!15298 add parallel operator for topk

From: @yangzhenzhang
Reviewed-by: @stsuteng,@kisnwang
Signed-off-by: @stsuteng

mindspore/ccsrc/frontend/parallel/auto_parallel/operator_costmodel.h

@@ -204,6 +204,7 @@ using RsqrtCost = SqrtCost;
 using AsinhCost = SqrtCost;
 using AcoshCost = SqrtCost;
 using ReLUV2Cost = SqrtCost;
+using TopKCost = SqrtCost;
 
 class ReLU6Cost : public CastCost {
  public:

mindspore/ccsrc/frontend/parallel/dynamic_creator.h

@@ -192,6 +192,7 @@ REGISTER(ConcatInfo);
 REGISTER(SplitInfo);
 REGISTER(UniqueInfo);
 REGISTER(GatherNdInfo);
+REGISTER(TopKInfo);
 }  // namespace parallel
 }  // namespace mindspore
 

mindspore/ccsrc/frontend/parallel/ops_info/ops_info_head_files.h

@@ -51,5 +51,6 @@
 #include "frontend/parallel/ops_info/uniform_candidate_sampler_info.h"
 #include "frontend/parallel/ops_info/reluv2_info.h"
 #include "frontend/parallel/ops_info/gathernd_info.h"
+#include "frontend/parallel/ops_info/topk_info.h"
 
 #endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_HEAD_FILES_H_

mindspore/ccsrc/frontend/parallel/ops_info/topk_info.cc

@@ -0,0 +1,233 @@
+/**
+ * Copyright 2021 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 "frontend/parallel/ops_info/topk_info.h"
+
+#include <algorithm>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "frontend/parallel/device_matrix.h"
+#include "frontend/parallel/strategy.h"
+#include "frontend/parallel/tensor_layout/tensor_redistribution.h"
+#include "pipeline/jit/resource.h"
+
+namespace mindspore {
+namespace parallel {
+Status TopKInfo::CheckStrategy(const StrategyPtr &strategy) {
+  MS_EXCEPTION_IF_NULL(strategy);
+  if (CheckStrategyValue(strategy, inputs_shape_) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy";
+    return FAILED;
+  }
+
+  std::vector<Dimensions> stra = strategy->GetInputDim();
+  if (stra.empty()) {
+    MS_LOG(ERROR) << name_ << ": The strategy is empty";
+    return FAILED;
+  }
+
+  if (stra[0].back() != 1) {
+    MS_LOG(ERROR) << name_ << ": Now we can not support to split last dimension";
+    return FAILED;
+  }
+
+  return SUCCESS;
+}
+
+Status TopKInfo::InferDevMatrixShape() {
+  MS_EXCEPTION_IF_NULL(strategy_);
+  std::vector<Dimensions> stra = strategy_->GetInputDim();
+  if (stra.empty()) {
+    MS_LOG(ERROR) << name_ << "The strategy is empty";
+    return FAILED;
+  }
+
+  dev_matrix_shape_ = stra[0];
+  return SUCCESS;
+}
+
+Status TopKInfo::InferTensorMap() {
+  TensorMap tensor_map;
+  if (inputs_shape_.empty()) {
+    MS_LOG(ERROR) << name_ << "The inputs shape is empty";
+    return FAILED;
+  }
+
+  // cannot use dev_matrix_shape_ replace inputs_shape_[0], because it may not be fully split in all devices.
+  int64_t size = SizeToLong(inputs_shape_[0].size());
+  for (int64_t i = 0; i < size; ++i) {
+    tensor_map.push_back(size - i - 1);
+  }
+
+  for (size_t i = 0; i < inputs_shape_.size(); ++i) {
+    inputs_tensor_map_.push_back(tensor_map);
+  }
+  outputs_tensor_map_.push_back(tensor_map);  // values
+  outputs_tensor_map_.push_back(tensor_map);  // indices
+  return SUCCESS;
+}
+
+Status TopKInfo::InferTensorInfo() {
+  if (inputs_shape_.empty() || outputs_shape_.empty() || inputs_tensor_map_.empty() || outputs_tensor_map_.empty()) {
+    MS_LOG(ERROR) << name_ << ": Invalid args";
+    return FAILED;
+  }
+
+  TensorLayout input_layout, output_layout;
+  for (size_t i = 0; i < inputs_shape_.size(); ++i) {
+    // infer tensor layout
+    if (input_layout.InitFromVector(dev_matrix_shape_, inputs_tensor_map_[i], inputs_shape_[i]) != SUCCESS) {
+      MS_LOG(ERROR) << name_ << ": Infer input tensor layout failed.";
+      return FAILED;
+    }
+    TensorInfo input_tensor_info(input_layout);
+    inputs_tensor_info_.push_back(input_tensor_info);
+  }
+
+  if (output_layout.InitFromVector(dev_matrix_shape_, outputs_tensor_map_[0], outputs_shape_[0]) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Infer output tensor layout failed.";
+    return FAILED;
+  }
+  TensorInfo output_tensor_info(output_layout);
+  outputs_tensor_info_.push_back(output_tensor_info);  // values
+  outputs_tensor_info_.push_back(output_tensor_info);  // indices
+  return SUCCESS;
+}
+
+Status TopKInfo::InferAsLossDivisor() {
+  if (outputs_tensor_map_.empty()) {
+    MS_LOG(ERROR) << name_ << ": The outputs tensor map is empty";
+    return FAILED;
+  }
+
+  MS_LOG(INFO) << name_ << " has two outputs, use output[0] to infer";
+  if (outputs_tensor_map_[0].empty()) {
+    as_loss_divisor_ = stage_device_size_;
+    MS_LOG(INFO) << name_ << ": The output is a scalar, use the dev size" << as_loss_divisor_ << " as loss divisor";
+    return SUCCESS;
+  }
+
+  as_loss_divisor_ = ComputeRepeatDeviceNumByTensorMap(dev_matrix_shape_, outputs_tensor_map_[0]);
+
+  std::string dev_matrix_shape_str = ShapeToString(dev_matrix_shape_);
+  std::string output_tensor_map_str = ShapeToString(outputs_tensor_map_[0]);
+  MS_LOG(INFO) << name_ << ": the dev matrix shape, the output tensor map, and loss divisor is " << dev_matrix_shape_str
+               << ", " << output_tensor_map_str << ", " << as_loss_divisor_;
+  return SUCCESS;
+}
+
+Status TopKInfo::InferMirrorOps() {
+  mirror_ops_.clear();
+  if (inputs_shape_.empty()) {
+    MS_LOG(INFO) << name_ << ": The inputs size is empty";
+    return SUCCESS;
+  }
+
+  if (inputs_tensor_map_.size() != inputs_shape_.size()) {
+    MS_LOG(ERROR) << name_ << ": The size of inputs tensor map is not equal to the size of inputs shape";
+    return FAILED;
+  }
+
+  bool group_is_empty = true;
+  for (size_t i = 0; i < inputs_tensor_map_.size(); ++i) {
+    std::vector<Group> group;
+    if (CreateGroupByTensorMap(inputs_tensor_map_[i], &group) != SUCCESS) {
+      MS_LOG(ERROR) << name_ << ": Create group failed, the input index is " << i;
+      mirror_ops_.clear();
+      return FAILED;
+    }
+
+    OperatorVector mirror_op;
+    if (group.empty()) {
+      MS_LOG(INFO) << name_ << ": The mirror group is empty, the input index is " << i;
+      mirror_ops_.push_back(mirror_op);
+      continue;
+    }
+
+    group_is_empty = false;
+    mirror_op = CreateMirrorOps(group[0].name(), group[0].GetDevNum());
+    mirror_ops_.push_back(mirror_op);
+  }
+
+  if (group_is_empty) {
+    mirror_ops_.clear();
+    MS_LOG(INFO) << name_ << ": No need to insert mirror ops";
+    return SUCCESS;
+  }
+
+  OperatorVector tmp_mirror_op;  // tmp mirror op for 'k'
+  mirror_ops_.push_back(tmp_mirror_op);
+  return SUCCESS;
+}
+
+Status TopKInfo::SetCostUnderStrategy(const StrategyPtr &strategy) { return SetCostUnderStrategyBase(strategy); }
+
+Status TopKInfo::GenerateStrategies(int64_t stage_id) {
+  if (InferAttrs() != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Infer attrs failed";
+    return FAILED;
+  }
+  if (inputs_shape_.empty()) {
+    MS_LOG(ERROR) << name_ << ": The inputs shape is empty";
+    return FAILED;
+  }
+
+  // to generate the first input's strategy
+  Shape input_split(inputs_shape_[0].size(), 1);
+  input_split.back() = 0;  // the last dimension can not be split
+  Shapes splittable_input = {input_split};
+  Shapes tmp_inputs_shape = {inputs_shape_[0]};
+
+  std::vector<StrategyPtr> sp_vector;
+  if (GenerateStrategiesForIndependentInputs(stage_id, tmp_inputs_shape, splittable_input, &sp_vector) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Generate strategies failed";
+    return FAILED;
+  }
+
+  size_t success = 0;
+  for (auto &sp : sp_vector) {
+    PrintStrategy(sp);
+    if (SetCostUnderStrategy(sp) == SUCCESS) {
+      success++;
+      MS_LOG(INFO) << name_ << ": Successfully generated " << success << " strategy.";
+      PrintStrategy(sp);
+    }
+  }
+  return SUCCESS;
+}
+
+Status TopKInfo::Init(const StrategyPtr &strategy) {
+  if (InitWithAutoRepeatCalc(strategy) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Init failed.";
+    return FAILED;
+  }
+  MS_LOG(INFO) << name_ << ": Init success.";
+  return SUCCESS;
+}
+
+Status TopKInfo::InitForCostModel(const StrategyPtr &strategy) {
+  if (InitForCostModelWithAutoRepeatCalc(strategy) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Init for cost model failed.";
+    return FAILED;
+  }
+
+  MS_LOG(INFO) << name_ << ": Init for cost model success.";
+  return SUCCESS;
+}
+}  // namespace parallel
+}  // namespace mindspore

mindspore/ccsrc/frontend/parallel/ops_info/topk_info.h

@@ -0,0 +1,60 @@
+/**
+ * Copyright 2021 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_FRONTEND_PARALLEL_OPS_INFO_TOPK_INFO_H_
+#define MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_TOPK_INFO_H_
+
+#include <string>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include "ir/value.h"
+#include "frontend/parallel/auto_parallel/operator_costmodel.h"
+#include "frontend/parallel/ops_info/operator_info.h"
+#include "frontend/parallel/strategy.h"
+
+namespace mindspore {
+namespace parallel {
+// the last dimension of input can not be split, other dimensions can be split
+class TopKInfo : public OperatorInfo {
+ public:
+  TopKInfo(const std::string &operator_name, const Shapes &inputs_shape, const Shapes &outputs_shape,
+           const PrimitiveAttrs &attrs)
+      : OperatorInfo(operator_name, inputs_shape, outputs_shape, attrs, std::make_shared<TopKCost>()) {}
+  ~TopKInfo() override = default;
+
+  Status Init(const StrategyPtr &strategy) override;
+  Status InitForCostModel(const StrategyPtr &strategy) override;
+  Status GenerateStrategies(int64_t) override;
+  Status SetCostUnderStrategy(const StrategyPtr &) override;
+
+ protected:
+  Status GetAttrs() override { return SUCCESS; }
+  Status CheckStrategy(const StrategyPtr &strategy) override;
+  Status InferForwardCommunication() override { return SUCCESS; }
+  Status InferMirrorOps() override;  // can not use OperatorInfo::InferMirrorOps(), since the 'k' of topk is scalar
+  Status InferTensorInfo() override;
+  Status InferDevMatrixShape() override;
+  Status InferTensorMap() override;
+  Status InferAsLossDivisor() override;
+};
+
+using TopKInfoPtr = std::shared_ptr<TopKInfo>;
+}  // namespace parallel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_TOPK_INFO_H_

mindspore/ccsrc/frontend/parallel/step_auto_parallel.cc

@@ -163,8 +163,7 @@ bool IsSplittableOperator(const std::string &op_name) {
      BESSELI0E, BESSELI1E, FLOORMOD, ASSIGN, ASSIGN_ADD, ATAN2, DIVNONAN, LOGICALAND, LOGICALOR, ELU, RELU6, RELUV2,
      SOFTPLUS, SOFTSIGN, GREATEREQUAL, LESSEQUAL, LESS, APPROXIMATEEQUAL, MOD, UNIQUE, UNSORTED_SEGMENT_SUM,
      UNSORTED_SEGMENT_MIN, REPEAT_ELEMENTS, TENSOR_DOT, RANGE, UNIFORM_CANDIDATE_SAMPLER, SLICE,
-     UNSORTED_SEGMENT_MAX, GATHER_ND};
-
+     UNSORTED_SEGMENT_MAX, GATHER_ND, TOPK};
   // clang-format on
 
   auto iter = splittable_op.find(op_name);

tests/ut/python/parallel/test_topk.py

@@ -0,0 +1,109 @@
+# Copyright 2021 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 pytest
+
+import mindspore as ms
+from mindspore import context, Tensor, Parameter
+from mindspore.nn import Cell, Momentum
+from mindspore.ops import operations as P
+from mindspore.train import Model
+from tests.dataset_mock import MindData
+
+
+class Dataset(MindData):
+    def __init__(self, predict, label, length=3):
+        super(Dataset, self).__init__(size=length)
+        self.predict = predict
+        self.label = label
+        self.index = 0
+        self.length = length
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        if self.index >= self.length:
+            raise StopIteration
+        self.index += 1
+        return self.predict, self.label
+
+    def reset(self):
+        self.index = 0
+
+
+class Net(Cell):
+    def __init__(self, w1, strategy1=None, strategy2=None):
+        super().__init__()
+        self.mul = P.Mul().shard(strategy1)
+        self.w1 = Parameter(w1, "w1")
+        self.topk = P.TopK().shard(strategy2)
+
+    def construct(self, x, b):
+        out = self.mul(x, self.w1)
+        out, _ = self.topk(out, 8)
+        return out
+
+
+_x = Tensor(np.ones([16, 64]), dtype=ms.float32)
+_b = Tensor(np.ones([16, 64]), dtype=ms.float32)
+_w1 = Tensor(np.ones([128, 64]), dtype=ms.float32)
+
+
+def compile_net(net):
+    context.set_context(save_graphs=True)
+    learning_rate = 0.1
+    momentum = 0.9
+    epoch_size = 2
+    dataset = Dataset(_x, _b)
+    opt = Momentum(net.trainable_params(), learning_rate, momentum)
+    model = Model(net, optimizer=opt)
+    model.train(epoch_size, dataset, dataset_sink_mode=False)
+    context.reset_auto_parallel_context()
+
+
+def test_topk_data_parallel():
+    context.set_auto_parallel_context(
+        parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((8, 1), (8, 1))
+    strategy2 = ((8, 1),)
+    net = Net(_w1, strategy1, strategy2)
+    compile_net(net)
+
+
+def test_topk_model_parallel():
+    context.set_auto_parallel_context(
+        parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((2, 4), (2, 4))
+    strategy2 = ((2, 1),)
+    net = Net(_w1, strategy1, strategy2)
+    compile_net(net)
+
+
+def test_topk_auto_parallel():
+    context.set_auto_parallel_context(
+        parallel_mode="auto_parallel", device_num=8, global_rank=0)
+    net = Net(_w1)
+    compile_net(net)
+
+
+def test_topk_strategy_error():
+    context.set_auto_parallel_context(
+        parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((8, 1), (8, 1))
+    strategy2 = ((1, 8),)
+    net = Net(_w1, strategy1, strategy2)
+    with pytest.raises(RuntimeError):
+        compile_net(net)