!1785 add dropout special kernel selected rules

Merge pull request !1785 from lianliguang/add-dropout-kernel-special-kernel-select-rules

mindspore/ccsrc/kernel/kernel.h

@@ -31,7 +31,7 @@ enum KernelType : int { UNKNOWN_KERNEL_TYPE = 0, AUTO_DIFF_KERNEL, AICPU_KERNEL,
 
 namespace kernel {
 
-enum Axis {
+enum Axis : int {
   N = 0,
   C,
   H,

mindspore/ccsrc/kernel/kernel_build_info.cc

@@ -167,5 +167,20 @@ void KernelBuildInfo::KernelBuildInfoBuilder::SetOpPattern(OpPattern pattern) {
   MS_EXCEPTION_IF_NULL(kernel_build_info_);
   kernel_build_info_->op_pattern_ = pattern;
 }
+void KernelBuildInfo::KernelBuildInfoBuilder::SetInputFormat(const std::string &format, size_t index) {
+  MS_EXCEPTION_IF_NULL(kernel_build_info_);
+  if (index >= kernel_build_info_->inputs_format_.size()) {
+    MS_LOG(EXCEPTION) << "index outof range!";
+  }
+  kernel_build_info_->inputs_format_[index] = format;
+}
+
+void KernelBuildInfo::KernelBuildInfoBuilder::SetOutputFormat(const std::string &format, size_t index) {
+  MS_EXCEPTION_IF_NULL(kernel_build_info_);
+  if (index >= kernel_build_info_->outputs_format_.size()) {
+    MS_LOG(EXCEPTION) << "index outof range!";
+  }
+  kernel_build_info_->outputs_format_[index] = format;
+}
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/kernel_build_info.h

@@ -131,6 +131,10 @@ class KernelBuildInfo::KernelBuildInfoBuilder {
 
   void SetOpPattern(OpPattern pattern);
 
+  void SetInputFormat(const std::string &format, size_t index);
+
+  void SetOutputFormat(const std::string &format, size_t index);
+
   std::shared_ptr<KernelBuildInfo> Build();
 
  private:

mindspore/ccsrc/kernel/kernel_query.cc

@@ -41,8 +41,16 @@ void FilterInvalidKernelInfo(const CNodePtr &kernel_node,
   } else {
     MS_LOG(WARNING) << "All kernel Info list does not match any kernel info ";
     for (size_t index = 0; index < kernel_info_list->size(); ++index) {
+      std::ostringstream buffer;
       MS_EXCEPTION_IF_NULL(kernel_info_list->at(index));
-      MS_LOG(WARNING) << "kernel [ " << index << " ] :" << kernel_info_list->at(index)->ToString();
+      if (AnfAlgo::GetOutputTensorNum(kernel_node) != kernel_info_list->at(index)->GetOutputNum()) {
+        buffer << "Kernel node's output size [" << AnfAlgo::GetOutputTensorNum(kernel_node) << "]"
+               << " cannot match the kernel's output size [" << kernel_info_list->at(index)->GetOutputNum() << "]";
+      } else {
+        buffer << "Kernel node's output size [" << AnfAlgo::GetInputTensorNum(kernel_node) << "]"
+               << " cannot match the kernel's output size [" << kernel_info_list->at(index)->GetInputNum() << "]";
+      }
+      MS_LOG(WARNING) << "kernel [ " << index << " ] :" << kernel_info_list->at(index)->ToString() << buffer.str();
     }
     kernel_info_list->clear();
     MS_LOG(WARNING) << "node" << kernel_node->DebugString() << "'s output size : ["

mindspore/ccsrc/operator/ops.cc

@@ -208,6 +208,7 @@ const PrimitivePtr kPrimLayerNormGrad = std::make_shared<Primitive>("LayerNormGr
 const PrimitivePtr kPrimLayerNormXBackprop = std::make_shared<Primitive>("LayerNormXBackprop");
 const PrimitivePtr kPrimLayerNormBetaGammaBackprop = std::make_shared<Primitive>("LayerNormBetaGammaBackprop");
 const PrimitivePtr kPrimDropoutGenMask = std::make_shared<Primitive>("DropoutGenMask");
+const PrimitivePtr kPrimDropoutDoMask = std::make_shared<Primitive>("DropoutDoMask");
 const PrimitivePtr kPrimOneHot = std::make_shared<Primitive>("OneHot");
 const PrimitivePtr kPrimGelu = std::make_shared<Primitive>("Gelu");
 const PrimitivePtr kPrimGeluGrad = std::make_shared<Primitive>("GeluGrad");

mindspore/ccsrc/operator/ops.h

@@ -214,6 +214,7 @@ extern const PrimitivePtr kPrimLayerNormGrad;
 extern const PrimitivePtr kPrimLayerNormXBackprop;
 extern const PrimitivePtr kPrimLayerNormBetaGammaBackprop;
 extern const PrimitivePtr kPrimDropoutGenMask;
+extern const PrimitivePtr kPrimDropoutDoMask;
 extern const PrimitivePtr kPrimOneHot;
 extern const PrimitivePtr kPrimGelu;
 extern const PrimitivePtr kPrimGeluGrad;

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -55,6 +55,7 @@
 #include "pre_activate/ascend/ir_fusion/confusion_mul_grad_fusion.h"
 #include "pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h"
 #include "pre_activate/ascend/format_type/insert_trans_op.h"
+#include "pre_activate/ascend/format_type/rectify_do_mask_kernel_info.h"
 #include "pre_activate/pass/getitem_tuple.h"
 #include "pre_activate/pass/optimize_dependence.h"
 #include "pre_activate/pass/erase_visit_attr.h"
@@ -82,7 +83,6 @@
 #include "pre_activate/ascend/format_type/deal_ref_trans_and_cast.h"
 #include "pre_activate/ascend/enhancer/add_memcpy_async.h"
 #include "pre_activate/ascend/enhancer/insert_pad_for_nms_with_mask.h"
-#include "pre_activate/ascend/format_type/insert_cast_for_runop.h"
 #include "pre_activate/ascend/format_type/insert_transdata_for_runop.h"
 #include "pre_activate/ascend/enhancer/getnext_memcpy_elimination.h"
 #include "pre_activate/ascend/ir_fission/addn_fission.h"
@@ -148,6 +148,7 @@ void RunOpAscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_g
   MS_EXCEPTION_IF_NULL(kernel_graph);
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto data_layout_pm = std::make_shared<PassManager>("pynative_transop_pm");
+  data_layout_pm->AddPass(std::make_shared<RectifyDoMaskKernelInfo>());
   data_layout_pm->AddPass(std::make_shared<RunOpInsertTransData>());
   data_layout_pm->AddPass(std::make_shared<GetitemTuple>());
   data_layout_pm->AddPass(std::make_shared<CommonSubexpressionElimination>());
@@ -160,30 +161,11 @@ void RunOpAscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_g
   kernel_graph->SetExecOrderByDefault();
 }
 
-void RunOpAscendMixPrecision(const std::shared_ptr<session::KernelGraph> &kernel_graph) {
-  MS_EXCEPTION_IF_NULL(kernel_graph);
-  auto optimizer = std::make_shared<GraphOptimizer>();
-  auto mixed_precision_pm = std::make_shared<PassManager>("pynative_transop_pm");
-  mixed_precision_pm->AddPass(std::make_shared<RunOpInsertCast>());
-  mixed_precision_pm->AddPass(std::make_shared<GetitemTuple>());
-  mixed_precision_pm->AddPass(std::make_shared<CommonSubexpressionElimination>());
-  mixed_precision_pm->AddPass(std::make_shared<EliminateRedundantOp>());
-  mixed_precision_pm->AddPass(std::make_shared<OptimizeDependence>());
-  mixed_precision_pm->AddPass(std::make_shared<EraseVisitAttr>());
-  mixed_precision_pm->AddPass(std::make_shared<DealRefTransAndCast>());
-  mixed_precision_pm->AddPass(std::make_shared<GetitemTuple>());
-  mixed_precision_pm->AddPass(std::make_shared<MergeCastToOp>());
-  mixed_precision_pm->AddPass(std::make_shared<LayerNormBetaGammaBackpropFusion>());
-  mixed_precision_pm->AddPass(std::make_shared<EraseVisitAttr>());
-  optimizer->AddPassManager(mixed_precision_pm);
-  (void)optimizer->Optimize(kernel_graph);
-  kernel_graph->SetExecOrderByDefault();
-}
-
 void AscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_graph) {
   MS_EXCEPTION_IF_NULL(kernel_graph);
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto data_layout_pm = std::make_shared<PassManager>("transop_pm");
+  data_layout_pm->AddPass(std::make_shared<RectifyDoMaskKernelInfo>());
   data_layout_pm->AddPass(std::make_shared<InsertTransOp>());
   data_layout_pm->AddPass(std::make_shared<GetitemTuple>());
   data_layout_pm->AddPass(std::make_shared<CommonSubexpressionElimination>());

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.h

@@ -20,7 +20,6 @@
 namespace mindspore {
 namespace opt {
 void RunOpAscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_graph);
-void RunOpAscendMixPrecision(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void RunOpAscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendMixPrecision(const std::shared_ptr<session::KernelGraph> &kernel_graph);

mindspore/ccsrc/pre_activate/ascend/ascend_helper.cc

@@ -65,7 +65,7 @@ AnfNodePtr AddTransOpNodeToGraph(const FuncGraphPtr &func_graph, const AnfNodePt
     dtype = AnfAlgo::GetInputDeviceDataType(cnode, insert_index);
     dst_format = AnfAlgo::GetInputFormat(cnode, insert_index);
     input_node = AnfAlgo::GetInputNode(cnode, insert_index);
-    padding_axis = AnfAlgo::GetInputReshapeType(node, 0);
+    padding_axis = AnfAlgo::GetInputReshapeType(node, insert_index);
   }
   bool need_padding = false;
   if (is_insert_input) {

mindspore/ccsrc/pre_activate/ascend/format_type/insert_cast_for_runop.cc

@@ -1,48 +0,0 @@
-/**
- * Copyright 2019 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 "pre_activate/ascend/format_type/insert_cast_for_runop.h"
-
-#include <memory>
-
-#include "device/kernel_info.h"
-#include "pre_activate/ascend/ascend_helper.h"
-#include "pre_activate/common/helper.h"
-#include "kernel/oplib/oplib.h"
-#include "session/anf_runtime_algorithm.h"
-#include "utils/utils.h"
-
-namespace mindspore {
-namespace opt {
-const BaseRef RunOpInsertCast::DefinePattern() const {
-  VarPtr V = std::make_shared<CondVar>(UnVisited);
-  VarPtr Xs = std::make_shared<SeqVar>();
-  return VectorRef({V, Xs});
-}
-
-const AnfNodePtr RunOpInsertCast::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
-                                          const EquivPtr &) const {
-  MS_EXCEPTION_IF_NULL(node);
-  if (!AnfAlgo::IsRealCNodeKernel(node) || func_graph == nullptr) {
-    return nullptr;
-  }
-  AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), node);
-  // process input
-  CNodePtr cnode = node->cast<CNodePtr>();
-  MS_EXCEPTION_IF_NULL(cnode);
-  return InsertCastForInput(func_graph, cnode);
-}
-}  // namespace opt
-}  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/format_type/insert_cast_for_runop.h

@@ -1,35 +0,0 @@
-/**
- * Copyright 2019 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_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_INSERT_CAST_FOR_RUNOP_H_
-#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_INSERT_CAST_FOR_RUNOP_H_
-#include <string>
-
-#include "pre_activate/common/optimizer.h"
-#include "pre_activate/common/pattern_engine.h"
-#include "ir/anf.h"
-namespace mindspore {
-namespace opt {
-class RunOpInsertCast : public PatternProcessPass {
- public:
-  explicit RunOpInsertCast(bool multigraph = true) : PatternProcessPass("insert_cast_for_runop", multigraph) {}
-  ~RunOpInsertCast() override = default;
-  const BaseRef DefinePattern() const override;
-  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
-};
-}  // namespace opt
-}  // namespace mindspore
-
-#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_INSERT_CAST_FOR_RUNOP_H_

mindspore/ccsrc/pre_activate/ascend/format_type/rectify_do_mask_kernel_info.cc

@@ -0,0 +1,154 @@
+/**
+ * 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 "pre_activate/ascend/format_type/rectify_do_mask_kernel_info.h"
+
+#include <vector>
+#include <map>
+#include <string>
+#include <memory>
+
+#include "session/anf_runtime_algorithm.h"
+#include "kernel/kernel_build_info.h"
+#include "utils/utils.h"
+#include "kernel/common_utils.h"
+#include "utils/context/ms_context.h"
+
+namespace mindspore {
+namespace opt {
+const BaseRef RectifyDoMaskKernelInfo::DefinePattern() const {
+  VarPtr X = std::make_shared<Var>();
+  VarPtr Xs = std::make_shared<SeqVar>();
+  return VectorRef({X, Xs});
+}
+
+const AnfNodePtr RectifyDoMaskKernelInfo::Process(const FuncGraphPtr &graph, const AnfNodePtr &node,
+                                                  const EquivPtr &) const {
+  if (node == nullptr || !node->isa<CNode>()) {
+    return nullptr;
+  }
+  auto cnode = node->cast<CNodePtr>();
+  auto ms_context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(ms_context);
+  if (ms_context->execution_mode() == kPynativeMode) {
+    if (AnfAlgo::GetCNodeName(cnode) != prim::kPrimDropoutDoMask->name()) {
+      return nullptr;
+    }
+    auto do_mask_input_format = AnfAlgo::GetInputFormat(node, 0);
+    if (do_mask_input_format != kOpFormat_DEFAULT) {
+      auto builder =
+        std::make_shared<kernel::KernelBuildInfo::KernelBuildInfoBuilder>(AnfAlgo::GetSelectKernelBuildInfo(node));
+      builder->SetInputFormat(kOpFormat_DEFAULT, 0);
+      builder->SetOutputFormat(kOpFormat_DEFAULT, 0);
+      AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), node.get());
+    }
+    return nullptr;
+  }
+  if (AnfAlgo::GetCNodeName(cnode) != prim::kPrimDropoutGenMask->name()) {
+    return nullptr;
+  }
+  std::vector<CNodePtr> do_mask_node_list;
+  auto manager = graph->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  auto node_map = manager->node_users();
+  auto iter = node_map.find(node);
+  if (iter == node_map.end()) {
+    MS_LOG(EXCEPTION) << "Cannot find the node " << node->DebugString() << " in the graph manager!";
+  }
+  auto gen_mask_output_nodes = iter->second;
+  for (const auto &output_node : gen_mask_output_nodes) {
+    if (AnfAlgo::GetCNodeName(output_node.first) == prim::kPrimDropoutDoMask->name()) {
+      auto output_cnode = output_node.first->cast<CNodePtr>();
+      do_mask_node_list.push_back(output_cnode);
+    }
+  }
+  std::vector<size_t> input_shape;
+  for (const auto &output_node : do_mask_node_list) {
+    if (input_shape.empty()) {
+      input_shape = AnfAlgo::GetPrevNodeOutputInferShape(output_node, 0);
+      continue;
+    }
+    auto shape = AnfAlgo::GetPrevNodeOutputInferShape(output_node, 0);
+    if (!kernel::IsSameShape(shape, input_shape)) {
+      MS_LOG(EXCEPTION) << "The DropOutGenMask connected with same genmask's shape must be equal!"
+                        << " GenMask " << node->DebugString();
+    }
+  }
+  RectifyKernelInfo(do_mask_node_list);
+  return nullptr;
+}
+
+void RectifyDoMaskKernelInfo::RectifyKernelInfo(const std::vector<CNodePtr> &do_mask_node_list) const {
+  std::map<std::string, size_t> format_counter;
+  std::string special_format;
+  std::string convert_format;
+  for (const auto &do_mask : do_mask_node_list) {
+    auto do_mask_data_format = AnfAlgo::GetInputFormat(do_mask, 0);
+    if (special_format.empty() && kNeedTransFormatSet.find(do_mask_data_format) != kNeedTransFormatSet.end()) {
+      special_format = do_mask_data_format;
+    }
+    if (format_counter.find(do_mask_data_format) == format_counter.end()) {
+      format_counter[do_mask_data_format] = 1;
+    } else {
+      format_counter[do_mask_data_format] = format_counter[do_mask_data_format] + 1;
+    }
+    // if has two or more special format we need change all domask's format to default that can avoid insert more
+    // transdata
+    if (format_counter.size() > 2) {
+      convert_format = kOpFormat_DEFAULT;
+      break;
+    }
+    if (kNeedTransFormatSet.find(do_mask_data_format) != kNeedTransFormatSet.end() &&
+        special_format != do_mask_data_format) {
+      convert_format = kOpFormat_DEFAULT;
+      break;
+    }
+  }
+  if (format_counter.size() == 1) {
+    return;
+  }
+  if (convert_format.empty()) {
+    convert_format = GetConvertFormat(format_counter);
+  }
+  RectifyDropOutDoMaskKernelInfo(do_mask_node_list, convert_format);
+}
+
+std::string RectifyDoMaskKernelInfo::GetConvertFormat(const std::map<std::string, size_t> &format_counter) const {
+  std::string convert_format;
+  size_t counter = 0;
+  for (const auto &iter : format_counter) {
+    if (counter < iter.second) {
+      convert_format = iter.first;
+    }
+    if (counter == iter.second && kNeedTransFormatSet.find(convert_format) == kNeedTransFormatSet.end()) {
+      convert_format = iter.first;
+    }
+  }
+  return convert_format;
+}
+void RectifyDoMaskKernelInfo::RectifyDropOutDoMaskKernelInfo(const std::vector<CNodePtr> &do_mask_node_list,
+                                                             const std::string &format) const {
+  for (const auto &do_mask : do_mask_node_list) {
+    auto builder =
+      std::make_shared<kernel::KernelBuildInfo::KernelBuildInfoBuilder>(AnfAlgo::GetSelectKernelBuildInfo(do_mask));
+    builder->SetInputFormat(format, 0);
+    builder->SetOutputFormat(format, 0);
+    AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), do_mask.get());
+  }
+}
+
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/format_type/rectify_do_mask_kernel_info.h

@@ -0,0 +1,41 @@
+/**
+ * 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_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_RECTIFY_DO_MASK_KERNEL_INFO_H
+#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_RECTIFY_DO_MASK_KERNEL_INFO_H
+#include <map>
+#include <string>
+#include <vector>
+
+#include "pre_activate/common/optimizer.h"
+namespace mindspore {
+namespace opt {
+class RectifyDoMaskKernelInfo : public PatternProcessPass {
+ public:
+  explicit RectifyDoMaskKernelInfo(bool multigraph = true)
+      : PatternProcessPass("batch_norm_bert_fission", multigraph) {}
+  ~RectifyDoMaskKernelInfo() override = default;
+  const BaseRef DefinePattern() const override;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+
+ private:
+  void RectifyKernelInfo(const std::vector<CNodePtr> &do_mask_node_list) const;
+  std::string GetConvertFormat(const std::map<std::string, size_t> &format_counter) const;
+  void RectifyDropOutDoMaskKernelInfo(const std::vector<CNodePtr> &do_mask_node_list, const std::string &format) const;
+};
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_RECTIFY_DO_MASK_KERNEL_INFO_H