 Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago  Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago  Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago  Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago  Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago  Add GraphKernelPassManager to manage the passes of GraphKernel
Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"
All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
1. Basic features, like cluster, splitter, and some preprocess, postprocess.
2. All stable features, mainly includes the optimization passes.
3. Experimental features, like stitch-fusion, parallel-fusion.
The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g. "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"
Others:
1. the pass "tensor_promotion" is not useful, remove it.
2. put the pass "InsertPadOps" before "ArithmeticSimplify".
4 years ago |
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081 |
- /**
- * Copyright 2019 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_COMMON_PASS_MANAGER_H_
- #define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_COMMON_PASS_MANAGER_H_
-
- #include <utility>
- #include <vector>
- #include <string>
- #include <memory>
- #include <map>
-
- #include "backend/optimizer/common/pass.h"
- #include "backend/optimizer/common/node_pass.h"
-
- namespace mindspore {
- namespace opt {
- class CacheManager {
- public:
- CacheManager() {}
- ~CacheManager() = default;
- void Update(const AnfNodePtr &node);
- TypeId GetOutputType(const AnfNodePtr &node, size_t index);
- std::vector<size_t> GetOutputShape(const AnfNodePtr &node, size_t index);
-
- private:
- std::map<AnfNodePtr, std::map<size_t, TypeId>> type_map_;
- std::map<AnfNodePtr, std::map<size_t, std::vector<size_t>>> shape_map_;
- };
- using CacheManagerPtr = std::shared_ptr<CacheManager>;
-
- // @brief For optimization passes management
- class PassManager {
- public:
- explicit PassManager(const std::string &name = "pm", bool run_only_once = true)
- : name_(name), passes_{}, run_only_once_(run_only_once), cache_manager_(std::make_shared<CacheManager>()) {}
- virtual ~PassManager() = default;
- // Get all the passes added by AddPass
- const std::vector<PassPtr> &Passes() const;
- // Add graph pass, the pass object will be freed when pass manager freed.
- virtual void AddPass(const PassPtr &pass);
- // Run passes added in pass manager on the input graph
- // @param [in out] graph The graph to be optimized
- // @return true, graph changed
- // @return false, graph not changed
- virtual bool Run(const FuncGraphPtr &func_graph) const;
- // Run the given graph passes on the input graph
- // @param [in out] graph The graph to be optimized
- // @param [in] passes The given graph passes
- // @return true, graph changed
- // @return false, graph not changed
- virtual bool Run(const FuncGraphPtr &func_graph, const std::vector<PassPtr> &passes) const;
- std::string name() const { return name_; }
-
- protected:
- virtual bool RunPass(const FuncGraphPtr &func_graph, size_t pass_id, const PassPtr &pass) const;
- virtual std::string GetPassFullname(size_t pass_id, const PassPtr &pass) const;
- virtual void DumpPassIR(const FuncGraphPtr &func_graph, const std::string &pass_fullname) const;
-
- const std::string name_;
- std::vector<PassPtr> passes_;
- bool run_only_once_;
- CacheManagerPtr cache_manager_;
- };
- using PassManagerPtr = std::shared_ptr<PassManager>;
- } // namespace opt
- } // namespace mindspore
-
- #endif // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_COMMON_PASS_MANAGER_H_
|