInsideSubgraphTeller
&
teller
);
bool
Initialize
(
Argument
*
argument
)
override
{
return
true
;
}
bool
Initialize
(
Argument
*
argument
)
override
{
argument_
=
argument
;
return
true
;
}
// This class get a sub-graph as input and determine whether to transform this
// sub-graph into TensorRT.
...
...
@@ -46,6 +49,7 @@ class TensorRTSubGraphPass : public DataFlowGraphPass {
private:
NodeInsideSubgraphTeller
node_inside_subgraph_teller_
;
Argument
*
argument_
;
};
}
// namespace analysis
...
...
paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc
浏览文件 @
94a57f1d
...
...
@@ -34,8 +34,6 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
bool
Init
(
const
std
::
shared_ptr
<
framework
::
Scope
>&
parent_scope
)
{
FLAGS_IA_enable_tensorrt_subgraph_engine
=
true
;
VLOG
(
3
)
<<
"Predictor::init()"
;
FLAGS_tensorrt_max_batch_size
=
config_
.
max_batch_size
;
FLAGS_tensorrt_workspace_size
=
config_
.
workspace_size
;
if
(
config_
.
use_gpu
)
{
place_
=
paddle
::
platform
::
CUDAPlace
(
config_
.
device
);
}
else
{
...
...
@@ -91,6 +89,16 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
void
OptimizeInferenceProgram
()
{
// Analyze inference_program
Argument
argument
;
int
*
minimum_subgraph_size
=
new
int
(
config_
.
minimun_subgraph_size
);
int
*
max_batch_size
=
new
int
(
config_
.
max_batch_size
);
int
*
workspace_size
=
new
int
(
config_
.
workspace_size
);
std
::
string
*
precision_mode
=
new
std
::
string
(
config_
.
precision_mode
);
argument
.
Set
<
int
>
(
"minimun_subgraph_size"
,
minimum_subgraph_size
);
argument
.
Set
<
int
>
(
"max_batch_size"
,
max_batch_size
);
argument
.
Set
<
int
>
(
"workspace_size"
,
workspace_size
);
argument
.
Set
<
std
::
string
>
(
"precision_mode"
,
precision_mode
);
if
(
!
config_
.
model_dir
.
empty
())
{
argument
.
fluid_model_dir
.
reset
(
new
std
::
string
(
config_
.
model_dir
));
}
else
{
...
...
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @
94a57f1d
...
...
@@ -150,6 +150,15 @@ struct TensorRTConfig : public NativeConfig {
// For workspace_size, refer it from here:
// https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting
int
workspace_size
{
1
<<
30
};
// 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
minimun_subgraph_size
=
3
;
// Reserved configuration
// We just support "FP32" now, "FP16" and "INT8" will be supported.
std
::
string
precision_mode
=
"FP32"
;
};
// NOTE WIP, not stable yet.
...
...
paddle/fluid/inference/tests/api/trt_models_tester.cc
浏览文件 @
94a57f1d
...
...
@@ -99,7 +99,7 @@ TEST(trt_models_test, main) {
std
::
vector
<
std
::
string
>
infer_models
=
{
"mobilenet"
,
"resnet50"
,
"resnext50"
};
for
(
auto
&
model_dir
:
infer_models
)
{
CompareTensorRTWithFluid
(
1
,
FLAGS_dirname
+
"/"
+
model_dir
);
CompareTensorRTWithFluid
(
5
,
FLAGS_dirname
+
"/"
+
model_dir
);
}
}
}
// namespace paddle
paddle/fluid/operators/tensorrt_engine_op.cc
浏览文件 @
94a57f1d
...
...
@@ -22,8 +22,6 @@
namespace
paddle
{
DEFINE_int32
(
tensorrt_engine_batch_size
,
1
,
"the batch_size of TensorRT"
);
DEFINE_int32
(
tensorrt_max_batch_size
,
1
,
"TensorRT maximum batch size"
);
DEFINE_int32
(
tensorrt_workspace_size
,
16
<<
20
,
"TensorRT workspace size"
);
namespace
operators
{
...
...
@@ -34,6 +32,8 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
AddOutput
(
"Ys"
,
"A list of outputs"
).
AsDuplicable
();
AddAttr
<
std
::
string
>
(
"subgraph"
,
"the subgraph."
);
AddAttr
<
std
::
string
>
(
"engine_uniq_key"
,
"unique key for the TRT engine."
);
AddAttr
<
int
>
(
"max_batch_size"
,
"the maximum batch size."
);
AddAttr
<
int
>
(
"workspace_size"
,
"the maximum batch size."
);
AddComment
(
"TensorRT engine operator."
);
}
};
...
...
paddle/fluid/operators/tensorrt_engine_op.h
浏览文件 @
94a57f1d
...
...
@@ -28,8 +28,6 @@
namespace
paddle
{
DECLARE_int32
(
tensorrt_engine_batch_size
);
DECLARE_int32
(
tensorrt_max_batch_size
);
DECLARE_int32
(
tensorrt_workspace_size
);
namespace
operators
{
...
...
@@ -92,14 +90,14 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
engine_name
=
context
.
Attr
<
std
::
string
>
(
"engine_uniq_key"
);
int
max_batch_size
=
context
.
Attr
<
int
>
(
"max_batch_size"
);
if
(
!
Singleton
<
TRT_EngineManager
>::
Global
().
HasEngine
(
engine_name
))
{
Prepare
(
context
);
}
auto
*
engine
=
Singleton
<
TRT_EngineManager
>::
Global
().
Get
(
engine_name
);
auto
input_names
=
context
.
op
().
Inputs
(
"Xs"
);
PADDLE_ENFORCE
(
!
input_names
.
empty
(),
"should pass more than one inputs"
);
PADDLE_ENFORCE_LE
(
FLAGS_tensorrt_engine_batch_size
,
FLAGS_tensorrt_max_batch_size
);
PADDLE_ENFORCE_LE
(
FLAGS_tensorrt_engine_batch_size
,
max_batch_size
);
std
::
vector
<
std
::
string
>
output_maps
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"output_name_mapping"
);
...
...
@@ -173,8 +171,9 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
// Get the ProgramDesc and pass to convert.
framework
::
proto
::
BlockDesc
block_desc
;
block_desc
.
ParseFromString
(
context
.
Attr
<
std
::
string
>
(
"subgraph"
));
int
max_batch
=
FLAGS_tensorrt_max_batch_size
;
auto
max_workspace
=
FLAGS_tensorrt_workspace_size
;
int
max_batch_size
=
context
.
Attr
<
int
>
(
"max_batch_size"
);
int
workspace_size
=
context
.
Attr
<
int
>
(
"workspace_size"
);
auto
params
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"parameters"
);
std
::
unordered_set
<
std
::
string
>
parameters
;
for
(
const
auto
&
param
:
params
)
{
...
...
@@ -186,7 +185,7 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
// TODO(Superjomn) replace this with a different stream
auto
*
engine
=
Singleton
<
TRT_EngineManager
>::
Global
().
Create
(
max_batch
,
max_workspac
e
,
nullptr
/*engine hold its own stream*/
,
max_batch
_size
,
workspace_siz
e
,
nullptr
/*engine hold its own stream*/
,
context
.
Attr
<
std
::
string
>
(
"engine_uniq_key"
),
boost
::
get
<
platform
::
CUDAPlace
>
(
context
.
GetPlace
()).
device
);
...
...
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