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
极客李华
Paddle
提交
59765362
P
Paddle
项目概览
极客李华
/
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 >>
未验证
提交
59765362
编写于
3月 29, 2022
作者:
W
weishengying
提交者:
GitHub
3月 29, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add rewrite pattern form paddle mlir to trt mlir (#41011)
上级
869287f8
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
203 addition
and
4 deletion
+203
-4
paddle/infrt/dialect/tensorrt/convert.h
paddle/infrt/dialect/tensorrt/convert.h
+86
-0
paddle/infrt/dialect/tensorrt/pd_lower_to_trt.td
paddle/infrt/dialect/tensorrt/pd_lower_to_trt.td
+18
-1
paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
+1
-0
paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
+3
-0
paddle/infrt/dialect/tensorrt/trt_ops.td
paddle/infrt/dialect/tensorrt/trt_ops.td
+95
-3
未找到文件。
paddle/infrt/dialect/tensorrt/convert.h
0 → 100644
浏览文件 @
59765362
// Copyright (c) 2022 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.
#pragma once
#include <mlir/IR/Builders.h>
#include <mlir/Transforms/DialectConversion.h>
#include "paddle/infrt/dialect/infrt/ir/infrt_dialect.h"
#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
namespace
infrt
{
namespace
trt
{
static
mlir
::
Value
createTRTConv2dOp
(
mlir
::
PatternRewriter
&
rewriter
,
mlir
::
Operation
*
op
)
{
::
mlir
::
Operation
::
operand_range
Input
(
op
->
getOperands
());
::
mlir
::
Operation
::
operand_range
Filter
(
op
->
getOperands
());
::
mlir
::
SmallVector
<::
mlir
::
Value
,
4
>
operands
;
auto
castedOp0
=
::
llvm
::
dyn_cast
<
infrt
::
pd
::
Conv2dOp
>
(
op
);
(
void
)
castedOp0
;
Input
=
castedOp0
.
getODSOperands
(
0
);
Filter
=
castedOp0
.
getODSOperands
(
1
);
operands
.
push_back
((
*
Input
.
begin
()));
operands
.
push_back
((
*
Input
.
begin
()));
::
mlir
::
SmallVector
<::
mlir
::
Type
,
4
>
resultTypes
;
for
(
auto
v
:
castedOp0
.
getODSResults
(
0
))
{
resultTypes
.
push_back
(
v
.
getType
());
}
::
mlir
::
SmallVector
<::
mlir
::
NamedAttribute
,
8
>
attributes
;
{
auto
tblgen_attr
=
rewriter
.
getSI32IntegerAttr
(
3
);
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"out_channel_num"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
rewriter
.
getI32ArrayAttr
({
3
,
3
});
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"kernel_size"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
ArrayAttr
>
(
"strides"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"strides"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
ArrayAttr
>
(
"paddings"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"paddings"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
StringAttr
>
(
"padding_algorithm"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"padding_mode"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
IntegerAttr
>
(
"groups"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"groups"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
ArrayAttr
>
(
"dilations"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"dilations"
),
tblgen_attr
);
}
{
auto
tblgen_attr
=
op
->
getAttrOfType
<::
mlir
::
StringAttr
>
(
"data_format"
);
(
void
)
tblgen_attr
;
attributes
.
emplace_back
(
rewriter
.
getStringAttr
(
"data_format"
),
tblgen_attr
);
}
return
rewriter
.
create
<
trt
::
ConvolutionOp
>
(
op
->
getLoc
(),
resultTypes
,
operands
,
attributes
);
}
}
// namespace trt
}
// namespace infrt
paddle/infrt/dialect/tensorrt/pd_lower_to_trt.td
浏览文件 @
59765362
...
...
@@ -14,7 +14,7 @@ def PD2TRT_Matmul_Lower : Pat<
(TRT_MatrixMultiplyOp $X, $transpose_X, $Y, $transpose_Y)>;
def PD2TRT_ElementwiseAdd_Lower : Pat<
(PD_Elementwise_addOp $X, $Y,
ConstantAttr<SI32Attr, "-1">
),
(PD_Elementwise_addOp $X, $Y,
$_
),
(TRT_ElementWiseOp $X, $Y, (TRT_createNvinferEnumAttr<"nvinfer1::ElementWiseOperation", "kSUM">))>;
def PD2TRT_Relu_Lower : Pat<
...
...
@@ -25,4 +25,21 @@ def PD2TRT_Relu6_Lower : Pat<
(PD_Relu6Op $X, $threshold),
(TRT_ActivationOp $X, (TRT_createNvinferEnumAttr<"nvinfer1::ActivationType", "kCLIP">), (INFRT_createF32Attr<"0.0">), $threshold)>;
def createTRTConv2dOp : NativeCodeCall<"createTRTConv2dOp($_builder, $0.getDefiningOp())">;
def PD2TRT_Conv2d_Lower : Pat<
(PD_Conv2dOp:$old_value $Input, $Filter, $strides, $paddings, $padding_algorithm, $groups, $dilations, $data_format),
(createTRTConv2dOp $old_value)>;
def PD2TRT_Pooling_Lower : Pat<
(PD_Pool2dOp $Input, $pooling_type, $ksize, $global_pooling, $strides, $paddings, $exclusive, $adaptive, $ceil_mode, $data_format, $padding_algorithm),
(TRT_PoolingOp $Input, (INFRT_createI32Attr<"0">)/*kmax*/, $ksize, $strides, $paddings, $padding_algorithm)>;
def PD2TRT_MatrixMultipl_Lower : Pat<
(PD_MulOp $Input1, $Input2, $x_num_col_dims, $y_num_col_dims),
(TRT_MatrixMultiplOp $Input1, (INFRT_createI32Attr<"0">)/*kNONE*/, $Input2, (INFRT_createI32Attr<"0">)/*kNONE*/)>;
def PD2TRT_SoftMax_Lower : Pat<
(PD_SoftmaxOp $Input, $axis, $_),
(TRT_SoftMaxOp $Input, $axis)>;
#endif // PD_LOWER_TO_TRT
paddle/infrt/dialect/tensorrt/trt_op_converter_pass.cc
浏览文件 @
59765362
...
...
@@ -21,6 +21,7 @@
#include "paddle/infrt/dialect/pd/ir/pd_ops.h"
#include "paddle/infrt/dialect/phi/ir/infrt_phi_tensor.h"
#include "paddle/infrt/dialect/phi/ir/phi_base.h"
#include "paddle/infrt/dialect/tensorrt/convert.h"
#include "paddle/infrt/dialect/tensorrt/trt_dialect_types.h"
#include "paddle/infrt/dialect/tensorrt/trt_ops.h"
...
...
paddle/infrt/dialect/tensorrt/trt_op_teller_pass.cc
浏览文件 @
59765362
...
...
@@ -20,6 +20,7 @@
#include "paddle/infrt/dialect/infrt/ir/basic_kernels.h"
#include "paddle/infrt/dialect/infrt/ir/infrt_dialect.h"
#include "paddle/infrt/dialect/pd/ir/pd_ops.h"
#include "paddle/infrt/dialect/phi/ir/infrt_phi_tensor.h"
namespace
infrt
{
namespace
trt
{
...
...
@@ -42,6 +43,8 @@ void TRTOpTellerPass::runOnFunction() {
if
(
::
llvm
::
dyn_cast_or_null
<
infrt
::
pd
::
FetchOp
>
(
op
))
continue
;
if
(
::
llvm
::
dyn_cast_or_null
<::
infrt
::
GraphOp
>
(
op
))
continue
;
if
(
::
llvm
::
dyn_cast_or_null
<::
infrt
::
ReturnOp
>
(
op
))
continue
;
if
(
::
llvm
::
dyn_cast_or_null
<::
infrt
::
phi
::
TensorMapGetTensorOp
>
(
op
))
continue
;
builder
.
setInsertionPoint
(
op
);
auto
loc
=
getFunction
().
getLoc
();
...
...
paddle/infrt/dialect/tensorrt/trt_ops.td
浏览文件 @
59765362
...
...
@@ -9,6 +9,7 @@ include "paddle/infrt/dialect/tensorrt/trt_op_base.td"
include "paddle/infrt/dialect/infrt/ir/infrt_base.td"
include "paddle/infrt/dialect/phi/ir/infrt_phi_base.td"
include "paddle/infrt/dialect/pd/ir/pd_op_base.td"
def TRT_CreateEngineOp : TRT_Op<"create_engine", [SingleBlockImplicitTerminator<"::infrt::ReturnOp">]> {
let summary = "trt CreateEngine Op";
...
...
@@ -16,7 +17,7 @@ def TRT_CreateEngineOp : TRT_Op<"create_engine", [SingleBlockImplicitTerminator<
Describe a tensorrt subgraph.
}];
let regions = (region SizedRegion<1>:$body);
let arguments = (ins Variadic<
Dense
Tensor>:$inputs, DefaultValuedAttr<BoolAttr, "true">:$run_once);
let arguments = (ins Variadic<
PD_
Tensor>:$inputs, DefaultValuedAttr<BoolAttr, "true">:$run_once);
let results = (outs TRT_EngineType:$engine);
}
...
...
@@ -75,9 +76,32 @@ def TRT_ConvolutionOp : TRT_Op<"Convolution", [NoSideEffect]> {
let arguments = (ins
DenseTensor:$input_tensor,
DenseTensor:$kernel_weights,
DenseTensor
:$bias_weights,
Optional<DenseTensor>
:$bias_weights,
SI32Attr:$out_channel_num,
I32ArrayAttr:$kernel_size
I32ArrayAttr:$kernel_size,
I32ArrayAttr:$strides,
I32ArrayAttr:$paddings,
StrAttr:$padding_mode,
SI32Attr:$groups,
I32ArrayAttr:$dilations
);
let results = (outs
DenseTensor:$output_tensor
);
}
def TRT_PoolingOp : TRT_Op<"Pooling", [NoSideEffect]> {
let summary = "TensorRT IPoolingLayer ";
let description = [{
TensorRT IPoolingLayer
}];
let arguments = (ins
DenseTensor:$input_tensor,
I32Attr:$pool_type,
I32ArrayAttr:$window_size,
I32ArrayAttr:$strides,
I32ArrayAttr:$paddings,
StrAttr:$padding_mode
);
let results = (outs
DenseTensor:$output_tensor
...
...
@@ -109,4 +133,72 @@ def TRT_MatrixMultiplyOp : TRT_Op<"MatrixMultiply", [NoSideEffect]> {
let results = (outs DenseTensor:$output);
}
def TRT_ScaleOp : TRT_Op<"scale", [NoSideEffect]> {
let summary = "TensorRT IScaleLayer";
let description = [{
TensorRT IScaleLayer
}];
let arguments = (ins
DenseTensor:$input_tensor,
DefaultValuedAttr<I32Attr, "0">:$mode,
DenseTensor:$shift,
DenseTensor:$scale,
DenseTensor:$power
);
let results = (outs DenseTensor:$Out);
}
def TRT_MatrixMultiplOp : TRT_Op<"MatrixMultiplOp", [NoSideEffect]> {
let summary = "TensorRT IMatrixMultiplyLayer";
let description = [{
TensorRT IMatrixMultiplyLayer
}];
let arguments = (ins
DenseTensor:$input1,
DefaultValuedAttr<I32Attr, "0">:$matrix_operation1,
DenseTensor:$input2,
DefaultValuedAttr<I32Attr, "0">:$matrix_operation2
);
let results = (outs DenseTensor:$Out);
}
def TRT_SoftMaxOp : TRT_Op<"SoftMaxOp", [NoSideEffect]> {
let summary = "TensorRT ISoftMaxLayer";
let description = [{
TensorRT ISoftMaxLayer
}];
let arguments = (ins
DenseTensor:$input_tensor,
SI32Attr:$axis
);
let results = (outs DenseTensor:$Out);
}
def TRT_ScaleNdOp : TRT_Op<"ScaleNd", [NoSideEffect]> {
let summary = "TensorRT IScaleLayer";
let description = [{
TensorRT IScaleLayer
}];
let arguments = (ins
DenseTensor:$input_tensor,
I32Attr:$mode,
DenseTensor:$shift,
DenseTensor:$scale,
DenseTensor:$power,
I32Attr:$axis
);
let results = (outs DenseTensor:$Out);
}
#endif // TRT_OPS
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