implement parallel Split

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

@@ -201,6 +201,8 @@ using TileCost = SoftmaxCost;
 using TileCostPtr = std::shared_ptr<TileCost>;
 using ConcatCost = TileCost;
 using ConcatCostPtr = std::shared_ptr<ConcatCost>;
+using SplitCost = TileCost;
+using SplitCostPtr = std::shared_ptr<SplitCost>;
 
 class TmpIdentityCost : public OperatorCost {
  public:

mindspore/ccsrc/frontend/parallel/dynamic_creator.h

@@ -179,6 +179,7 @@ REGISTER(TileInfo);
 REGISTER(StridedSliceInfo);
 REGISTER(DropoutInfo);
 REGISTER(ConcatInfo);
+REGISTER(SplitInfo);
 }  // namespace parallel
 }  // namespace mindspore
 

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

@@ -40,5 +40,6 @@
 #include "frontend/parallel/ops_info/tile_info.h"
 #include "frontend/parallel/ops_info/strided_slice_info.h"
 #include "frontend/parallel/ops_info/concat_info.h"
+#include "frontend/parallel/ops_info/split_info.h"
 
 #endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_HEAD_FILES_H_

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

@@ -0,0 +1,294 @@
+/**
+ * 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 "frontend/parallel/ops_info/split_info.h"
+
+#include <string>
+#include <memory>
+#include <vector>
+
+#include "frontend/parallel/device_matrix.h"
+#include "frontend/parallel/strategy.h"
+#include "frontend/parallel/tensor_layout/tensor_redistribution.h"
+#include "frontend/parallel/context.h"
+#include "pipeline/jit/resource.h"
+
+namespace mindspore {
+namespace parallel {
+Status SplitInfo::GetAttrs() {
+  int axis = 0;
+  int output_num = 0;
+
+  auto axis_iter = attrs_.find(AXIS);
+  if (axis_iter != attrs_.end()) {
+    MS_EXCEPTION_IF_NULL(axis_iter->second);
+    if (axis_iter->second->isa<Int32Imm>()) {
+      axis = axis_iter->second->cast<Int32ImmPtr>()->value();
+    } else {
+      MS_LOG(ERROR) << name_ << ": The value of axis is not int";
+      return FAILED;
+    }
+  } else {
+    MS_LOG(ERROR) << name_ << ": Can not find the axis attr";
+    return FAILED;
+  }
+
+  if (inputs_shape_.empty()) {
+    MS_LOG(ERROR) << name_ << ": The inputs shape is empty";
+    return FAILED;
+  }
+  int dim = SizeToInt(inputs_shape_[0].size());
+  if (axis < 0) {
+    axis = axis + dim;
+  }
+  axis_ = SizeToInt(axis);
+
+  auto output_num_iter = attrs_.find(OUTPUT_NUM);
+  if (output_num_iter != attrs_.end()) {
+    MS_EXCEPTION_IF_NULL(output_num_iter->second);
+    if (output_num_iter->second->isa<Int32Imm>()) {
+      output_num = output_num_iter->second->cast<Int32ImmPtr>()->value();
+    } else {
+      MS_LOG(ERROR) << name_ << ": The value of output_num is not int";
+      return FAILED;
+    }
+  } else {
+    MS_LOG(ERROR) << name_ << ": Can not find the output_num attr";
+    return FAILED;
+  }
+  output_num_ = output_num;
+
+  return SUCCESS;
+}
+
+Status SplitInfo::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 (axis_ >= stra[0].size()) {
+    MS_LOG(ERROR) << name_ << ": The axis is out of range, the axis is " << axis_;
+    return FAILED;
+  }
+
+  if (stra[0][axis_] != 1) {
+    MS_LOG(ERROR) << name_ << ": The axis can not be split";
+    return FAILED;
+  }
+
+  return SUCCESS;
+}
+
+Status SplitInfo::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 SplitInfo::InferTensorMap() {
+  TensorMap tensor_map;
+  if (inputs_shape_.empty()) {
+    MS_LOG(ERROR) << name_ << "The inputs shape is empty";
+    return FAILED;
+  }
+
+  int32_t size = SizeToInt(inputs_shape_[0].size());
+  for (int i = 0; i < size; ++i) {
+    tensor_map.push_back(size - i - 1);
+  }
+
+  inputs_tensor_map_.push_back(tensor_map);
+
+  for (size_t i = 0; i < outputs_shape_.size(); ++i) {
+    outputs_tensor_map_.push_back(tensor_map);
+  }
+
+  return SUCCESS;
+}
+
+Status SplitInfo::InferMirrorOps() {
+  mirror_ops_.clear();
+  if (inputs_tensor_map_.empty()) {
+    MS_LOG(ERROR) << name_ << ": The inputs tensor map is empty";
+    return FAILED;
+  }
+
+  Shape input_tensor_map = inputs_tensor_map_[0];
+  std::vector<Group> group;
+  if (CreateGroupByTensorMap(input_tensor_map, &group) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Create group for input failed.";
+    return FAILED;
+  }
+
+  OperatorVector mirror_op;
+  if (group.empty()) {
+    MS_LOG(INFO) << name_ << ": The mirror group is empty.";
+    return SUCCESS;
+  } else {
+    mirror_op = CreateMirrorOps(group[0].name(), group[0].GetDevNum());
+    mirror_ops_.push_back(mirror_op);
+    std::string group_name = group[0].name();
+    MS_LOG(INFO) << name_ << " : Create the mirror ops success, the group name is " << group_name;
+  }
+
+  return SUCCESS;
+}
+
+Status SplitInfo::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;
+  // infer tensor layout
+  if (input_layout.InitFromVector(dev_matrix_shape_, inputs_tensor_map_[0], inputs_shape_[0]) != 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;
+  }
+  for (size_t i = 0; i < outputs_shape_.size(); ++i) {
+    TensorInfo output_tensor_info(output_layout);
+    outputs_tensor_info_.push_back(output_tensor_info);
+  }
+
+  return SUCCESS;
+}
+
+Status SplitInfo::SetCostUnderStrategy(const StrategyPtr &strategy) { return SetCostUnderStrategyBase(strategy); }
+
+Status SplitInfo::GenerateStrategies(int32_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;
+  }
+  Shape input_split;
+  for (size_t i = 0; i < inputs_shape_[0].size(); ++i) {
+    if (i == axis_) {
+      input_split.push_back(0);
+    } else {
+      input_split.push_back(1);
+    }
+  }
+
+  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;
+}
+
+std::shared_ptr<Strategys> SplitInfo::GenerateBatchStrategies() {
+  if (GetAttrs() != SUCCESS) {
+    MS_LOG(EXCEPTION) << name_ << ": Get attr failed";
+  }
+  CheckGlobalDeviceManager();
+  size_t dev_num = g_device_manager->GetDeviceListByStageId(0).size();
+  Dimensions input_strategy(inputs_shape_[0].size(), 1);
+  // axis can't split
+  if (inputs_shape_[0].size() > 1) {
+    if (axis_ == 0) {
+      input_strategy[1] = dev_num;
+    } else {
+      input_strategy[0] = dev_num;
+    }
+  }
+  Strategys strategy_v = {input_strategy};
+  return std::make_shared<Strategys>(strategy_v);
+}
+
+Status SplitInfo::InferAsLossDivisor() {
+  if (!ParallelContext::GetInstance()->loss_repeated_mean()) {
+    as_loss_divisor_ = 1;
+    return SUCCESS;
+  }
+
+  if (outputs_tensor_map_.empty()) {
+    MS_LOG(ERROR) << name_ << ": The outputs tensor map is empty.";
+    return FAILED;
+  }
+
+  if (outputs_tensor_map_[0].empty()) {
+    as_loss_divisor_ = SizeToInt(global_device_list_.size());
+    MS_LOG(INFO) << name_ << ": The output is a scalar, use the dev size " << as_loss_divisor_ << ", loss divisor.";
+    return SUCCESS;
+  }
+
+  as_loss_divisor_ = ComputeRepeatDeviceNumByTensorMap(dev_matrix_shape_, outputs_tensor_map_[0]);
+  MS_LOG(INFO) << name_ << ": the dev matrix shape is " << ShapeToString(dev_matrix_shape_)
+               << ", the output tensor map is " << ShapeToString(outputs_tensor_map_[0]) << ", loss divisor is "
+               << as_loss_divisor_;
+  return SUCCESS;
+}
+
+Status SplitInfo::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 SplitInfo::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/split_info.h

@@ -0,0 +1,60 @@
+/**
+ * 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_FRONTEND_PARALLEL_OPS_INFO_SPLIT_INFO_H_
+#define MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_SPLIT_INFO_H_
+
+#include <string>
+#include <memory>
+
+#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 {
+class SplitInfo : public OperatorInfo {
+ public:
+  SplitInfo(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<ConcatCost>(false)) {}
+  ~SplitInfo() override = default;
+
+  Status Init(const StrategyPtr &strategy) override;
+  Status InitForCostModel(const StrategyPtr &strategy) override;
+  Status GenerateStrategies(int32_t) override;
+  std::shared_ptr<Strategys> GenerateBatchStrategies() override;
+  Status SetCostUnderStrategy(const StrategyPtr &) override;
+
+ protected:
+  Status GetAttrs() override;
+  Status CheckStrategy(const StrategyPtr &strategy) override;
+  Status InferMirrorOps() override;
+  Status InferForwardCommunication() override { return SUCCESS; }
+  Status InferTensorInfo() override;
+  Status InferDevMatrixShape() override;
+  Status InferTensorMap() override;
+  Status InferAsLossDivisor() override;
+
+ private:
+  size_t axis_ = 0;
+  size_t output_num_ = 0;
+};
+}  // namespace parallel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_SPLIT_INFO_H_

mindspore/ccsrc/frontend/parallel/step_auto_parallel.cc

@@ -263,7 +263,7 @@ bool IsSplittableOperator(const std::string &op_name) {
      LOG, REDUCE_MEAN, REAL_DIV, SIGMOID, POW, MAXIMUM, MINIMUM, EQUAL, NOT_EQUAL, LOGICALNOT, GATHERV2, SQRT, CONCAT,
      STRIDEDSLICE, GET_NEXT, CAST, NEG, SQUARE, BATCH_MATMUL, EXPAND_DIMS, SQUEEZE, SPARSE_GATHERV2, TILE, DROPOUT,
      SOFTMAX_CROSS_ENTROPY_WITH_LOGITS, SIGMOID_CROSS_ENTROPY_WITH_LOGITS, SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS,
-     EMBEDDING_LOOKUP, FUSE_BATCH_NORM_EX};
+     EMBEDDING_LOOKUP, FUSE_BATCH_NORM_EX, SPLIT};
   // clang-format on
 
   auto iter = splittable_op.find(op_name);

tests/ut/python/parallel/test_split.py

@@ -0,0 +1,147 @@
+# 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 as ms
+import mindspore.context as context
+from mindspore import Tensor, Parameter
+import mindspore.nn as nn
+from mindspore.common.api import _executor
+from mindspore.nn import TrainOneStepCell, Momentum
+from mindspore.ops import operations as P
+
+
+class Net(nn.Cell):
+    def __init__(self, mul_weight, axis=0, out_nums=1, strategy1=None, strategy2=None, strategy3=None):
+        super(Net, self).__init__()
+        self.split = P.Split(axis, out_nums).shard(strategy1)
+        self.mul = P.Mul().shard(strategy2)
+        self.matmul = P.MatMul(transpose_b=True).shard(strategy2)
+        self.matmul2 = P.MatMul().shard(strategy3)
+        self.weight = Parameter(mul_weight, "w1")
+
+    def construct(self, x):
+        out = self.mul(x, self.weight)
+        out1, out2, out3 = self.split(out)
+        out = self.matmul(out1, out2)
+        out = self.matmul2(out, out3)
+        return out
+
+
+class Net1(nn.Cell):
+    def __init__(self, mul_weight, axis=0, out_nums=1, strategy1=None, strategy2=None):
+        super(Net1, self).__init__()
+        self.split = P.Split(axis, out_nums).shard(strategy1)
+        self.mul = P.Mul().shard(strategy2)
+        self.weight = Parameter(mul_weight, "w1")
+
+    def construct(self, x):
+        out1, out2 = self.split(self.weight)
+        out = self.mul(x, out1)
+        out = self.mul(out, out2)
+        return out
+
+
+class Net2(nn.Cell):
+    def __init__(self, mul_weight, axis=0, out_nums=1, strategy1=None, strategy2=None):
+        super(Net2, self).__init__()
+        self.split = P.Split(axis, out_nums).shard(strategy1)
+        self.mul = P.Mul().shard(strategy2)
+        self.weight = Parameter(mul_weight, "w1")
+
+    def construct(self, x):
+        out = self.mul(x, self.weight)
+        out1, _ = self.split(out)
+        return out1
+
+
+_w = Tensor(np.ones([48, 64]), dtype=ms.float32)
+_x = Tensor(np.ones([48, 64]), dtype=ms.float32)
+
+_w1 = Tensor(np.ones([96, 64, 32]), dtype=ms.float32)
+_x1 = Tensor(np.ones([48, 64, 32]), dtype=ms.float32)
+
+_w2 = Tensor(np.ones([48, 64, 32]), dtype=ms.float32)
+
+def compile_net(net):
+    context.set_context(mode=context.GRAPH_MODE, save_graphs=True)
+    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
+    train_net = TrainOneStepCell(net, optimizer)
+    train_net.set_auto_parallel()
+    _executor.compile(train_net, _x)
+    context.reset_auto_parallel_context()
+
+
+def compile_net1(net):
+    context.set_context(mode=context.GRAPH_MODE, save_graphs=True)
+    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
+    train_net = TrainOneStepCell(net, optimizer)
+    train_net.set_auto_parallel()
+    _executor.compile(train_net, _x1)
+    context.reset_auto_parallel_context()
+
+
+def test_split_parameter():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((1, 4, 2),)
+    strategy2 = ((1, 4, 2), (1, 4, 2))
+    net = Net1(_w1, 0, 2, strategy1, strategy2)
+    compile_net1(net)
+
+
+def test_split_parameter_no_full_split():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((1, 2, 2),)
+    strategy2 = ((1, 4, 2), (1, 4, 2))
+    net = Net1(_w1, 0, 2, strategy1, strategy2)
+    compile_net1(net)
+
+
+def test_split_tensor():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((1, 8),)
+    strategy2 = ((1, 8), (1, 8))
+    strategy3 = ((1, 1), (1, 8))
+    net = Net(_w, 0, 3, strategy1, strategy2, strategy3)
+    compile_net(net)
+
+
+def test_split_output():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((1, 4, 2),)
+    strategy2 = ((1, 4, 2), (1, 4, 2))
+    net = Net2(_w2, 0, 2, strategy1, strategy2)
+    compile_net1(net)
+
+
+def test_split_output_no_full_split():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((1, 2, 2),)
+    strategy2 = ((1, 4, 2), (1, 4, 2))
+    net = Net2(_w2, 0, 2, strategy1, strategy2)
+    compile_net1(net)
+
+
+def test_split_no_strategy():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = None
+    strategy2 = ((1, 4, 2), (1, 4, 2))
+    net = Net2(_w2, 0, 2, strategy1, strategy2)
+    compile_net1(net)
+
+
+def test_split_auto_parallel():
+    context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=8, global_rank=0)
+    net = Net2(_w2, 0, 2)
+    compile_net1(net)