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
c407dfa3
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
体验新版 GitCode,发现更多精彩内容 >>
You need to sign in or sign up before continuing.
提交
c407dfa3
编写于
3月 20, 2019
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
cherry-pick from feature/anakin-engine: refine paddle-anakin to new interface. #16276
上级
a25331bc
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
56 addition
and
42 deletion
+56
-42
paddle/fluid/framework/ir/CMakeLists.txt
paddle/fluid/framework/ir/CMakeLists.txt
+1
-1
paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.cc
.../fluid/framework/ir/transpose_flatten_concat_fuse_pass.cc
+6
-0
paddle/fluid/inference/anakin/convert/op_converter.h
paddle/fluid/inference/anakin/convert/op_converter.h
+4
-17
paddle/fluid/inference/anakin/engine.cc
paddle/fluid/inference/anakin/engine.cc
+1
-2
paddle/fluid/inference/anakin/engine.h
paddle/fluid/inference/anakin/engine.h
+7
-4
paddle/fluid/inference/api/paddle_pass_builder.cc
paddle/fluid/inference/api/paddle_pass_builder.cc
+1
-1
paddle/fluid/operators/anakin/anakin_engine_op.h
paddle/fluid/operators/anakin/anakin_engine_op.h
+36
-17
未找到文件。
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @
c407dfa3
...
...
@@ -79,7 +79,7 @@ pass_library(anakin_fillconstant_elementwisemul_fuse inference)
# be detected by our pass. The index here represents the number of structures in the
# pattern. We use index 3 ~ 6, because these quantities of structures are
# common in the models.
foreach
(
index RANGE
3
6
)
foreach
(
index RANGE
2
6
)
file
(
APPEND
${
pass_file
}
"USE_PASS(transpose_flatten
${
index
}
_concat_fuse_pass);
\n
"
)
endforeach
()
...
...
paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.cc
浏览文件 @
c407dfa3
...
...
@@ -12,7 +12,9 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <memory>
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
...
...
@@ -123,6 +125,7 @@ std::unique_ptr<ir::Graph> TransposeFlattenConcatFusePass<times>::ApplyImpl(
}
template
class
TransposeFlattenConcatFusePass
<
1
>;
template
class
TransposeFlattenConcatFusePass
<
2
>;
template
class
TransposeFlattenConcatFusePass
<
3
>;
template
class
TransposeFlattenConcatFusePass
<
4
>;
template
class
TransposeFlattenConcatFusePass
<
5
>;
...
...
@@ -135,6 +138,9 @@ template class TransposeFlattenConcatFusePass<6>;
REGISTER_PASS
(
transpose_flatten_concat_fuse_pass
,
paddle
::
framework
::
ir
::
TransposeFlattenConcatFusePass
<
1
>
);
REGISTER_PASS
(
transpose_flatten2_concat_fuse_pass
,
paddle
::
framework
::
ir
::
TransposeFlattenConcatFusePass
<
2
>
);
REGISTER_PASS
(
transpose_flatten3_concat_fuse_pass
,
paddle
::
framework
::
ir
::
TransposeFlattenConcatFusePass
<
3
>
);
...
...
paddle/fluid/inference/anakin/convert/op_converter.h
浏览文件 @
c407dfa3
...
...
@@ -117,27 +117,14 @@ class AnakinOpConverter {
}
temp_max_input_shape
[
input
]
=
input_shape
;
engine
->
SetInputShape
(
input
,
input_shape
);
// engine->Graph()->RegistVar(input);
// For share from data.
engine
->
Graph
()
->
RegistVar
(
input
);
// For share from data.
}
engine
->
SetMaxInputShape
(
temp_max_input_shape
);
engine
->
Optimize
();
// For anakin share with fluid tensor.
engine
->
AllocTmpMem
();
engine
->
InitGraph
();
/*
for(auto& input : inputs) {
platform::CUDAPlace gpu_place(engine->GetDevice());
auto input_var = scope->Var();
auto input_tensor = input_var->GetMutable<framework::LoDTensor>();
auto input_max_shape = temp_max_input_shape[input];
input_tensor->Resize(framework::make_ddim(input_max_shape));
auto input_data = input_tensor->mutable_data<float>(gpu_place);
auto* anakin_input = engine->Net()->get_in(input);
::anakin::saber::Tensor<::anakin::saber::NV> tmp_anakin_tensor(input_data,
::anakin::saber::NV(), 0, input_max_shape);
anakin_input->share_from(tmp_anakin_tensor);
}
*/
}
void
SetEngine
(
AnakinNvEngine
*
engine
)
{
engine_
=
engine
;
}
...
...
paddle/fluid/inference/anakin/engine.cc
浏览文件 @
c407dfa3
...
...
@@ -97,15 +97,14 @@ void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
anakin_input = net_->get_in(input.first);
}
*/
anakin_input
->
reshape
(
fluid_input_shape
);
::
anakin
::
saber
::
Tensor
<
TargetT
>
tmp_anakin_tensor
(
data
,
TargetT
(),
0
,
fluid_input_shape
);
anakin_input
->
copy_from
(
tmp_anakin_tensor
);
}
cudaDeviceSynchronize
();
net_
->
prediction
();
cudaDeviceSynchronize
();
for
(
const
auto
&
output
:
outputs
)
{
platform
::
CUDAPlace
gpu_place
(
device_
);
auto
*
tensor
=
output
.
second
;
...
...
paddle/fluid/inference/anakin/engine.h
浏览文件 @
c407dfa3
...
...
@@ -84,17 +84,20 @@ class AnakinEngine {
int
GetMaxBatchSize
()
{
return
max_batch_size_
;
}
void
Freeze
();
void
Optimize
();
void
AllocTmpMem
()
{
PADDLE_ENFORCE
(
net_
->
alloc_memory_first
(
*
graph_
),
"anakin alloc temp memory first failed"
);
}
void
Save
(
std
::
string
path
)
{
graph_
->
save
(
path
);
}
bool
IsInit
()
{
return
initialized_
;
}
int
GetDevice
()
{
return
device_
;
}
// void SaveSerializedData(std::string& data) { graph_->save_to_string(data);
// }
// void LoadSerializedData(const std::string& data) {
// graph_->load_from_string(data); }
void
Execute
(
const
std
::
map
<
std
::
string
,
framework
::
LoDTensor
*>
&
inputs
,
const
std
::
map
<
std
::
string
,
framework
::
LoDTensor
*>
&
outputs
,
cudaStream_t
stream
);
private:
bool
initialized_
{
false
};
int
max_batch_size_
;
std
::
map
<
std
::
string
,
std
::
vector
<
int
>>
max_input_shape_
;
int
device_
;
...
...
paddle/fluid/inference/api/paddle_pass_builder.cc
浏览文件 @
c407dfa3
...
...
@@ -99,7 +99,7 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
#endif
});
for
(
int
i
=
6
;
i
>=
3
;
i
--
)
{
for
(
int
i
=
6
;
i
>=
2
;
i
--
)
{
passes_
.
push_back
(
"transpose_flatten"
+
std
::
to_string
(
i
)
+
"_concat_fuse_pass"
);
}
...
...
paddle/fluid/operators/anakin/anakin_engine_op.h
浏览文件 @
c407dfa3
...
...
@@ -97,23 +97,7 @@ class AnakinEngineOp : public framework::OperatorBase {
if
(
param_names_
.
count
(
x
))
continue
;
auto
&
t
=
inference
::
analysis
::
GetFromScope
<
framework
::
LoDTensor
>
(
scope
,
x
);
/*
auto t_shape = framework::vectorize(t.dims());
auto *anakin_input = engine->Net()->get_in(x);
auto net_shape = anakin_input->shape();
size_t anakin_net_input_size = net_shape.count() * sizeof(float);
size_t fluid_input_size = t.memory_size();
if (fluid_input_size < anakin_net_input_size) {
framework::LoDTensor temp_t;
auto t_dims = t.dims();
temp_t.Resize(t_dims);
TensorCopySync(t, dev_place, &temp_t);
t.Resize(framework::make_ddim(net_shape));
t.mutable_data<float>(dev_place);
TensorCopySync(temp_t, dev_place, &t);
}
*/
inputs
.
insert
({
x
,
&
t
});
}
...
...
@@ -136,6 +120,41 @@ class AnakinEngineOp : public framework::OperatorBase {
inference
::
Singleton
<
inference
::
anakin
::
AnakinEngineManager
>::
Global
()
.
Get
(
engine_key_
);
}
// BUG here, detect that the tensor data pointer here will change sometime.
// Will fix it later.
/*
// For share with the tensor from fluid, We do the net init in the first net
precit.
if (!anakin_engine_->IsInit()) {
auto temp_max_input_shape = anakin_engine_->GetMaxInputShape();
anakin_engine_->AllocTmpMem();
for(auto& input : Inputs("Xs")) {
if (param_names_.count(input)) continue;
platform::CUDAPlace
gpu_place(boost::get<platform::CUDAPlace>(dev_place).device);
auto *input_var = scope.FindVar(input);
auto input_tensor = input_var->GetMutable<framework::LoDTensor>();
auto input_max_shape = temp_max_input_shape[input];
framework::LoDTensor temp_t;
auto t_dims = input_tensor->dims();
temp_t.Resize(t_dims);
TensorCopySync(*input_tensor, dev_place, &temp_t);
input_tensor->Resize(framework::make_ddim(input_max_shape));
input_tensor->mutable_data<float>(dev_place);
TensorCopySync(temp_t, dev_place, input_tensor);
auto* input_data = input_tensor->mutable_data<float>(gpu_place);
auto* anakin_input = anakin_engine_->Net()->get_in(input);
::anakin::saber::Tensor<::anakin::saber::NV>
tmp_anakin_tensor(input_data,
::anakin::saber::NV(), 0, input_max_shape);
anakin_input->share_from(tmp_anakin_tensor);
}
anakin_engine_->InitGraph();
}
*/
return
anakin_engine_
;
}
...
...
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