pj3fizlut
/
mindspore
forked from OSSInnovation/mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

!3709 GPU update bng pass 

Merge pull request !3709 from VectorSL/update-bng-pass
This commit is contained in:
mindspore-ci-bot 2020-07-30 19:21:13 +08:00 committed by Gitee
parent 6875095b3a 9b0602747c
commit 0df4b11487
4 changed files with 36 additions and 170 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

88
mindspore/ccsrc/backend/optimizer/gpu/replace_bn_grad_cast2_fusion.cc
View File

@ -1,88 +0,0 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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.
*/
#include "backend/optimizer/gpu/replace_bn_grad_cast2_fusion.h"
#include <memory>
#include <vector>
#include <string>
#include "backend/session/anf_runtime_algorithm.h"
#include "ir/primitive.h"
#include "utils/utils.h"
#include "backend/optimizer/common/helper.h"
namespace mindspore {
namespace opt {
const BaseRef ReplaceBNGradCast2Fusion::DefinePattern() const {
VectorRef fbn2g = VectorRef({prim::kPrimFusedBatchNormGrad, dy_, x_, scale_, mean_, var_});
VectorRef tupleget = VectorRef({prim::kPrimTupleGetItem, fbn2g, index_});
VectorRef out_cast = VectorRef({prim::kPrimCast, tupleget});
return out_cast;
}
const AnfNodePtr ReplaceBNGradCast2Fusion::Process(const FuncGraphPtr &graph, const AnfNodePtr &node,
const EquivPtr &equiv) const {
MS_EXCEPTION_IF_NULL(graph);
MS_EXCEPTION_IF_NULL(node);
MS_EXCEPTION_IF_NULL(equiv);
auto tuple = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(node), 0);
auto index_node = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(tuple), 1);
MS_EXCEPTION_IF_NULL(index_node);
auto value_node = index_node->cast<ValueNodePtr>();
MS_EXCEPTION_IF_NULL(value_node);
int item_idx = GetValue<int>(value_node->value());
if (item_idx != 0) {
return nullptr;
}
auto fbn2g = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(tuple), 0);
auto dy_ = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 0);
auto x_ = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 1);
auto scale = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 2);
auto mean = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 3);
auto var = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 4);
MS_EXCEPTION_IF_NULL(fbn2g);
MS_EXCEPTION_IF_NULL(dy_);
MS_EXCEPTION_IF_NULL(scale);
MS_EXCEPTION_IF_NULL(x_);
MS_EXCEPTION_IF_NULL(mean);
MS_EXCEPTION_IF_NULL(var);
auto manager = graph->manager();
MS_EXCEPTION_IF_NULL(manager);
manager->Replace(utils::cast<CNodePtr>(node), utils::cast<CNodePtr>(tuple));
std::vector<TypeId> outputs_type;
std::vector<std::vector<size_t>> outputs_shape;
auto output_num = AnfAlgo::GetOutputTensorNum(fbn2g);
for (size_t i = 0; i < output_num; i++) {
outputs_type.push_back(AnfAlgo::GetOutputInferDataType(fbn2g, i));
outputs_shape.push_back(AnfAlgo::GetOutputInferShape(fbn2g, i));
}
outputs_type[0] = AnfAlgo::GetPrevNodeOutputInferDataType(fbn2g, 0);
AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, fbn2g.get());
outputs_type.clear();
outputs_shape.clear();
outputs_type.push_back(AnfAlgo::GetPrevNodeOutputInferDataType(fbn2g, 0));
outputs_shape.push_back(AnfAlgo::GetOutputInferShape(tuple, 0));
AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, tuple.get());
return tuple;
}
} // namespace opt
} // namespace mindspore

54
mindspore/ccsrc/backend/optimizer/gpu/replace_bn_grad_cast2_fusion.h
View File

@ -1,54 +0,0 @@
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_REPLACE_BN_GRAD_CAST2_FUSION_H_
#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_REPLACE_BN_GRAD_CAST2_FUSION_H_
#include <memory>
#include "backend/optimizer/common/optimizer.h"
namespace mindspore {
namespace opt {
class ReplaceBNGradCast2Fusion : public PatternProcessPass {
public:
explicit ReplaceBNGradCast2Fusion(bool multigraph = true) : PatternProcessPass("replace_grad_cast2", multigraph) {
dy_ = std::make_shared<Var>();
x_ = std::make_shared<Var>();
scale_ = std::make_shared<Var>();
mean_ = std::make_shared<Var>();
var_ = std::make_shared<Var>();
dx_ = std::make_shared<Var>();
bn_scale_ = std::make_shared<Var>();
bn_bias_ = std::make_shared<Var>();
index_ = std::make_shared<Var>();
}
~ReplaceBNGradCast2Fusion() override = default;
const BaseRef DefinePattern() const override;
const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
private:
VarPtr dy_;
VarPtr x_;
VarPtr scale_;
VarPtr mean_;
VarPtr var_;
VarPtr dx_;
VarPtr bn_scale_;
VarPtr bn_bias_;
VarPtr index_;
};
} // namespace opt
} // namespace mindspore
#endif // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_REPLACE_BN_GRAD_CAST2_FUSION_H_

50
mindspore/ccsrc/backend/optimizer/gpu/replace_bn_grad_cast_fusion.cc
View File

@ -45,7 +45,7 @@ const AnfNodePtr ReplaceBNGradCastFusion::Process(const FuncGraphPtr &graph, con
auto dy_before = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(dy_after), 0); auto dy_before = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(dy_after), 0);
auto x_ = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 1); auto x_ = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(fbn2g), 1);
auto x_type = AnfAlgo::GetOutputInferDataType(x_, 0); auto x_type = AnfAlgo::GetOutputInferDataType(x_, 0);
// if x_type is fp32, the cast is nessery. // if x_type is fp32, the cast is necessary.
if (x_type == kNumberTypeFloat32) { if (x_type == kNumberTypeFloat32) {
return nullptr; return nullptr;
} }
@ -65,35 +65,45 @@ const AnfNodePtr ReplaceBNGradCastFusion::Process(const FuncGraphPtr &graph, con
auto manager = graph->manager(); auto manager = graph->manager();
MS_EXCEPTION_IF_NULL(manager); MS_EXCEPTION_IF_NULL(manager);
auto outlist = GetRealNodeUsedList(graph, fbn2g); // 1. get all of the fusedbatchnormgrad nodes connected after dy_after.
for (size_t i = 0; i < outlist->size(); i++) { auto fbn2g_all = GetRealNodeUsedList(graph, dy_after);
auto index_node = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(outlist->at(i).first), 1); for (size_t i = 0; i < fbn2g_all->size(); i++) {
outputs_type.clear();
outputs_shape.clear();
auto kernel = utils::cast<CNodePtr>(fbn2g_all->at(i).first);
auto kernel_name = AnfAlgo::GetCNodeName(kernel);
// 2. deal all of the fusedbatchnormgrad, change the data type.
if (kernel_name == AnfAlgo::GetCNodeName(utils::cast<CNodePtr>(fbn2g))) {
auto output_num = AnfAlgo::GetOutputTensorNum(kernel);
for (size_t j = 0; j < output_num; j++) {
outputs_type.push_back(AnfAlgo::GetOutputInferDataType(kernel, j));
outputs_shape.push_back(AnfAlgo::GetOutputInferShape(kernel, j));
}
outputs_type[0] = kNumberTypeFloat16;
AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, kernel.get());
}
// 3. handle the output of fusedbatchnormgrad: tuplegetitem
auto outlist = GetRealNodeUsedList(graph, kernel);
for (size_t j = 0; j < outlist->size(); j++) {
outputs_type.clear();
outputs_shape.clear();
auto index_node = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(outlist->at(j).first), 1);
auto value_node = index_node->cast<ValueNodePtr>(); auto value_node = index_node->cast<ValueNodePtr>();
MS_EXCEPTION_IF_NULL(value_node); MS_EXCEPTION_IF_NULL(value_node);
int item_idx = GetValue<int>(value_node->value()); int item_idx = GetValue<int>(value_node->value());
if (item_idx == 0) { if (item_idx == 0) {
auto cast = GetRealNodeUsedList(graph, outlist->at(i).first); auto cast = GetRealNodeUsedList(graph, outlist->at(j).first);
if (AnfAlgo::GetCNodeName(cast->at(0).first) != "Cast") { if (AnfAlgo::GetCNodeName(cast->at(0).first) != "Cast") {
return nullptr; continue;
} }
manager->Replace(utils::cast<CNodePtr>(cast->at(0).first), utils::cast<CNodePtr>(outlist->at(i).first)); manager->Replace(utils::cast<CNodePtr>(cast->at(0).first), utils::cast<CNodePtr>(outlist->at(j).first));
outputs_type.push_back(kNumberTypeFloat16); outputs_type.push_back(kNumberTypeFloat16);
outputs_shape.push_back(AnfAlgo::GetOutputInferShape(outlist->at(i).first, 0)); outputs_shape.push_back(AnfAlgo::GetOutputInferShape(outlist->at(j).first, 0));
AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, outlist->at(i).first.get()); AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, outlist->at(j).first.get());
}
} }
} }
outputs_type.clear();
outputs_shape.clear();
manager->Replace(utils::cast<CNodePtr>(dy_after), utils::cast<CNodePtr>(dy_before)); manager->Replace(utils::cast<CNodePtr>(dy_after), utils::cast<CNodePtr>(dy_before));
auto output_num = AnfAlgo::GetOutputTensorNum(fbn2g);
for (size_t i = 0; i < output_num; i++) {
outputs_type.push_back(AnfAlgo::GetOutputInferDataType(fbn2g, i));
outputs_shape.push_back(AnfAlgo::GetOutputInferShape(fbn2g, i));
}
outputs_type[0] = kNumberTypeFloat16;
AnfAlgo::SetOutputInferTypeAndShape(outputs_type, outputs_shape, fbn2g.get());
return node; return node;
} }
} // namespace opt } // namespace opt

2
mindspore/ccsrc/backend/session/gpu_session.cc
View File

@ -28,7 +28,6 @@
#include "backend/optimizer/gpu/adam_fusion.h" #include "backend/optimizer/gpu/adam_fusion.h"
#include "backend/optimizer/gpu/replace_bn_cast_fusion.h" #include "backend/optimizer/gpu/replace_bn_cast_fusion.h"
#include "backend/optimizer/gpu/replace_bn_grad_cast_fusion.h" #include "backend/optimizer/gpu/replace_bn_grad_cast_fusion.h"
#include "backend/optimizer/gpu/replace_bn_grad_cast2_fusion.h"
#include "backend/optimizer/gpu/replace_momentum_cast_fusion.h" #include "backend/optimizer/gpu/replace_momentum_cast_fusion.h"
#include "backend/optimizer/gpu/replace_addn_fusion.h" #include "backend/optimizer/gpu/replace_addn_fusion.h"
#include "runtime/device/kernel_runtime_manager.h" #include "runtime/device/kernel_runtime_manager.h"
@ -68,7 +67,6 @@ void GPUSession::Optimize(const std::shared_ptr<KernelGraph> &kernel_graph) {
pm->AddPass(std::make_shared<opt::AdamFusion>()); pm->AddPass(std::make_shared<opt::AdamFusion>());
pm->AddPass(std::make_shared<opt::ReplaceBNCastFusion>()); pm->AddPass(std::make_shared<opt::ReplaceBNCastFusion>());
pm->AddPass(std::make_shared<opt::ReplaceBNGradCastFusion>()); pm->AddPass(std::make_shared<opt::ReplaceBNGradCastFusion>());
pm->AddPass(std::make_shared<opt::ReplaceBNGradCast2Fusion>());
pm->AddPass(std::make_shared<opt::ReplaceMomentumCastFusion>()); pm->AddPass(std::make_shared<opt::ReplaceMomentumCastFusion>());
pm->AddPass(std::make_shared<opt::ReplaceAddNFusion>()); pm->AddPass(std::make_shared<opt::ReplaceAddNFusion>());
optimizer->AddPassManager(pm); optimizer->AddPassManager(pm);