!2557 Add UnsortedSegmentSum fission pass

Merge pull request !2557 from huanghui/unsorted-segment-sum-fission-pass

mindspore/ccsrc/backend/optimizer/ascend/ascend_backend_optimization.cc

@@ -26,6 +26,7 @@
 #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/ascend/ir_fission/unsorted_segment_sum_fission.h"
 #include "backend/optimizer/pass/communication_op_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/square_sum_fusion.h"
 #include "backend/optimizer/ascend/ir_fusion/clip_by_norm_no_div_square_sum_fusion.h"
@@ -172,6 +173,7 @@ void AddAscendIRFusionPass(PassManager *ir_fusion_pm) {
   ir_fusion_pm->AddPass(std::make_shared<PackFission>());
   ir_fusion_pm->AddPass(std::make_shared<ConcatFission>());
   ir_fusion_pm->AddPass(std::make_shared<ReduceMinFission>());
+  ir_fusion_pm->AddPass(std::make_shared<UnsortSegmentSumFission>());
 }
 }  // namespace
 

mindspore/ccsrc/backend/optimizer/ascend/ir_fission/unsorted_segment_sum_fission.cc

@@ -0,0 +1,118 @@
+/**
+ * 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/unsorted_segment_sum_fission.h"
+#include <memory>
+#include <vector>
+#include "backend/session/anf_runtime_algorithm.h"
+#include "ir/primitive.h"
+#include "utils/utils.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+CNodePtr CreatePadding(const FuncGraphPtr &graph, const CNodePtr &origin_node, const size_t &pad_dim_size) {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(origin_node);
+  std::vector<AnfNodePtr> padding_inputs = {NewValueNode(std::make_shared<Primitive>(kPaddingOpName)),
+                                            origin_node->input(1)};
+  auto padding = graph->NewCNode(padding_inputs);
+  MS_EXCEPTION_IF_NULL(padding);
+  padding->set_scope(origin_node->scope());
+  auto shape = AnfAlgo::GetPrevNodeOutputInferShape(origin_node, 0);
+  shape[shape.size() - 1] = pad_dim_size;
+  AnfAlgo::SetOutputInferTypeAndShape({AnfAlgo::GetPrevNodeOutputInferDataType(origin_node, 0)}, {shape},
+                                      padding.get());
+  AnfAlgo::SetNodeAttr(kAttrPadDimSize, MakeValue(SizeToInt(pad_dim_size)), padding);
+  return padding;
+}
+
+CNodePtr CreateUnsortedSegmentSum(const FuncGraphPtr &graph, const CNodePtr &origin_node, const CNodePtr &padding,
+                                  const size_t &pad_dim_size) {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(origin_node);
+  MS_EXCEPTION_IF_NULL(padding);
+  std::vector<AnfNodePtr> unsorted_segment_sum8_inputs = {
+    NewValueNode(std::make_shared<Primitive>(prim::kPrimUnsortedSegmentSum->name())), padding, origin_node->input(2)};
+  auto unsorted_segment_sum = graph->NewCNode(unsorted_segment_sum8_inputs);
+  MS_EXCEPTION_IF_NULL(unsorted_segment_sum);
+  unsorted_segment_sum->set_scope(origin_node->scope());
+  auto shape = AnfAlgo::GetOutputInferShape(origin_node, 0);
+  shape[shape.size() - 1] = pad_dim_size;
+  AnfAlgo::SetOutputInferTypeAndShape({AnfAlgo::GetOutputInferDataType(origin_node, 0)}, {shape},
+                                      unsorted_segment_sum.get());
+  AnfAlgo::SetNodeAttr(kAttrNumSegments, MakeValue(SizeToInt(shape[0])), unsorted_segment_sum);
+  return unsorted_segment_sum;
+}
+
+CNodePtr CreateSlice(const FuncGraphPtr &graph, const CNodePtr &unsort_segment_sum,
+                     const CNodePtr &unsorted_segment_sum8) {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(unsort_segment_sum);
+  MS_EXCEPTION_IF_NULL(unsorted_segment_sum8);
+  std::vector<AnfNodePtr> slice_inputs = {NewValueNode(std::make_shared<Primitive>(kSliceOpName)),
+                                          unsorted_segment_sum8};
+  auto slice = graph->NewCNode(slice_inputs);
+  MS_EXCEPTION_IF_NULL(slice);
+  slice->set_scope(unsort_segment_sum->scope());
+  slice->set_abstract(unsort_segment_sum->abstract());
+  auto unsort_segment_sum_shape = AnfAlgo::GetOutputInferShape(unsort_segment_sum, 0);
+  std::vector<size_t> offsets(unsort_segment_sum_shape.size(), 0);
+  AnfAlgo::SetNodeAttr(kAttrBegin, MakeValue(Convert2Int(offsets)), slice);
+  AnfAlgo::SetNodeAttr(kAttrSize, MakeValue(Convert2Int(unsort_segment_sum_shape)), slice);
+  return slice;
+}
+}  // namespace
+
+const BaseRef UnsortSegmentSumFission::DefinePattern() const {
+  VarPtr Xs = std::make_shared<SeqVar>();
+  VectorRef pattern({prim::kPrimUnsortedSegmentSum, Xs});
+  return pattern;
+}
+
+const AnfNodePtr UnsortSegmentSumFission::Process(const FuncGraphPtr &graph, const AnfNodePtr &node,
+                                                  const EquivPtr &) const {
+  MS_EXCEPTION_IF_NULL(graph);
+  MS_EXCEPTION_IF_NULL(node);
+  auto origin_node = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(origin_node);
+  if (origin_node->size() != kUnsortedSegmentSumInputNum + 1) {
+    MS_LOG(INFO) << "UnsortedSegmentSum has wrong inputs num, not equal " << kUnsortedSegmentSumInputNum
+                 << ". CNode= " << origin_node->DebugString();
+    return nullptr;
+  }
+  auto input0_shape = AnfAlgo::GetPrevNodeOutputInferShape(origin_node, 0);
+  if (input0_shape[input0_shape.size() - 1] != 1) {
+    MS_LOG(INFO) << "UnsortedSegmentSum is not need fission. The last value of input0's shape is "
+                 << input0_shape[input0_shape.size() - 1];
+    return nullptr;
+  }
+  size_t pad_dim_size;
+  auto input_dtype = AnfAlgo::GetPrevNodeOutputInferDataType(origin_node, 0);
+  if (input_dtype == kNumberTypeFloat32) {
+    pad_dim_size = 8;
+  } else if (input_dtype == kNumberTypeFloat16) {
+    pad_dim_size = 16;
+  } else {
+    MS_LOG(INFO) << "UnsortedSegmentSum data type not in (float21, float16), no need change";
+    return nullptr;
+  }
+
+  auto padding = CreatePadding(graph, origin_node, pad_dim_size);
+  auto unsorted_segment_sum8 = CreateUnsortedSegmentSum(graph, origin_node, padding, pad_dim_size);
+  return CreateSlice(graph, origin_node, unsorted_segment_sum8);
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/backend/optimizer/ascend/ir_fission/unsorted_segment_sum_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_BACKEND_OPTIMIZER_ASCEND_IR_FISSION_UNSORTED_SEGMENT_SUM_FISSION_H_
+#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FISSION_UNSORTED_SEGMENT_SUM_FISSION_H_
+
+#include <vector>
+#include <memory>
+#include "backend/optimizer/common/optimizer.h"
+#include "backend/optimizer/common/helper.h"
+#include "backend/optimizer/ascend/ascend_helper.h"
+
+namespace mindspore {
+namespace opt {
+class UnsortSegmentSumFission : public PatternProcessPass {
+ public:
+  explicit UnsortSegmentSumFission(bool multigraph = true)
+      : PatternProcessPass("unsorted_segment_sum_fission", multigraph) {}
+  ~UnsortSegmentSumFission() 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_UNSORTED_SEGMENT_SUM_FISSION_H_

mindspore/ccsrc/backend/optimizer/common/helper.h

@@ -98,6 +98,7 @@ constexpr size_t kTopkInputNum = 3;
 constexpr size_t kLarsV2InputNum = 5;
 constexpr size_t kFusedMulApplyMomentumOutputNum = 2;
 constexpr size_t kSplitInputNum = 2;
+constexpr size_t kUnsortedSegmentSumInputNum = 2;
 
 enum FusedBatchNormInput {
   kX = 1,

mindspore/ccsrc/utils/utils.h

@@ -182,6 +182,7 @@ constexpr auto kPushOpName = "Push";
 constexpr auto kPullOpName = "Pull";
 constexpr auto kEmbeddingLookupOpName = "EmbeddingLookup";
 constexpr auto kEmbeddingLookupProxyOpName = "EmbeddingLookupProxy";
+constexpr auto kPaddingOpName = "Padding";
 
 // attr key name
 constexpr auto kAttrInputNames = "input_names";
@@ -253,6 +254,10 @@ constexpr auto kAttrInputNums = "inputNums";
 constexpr auto kAttrT = "T";
 constexpr auto kAttrNum = "num";
 constexpr auto kAttrRankSize = "rank_size";
+constexpr auto kAttrPadDimSize = "pad_dim_size";
+constexpr auto kAttrNumSegments = "num_segments";
+constexpr auto kAttrBegin = "begin";
+constexpr auto kAttrSize = "size";
 
 // attr value
 constexpr auto kValueTargetSwitch = "target_switch";

tests/st/fusion/test_unsorted_segment_sum_fission.py

@@ -0,0 +1,47 @@
+# 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.
+# ============================================================================
+import numpy as np
+
+import mindspore
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.ops import operations as P
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+context.set_context(save_graphs=True)
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.unsorted_segment_sum = P.UnsortedSegmentSum()
+        self.num_segments = 3
+
+    def construct(self, x, segment_ids):
+        x = self.unsorted_segment_sum(x, segment_ids, self.num_segments)
+        return x
+
+
+def test_net():
+    input_x = np.random.randn(3, 39, 1).astype(np.float32)
+    segment_ids = Tensor([0, 1, 2], mindspore.int32)
+    net = Net()
+    output = net(Tensor(input_x), segment_ids)
+    print("result", output.asnumpy())
+
+
+if __name__ == "__main__":
+    test_net()

tests/ut/cpp/pre_activate/ascend/ir_fission/unsorted_segment_sum_fission_test.cc

@@ -0,0 +1,68 @@
+/**
+ * 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/unsorted_segment_sum_fission.h"
+#include "common/backend_common_test.h"
+#include "common/py_func_graph_fetcher.h"
+#include "debug/anf_ir_dump.h"
+
+namespace mindspore {
+namespace opt {
+class TestHWUnsortedSegmentSumFission : public BackendCommon {
+ public:
+  TestHWUnsortedSegmentSumFission() : get_py_fun_("gtest_input.pre_activate.unsorted_segment_sum_fission", true) {}
+  ~TestHWUnsortedSegmentSumFission() override = default;
+
+  UT::PyFuncGraphFetcher get_py_fun_;
+};
+
+TEST_F(TestHWUnsortedSegmentSumFission, test_fission) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_unsorted_segment_sum_fission", "before1");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp_x{16, 1};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
+  AbstractBasePtrList args_spec_list{x_abstract, x_abstract};
+  auto kg = GetKernelGraph(g, args_spec_list);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::UnsortSegmentSumFission>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(kg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_unsorted_segment_sum_fission", "after1");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
+
+TEST_F(TestHWUnsortedSegmentSumFission, test_no_fission) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_unsorted_segment_sum_fission", "before2");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp_x{16, 2};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp_x);
+  AbstractBasePtrList args_spec_list{x_abstract, x_abstract};
+  auto kg = GetKernelGraph(g, args_spec_list);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::UnsortSegmentSumFission>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(kg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_unsorted_segment_sum_fission", "after2");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
+}  // namespace opt
+}  // namespace mindspore

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

@@ -0,0 +1,63 @@
+# 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')
+unsorted_segment_sum = P.UnsortedSegmentSum()
+num_segments = 4
+padding = Primitive('Padding')
+op_slice = Primitive('Slice')
+op_unsorted_segment_sum = Primitive('UnsortedSegmentSum')
+
+
+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_unsorted_segment_sum_fission(tag):
+    fns = FnDict()
+
+    @fns
+    def before1(input0, input1):
+        x = unsorted_segment_sum(input0, input1, num_segments)
+        return x
+
+    @fns
+    def after1(input0, input1):
+        x = padding(input0)
+        x = op_unsorted_segment_sum(x, input1)
+        x = op_slice(x)
+        return make_tuple(x)
+
+    @fns
+    def before2(input0, input1):
+        x = unsorted_segment_sum(input0, input1, num_segments)
+        return x
+
+    @fns
+    def after2(input0, input1):
+        x = op_unsorted_segment_sum(input0, input1)
+        return make_tuple(x)
+
+    return fns[tag]