!22096 add parallel sparse attention ops

Merge pull request !22096 from yao_yf/parallel_sparse_attention_ops

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

@@ -1744,5 +1744,35 @@ void VirtualDatasetCost::CalculateInputsInMemory(const std::map<size_t, bool> &)
     is_inputs_should_in_memory_[i] = is_parameter_[i];
   }
 }
+
+double MatmulDDSCost::GetForwardComputationCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &,
+                                                int64_t) const {
+  double result = 0.0;
+  if (inputs_type_lengths_.size() != inputs.size()) {
+    MS_LOG(EXCEPTION) << "Invalid inputs type size " << inputs_type_lengths_.size() << " for layer norm cost";
+  }
+
+  for (size_t index = 0; index < inputs.size(); ++index) {
+    TensorInfo tensor_info = inputs[index];
+    Shape slice_shape = tensor_info.slice_shape();
+    result += ListProduct(slice_shape) * static_cast<double>(inputs_type_lengths_[index]);
+  }
+  return result;
+}
+
+// Not taking account of output
+void MatmulDDSCost::CalculateOutputInMemory() {
+  is_output_should_in_memory_ =
+    (std::find(is_parameter_involve_.begin(), is_parameter_involve_.end(), true) != is_parameter_involve_.end());
+}
+
+// Taking account of input
+void MatmulDDSCost::CalculateInputsInMemory(const std::map<size_t, bool> &prev_output_in_mem) {
+  bool keep_mem =
+    (std::find(is_parameter_.begin(), is_parameter_.end(), true) != is_parameter_.end()) ||
+    (std::find(is_parameter_involve_.begin(), is_parameter_involve_.end(), true) != is_parameter_involve_.end());
+  std::fill(is_inputs_should_in_memory_.begin(), is_inputs_should_in_memory_.end(), keep_mem);
+}
+
 }  // namespace parallel
 }  // namespace mindspore

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

@@ -1051,6 +1051,44 @@ class GatherV2PCost : public GatherV2Cost {
   int64_t axis_;
   Shape strategy_;
 };
+
+class MatmulDDSCost : public OperatorCost {
+ public:
+  MatmulDDSCost() : OperatorCost() {}
+  ~MatmulDDSCost() override = default;
+
+  // per device communication cost
+  double GetCommCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                     int64_t stage_id) const override {
+    return GetForwardCommCost(inputs, outputs, stage_id) + GetBackwardCommCost(inputs, outputs, stage_id);
+  }
+  double GetForwardCommCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                            int64_t stage_id) const override {
+    return 0.0;
+  };
+  double GetBackwardCommCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                             int64_t stage_id) const override {
+    return 0.0;
+  };
+
+  // per device computation cost
+  double GetComputationCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                            int64_t stage_id) const override {
+    return GetForwardComputationCost(inputs, outputs, stage_id) + GetBackwardComputationCost(inputs, outputs, stage_id);
+  }
+  double GetForwardComputationCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                                   int64_t stage_id) const override;
+  double GetBackwardComputationCost(const std::vector<TensorInfo> &inputs, const std::vector<TensorInfo> &outputs,
+                                    int64_t stage_id) const override {
+    return 0.0;
+  };
+  // Not taking account of output
+  void CalculateOutputInMemory() override;
+  // Taking account of input
+  void CalculateInputsInMemory(const std::map<size_t, bool> &prev_output_in_mem) override;
+};
+using MatmulDDSCostPtr = std::shared_ptr<MatmulDDSCost>;
+
 }  // namespace parallel
 }  // namespace mindspore
 #endif  // PARALLEL_AUTO_PARALLEL_OPERATOR_COSTMODEL_H_

mindspore/ccsrc/frontend/parallel/dynamic_creator.h

@@ -204,6 +204,7 @@ REGISTER(MaxPoolInfo);
 REGISTER(AvgPoolInfo);
 REGISTER(GatherDInfo);
 REGISTER(ReduceAnyInfo);
+REGISTER(MatmulDDSInfo);
 }  // namespace parallel
 }  // namespace mindspore
 

mindspore/ccsrc/frontend/parallel/graph_util/generate_graph.cc

@@ -165,7 +165,7 @@ AnfNodePtr GenerateGraph::PushBack(const std::vector<AnfNodePtr> &inputs) {
   if (inputs.size() < 2) {
     MS_LOG(EXCEPTION) << "inputs.size() must be more than 1";
   }
-  (void)manager_->Replace(inputs.at(1), cnode);  // using Replace function to insert cnode after inputs[0]
+  (void)manager_->Replace(inputs.at(1), cnode);  // using Replace function to insert cnode after inputs[1]
   auto new_anf_node_ptr = cnode->cast<AnfNodePtr>();
   MS_EXCEPTION_IF_NULL(new_anf_node_ptr);
   return new_anf_node_ptr;

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

@@ -0,0 +1,320 @@
+/**
+ * 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/matmul_dds_info.h"
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "frontend/parallel/device_manager.h"
+#include "frontend/parallel/device_matrix.h"
+#include "frontend/parallel/step_parallel.h"
+#include "frontend/parallel/graph_util/generate_graph.h"
+#include "utils/log_adapter.h"
+
+namespace mindspore {
+namespace parallel {
+/*
+ * MatmulDDS has 4 input
+ *  q, k: A 4D float used in transformer model,
+ *  the shape is [num_heads * size_per_head // 16, bs * seq_len // 16, 16, 16], num_heads*size_per_head = embedding_size
+ *  The shape is reshaped for cube. origin shape is
+ *  (bs*seq_len, embedding_size) <=> (bs, num_heads, seq_len, size_per_head)
+ *  local_mask: Local mask in sparse attention, the shape is
+ *  (seq_len // 16, bs * block_size // 16, 16, 16).
+ *  block_num = seq_len // block_size, block_size = 64, always.
+ *  global_mask: Global mask in sparse attention,
+ *  the shape is (bs * global_size // 16, seq_len // 16, 16, 16)
+ *  seq_len = 1024, global_size = 256, always.
+ *  Only bs and num_heads can be splited, thus the q[0] should at least be size_per_head,
+ *  q[1] should at least be seq_len // 16. The strategy check can use bs/head from attrs.
+ */
+Status MatmulDDSInfo::CheckStrategy(const StrategyPtr &strategy) {
+  if (CheckStrategyValue(strategy, inputs_shape_) != SUCCESS) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy.";
+    return FAILED;
+  }
+  Strategys stras = strategy->GetInputDim();
+  if (stras.size() != MATMUL_DDS_INPUTS_SIZE) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys size should be 4.";
+    return FAILED;
+  }
+  for (auto stra : stras) {
+    if (stra.size() != MATMUL_DDS_STRATEGY_SIZE) {
+      MS_LOG(ERROR) << name_
+                    << ": Invalid strategy. The strategy size should be 4, but in current dim, "
+                       "the strategy is"
+                    << stra;
+      return FAILED;
+    }
+  }
+  if (stras[0][0] != stras[1][0] || num_heads_ % stras[0][0] != 0) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[0][0]:" << stras[0][0]
+                  << " should be equal to strategys[1][0]:" << stras[1][0]
+                  << " ,and should be divisible by num_heads: " << num_heads_;
+    return FAILED;
+  }
+  if (stras[0][1] != stras[1][1] || stras[0][1] != stras[2][1] || stras[0][1] != stras[3][0]) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[0][1]:" << stras[0][1]
+                  << ", strategys[1][1]:" << stras[1][1] << ", strategys[2][1]:" << stras[2][1]
+                  << ", strategys[3][0]:" << stras[3][0] << " should be the same.";
+    return FAILED;
+  }
+  if (batch_size_ % stras[0][1] != 0) {
+    MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[0][1]:" << stras[0][1]
+                  << " should be divisible by batch_sizes:" << batch_size_;
+    return FAILED;
+  }
+  for (size_t i = 2; i < stras[0].size(); ++i) {
+    if (stras[0][i] != 1) {
+      MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[0][" << i << "] only support 1";
+      return FAILED;
+    }
+  }
+  for (size_t i = 2; i < stras[1].size(); ++i) {
+    if (stras[1][i] != 1) {
+      MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[1][" << i << "] only support 1";
+      return FAILED;
+    }
+  }
+  for (size_t i = 0; i < stras[2].size(); ++i) {
+    if (i != 1 && stras[2][i] != 1) {
+      MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[2][" << i << "] only support 1";
+      return FAILED;
+    }
+  }
+  for (size_t i = 1; i < stras[3].size(); ++i) {
+    if (stras[3][i] != 1) {
+      MS_LOG(ERROR) << name_ << ": Invalid strategy. The strategys[3][" << i << "] only support 1";
+      return FAILED;
+    }
+  }
+  return SUCCESS;
+}
+
+/*
+ * device matrix is extended by the strategy0.
+ */
+Status MatmulDDSInfo::InferDevMatrixShape() {
+  Strategys stra = strategy_->GetInputDim();
+  Dimensions input_strategy = stra.at(0);
+  input_strategy_ = input_strategy;
+  dev_matrix_shape_ = input_strategy;
+  dev_matrix_shape_.push_back(1);
+  dev_matrix_shape_.push_back(1);
+  dev_matrix_shape_.push_back(1);
+  return SUCCESS;
+}
+
+/*
+ * q: [num_heads * size_per_head // 16, bs * seq_len // 16, 16, 16]
+ * k: [num_heads * size_per_head // 16, bs * seq_len // 16, 16, 16]
+ * local_mask: (block_num * block_size // 16, bs * block_size // 16, 16, 16)
+ * global_mask: (bs * global_size // 16, seq_len // 16, 16, 16)
+ * local_prob: (bs, num_heads, block_num, block_size // 16, block_size // 16, 16, 16)
+ * global_prob: (bs, num_heads, block_num, global_size // 16, block_size // 16, 16, 16)
+ * device_matrix: [num_heads_stra, bs_stra, 1, 1, 1, 1, 1]
+ */
+Status MatmulDDSInfo::InferTensorMap() {
+  TensorMap input_tensor_map_q;
+  // input_tensor_map_q [6, 5, -1, -1]
+  for (size_t i = 0; i < inputs_shape_[0].size(); ++i) {
+    if (i <= 1) {
+      input_tensor_map_q.push_back((int64_t)(inputs_shape_[0].size() + 3 - i - 1));
+    } else {
+      input_tensor_map_q.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  TensorMap input_tensor_map_k;
+  // input_tensor_map_k [6, 5, -1, -1]
+  for (size_t i = 0; i < inputs_shape_[1].size(); ++i) {
+    if (i <= 1) {
+      input_tensor_map_k.push_back((int64_t)(inputs_shape_[1].size() + 3 - i - 1));
+    } else {
+      input_tensor_map_k.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  TensorMap input_tensor_map_local_mask;
+  // input_tensor_map_local_mask [-1, 5, -1, -1]
+  for (size_t i = 0; i < inputs_shape_[2].size(); ++i) {
+    if (i == 1) {
+      input_tensor_map_local_mask.push_back((int64_t)(inputs_shape_[2].size() + 3 - 2));
+    } else {
+      input_tensor_map_local_mask.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  TensorMap input_tensor_map_global_mask;
+  // input_tensor_map_local_mask [5, -1, -1, -1]
+  for (size_t i = 0; i < inputs_shape_[3].size(); ++i) {
+    if (i == 0) {
+      input_tensor_map_global_mask.push_back((int64_t)(inputs_shape_[3].size() + 3 - 2));
+    } else {
+      input_tensor_map_global_mask.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  TensorMap output_tensor_map_local_prob;
+  // output_tensor_map_local_prob [5, 6, -1, -1, -1, -1, -1]
+  for (size_t i = 0; i < dev_matrix_shape_.size(); ++i) {
+    if (i == 0) {
+      output_tensor_map_local_prob.push_back((int64_t)(dev_matrix_shape_.size() - 2));
+    } else if (i == 1) {
+      output_tensor_map_local_prob.push_back((int64_t)(dev_matrix_shape_.size() - 1));
+    } else {
+      output_tensor_map_local_prob.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  TensorMap output_tensor_map_global_prob;
+  // output_tensor_map_global_prob [5, 6, -1, -1, -1, -1, -1]
+  for (size_t i = 0; i < dev_matrix_shape_.size(); ++i) {
+    if (i == 0) {
+      output_tensor_map_global_prob.push_back((int64_t)(dev_matrix_shape_.size() - 2));
+    } else if (i == 1) {
+      output_tensor_map_global_prob.push_back((int64_t)(dev_matrix_shape_.size() - 1));
+    } else {
+      output_tensor_map_global_prob.push_back((int64_t)(MAP_NONE));
+    }
+  }
+  inputs_tensor_map_.push_back(input_tensor_map_q);
+  inputs_tensor_map_.push_back(input_tensor_map_k);
+  inputs_tensor_map_.push_back(input_tensor_map_local_mask);
+  inputs_tensor_map_.push_back(input_tensor_map_global_mask);
+  outputs_tensor_map_.push_back(output_tensor_map_local_prob);
+  outputs_tensor_map_.push_back(output_tensor_map_global_prob);
+  return SUCCESS;
+}
+
+Status MatmulDDSInfo::GetAttrs() {
+  if ((inputs_shape_.size() != MATMUL_DDS_INPUTS_SIZE) || (outputs_shape_.size() != MATMUL_DDS_OUTPUTS_SIZE)) {
+    MS_LOG(ERROR) << name_ << ": Inputs shape size " << inputs_shape_.size() << " or outputs shape size "
+                  << outputs_shape_.size() << " is wrong.";
+    return FAILED;
+  }
+  auto iter = attrs_.find(BS);
+  if (iter != attrs_.end()) {
+    MS_EXCEPTION_IF_NULL(iter->second);
+    if (iter->second->isa<Int64Imm>()) {
+      batch_size_ = iter->second->cast<Int64ImmPtr>()->value();
+    } else {
+      MS_LOG(ERROR) << name_ << ": The value of axis is not int64_t.";
+      return FAILED;
+    }
+  }
+  iter = attrs_.find(HEADS);
+  if (iter != attrs_.end()) {
+    MS_EXCEPTION_IF_NULL(iter->second);
+    if (iter->second->isa<Int64Imm>()) {
+      num_heads_ = iter->second->cast<Int64ImmPtr>()->value();
+    } else {
+      MS_LOG(ERROR) << name_ << ": The value of axis is not int64_t.";
+      return FAILED;
+    }
+  }
+  return SUCCESS;
+}
+
+Status MatmulDDSInfo::ComputeReplaceGraph(const CNodePtr &cnode) {
+  int64_t new_batch_size;
+  int64_t new_num_heads;
+  int64_t batch_shard_num = strategy_->GetInputDim()[0][1];
+  int64_t heads_shard_num = strategy_->GetInputDim()[0][0];
+  new_batch_size = batch_size_ / batch_shard_num;
+  new_num_heads = num_heads_ / heads_shard_num;
+  ValuePtr new_bs_value = MakeValue(new_batch_size);
+  ValuePtr new_heads_value = MakeValue(new_num_heads);
+  Attr attr_batch_size = std::make_pair(BS, new_bs_value);
+  Attr attr_num_heads = std::make_pair(HEADS, new_heads_value);
+  OperatorAttrs attrs = {attr_batch_size, attr_num_heads};
+  GenerateGraph gen_g = GenerateGraph(attrs_);
+  if (gen_g.Init(cnode) != SUCCESS) {
+    MS_LOG(ERROR) << "GenerateGraph Init failed";
+    return FAILED;
+  }
+  auto matmul_dds_node =
+    gen_g.PushBack({gen_g.NewOpInst(MATMUL_DDS, attrs), gen_g.virtual_input_node(), gen_g.virtual_input_node(),
+                    gen_g.virtual_input_node(), gen_g.virtual_input_node()});
+  std::vector<std::pair<AnfNodePtr, int64_t>> input_nodes = {
+    std::make_pair(matmul_dds_node, 1),
+    std::make_pair(matmul_dds_node, 2),
+    std::make_pair(matmul_dds_node, 3),
+    std::make_pair(matmul_dds_node, 4),
+  };
+  replace_graph_ = std::make_shared<std::pair<std::vector<std::pair<AnfNodePtr, int64_t>>, AnfNodePtr>>(
+    std::make_pair(input_nodes, matmul_dds_node));
+  return SUCCESS;
+}
+
+ReplaceGraphPtr MatmulDDSInfo::replace_graph(const CNodePtr &cnode) {
+  if (ComputeReplaceGraph(cnode) != SUCCESS) {
+    MS_LOG(EXCEPTION) << name_ << ": ComputeReplaceGraph failed.";
+  }
+  return replace_graph_;
+}
+
+Status MatmulDDSInfo::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 MatmulDDSInfo::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;
+}
+
+std::vector<StrategyPtr> MatmulDDSInfo::GenerateOpStrategies(int64_t stage_id) {
+  // to generate the first input's strategy
+  Shape input0_split = {1, 1, 0, 0};
+  Shapes splittable_input = {input0_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(EXCEPTION) << name_ << ": Generate strategies failed";
+  }
+
+  // the others strategies are set by the first input's strategy
+  for (auto &sp : sp_vector) {
+    if ((sp == nullptr) || sp->GetInputDim().empty()) {
+      MS_LOG(EXCEPTION) << name_ << ": The strategy is null or empty";
+    }
+    Strategys tmp_strategy;
+    Dimensions q_strategy = sp->GetInputDim()[0];
+    Dimensions k_strategy = q_strategy;
+    Dimensions local_mask_strategy = {1, q_strategy[0], 1, 1};
+    Dimensions global_mask_strategy = {q_strategy[0], 1, 1, 1};
+
+    tmp_strategy.push_back(q_strategy);            // q
+    tmp_strategy.push_back(k_strategy);            // k
+    tmp_strategy.push_back(local_mask_strategy);   // local_mask
+    tmp_strategy.push_back(global_mask_strategy);  // global_mask
+    sp->ResetInputs(tmp_strategy);
+  }
+  return sp_vector;
+}
+
+Status MatmulDDSInfo::SetCostUnderStrategy(const StrategyPtr &strategy) { return SetCostUnderStrategyBase(strategy); }
+
+}  // namespace parallel
+}  // namespace mindspore

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

@@ -0,0 +1,64 @@
+/**
+ * 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_MATMUL_DDS_INFO_H_
+#define MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_MATMUL_DDS_INFO_H_
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "ir/value.h"
+#include "frontend/parallel/ops_info/operator_info.h"
+#include "frontend/parallel/strategy.h"
+
+namespace mindspore {
+namespace parallel {
+/*
+ * parallel class for MatmulDDS Primitive
+ */
+class MatmulDDSInfo : public OperatorInfo {
+ public:
+  MatmulDDSInfo(const std::string &name, const Shapes &inputs_shape, const Shapes &outputs_shape,
+                const PrimitiveAttrs &attrs)
+      : OperatorInfo(name, inputs_shape, outputs_shape, attrs, std::make_shared<MatmulDDSCost>()) {}
+  ~MatmulDDSInfo() override = default;
+  Status Init(const StrategyPtr &strategy) override;
+  Status InitForCostModel(const StrategyPtr &strategy) override;
+
+  std::vector<StrategyPtr> GenerateOpStrategies(int64_t stage_id) override;
+  Status SetCostUnderStrategy(const StrategyPtr &strategy) override;
+  ReplaceGraphPtr replace_graph(const CNodePtr &cnode) override;
+
+ protected:
+  Status CheckStrategy(const StrategyPtr &strategy) override;
+  Status InferForwardCommunication() override { return SUCCESS; }
+  Status InferDevMatrixShape() override;
+  Status InferTensorMap() override;
+  Status GetAttrs() override;
+  Status InferAsLossDivisor() override { return SUCCESS; }
+  Status ComputeReplaceGraph(const CNodePtr &cnode);
+
+ private:
+  Dimensions input_strategy_;
+  int64_t batch_size_ = 0;
+  int64_t num_heads_ = 0;
+};
+}  // namespace parallel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_MATMUL_DDS_INFO_H_

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

@@ -59,5 +59,6 @@
 #include "frontend/parallel/ops_info/batchnorm_info.h"
 #include "frontend/parallel/ops_info/maxpool_info.h"
 #include "frontend/parallel/ops_info/gatherd_info.h"
+#include "frontend/parallel/ops_info/matmul_dds_info.h"
 
 #endif  // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_HEAD_FILES_H_

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

@@ -74,6 +74,9 @@ constexpr size_t TRANSFER_CONCAT_TENSOR_DIM_INDEX = 0;
 constexpr size_t TRANSFER_CONCAT_DEV_DIM_INDEX = 1;
 constexpr size_t TRANSFER_CONCAT_SPLIT_COUNT_INDEX = 2;
 constexpr size_t TRANSFER_SPLIT_ARGS_SIZE = 3;
+constexpr size_t MATMUL_DDS_INPUTS_SIZE = 4;
+constexpr size_t MATMUL_DDS_OUTPUTS_SIZE = 2;
+constexpr size_t MATMUL_DDS_STRATEGY_SIZE = 4;
 constexpr double EPS = 1e-6;
 constexpr double INF = 1e20;
 constexpr double COST_FACTOR = 2.0;
@@ -142,6 +145,8 @@ constexpr char RANGE_MAX[] = "range_max";
 constexpr char REMOVE_ACCIDENTAL_HITS[] = "remove_accidental_hits";
 constexpr char UNIQUE_STRING[] = "unique";
 constexpr char AXIS[] = "axis";
+constexpr char BS[] = "bs";
+constexpr char HEADS[] = "heads";
 constexpr char AXES[] = "axes";
 constexpr char START[] = "start";
 constexpr char LIMIT[] = "limit";
@@ -319,6 +324,7 @@ constexpr char BROADCAST_TO[] = "BroadcastTo";
 constexpr char SQRT[] = "Sqrt";
 constexpr char ASSIGN[] = "Assign";
 constexpr char GET_NEXT[] = "GetNext";
+constexpr char MATMUL_DDS[] = "MatmulDDS";
 constexpr char SQUEEZE[] = "Squeeze";
 constexpr char NEG[] = "Neg";
 constexpr char ABS[] = "Abs";

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

@@ -45,7 +45,7 @@ class PReLUInfo : public OperatorInfo {
 
  protected:
   Status CheckStrategy(const StrategyPtr &strategy) override;
-  Status InferForwardCommunication() { return SUCCESS; }
+  Status InferForwardCommunication() override { return SUCCESS; }
   Status InferDevMatrixShape() override;
   Status InferTensorMap() override;
   Status GetAttrs() override;

mindspore/ccsrc/frontend/parallel/step_auto_parallel.cc

@@ -170,7 +170,8 @@ 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, SELECT, GATHERD,
-     UNSORTED_SEGMENT_MAX, GATHER_ND, TOPK, SCATTER_UPDATE, VIRTUAL_OUTPUT, CONV2D_BACK_PROP_INPUT, CONV2D_TRANSPOSE};
+     UNSORTED_SEGMENT_MAX, GATHER_ND, TOPK, SCATTER_UPDATE, VIRTUAL_OUTPUT, CONV2D_BACK_PROP_INPUT, CONV2D_TRANSPOSE,
+     MATMUL_DDS};
   // clang-format on
 
   auto iter = splittable_op.find(op_name);

tests/ut/python/parallel/test_cus_matmul_dds.py

@@ -0,0 +1,169 @@
+# 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 mindspore as ms
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore import context
+from mindspore.common.api import _cell_graph_executor
+from mindspore.ops import composite as C
+from mindspore.ops import operations as P
+from mindspore.parallel import set_algo_parameters
+from mindspore.ops.operations._inner_ops import MatmulDDS
+from tests.ut.python.ops.test_math_ops import VirtualLoss
+
+context.set_context(mode=context.GRAPH_MODE)
+
+grad_all = C.GradOperation(get_all=True)
+
+# q: (num_heads * size_per_head // 16, bs * seq_len // 16, 16, 16)
+# k: (num_heads * size_per_head // 16, bs * seq_len // 16, 16, 16)
+# local_mask: (block_num * block_size // 16, bs * block_size // 16, 16, 16)
+# global_mask: (bs * global_size // 16, seq_len // 16, 16, 16)
+# local_prob: (bs, num_heads, block_num, block_size // 16, block_size // 16, 16, 16)
+# global_prob: (bs, num_heads, block_num, global_size // 16, block_size // 16, 16, 16)
+# x: (bs*seq_len, num_heads*size_per_head)
+class Net(nn.Cell):
+    def __init__(self, batch_size, num_heads, dp, mp, shard=True):
+        super(Net, self).__init__()
+        self.batch_size = batch_size
+        self.num_heads = num_heads
+        self.size_per_head = 128
+        self.seq_len = 1024
+        self.block_size = 64
+        self.block_num = self.seq_len // self.block_size
+        self.global_size = 256
+        self.embedding_size = num_heads * self.size_per_head
+        self.cus_matmul = MatmulDDS(batch_size, num_heads)
+        self.reduce_sum = P.ReduceSum()
+        self.global_mask = Tensor(np.ones((batch_size * self.global_size // 16, self.seq_len // 16, 16, 16)))
+        self.local_mask = Tensor(np.ones((self.block_num * self.block_size // 16,
+                                          batch_size * self.block_size // 16, 16, 16)))
+        self.dense1 = nn.Dense(self.embedding_size, self.embedding_size, has_bias=False)
+        self.dense2 = nn.Dense(self.embedding_size, self.embedding_size, has_bias=False)
+        self.reshape = P.Reshape()
+        self.transpose = P.Transpose()
+        self.add = P.Add()
+        if shard:
+            self.cus_matmul.shard(((mp, dp, 1, 1), (mp, dp, 1, 1), (1, dp, 1, 1), (dp, 1, 1, 1)))
+            self.dense1.matmul.shard(((dp, 1), (mp, 1)))
+            self.dense2.matmul.shard(((dp, 1), (mp, 1)))
+            self.transpose.shard(((dp, 1, mp, 1),))
+
+
+    def construct(self, x):
+        q = self.dense1(x)
+        k = self.dense2(x)
+        q = self.transpose(self.reshape(q, (-1, 16, self.embedding_size // 16, 16)), (2, 0, 1, 3))
+        k = self.transpose(self.reshape(k, (-1, 16, self.embedding_size // 16, 16)), (2, 0, 1, 3))
+        local_prob, global_prob = self.cus_matmul(q, k, self.local_mask, self.global_mask)
+        local_prob = self.reshape(local_prob, (self.batch_size, self.num_heads, -1))
+        global_prob = self.reshape(global_prob, (self.batch_size, self.num_heads, -1))
+        local_prob_reduce = self.reduce_sum(local_prob, 2)
+        global_prob_reduce = self.reduce_sum(global_prob, 2)
+        result = self.add(local_prob_reduce, global_prob_reduce)
+        return result
+
+
+class GradWrap(nn.Cell):
+    def __init__(self, network):
+        super(GradWrap, self).__init__()
+        self.network = network
+
+    def construct(self, x):
+        return grad_all(self.network)(x)
+
+
+class NetWithLoss(nn.Cell):
+    def __init__(self, network):
+        super(NetWithLoss, self).__init__()
+        self.network = network
+        self.loss = VirtualLoss()
+
+    def construct(self, x):
+        predict = self.network(x)
+        return self.loss(predict)
+
+
+def compile_graph(batch_size, num_heads, dp, mp, auto=False, shard=True):
+    if auto:
+        context.set_auto_parallel_context(parallel_mode="auto_parallel")
+    else:
+        context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
+    x = Tensor(np.ones((batch_size * 1024, num_heads * 128)), ms.float32)
+    net = GradWrap(NetWithLoss(Net(batch_size, num_heads, dp, mp, shard=shard)))
+    net.set_auto_parallel()
+    net.set_train()
+    _cell_graph_executor.compile(net, x)
+
+def test_cus_matmul_dds_model_parallel_mix():
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 2
+    mp = 8
+    compile_graph(batch_size, num_heads, dp, mp)
+
+def test_cus_matmul_dds_model_parallel_dp():
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 16
+    mp = 1
+    compile_graph(batch_size, num_heads, dp, mp)
+
+def test_cus_matmul_dds_model_parallel_mp():
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 1
+    mp = 16
+    compile_graph(batch_size, num_heads, dp, mp)
+
+def test_cus_matmul_dds_model_parallel_mix_auto():
+    set_algo_parameters(fully_use_devices=False)
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 2
+    mp = 8
+    compile_graph(batch_size, num_heads, dp, mp, auto=True)
+
+def test_cus_matmul_dds_model_parallel_dp_auto():
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 16
+    mp = 1
+    compile_graph(batch_size, num_heads, dp, mp, auto=True)
+
+def test_cus_matmul_dds_model_parallel_mp_auto():
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 1
+    mp = 16
+    compile_graph(batch_size, num_heads, dp, mp, auto=True)
+
+def test_cus_matmul_dds_model_parallel_auto():
+    set_algo_parameters(fully_use_devices=False)
+    context.set_auto_parallel_context(device_num=16, global_rank=0)
+    batch_size = 128
+    num_heads = 32
+    dp = 1
+    mp = 16
+    compile_graph(batch_size, num_heads, dp, mp, auto=True, shard=False)