!2865 asymmetric param row split support for GatherV2

Merge pull request !2865 from yihuaijie/dev

mindspore/ccsrc/parallel/ops_info/gather_v2_p_info.cc

@@ -20,6 +20,7 @@
 #include <numeric>
 #include <functional>
 #include <utility>
+#include <algorithm>
 
 #include "parallel/device_matrix.h"
 #include "parallel/graph_util/generate_graph.h"
@@ -62,6 +63,55 @@ Status GatherV2PInfo::GetAttrs() {
     return FAILED;
   }
 
+  auto manual_split_iter = attrs_.find("manual_split");
+  if (manual_split_iter != attrs_.end()) {
+    param_split_shapes_.clear();
+    manual_split_ = true;
+    auto var = manual_split_iter->second->cast<ValueTuplePtr>();
+    MS_LOG(DEBUG) << "Extract manual split strategy " << manual_split_iter->second->ToString();
+
+    if (var->size() > 0) {
+      std::vector<ValuePtr> elements = var->value();
+      for (auto &ele : elements) {
+        if (ele->isa<ValueSequeue>()) {
+          auto value_tuple = ele->cast<ValueTuplePtr>();
+          std::vector<ValuePtr> value_vector = value_tuple->value();
+          if (value_vector.size() != 2) {
+            MS_LOG(ERROR) << "Failure: Size of manual_split element must be 2.";
+            return FAILED;
+          }
+          param_split_shapes_.push_back(static_cast<int32_t>(GetValue<int>(value_vector[0])));
+          index_offsets_.push_back(static_cast<int32_t>(GetValue<int>(value_vector[1])));
+        } else {
+          MS_LOG(ERROR) << "Failure: Manual split strategy's format is wrong! Need ValueSequeue";
+          return FAILED;
+        }
+      }
+
+      if (param_split_shapes_.empty()) {
+        MS_LOG(ERROR) << "Failed to extract param split strategy.";
+        return FAILED;
+      }
+    }
+  }
+
+  return SUCCESS;
+}
+
+Status GatherV2PInfo::CheckManualSplit() {
+  auto param_shape = inputs_shape_.at(0);
+  int32_t split_shape_sum = std::accumulate(param_split_shapes_.begin(), param_split_shapes_.end(), 0,
+                                            [](int32_t s, int32_t shape) { return s + shape; });
+  if (split_shape_sum < param_shape.at(0)) {
+    MS_LOG(ERROR) << "Failure: Sum of splited shapes should not be smaller than param_shape.";
+    return FAILED;
+  }
+
+  if (std::any_of(index_offsets_.begin(), index_offsets_.end(), [](const int32_t &offset) { return offset < 0; })) {
+    MS_LOG(ERROR) << "Failure: Index offset must not less than 0.";
+    return FAILED;
+  }
+
   return SUCCESS;
 }
 
@@ -103,6 +153,14 @@ Status GatherV2PInfo::CheckStrategy(const StrategyPtr &strategy) {
     return FAILED;
   }
 
+  if (manual_split_) {
+    if (CheckManualSplit() != SUCCESS) {
+      return FAILED;
+    }
+    // when using manual_split, no need to check belowings.
+    return SUCCESS;
+  }
+
   // axis != 0, param_shape(0)%(param_strategy(0)*param_strategy(axis)) must be 0
   if (axis_ != 0 && param_shape.at(0) % (param_strategy.at(0) * param_strategy.at(IntToSize(axis_))) != 0) {
     MS_LOG(DEBUG) << name_ << ": index_shape(0) can't be divided by (param_strategy(0)*param_strategy(axis)).";
@@ -130,6 +188,11 @@ Status GatherV2PInfo::CheckStrategy(const StrategyPtr &strategy) {
 }
 
 Status GatherV2PInfo::InferMirrorOps() {
+  // There is no mirror operators for manual split
+  if (manual_split_) {
+    return SUCCESS;
+  }
+
   mirror_ops_.clear();
   Shape input_a_tensor_map = inputs_tensor_map_.at(0);
   std::vector<Group> input_a_group;
@@ -160,6 +223,13 @@ Status GatherV2PInfo::InferDevMatrixShape() {
   // infer input dev_matrix_shape
   auto param_strategy = strategy_->GetInputDim().at(0);
   auto index_strategy = strategy_->GetInputDim().at(1);
+
+  if (manual_split_) {
+    dev_matrix_shape_ = param_strategy;
+    out_dev_matrix_shape_ = dev_matrix_shape_;
+    return SUCCESS;
+  }
+
   dev_matrix_shape_ = param_strategy;
 
   // param_strategy(axis)!=1,
@@ -195,6 +265,12 @@ Status GatherV2PInfo::InferDevMatrixShape() {
 }
 
 Status GatherV2PInfo::InferTensorMap() {
+  if (manual_split_) {
+    inputs_tensor_map_.push_back({1, 0});
+    inputs_tensor_map_.push_back({-1, 1});
+    outputs_tensor_map_.push_back({-1, 1, 0});
+    return SUCCESS;
+  }
   // infer input tensor map
   // param_strategy(axis) != 1
   size_t param_size = inputs_shape_.at(0).size();
@@ -261,8 +337,13 @@ Status GatherV2PInfo::InferTensorInfo() {
   Shape input_shape = inputs_shape_.at(0);
   Shape input_index_shape = inputs_shape_.at(1);
   Shape output_shape = outputs_shape_.at(0);
+  int32_t rank = g_device_manager->global_rank();
   // infer tensor layout
   TensorLayout input_tensor_layout, input_index_layout, output_tensor_layout;
+  if (manual_split_) {
+    input_shape[0] = param_split_shapes_[rank / dev_matrix_shape_[1]];
+    input_shape[0] = input_shape[0] * dev_matrix_shape_[0];
+  }
   if ((input_tensor_layout.InitFromVector(dev_matrix_shape_, inputs_tensor_map_.at(0), input_shape) != SUCCESS) ||
       (input_index_layout.InitFromVector(dev_matrix_shape_, inputs_tensor_map_.at(1), input_index_shape) != SUCCESS) ||
       (output_tensor_layout.InitFromVector(out_dev_matrix_shape_, outputs_tensor_map_.at(0), output_shape) !=
@@ -274,6 +355,9 @@ Status GatherV2PInfo::InferTensorInfo() {
   TensorInfo input_index_info(input_index_layout);
   TensorInfo output_tensor_info(output_tensor_layout);
 
+  Shape slice_shape = input_tensor_info.slice_shape();
+  MS_LOG(DEBUG) << "The fake slice shape is: " << ShapeToString(slice_shape);
+
   inputs_tensor_info_.push_back(input_tensor_info);
   inputs_tensor_info_.push_back(input_index_info);
   outputs_tensor_info_.push_back(output_tensor_info);
@@ -312,6 +396,19 @@ Status GatherV2PInfo::InferBias() {
   return FAILED;
 }
 
+Status GatherV2PInfo::InferOffset() {
+  CheckGlobalDeviceManager();
+  size_t rank = g_device_manager->global_rank();
+  if (rank < index_offsets_.size()) {
+    index_offset_ = index_offsets_.at(rank);
+    MS_LOG(DEBUG) << name_ << ": Device rank " << rank << ", Index Offset: " << index_offset_;
+    return SUCCESS;
+  }
+
+  MS_LOG(ERROR) << name_ << ": Get index offset failed, index offset size is" << index_offsets_.size();
+  return FAILED;
+}
+
 Status GatherV2PInfo::InferGroup() {
   auto param_strategy = strategy_->GetInputDim().at(0);
   size_t dim = IntToSize(axis_);
@@ -410,6 +507,19 @@ Status GatherV2PInfo::ComputeReplaceGraph(const CNodePtr &cnode) {
     MS_LOG(ERROR) << "GenerateGraph Init failed";
     return FAILED;
   }
+  if (manual_split_) {
+    if (InferOffset() != SUCCESS) {
+      MS_LOG(ERROR) << name_ << ": Infer Bias failed.";
+      return FAILED;
+    }
+    auto sub = gen_g.PushBack({gen_g.NewOpInst(SUB), gen_g.virtual_input_node(), CreateInt32Tensor(index_offset_)});
+    auto gather_v2 =
+      gen_g.PushBack({gen_g.NewOpInst(replace_op_name_), gen_g.virtual_input_node(), sub, CreatInt32Imm(axis_)});
+    std::vector<std::pair<AnfNodePtr, int>> input_nodes = {std::make_pair(sub, 2), std::make_pair(gather_v2, 1)};
+    replace_graph_ = std::make_shared<std::pair<std::vector<std::pair<AnfNodePtr, int>>, AnfNodePtr>>(
+      std::make_pair(input_nodes, gather_v2));
+    return SUCCESS;
+  }
   if (InferBias() != SUCCESS) {
     MS_LOG(ERROR) << name_ << ": Infer Bias failed.";
     return FAILED;
@@ -444,6 +554,14 @@ Status GatherV2PInfo::ComputeReplaceGraph(const CNodePtr &cnode) {
 }
 
 ReplaceGraphPtr GatherV2PInfo::replace_graph(const CNodePtr &cnode) {
+  if (manual_split_) {
+    if (ComputeReplaceGraph(cnode) != SUCCESS) {
+      MS_LOG(ERROR) << name_ << ": ComputeReplaceGraph failed.";
+      return nullptr;
+    }
+    return replace_graph_;
+  }
+
   auto param_strategy = strategy_->GetInputDim().at(0);
   // target_ == CPU, no need to raplace graph
   if (target_ == CPU) {

mindspore/ccsrc/parallel/ops_info/gather_v2_p_info.h

@@ -36,6 +36,7 @@ class GatherV2PInfo : public OperatorInfo {
       : OperatorInfo(name, inputs_shape, outputs_shape, attrs, std::make_shared<GatherV2PCost>()),
         axis_(0),
         bias_(0),
+        index_offset_(0),
         slice_size_(0) {}
   ~GatherV2PInfo() override = default;
   Status Init(const StrategyPtr &strategy) override;
@@ -57,20 +58,26 @@ class GatherV2PInfo : public OperatorInfo {
 
  private:
   Status ComputeReplaceGraph(const CNodePtr &cnode);
+  Status CheckManualSplit();
   Status ComputeReplaceOp();
   Status InferBias();
+  Status InferOffset();
   Status InferGroup();
 
   int32_t axis_;
   std::string target_;
   std::string replace_op_name_ = GATHERV2;
   int32_t bias_;
+  int32_t index_offset_;
   int32_t slice_size_;
   Shape out_dev_matrix_shape_;
   Group group_;
   bool reduce_scatter_flag_ = false;
   int32_t split_num_ = 1;
   bool host_reduce_scatter_ = false;
+  bool manual_split_ = false;
+  std::vector<int32_t> param_split_shapes_;
+  std::vector<int32_t> index_offsets_;
 };
 
 class SparseGatherV2Info : public GatherV2PInfo {

tests/ut/python/parallel/test_manual_gatherv2.py

+# 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
+from mindspore import context, Tensor, Parameter
+from mindspore.common.api import _executor
+from mindspore.nn import Cell, TrainOneStepCell, Momentum
+from mindspore.ops import operations as P
+from mindspore.common.initializer import initializer
+
+class Net(Cell):
+    def __init__(self, strategy1=None, strategy2=None, strategy3=None):
+        super().__init__()
+        self.gatherv2 = P.GatherV2().set_strategy(strategy1)
+        self.gatherv2.add_prim_attr("manual_split", ((1, 0), (7, 1)))
+        self.mul = P.Mul().set_strategy(strategy2)
+        self.reshape = P.Reshape()
+        self.matmul = P.MatMul().set_strategy(strategy3)
+        self.matmul.add_prim_attr("forward_reduce_scatter", True)
+        self.param = Parameter(initializer("ones", (8, 64), ms.float32), name="gatherv2_param")
+        self.mul_weight = Parameter(initializer("ones", (2, 4, 64), ms.float32), name="mul_weight")
+        self.matmul_weight = Parameter(initializer("ones", (256, 16), ms.float32), name="matmul_weight")
+
+    def construct(self, x, b):
+        out = self.gatherv2(self.param, x, 0)
+        out = self.mul(out, self.mul_weight)
+        out = self.reshape(out, (2, 256))
+        out = self.matmul(out, self.matmul_weight)
+        return out
+
+_x = Tensor(np.ones([2, 4]), dtype=ms.int32)
+_b = Tensor(np.ones([64, 8]), dtype=ms.float32)
+
+def compile_net(net):
+    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, _b)
+    context.reset_auto_parallel_context()
+
+def test_neg_data_parallel():
+    context.set_context(save_graphs=True)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
+    strategy1 = ((2, 1), (1, 2))
+    strategy2 = ((1, 2, 1), (1, 2, 1))
+    strategy3 = ((1, 2), (2, 1))
+    net = Net(strategy1, strategy2, strategy3)
+    compile_net(net)