diff --git a/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.cc b/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.cc new file mode 100644 index 00000000000..479e00e4c0e --- /dev/null +++ b/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.cc @@ -0,0 +1,91 @@ +/** + * 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 "pre_activate/ascend/ir_fission/lars_v2_fission.h" +#include +#include +#include "session/anf_runtime_algorithm.h" +#include "pre_activate/common/helper.h" +#include "utils/utils.h" + +namespace mindspore { +namespace opt { +namespace { +void CreateOutputsOfSquareSumAll(const FuncGraphPtr &graph, const CNodePtr &lars_v2, + std::vector *square_sum_all_outputs) { + MS_EXCEPTION_IF_NULL(graph); + MS_EXCEPTION_IF_NULL(lars_v2); + if (lars_v2->size() != kLarsV2InputNum) { + MS_LOG(EXCEPTION) << "Op lars_v2's input not equal " << kLarsV2InputNum; + } + + std::vector inputs = {NewValueNode(std::make_shared(kSquareSumAllOpName))}; + inputs.push_back(lars_v2->input(1)); + inputs.push_back(lars_v2->input(2)); + auto square_sum_all = graph->NewCNode(inputs); + MS_EXCEPTION_IF_NULL(square_sum_all); + square_sum_all->set_scope(lars_v2->scope()); + + auto types = {kNumberTypeFloat32, kNumberTypeFloat32}; + std::vector shape; + auto shapes = {shape, shape}; + AnfAlgo::SetOutputInferTypeAndShape(types, shapes, square_sum_all.get()); + + CreateMultipleOutputsOfAnfNode(graph, square_sum_all, 2, square_sum_all_outputs); +} + +CNodePtr CreateLarsV2Update(const FuncGraphPtr &graph, const CNodePtr &lars_v2, + const std::vector &square_sum_all_outputs) { + MS_EXCEPTION_IF_NULL(graph); + MS_EXCEPTION_IF_NULL(lars_v2); + if (square_sum_all_outputs.size() != 2) { + MS_LOG(EXCEPTION) << "square_sum_all_outputs' size not equal 2"; + } + if (lars_v2->size() != kLarsV2InputNum) { + MS_LOG(EXCEPTION) << "Op lars_v2's input not equal " << kLarsV2InputNum; + } + std::vector inputs = {NewValueNode(std::make_shared(kLarsV2UpdateOpName))}; + inputs.push_back(lars_v2->input(1)); + inputs.push_back(lars_v2->input(2)); + inputs.push_back(square_sum_all_outputs[0]); + inputs.push_back(square_sum_all_outputs[1]); + inputs.push_back(lars_v2->input(3)); + inputs.push_back(lars_v2->input(4)); + auto lars_v2_update = graph->NewCNode(inputs); + MS_EXCEPTION_IF_NULL(lars_v2_update); + lars_v2_update->set_scope(lars_v2->scope()); + lars_v2_update->set_abstract(lars_v2->abstract()); + return lars_v2_update; +} +} // namespace + +const BaseRef LarsV2Fission::DefinePattern() const { + VarPtr Xs = std::make_shared(); + auto lars_v2_prim = std::make_shared(kLarsV2OpName); + return VectorRef({lars_v2_prim, Xs}); +} + +const AnfNodePtr LarsV2Fission::Process(const FuncGraphPtr &graph, const AnfNodePtr &node, const EquivPtr &) const { + MS_EXCEPTION_IF_NULL(graph); + MS_EXCEPTION_IF_NULL(node); + auto lars_v2 = node->cast(); + MS_EXCEPTION_IF_NULL(lars_v2); + + std::vector square_sum_all_outputs; + CreateOutputsOfSquareSumAll(graph, lars_v2, &square_sum_all_outputs); + return CreateLarsV2Update(graph, lars_v2, square_sum_all_outputs); +} +} // namespace opt +} // namespace mindspore diff --git a/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.h b/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.h new file mode 100644 index 00000000000..846d221c538 --- /dev/null +++ b/mindspore/ccsrc/pre_activate/ascend/ir_fission/lars_v2_fission.h @@ -0,0 +1,32 @@ +/** + * 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_PRE_ACTIVATE_ASCEND_IR_FISSION_LARS_V2_FISSION_H_ +#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FISSION_LARS_V2_FISSION_H_ + +#include "pre_activate/common/optimizer.h" + +namespace mindspore { +namespace opt { +class LarsV2Fission : public PatternProcessPass { + public: + explicit LarsV2Fission(bool multigraph = true) : PatternProcessPass("lars_v2_fission", multigraph) {} + ~LarsV2Fission() override = default; + const BaseRef DefinePattern() const override; + const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override; +}; +} // namespace opt +} // namespace mindspore +#endif // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FISSION_LARS_V2_FISSION_H_ diff --git a/mindspore/ccsrc/pre_activate/common/helper.h b/mindspore/ccsrc/pre_activate/common/helper.h index f244baa4a19..78c7e11eca2 100644 --- a/mindspore/ccsrc/pre_activate/common/helper.h +++ b/mindspore/ccsrc/pre_activate/common/helper.h @@ -91,6 +91,7 @@ constexpr size_t kBackendTransDataInputNum = 2; constexpr size_t kApplyMomentumInputNum = 6; constexpr size_t kBiasAddInputNum = 3; constexpr size_t kTopkInputNum = 3; +constexpr size_t kLarsV2InputNum = 5; enum FusedBatchNormInput { kX = 1, diff --git a/mindspore/ccsrc/utils/utils.h b/mindspore/ccsrc/utils/utils.h index 6801a6fe324..36f476473f6 100644 --- a/mindspore/ccsrc/utils/utils.h +++ b/mindspore/ccsrc/utils/utils.h @@ -144,6 +144,9 @@ constexpr auto kBNInferGradOpName = "BNInferGrad"; constexpr auto kCallOpName = "call"; constexpr auto kPartialOpName = "partial"; constexpr auto kSwitchOpName = "switch"; +constexpr auto kLarsV2OpName = "LarsV2"; +constexpr auto kLarsV2UpdateOpName = "LarsV2Update"; +constexpr auto kSquareSumAllOpName = "SquareSumAll"; // attr key name constexpr auto kAttrInputNames = "input_names"; diff --git a/tests/ut/cpp/pre_activate/ascend/ir_fission/lars_v2_fission_test.cc b/tests/ut/cpp/pre_activate/ascend/ir_fission/lars_v2_fission_test.cc new file mode 100644 index 00000000000..c0a0cc455eb --- /dev/null +++ b/tests/ut/cpp/pre_activate/ascend/ir_fission/lars_v2_fission_test.cc @@ -0,0 +1,56 @@ +/** + * 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 "common/backend_common_test.h" +#include "common/py_func_graph_fetcher.h" +#include "pre_activate/ascend/ir_fission/lars_v2_fission.h" + +namespace mindspore { +namespace opt { +class TestHWLarsV2Fission : public BackendCommon { + public: + TestHWLarsV2Fission() : get_py_fun_("gtest_input.pre_activate.lars_v2_fission_test", true) {} + ~TestHWLarsV2Fission() override = default; + + UT::PyFuncGraphFetcher get_py_fun_; +}; + +TEST_F(TestHWLarsV2Fission, test_fission) { + FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_lars_v2_fission", "before"); + EXPECT_NE(g, nullptr); + + // set abstract for all nodes in g + std::vector shp{2, 32, 224, 224}; + auto x_abstract = std::make_shared(kFloat32, shp); + g->get_return()->input(1)->set_abstract(x_abstract); + for (auto &p: g->parameters()){ + p->set_abstract(x_abstract); + } + AbstractBasePtrList args_spec_list; + auto kg = GetKernelGraph(g, args_spec_list, false); + + auto optimizer = std::make_shared(); + auto pm = std::make_shared(); + pm->AddPass(std::make_shared()); + optimizer->AddPassManager(pm); + FuncGraphPtr new_graph = optimizer->Optimize(kg); + + FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_lars_v2_fission", "after"); + EXPECT_NE(g_after, nullptr); + EXPECT_TRUE(CheckEqualGraph(g_after, new_graph)); +} +} // namespace opt +} // namespace mindspore diff --git a/tests/ut/cpp/python_input/gtest_input/pre_activate/lars_v2_fission_test.py b/tests/ut/cpp/python_input/gtest_input/pre_activate/lars_v2_fission_test.py new file mode 100644 index 00000000000..75fcb5be531 --- /dev/null +++ b/tests/ut/cpp/python_input/gtest_input/pre_activate/lars_v2_fission_test.py @@ -0,0 +1,50 @@ +# 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. +# ============================================================================ + +from mindspore.ops import Primitive + +lars_v2 = Primitive('LarsV2') +square_sum_all = Primitive('SquareSumAll') +lars_v2_update = Primitive('LarsV2Update') +make_tuple = Primitive('make_tuple') +tuple_getitem = Primitive('tuple_getitem') + +class FnDict: + def __init__(self): + self.fnDict = {} + + def __call__(self, fn): + self.fnDict[fn.__name__] = fn + + def __getitem__(self, name): + return self.fnDict[name] + +def test_lars_v2_fission(tag): + fns = FnDict() + + @fns + def before(input0, input1, input2, input3): + res = lars_v2(input0, input1, input2, input3) + return res + + @fns + def after(input0, input1, input2, input3): + res = square_sum_all(input0, input1) + item0 = tuple_getitem(res, 0) + item1 = tuple_getitem(res, 1) + res = lars_v2_update(input0, input1, item0, item1, input2, input3) + return make_tuple(res) + + return fns[tag]