add ReduceMin fission pass

This commit is contained in:
huanghui 2020-08-04 21:16:18 +08:00
parent cdc5131869
commit 30000fdb52
5 changed files with 310 additions and 22 deletions

View File

@ -16,7 +16,6 @@
#include "backend/optimizer/ascend/ascend_backend_optimization.h"
#include <memory>
#include <string>
#include <set>
#include "backend/optimizer/common/optimizer.h"
#include "backend/optimizer/ascend/ir_fission/bn_split.h"
#include "backend/optimizer/ascend/ir_fission/bn_grad_split.h"
@ -24,6 +23,7 @@
#include "backend/optimizer/ascend/ir_fission/batch_norm_bert_fission.h"
#include "backend/optimizer/ascend/ir_fission/single_batch_norm_fission.h"
#include "backend/optimizer/ascend/ir_fission/tensor_scatter_update_fission.h"
#include "backend/optimizer/ascend/ir_fission/reduce_min_fission.h"
#include "backend/optimizer/ascend/ir_fusion/fused_batch_norm_fusion.h"
#include "backend/optimizer/ascend/ir_fission/layer_norm_grad_split.h"
#include "backend/optimizer/pass/communication_op_fusion.h"
@ -111,18 +111,9 @@
namespace mindspore {
namespace opt {
namespace {
void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
void AddAscendIRFusionRulesPass(PassManager *ir_fusion_pm) {
MS_EXCEPTION_IF_NULL(ir_fusion_pm);
ir_fusion_pm->AddPass(std::make_shared<BatchNormBertFission>());
ir_fusion_pm->AddPass(std::make_shared<SingleBatchNormFission>());
ir_fusion_pm->AddPass(std::make_shared<SquareSumFusion>());
ir_fusion_pm->AddPass(std::make_shared<ClipByNormNoDivSquareSumFusion>());
ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLRRuleFusion>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusion>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV2>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV3>());
ir_fusion_pm->AddPass(std::make_shared<ConfusionMulGradFusion>());
ir_fusion_pm->AddPass(std::make_shared<ConfusionSoftmaxGradRule>());
ir_fusion_pm->AddPass(std::make_shared<LambNextMVWithDecayRuleCond1>());
ir_fusion_pm->AddPass(std::make_shared<LambNextMVWithDecayRuleCond2>());
ir_fusion_pm->AddPass(std::make_shared<LambNextMVWithDecayRuleCond3>());
@ -133,10 +124,6 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
ir_fusion_pm->AddPass(std::make_shared<LambNextMVRuleCond4>());
ir_fusion_pm->AddPass(std::make_shared<LambNextRightRule>());
ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLrV2>());
ir_fusion_pm->AddPass(std::make_shared<ReshapeTransposeFusion>());
ir_fusion_pm->AddPass(std::make_shared<TransposeReshapeFusion>());
ir_fusion_pm->AddPass(std::make_shared<ClipByValueFusion>());
ir_fusion_pm->AddPass(std::make_shared<TopKSplit>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond1Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond2Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond3Fusion>());
@ -146,6 +133,27 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneWithDecayRuleCond3>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneWithDecayRuleCond4>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneWithDecayRuleCond5>());
ir_fusion_pm->AddPass(std::make_shared<ClipByNormNoDivSquareSumFusion>());
ir_fusion_pm->AddPass(std::make_shared<SquareSumFusion>());
ir_fusion_pm->AddPass(std::make_shared<ClipByValueFusion>());
}
void AddAscendIRFusionPass(PassManager *ir_fusion_pm) {
MS_EXCEPTION_IF_NULL(ir_fusion_pm);
ir_fusion_pm->AddPass(std::make_shared<BatchNormBertFission>());
ir_fusion_pm->AddPass(std::make_shared<SingleBatchNormFission>());
ir_fusion_pm->AddPass(std::make_shared<BatchNorm2BNInfer>());
ir_fusion_pm->AddPass(std::make_shared<BatchNormGrad2BNInferGrad>());
ir_fusion_pm->AddPass(std::make_shared<BatchNormGradInferFission>());
ir_fusion_pm->AddPass(std::make_shared<GetitemTuple>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusion>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV2>());
ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV3>());
ir_fusion_pm->AddPass(std::make_shared<ConfusionMulGradFusion>());
ir_fusion_pm->AddPass(std::make_shared<ConfusionSoftmaxGradRule>());
ir_fusion_pm->AddPass(std::make_shared<ReshapeTransposeFusion>());
ir_fusion_pm->AddPass(std::make_shared<TransposeReshapeFusion>());
ir_fusion_pm->AddPass(std::make_shared<TopKSplit>());
ir_fusion_pm->AddPass(std::make_shared<MomentumLossscaleFusion>());
ir_fusion_pm->AddPass(std::make_shared<MulAddFusion>());
ir_fusion_pm->AddPass(std::make_shared<MulAddNFusion>());
@ -153,15 +161,12 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
ir_fusion_pm->AddPass(std::make_shared<AddnFission>());
ir_fusion_pm->AddPass(std::make_shared<DereluFusion>());
ir_fusion_pm->AddPass(std::make_shared<TransposeTransDataFusion>());
ir_fusion_pm->AddPass(std::make_shared<GetitemTuple>());
ir_fusion_pm->AddPass(std::make_shared<BatchNorm2BNInfer>());
ir_fusion_pm->AddPass(std::make_shared<BatchNormGrad2BNInferGrad>());
ir_fusion_pm->AddPass(std::make_shared<BatchNormGradInferFission>());
ir_fusion_pm->AddPass(std::make_shared<SplitFission>());
ir_fusion_pm->AddPass(std::make_shared<TensorScatterUpdateFission>());
ir_fusion_pm->AddPass(std::make_shared<GetitemTuple>());
ir_fusion_pm->AddPass(std::make_shared<PackFission>());
ir_fusion_pm->AddPass(std::make_shared<ConcatFission>());
ir_fusion_pm->AddPass(std::make_shared<ReduceMinFission>());
}
} // namespace
@ -265,9 +270,8 @@ void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGrap
ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion1>());
}
ir_fusion_pm->AddPass(std::make_shared<InsertPadForNMSWithMask>());
if (context_ptr->ir_fusion_flag()) {
AddAscendBackendOptionalIRFusion(ir_fusion_pm.get());
}
AddAscendIRFusionRulesPass(ir_fusion_pm.get());
AddAscendIRFusionPass(ir_fusion_pm.get());
if (context_ptr->enable_task_sink() && context_ptr->loop_sink_flag() && ConfigManager::GetInstance().iter_num() > 1) {
ir_fusion_pm->AddPass(std::make_shared<InsertMemcpyAsyncForGetNext>());

View File

@ -0,0 +1,144 @@
/**
* 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/ascend/ir_fission/reduce_min_fission.h"
#include <memory>
#include <vector>
#include "backend/session/anf_runtime_algorithm.h"
namespace mindspore {
namespace opt {
namespace {
CNodePtr CreateReduceMin(const FuncGraphPtr &graph, const AnfNodePtr &input, const CNodePtr &old_node) {
MS_EXCEPTION_IF_NULL(graph);
MS_EXCEPTION_IF_NULL(input);
MS_EXCEPTION_IF_NULL(old_node);
std::vector<AnfNodePtr> inputs = {NewValueNode(std::make_shared<Primitive>(prim::kPrimReduceMin->name())), input};
CNodePtr reduce_min = graph->NewCNode(inputs);
reduce_min->set_scope(old_node->scope());
AnfAlgo::CopyNodeAttr(kAttrKeepDims, old_node, reduce_min);
return reduce_min;
}
bool NeedOptmize(const TypeId &dtype, const std::vector<size_t> &shape, const std::vector<int> &axis) {
if (dtype != kNumberTypeFloat32) {
MS_LOG(INFO) << "ReduceMin's input Dtype is not float32, no need optimize!";
return false;
}
if (shape.size() == 0 || shape.size() == 1) {
MS_LOG(INFO) << "ReduceMin's input shape size is " << shape.size() << ", no need optimize!";
return false;
}
if (axis.size() == 1) {
MS_LOG(INFO) << "ReduceMin axis size is 1, no need optimize!";
return false;
}
int last_dim = SizeToInt(shape.size() - 1);
if (std::find(axis.begin(), axis.end(), -1) == axis.end() &&
std::find(axis.begin(), axis.end(), last_dim) == axis.end()) {
MS_LOG(INFO) << "Attribute of axis does not contain the last axis, not match!";
return false;
}
return true;
}
std::vector<int> CalFirstAxis(const std::vector<size_t> &shape, const std::vector<int> &axis) {
std::vector<int> axis_fisrt;
int last_dim = SizeToInt(shape.size() - 1);
std::copy_if(axis.begin(), axis.end(), std::back_inserter(axis_fisrt),
[&last_dim](int v) { return v != -1 && v != last_dim; });
int dim_size = SizeToInt(shape.size());
if (axis_fisrt.empty()) {
for (int i = 0; i < dim_size - 1; ++i) {
axis_fisrt.push_back(i);
}
}
for (size_t i = 0; i < axis_fisrt.size(); ++i) {
if (axis_fisrt[i] < -dim_size || axis_fisrt[i] > dim_size - 1) {
MS_LOG(EXCEPTION) << "The axis of ReduceMin verify failed, quit optimizing";
}
if (axis_fisrt[i] < 0) {
axis_fisrt[i] = dim_size + axis_fisrt[i];
}
}
return axis_fisrt;
}
std::vector<size_t> GetInferShape(const std::vector<size_t> &shape, const std::vector<int> &axis_first,
bool keep_dims) {
std::vector<size_t> shape_first;
for (size_t item = 0; item < shape.size(); ++item) {
if (axis_first.end() != std::find(axis_first.begin(), axis_first.end(), item)) {
if (keep_dims) {
// If keep_dims is true, curretn dimesion set to 1
shape_first.push_back(1);
}
} else {
// item is not in ConstValueAxis
shape_first.push_back(shape[item]);
}
}
return shape_first;
}
} // namespace
const BaseRef ReduceMinFission::DefinePattern() const {
VarPtr X = std::make_shared<Var>();
return VectorRef({prim::kPrimReduceMin, X});
}
const AnfNodePtr ReduceMinFission::Process(const FuncGraphPtr &graph, const AnfNodePtr &node, const EquivPtr &) const {
if (graph == nullptr || node == nullptr) {
return nullptr;
}
auto cnode = node->cast<CNodePtr>();
MS_EXCEPTION_IF_NULL(cnode);
CheckCNodeInputSize(cnode, 2);
auto shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, 0);
auto dtype = AnfAlgo::GetPrevNodeOutputInferDataType(cnode, 0);
if (!AnfAlgo::HasNodeAttr(kAttrAxis, cnode)) {
MS_LOG(INFO) << "ReduceMin has no axis, no need optimize!";
return nullptr;
}
auto axis = AnfAlgo::GetNodeAttr<std::vector<int>>(cnode, kAttrAxis);
if (!AnfAlgo::HasNodeAttr(kAttrKeepDims, cnode)) {
MS_LOG(INFO) << "ReduceMin has no keep_dims, no need optimize!";
return nullptr;
}
auto keep_dims = AnfAlgo::GetNodeAttr<bool>(cnode, kAttrKeepDims);
if (!NeedOptmize(dtype, shape, axis)) {
MS_LOG(INFO) << "No need optimize for this ReduceMin. " << cnode->DebugString();
return nullptr;
}
// Create reduce_min1
CNodePtr reduce_min1 = CreateReduceMin(graph, cnode->input(1), cnode);
std::vector<int> axis_fisrt = CalFirstAxis(shape, axis);
std::vector<size_t> shape_first = GetInferShape(shape, axis_fisrt, keep_dims);
AnfAlgo::SetOutputInferTypeAndShape({dtype}, {shape_first}, reduce_min1.get());
AnfAlgo::SetNodeAttr(kAttrAxis, MakeValue(axis_fisrt), reduce_min1);
// Create reduce_min2
CNodePtr reduce_min2 = CreateReduceMin(graph, reduce_min1, cnode);
reduce_min2->set_abstract(cnode->abstract());
std::vector<int> axis_last = {-1};
AnfAlgo::SetNodeAttr(kAttrAxis, MakeValue(axis_last), reduce_min2);
return reduce_min2;
}
} // namespace opt
} // namespace mindspore

View File

@ -0,0 +1,33 @@
/**
* 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_ASCEND_IR_FISSION_REDUCE_MIN_FISSION_H_
#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FISSION_REDUCE_MIN_FISSION_H_
#include "backend/optimizer/common/optimizer.h"
#include "backend/optimizer/common/helper.h"
namespace mindspore {
namespace opt {
class ReduceMinFission : public PatternProcessPass {
public:
explicit ReduceMinFission(bool multigraph = true) : PatternProcessPass("reduce_min_fission", multigraph) {}
~ReduceMinFission() 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_BACKEND_OPTIMIZER_ASCEND_IR_FISSION_REDUCE_MIN_FISSION_H_

View File

@ -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 "debug/anf_ir_dump.h"
#define private public
#define protected public
#include "backend/optimizer/ascend/ir_fission/reduce_min_fission.h"
#undef private
#undef protected
namespace mindspore {
namespace opt {
class TestHWOptReduceMinFission : public BackendCommon {
public:
TestHWOptReduceMinFission() : get_py_fun_("gtest_input.pre_activate.reduce_min_fission_test", true) {}
~TestHWOptReduceMinFission() override = default;
UT::PyFuncGraphFetcher get_py_fun_;
};
TEST_F(TestHWOptReduceMinFission, test_fission) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_reduce_min_fission", "before");
EXPECT_NE(g, nullptr);
std::vector<int> shp{32, 32, 32, 32};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
args_spec_list.push_back(x_abstract);
auto kg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
auto split_fission = std::make_shared<opt::ReduceMinFission>();
pm->AddPass(split_fission);
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(kg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_reduce_min_fission", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
} // namespace opt
} // namespace mindspore

View File

@ -0,0 +1,51 @@
# 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
from mindspore.ops import operations as P
make_tuple = Primitive('make_tuple')
tuple_getitem = Primitive('tuple_getitem')
reduce_min = P.ReduceMin(keep_dims=False)
reduce_min1 = Primitive('ReduceMin')
reduce_min2 = Primitive('ReduceMin')
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_reduce_min_fission(tag):
fns = FnDict()
@fns
def before(x):
res = reduce_min(x, (2, 3))
return res
@fns
def after(x):
res = reduce_min1(x)
res = reduce_min2(res)
return make_tuple(res)
return fns[tag]