!2384 Add split fission pass

Merge pull request !2384 from YuJianfeng/split

mindspore/ccsrc/kernel/tbe/tbe_adapter.cc

@@ -90,6 +90,7 @@ static std::map<string, string> tbe_func_adapter_map = {
   {"lamb_next_mv_with_decay", "lamb_next_m_v_with_decay"},
   {"lamb_next_mv", "lamb_next_m_v"},
   {"split", "split_d"},
+  {"split_v", "split_v_d"},
   {"resize_nearest_neighbor", "resize_nearest_neighbor_v2_d"},
   {"resize_nearest_neighbor_grad", "resize_nearest_neighbor_v2_grad_d"},
   {"pad", "pad_d"},

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -87,6 +87,7 @@
 #include "pre_activate/ascend/ir_fission/addn_fission.h"
 #include "pre_activate/ascend/enhancer/insert_memcpy_async_for_getnext.h"
 #include "pre_activate/ascend/ir_fission/batch_norm_grad_infer_fission.h"
+#include "pre_activate/ascend/ir_fission/split_fission.h"
 #include "utils/context/ms_context.h"
 #include "utils/config_manager.h"
 #include "debug/anf_ir_dump.h"
@@ -141,6 +142,8 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
   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<GetitemTuple>());
 }
 }  // namespace
 

mindspore/ccsrc/pre_activate/ascend/ir_fission/split_fission.cc

@@ -0,0 +1,191 @@
+/**
+ * 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/split_fission.h"
+#include <memory>
+#include <vector>
+#include "session/anf_runtime_algorithm.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+CNodePtr CreateSplitVNode(const FuncGraphPtr &func_graph, const AnfNodePtr &input_node) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  MS_EXCEPTION_IF_NULL(input_node);
+  std::vector<AnfNodePtr> splitv_inputs{NewValueNode(std::make_shared<Primitive>(kSplitVOpName)), input_node};
+  CNodePtr splitv = func_graph->NewCNode(splitv_inputs);
+  MS_EXCEPTION_IF_NULL(splitv);
+  splitv->set_scope(input_node->scope());
+  return splitv;
+}
+
+CNodePtr CreateBaseSplitVNode(const FuncGraphPtr &func_graph, const CNodePtr &origin_cnode) {
+  MS_EXCEPTION_IF_NULL(origin_cnode);
+  if (origin_cnode->inputs().size() < kSplitInputNum) {
+    MS_LOG(EXCEPTION) << "The input number of split: " << origin_cnode->DebugString() << " should be "
+                      << kSplitInputNum - 1;
+  }
+  return CreateSplitVNode(func_graph, origin_cnode->input(1));
+}
+
+void SetAttrForSplitVNode(const AnfNodePtr &splitv, const std::vector<int> &size_splits, int split_dim, int num_split) {
+  AnfAlgo::SetNodeAttr(kAttrSizeSplits, MakeValue(size_splits), splitv);
+  AnfAlgo::SetNodeAttr(kAttrSplitDim, MakeValue(split_dim), splitv);
+  AnfAlgo::SetNodeAttr(kAttrNumSplit, MakeValue(num_split), splitv);
+}
+
+size_t GetSmallSplitSize(const AnfNodePtr &split_node, int split_dim, int num_split) {
+  auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(split_node, 0);
+  if (split_dim < 0) {
+    split_dim += input_shape.size();
+  }
+  if (IntToSize(split_dim) >= input_shape.size()) {
+    MS_LOG(EXCEPTION) << "The split_dim value should be less than the shape size of input 0";
+  }
+  return input_shape[split_dim] / num_split;
+}
+
+void AddNewOutputs(const FuncGraphPtr &func_graph, const AnfNodePtr &new_splitv, int outputs_num,
+                   std::vector<AnfNodePtr> *inputs) {
+  MS_EXCEPTION_IF_NULL(inputs);
+  std::vector<AnfNodePtr> new_splitv_output;
+  CreateMultipleOutputsOfAnfNode(func_graph, new_splitv, outputs_num, &new_splitv_output);
+  inputs->insert(inputs->end(), new_splitv_output.begin(), new_splitv_output.end());
+}
+
+AnfNodePtr CreateTupleGetItem(const FuncGraphPtr &func_graph, const AnfNodePtr &input, size_t index) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  auto idx = NewValueNode(SizeToInt(index));
+  MS_EXCEPTION_IF_NULL(idx);
+  auto imm = std::make_shared<Int32Imm>(SizeToInt(index));
+  auto abstract_scalar = std::make_shared<abstract::AbstractScalar>(imm);
+  idx->set_abstract(abstract_scalar);
+  auto tuple_getitem = func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), input, idx});
+  return tuple_getitem;
+}
+
+void CreateOutputShapeAndTypeId(const CNodePtr &origin_cnode, int split_dim, int split_size, int num_split,
+                                std::vector<TypeId> *new_type_ids,
+                                std::vector<std::vector<size_t>> *new_output_shapes) {
+  MS_EXCEPTION_IF_NULL(new_type_ids);
+  MS_EXCEPTION_IF_NULL(new_output_shapes);
+  auto output_shape = AnfAlgo::GetOutputInferShape(origin_cnode, 0);
+  output_shape[split_dim] = split_size;
+  TypeId type_id = AnfAlgo::GetOutputInferDataType(origin_cnode, 0);
+  for (int i = 0; i < num_split; ++i) {
+    new_type_ids->emplace_back(type_id);
+    new_output_shapes->emplace_back(output_shape);
+  }
+}
+
+void SetAttrAndAbstractForBaseSplitv(const CNodePtr &origin_cnode, const CNodePtr &base_splitv,
+                                     const std::vector<int> &size_splits_base, int split_dim, int num_split) {
+  SetAttrForSplitVNode(base_splitv, size_splits_base, split_dim, num_split);
+  std::vector<TypeId> base_type_ids;
+  std::vector<std::vector<size_t>> base_output_shapes_base;
+  auto output_shape = AnfAlgo::GetOutputInferShape(origin_cnode, 0);
+  TypeId type_id = AnfAlgo::GetOutputInferDataType(origin_cnode, 0);
+  for (int i = 0; i < num_split; ++i) {
+    output_shape[split_dim] = size_splits_base[i];
+    base_output_shapes_base.emplace_back(output_shape);
+    base_type_ids.emplace_back(type_id);
+  }
+  AnfAlgo::SetOutputInferTypeAndShape(base_type_ids, base_output_shapes_base, base_splitv.get());
+}
+
+AnfNodePtr DoFission(const FuncGraphPtr &func_graph, const CNodePtr &cnode, int num_split, int divisor) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  auto split_dim = AnfAlgo::GetNodeAttr<int>(cnode, kAttrAxis);
+  CNodePtr base_splitv = CreateBaseSplitVNode(func_graph, cnode);
+
+  // Create new size_splits for "size_splits" attr of each new Splitv node which has full inputs.
+  auto small_split_size = SizeToInt(GetSmallSplitSize(cnode, split_dim, num_split));
+  std::vector<int> size_splits_new;
+  for (int i = 0; i < divisor; ++i) {
+    size_splits_new.emplace_back(small_split_size);
+  }
+  // Create new output shape and new output type id for each new Splitv node which has full inputs.
+  std::vector<TypeId> new_type_ids;
+  std::vector<std::vector<size_t>> new_output_shapes;
+  CreateOutputShapeAndTypeId(cnode, split_dim, small_split_size, divisor, &new_type_ids, &new_output_shapes);
+
+  // Create make_tuple input to create a make_tuple for replacing the old Split node.
+  std::vector<AnfNodePtr> make_tuple_inputs{NewValueNode(prim::kPrimMakeTuple)};
+  // Start to divide the outputs of Split.
+  std::vector<int> size_splits_base;
+  const auto base_split_size = divisor * small_split_size;
+  int nodes_num = 0;
+  int cur_output_index = 0;
+  while (num_split - cur_output_index > divisor) {
+    CNodePtr new_splitv = CreateSplitVNode(func_graph, CreateTupleGetItem(func_graph, base_splitv, nodes_num));
+    SetAttrForSplitVNode(new_splitv, size_splits_new, split_dim, divisor);
+    AnfAlgo::SetOutputInferTypeAndShape(new_type_ids, new_output_shapes, new_splitv.get());
+    AddNewOutputs(func_graph, new_splitv, divisor, &make_tuple_inputs);
+    cur_output_index += divisor;
+    size_splits_base.emplace_back(base_split_size);
+    nodes_num++;
+  }
+  if (cur_output_index < num_split) {
+    auto last_node_num_split = num_split - cur_output_index;
+    if (last_node_num_split > 1) {
+      CNodePtr new_splitv = CreateSplitVNode(func_graph, CreateTupleGetItem(func_graph, base_splitv, nodes_num));
+      std::vector<int> size_splits_new_last;
+      for (int i = 0; i < last_node_num_split; ++i) {
+        size_splits_new_last.emplace_back(small_split_size);
+      }
+      SetAttrForSplitVNode(new_splitv, size_splits_new_last, split_dim, last_node_num_split);
+      // Create new output shape and new output type id for the last Splitv node
+      std::vector<TypeId> last_new_type_ids;
+      std::vector<std::vector<size_t>> last_new_output_shapes;
+      CreateOutputShapeAndTypeId(cnode, split_dim, small_split_size, last_node_num_split, &last_new_type_ids,
+                                 &last_new_output_shapes);
+      AnfAlgo::SetOutputInferTypeAndShape(last_new_type_ids, last_new_output_shapes, new_splitv.get());
+      AddNewOutputs(func_graph, new_splitv, last_node_num_split, &make_tuple_inputs);
+      size_splits_base.emplace_back(last_node_num_split * small_split_size);
+    } else {
+      make_tuple_inputs.emplace_back(CreateTupleGetItem(func_graph, base_splitv, nodes_num));
+      size_splits_base.emplace_back(small_split_size);
+    }
+    nodes_num++;
+  }
+  // Set Attr and abstract for the base splitv
+  SetAttrAndAbstractForBaseSplitv(cnode, base_splitv, size_splits_base, split_dim, nodes_num);
+  AnfNodePtr make_tuple = func_graph->NewCNode(make_tuple_inputs);
+  return make_tuple;
+}
+}  // namespace
+
+const BaseRef SplitFission::DefinePattern() const {
+  VarPtr Xs = std::make_shared<SeqVar>();
+  auto split_prim = std::make_shared<Primitive>(kSplitOpName);
+  return VectorRef({split_prim, Xs});
+}
+
+const AnfNodePtr SplitFission::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node, const EquivPtr &) const {
+  MS_EXCEPTION_IF_NULL(node);
+  auto cnode = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  // Check output num
+  if (!AnfAlgo::HasNodeAttr(kAttrOutputNum, cnode)) {
+    return nullptr;
+  }
+  auto num_split = AnfAlgo::GetNodeAttr<int>(cnode, kAttrOutputNum);
+  if (num_split <= outputs_divisor_) {
+    return nullptr;
+  }
+  return DoFission(func_graph, cnode, num_split, outputs_divisor_);
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/ir_fission/split_fission.h

@@ -0,0 +1,37 @@
+/**
+ * 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_SPLIT_FISSION_H_
+#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FISSION_SPLIT_FISSION_H_
+
+#include "pre_activate/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+constexpr int kSplitOutputsDivisor = 63;
+class SplitFission : public PatternProcessPass {
+ public:
+  explicit SplitFission(bool multigraph = true)
+      : PatternProcessPass("split_fission", multigraph), outputs_divisor_(kSplitOutputsDivisor) {}
+  ~SplitFission() override = default;
+  const BaseRef DefinePattern() const override;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+
+ private:
+  int outputs_divisor_;
+};
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FISSION_SPLIT_FISSION_H_

mindspore/ccsrc/pre_activate/common/helper.h

@@ -97,6 +97,7 @@ constexpr size_t kBiasAddInputNum = 3;
 constexpr size_t kTopkInputNum = 3;
 constexpr size_t kLarsV2InputNum = 5;
 constexpr size_t kFusedMulApplyMomentumOutputNum = 2;
+constexpr size_t kSplitInputNum = 2;
 
 enum FusedBatchNormInput {
   kX = 1,

mindspore/ccsrc/utils/utils.h

@@ -72,6 +72,7 @@ constexpr auto kUnsortedSegmentMinOpName = "UnsortedSegmentMin";
 constexpr auto kFlattenGradOpName = "FlattenGrad";
 constexpr auto kExpandDimsOpName = "ExpandDims";
 constexpr auto kSplitOpName = "Split";
+constexpr auto kSplitVOpName = "SplitV";
 constexpr auto kSparseApplyAdagradOpName = "SparseApplyAdagrad";
 constexpr auto kMomentumOpName = "Momentum";
 constexpr auto kApplyMomentumOpName = "ApplyMomentum";
@@ -211,6 +212,10 @@ constexpr auto kAttrWaitEvent = "wait_event";
 constexpr auto kAttrRecordEventStream = "record_event_stream";
 constexpr auto kAttrWaitEventStream = "wait_event_stream";
 constexpr auto kAttrIndex = "index";
+constexpr auto kAttrSplitDim = "split_dim";
+constexpr auto kAttrNumSplit = "num_split";
+constexpr auto kAttrOutputNum = "output_num";
+constexpr auto kAttrSizeSplits = "size_splits";
 
 // attr value
 constexpr auto kValueTargetSwitch = "target_switch";

tests/ut/cpp/pre_activate/ascend/ir_fission/split_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"
+#define private public
+#define protected public
+#include "pre_activate/ascend/ir_fission/split_fission.h"
+#undef private
+#undef protected
+
+namespace mindspore {
+namespace opt {
+class TestHWSplitFission : public BackendCommon {
+ public:
+  TestHWSplitFission() : get_py_fun_("gtest_input.pre_activate.split_fission_test", true) {}
+  ~TestHWSplitFission() override = default;
+
+  UT::PyFuncGraphFetcher get_py_fun_;
+};
+
+TEST_F(TestHWSplitFission, test_split_fission_divided_by_3) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_split_fission", "before");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp{512, 3, 1};
+  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::SplitFission>();
+  split_fission->outputs_divisor_ = 3;
+  pm->AddPass(split_fission);
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(kg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_split_fission", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
+}  // namespace opt
+}  // namespace mindspore

tests/ut/cpp/python_input/gtest_input/pre_activate/split_fission_test.py

@@ -0,0 +1,58 @@
+# 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
+
+split = P.Split(0, 8)
+make_tuple = Primitive('make_tuple')
+tuple_getitem = Primitive('tuple_getitem')
+splitv = Primitive('SplitV')
+
+
+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_split_fission(tag):
+    """ test_adam_apply_one_with_decay_rule """
+    fns = FnDict()
+
+    @fns
+    def before(x):
+        return split(x)
+
+    @fns
+    def after(x):
+        splitv0 = splitv(x)
+        splitv1 = splitv(tuple_getitem(splitv0, 0))
+        splitv2 = splitv(tuple_getitem(splitv0, 1))
+        splitv3 = splitv(tuple_getitem(splitv0, 2))
+        make_tuple0 = make_tuple(tuple_getitem(splitv1, 0), tuple_getitem(splitv1, 1), tuple_getitem(splitv1, 2),
+                                 tuple_getitem(splitv2, 0), tuple_getitem(splitv2, 1), tuple_getitem(splitv2, 2),
+                                 tuple_getitem(splitv3, 0), tuple_getitem(splitv3, 1))
+        return make_tuple(
+            make_tuple(tuple_getitem(make_tuple0, 0), tuple_getitem(make_tuple0, 1), tuple_getitem(make_tuple0, 2),
+                       tuple_getitem(make_tuple0, 3), tuple_getitem(make_tuple0, 4), tuple_getitem(make_tuple0, 5),
+                       tuple_getitem(make_tuple0, 6), tuple_getitem(make_tuple0, 7)))
+
+    return fns[tag]