implement parallel BroadcastTo

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

@@ -172,6 +172,8 @@ using TransposeCost = ActivationCost;
 using TransposeCostPtr = std::shared_ptr<TransposeCost>;
 using StridedSliceCost = ActivationCost;
 using StridedSliceCostPtr = std::shared_ptr<StridedSliceCost>;
+using SplitCost = ActivationCost;
+using SplitCostPtr = std::shared_ptr<SplitCost>;
 
 class SoftmaxCost : public OperatorCost {
  public:
@@ -203,8 +205,8 @@ using PackCost = TileCost;
 using PackCostPtr = std::shared_ptr<PackCost>;
 using ConcatCost = TileCost;
 using ConcatCostPtr = std::shared_ptr<ConcatCost>;
-using SplitCost = TileCost;
+using BroadcastToCost = SoftmaxCost;
-using SplitCostPtr = std::shared_ptr<SplitCost>;
+using BroadcastToCostPtr = std::shared_ptr<BroadcastToCost>;
 
 class TmpIdentityCost : public OperatorCost {
  public:

mindspore/ccsrc/frontend/parallel/dynamic_creator.h

@@ -176,6 +176,7 @@ REGISTER(SquareInfo);
 REGISTER(GatherV2PInfo);
 REGISTER(EmbeddingLookupInfo);
 REGISTER(TileInfo);
+REGISTER(BroadcastToInfo);
 REGISTER(StridedSliceInfo);
 REGISTER(DropoutInfo);
 REGISTER(PackInfo);

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

@@ -0,0 +1,64 @@
+/**
+ * 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_BROADCAST_TO_INFO_H_
+#define MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_BROADCAST_TO_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 {
+/*
+ * Limitation: Dimensions with size 1 can't be splited.
+ */
+class BroadcastToInfo : public OperatorInfo {
+ public:
+  BroadcastToInfo(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<BroadcastToCost>(false)) {}
+  ~BroadcastToInfo() override = default;
+
+  Status Init(const StrategyPtr &strategy) override;
+  Status InitForCostModel(const StrategyPtr &strategy) override;
+  Status GenerateStrategies(int32_t) override;
+  Status SetCostUnderStrategy(const StrategyPtr &) override;
+  ReplaceGraphPtr replace_graph(const CNodePtr &cnode) 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 ComputeReplaceGraph(const CNodePtr &cnode);
+
+ private:
+  Shape out_shape_;
+};
+}  // namespace parallel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_BROADCAST_TO_INFO_H_

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

@@ -42,5 +42,6 @@
 #include "frontend/parallel/ops_info/concat_info.h"
 #include "frontend/parallel/ops_info/split_info.h"
 #include "frontend/parallel/ops_info/pack_info.h"
+#include "frontend/parallel/ops_info/broadcast_to_info.h"
 
 #endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_HEAD_FILES_H_

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

@@ -222,6 +222,7 @@ constexpr char GATHERV2[] = "GatherV2";
 constexpr char SPARSE_GATHERV2[] = "SparseGatherV2";
 constexpr char STRIDEDSLICE[] = "StridedSlice";
 constexpr char BROADCAST[] = "Broadcast";
+constexpr char BROADCAST_TO[] = "BroadcastTo";
 constexpr char SQRT[] = "Sqrt";
 constexpr char ASSIGN[] = "Assign";
 constexpr char GET_NEXT[] = "GetNext";

mindspore/ccsrc/frontend/parallel/step_auto_parallel.cc

@@ -265,7 +265,7 @@ 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, SPLIT};
+     EMBEDDING_LOOKUP, FUSE_BATCH_NORM_EX, SPLIT, BROADCAST_TO};
   // clang-format on
 
   auto iter = splittable_op.find(op_name);

tests/ut/python/parallel/test_broadcast_to.py

@@ -0,0 +1,112 @@
+# 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, weight1, strategy1=None, strategy2=None, is_parameter=True):
+        super(Net, self).__init__()
+        self.shape = (8, 48, 64)
+        self.broadcast = P.BroadcastTo(self.shape).shard(strategy1)
+        self.mul = P.Mul().shard(strategy2)
+        if is_parameter:
+            self.weight1 = Parameter(weight1, "w1")
+        else:
+            self.weight1 = weight1
+
+    def construct(self, x):
+        out = self.broadcast(self.weight1)
+        out = self.mul(x, out)
+        return out
+
+
+class MatMulNet(nn.Cell):
+    def __init__(self, weight1, strategy1=None, strategy2=None, strategy3=None, is_parameter=True):
+        super(MatMulNet, self).__init__()
+        self.shape = (8, 64, 64)
+        self.broadcast = P.BroadcastTo(self.shape).shard(strategy1)
+        self.matmul = P.BatchMatMul().shard(strategy2)
+        self.mul = P.Mul().shard(strategy3)
+        if is_parameter:
+            self.weight1 = Parameter(weight1, "w1")
+        else:
+            self.weight1 = weight1
+
+    def construct(self, x1, x2):
+        out = self.broadcast(x2)
+        out = self.matmul(x1, out)
+        out = self.mul(out, self.weight1)
+        return out
+
+
+_w1 = Tensor(np.ones([1, 48, 64]), dtype=ms.float32)
+_x1 = Tensor(np.ones([8, 48, 64]), dtype=ms.float32)
+_x2 = Tensor(np.ones([64, 64]), 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, _x1)
+    context.reset_auto_parallel_context()
+
+
+def compile_net2(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, _x2)
+    context.reset_auto_parallel_context()
+
+
+def test_BroadcastTo_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 = Net(_w1, strategy1, strategy2)
+    compile_net(net)
+
+
+def test_BroadcastTo_parameter_no_full():
+    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 = Net(_w1, strategy1, strategy2)
+    compile_net(net)
+
+
+def test_BroadcastTo_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_BroadcastTo_matmul():
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    strategy1 = ((2, 4),)
+    strategy2 = ((1, 1, 2), (1, 2, 4))
+    strategy3 = ((1, 2, 4), (1, 2, 4))
+    net = MatMulNet(_w1, strategy1, strategy2, strategy3)
+    compile_net2(net)