!14513 executor and subgraph optimize

From: @ling_qiao_min
Reviewed-by: @zhang_xue_tong,@zhanghaibo5
Signed-off-by: @zhang_xue_tong

mindspore/lite/src/common/tensor_util.cc

@@ -215,5 +215,26 @@ int GenerateInTensorC(const OpParameter *const parameter, const std::vector<lite
   return ret;
 }
 
+int CheckTensorsInvalid(const std::vector<Tensor *> &tensors) {
+  for (auto &tensor : tensors) {
+    if (tensor == nullptr) {
+      MS_LOG(ERROR) << "check tensor is nullptr";
+      return RET_ERROR;
+    }
+    if (tensor->data_type() != kObjectTypeTensorType && tensor->data_c() == nullptr) {
+      MS_LOG(ERROR) << "check tensor data is nullptr " << tensors;
+      return RET_ERROR;
+    }
+    auto shape = tensor->shape();
+    bool valid = all_of(shape.begin(), shape.end(), [](int i) { return i >= 0; });
+    if (!valid) {
+      MS_LOG(ERROR) << "The shape of tensor contains negative dimension,"
+                    << "check the model and assign the input shape with method Resize().";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/common/tensor_util.h

@@ -39,6 +39,8 @@ int GenerateInTensorC(const OpParameter *const parameter, const std::vector<lite
                       std::vector<lite::Tensor *> *outputs, std::vector<TensorC *> *in_tensor_c);
 int GenerateOutTensorC(const OpParameter *const parameter, const std::vector<lite::Tensor *> &inputs,
                        std::vector<lite::Tensor *> *outputs, std::vector<TensorC *> *out_tensor_c);
+
+int CheckTensorsInvalid(const std::vector<Tensor *> &tensors);
 }  // namespace lite
 }  // namespace mindspore
 

mindspore/lite/src/executor.cc

@@ -17,34 +17,14 @@
 #include "src/executor.h"
 #include <queue>
 #include "include/errorcode.h"
+#include "src/common/tensor_util.h"
 
 namespace mindspore::lite {
-int Executor::CheckInputs(const std::vector<Tensor *> &in_tensors) {
-  for (auto &inTensor : in_tensors) {
-    if (inTensor == nullptr) {
-      MS_LOG(ERROR) << "Graph input tensor is nullptr";
-      return RET_ERROR;
-    }
-    if (inTensor->data_type() != kObjectTypeTensorType && inTensor->data_c() == nullptr) {
-      MS_LOG(ERROR) << "Graph input tensor data is nullptr " << in_tensors;
-      return RET_ERROR;
-    }
-    auto shape = inTensor->shape();
-    bool valid = all_of(shape.begin(), shape.end(), [](int i) { return i >= 0; });
-    if (!valid) {
-      MS_LOG(ERROR) << "The shape of input tensor contains negative dimension,"
-                    << "check the model and assign the input shape with method Resize().";
-      return RET_ERROR;
-    }
-  }
-  return RET_OK;
-}
-
 int Executor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
                   const std::vector<kernel::LiteKernel *> &kernels, mindspore::Allocator *allocator,
                   const KernelCallBack &before, const KernelCallBack &after) {
   MS_ASSERT(nullptr != allocator);
-  auto ret = this->CheckInputs(in_tensors);
+  auto ret = CheckTensorsInvalid(in_tensors);
   if (RET_OK != ret) {
     MS_LOG(ERROR) << "CheckInputs failed";
     return ret;
@@ -88,42 +68,4 @@ int Executor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<Ten
   }
   return RET_OK;
 }
-
-int CpuExecutor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
-                     const std::vector<kernel::LiteKernel *> &kernels, Allocator *allocator,
-                     const KernelCallBack &before, const KernelCallBack &after) {
-  MS_ASSERT(nullptr != allocator);
-  //  not check input for merge. too hard
-  if (kernels.front()->Type() != schema::PrimitiveType_Merge) {
-    auto ret = this->CheckInputs(in_tensors);
-    if (ret != RET_OK) {
-      MS_LOG(ERROR) << "CheckInputs failed";
-      return ret;
-    }
-  }
-#ifdef SUPPORT_TRAIN
-  for (auto out_tensor : out_tensors) {  // increase RefCount of output tensors, such that Run will not free them
-    out_tensor->set_ref_count(out_tensor->ref_count() + 1);
-  }
-#endif
-  for (auto *kernel : kernels) {
-    MS_ASSERT(nullptr != kernel);
-    auto ret = kernel->PreProcess();
-    if (RET_OK != ret) {
-      MS_LOG(ERROR) << "PreProcess kernel failed, name: " << kernel->name();
-      return ret;
-    }
-    ret = kernel->Run(before, after);
-    if (RET_OK != ret) {
-      MS_LOG(ERROR) << "run kernel failed, name: " << kernel->name();
-      return ret;
-    }
-    ret = kernel->PostProcess();
-    if (RET_OK != ret) {
-      MS_LOG(ERROR) << "PostProcess kernel failed, name: " << kernel->name();
-      return ret;
-    }
-  }
-  return RET_OK;
-}
 }  // namespace mindspore::lite

mindspore/lite/src/executor.h

@@ -33,20 +33,6 @@ class Executor {
   virtual int Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
                   const std::vector<kernel::LiteKernel *> &kernels, mindspore::Allocator *allocator = nullptr,
                   const KernelCallBack &before = nullptr, const KernelCallBack &after = nullptr);
-
- protected:
-  static int CheckInputs(const std::vector<Tensor *> &in_tensors);
-};
-
-class CpuExecutor : public Executor {
- public:
-  CpuExecutor() = default;
-  virtual ~CpuExecutor() = default;
-
-  int Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
-          const std::vector<kernel::LiteKernel *> &kernels, mindspore::Allocator *allocator = nullptr,
-          const KernelCallBack &before = nullptr, const KernelCallBack &after = nullptr) override;
 };
-
 }  // namespace mindspore::lite
 #endif

mindspore/lite/src/mindrt_executor.cc

@@ -18,6 +18,7 @@
 #include "src/mindrt_executor.h"
 #include "src/lite_mindrt.h"
 #include "include/errorcode.h"
+#include "src/common/tensor_util.h"
 
 namespace mindspore::lite {
 
@@ -62,7 +63,7 @@ int MindrtExecutor::Run(const std::vector<Tensor *> &in_tensors, const std::vect
                         const KernelCallBack &before, const KernelCallBack &after) {
   MS_ASSERT(nullptr != allocator);
   if (kernels.front()->Type() != schema::PrimitiveType_Merge) {
-    auto ret = this->CheckInputs(in_tensors);
+    auto ret = CheckTensorsInvalid(in_tensors);
     if (RET_OK != ret) {
       MS_LOG(ERROR) << "CheckInputs failed";
       return ret;

mindspore/lite/src/sub_graph_kernel.cc

@@ -22,6 +22,7 @@
 #endif
 #include "src/common/version_manager.h"
 #include "src/runtime/infer_manager.h"
+#include "src/common/tensor_util.h"
 
 namespace mindspore::kernel {
 using mindspore::lite::RET_ERROR;
@@ -70,19 +71,6 @@ std::string SubGraphKernel::ToString() const {
   return oss.str();
 }
 
-int SubGraphKernel::Run() {
-  if (this->executor_ == nullptr) {
-    MS_LOG(ERROR) << "executor is nullptr";
-    return RET_ERROR;
-  }
-  auto ret = executor_->Run(this->in_tensors_, this->out_tensors_, this->nodes_, this->context_->allocator.get());
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "Run sub graph failed: " << ret;
-    return ret;
-  }
-  return RET_OK;
-}
-
 int SubGraphKernel::Run(const KernelCallBack &before, const KernelCallBack &after) {
   if (this->executor_ == nullptr) {
     MS_LOG(ERROR) << "executor is nullptr";
@@ -165,15 +153,42 @@ int CpuSubGraph::Prepare() {
       tensor->set_allocator(this->context_->allocator.get());
     }
   }
-  this->executor_ = new (std::nothrow) mindspore::lite::CpuExecutor;
-  if (this->executor_ == nullptr) {
-    MS_LOG(ERROR) << "new CpuExecutor failed";
-    return RET_ERROR;
+  return RET_OK;
+}
+
+int CpuSubGraph::Run(const KernelCallBack &before, const KernelCallBack &after) {
+  MS_ASSERT(nullptr != this->context_->allocator.get());
+  if (nodes_.front()->Type() != schema::PrimitiveType_Merge) {
+    auto ret = CheckTensorsInvalid(in_tensors_);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "CheckInputs failed";
+      return ret;
+    }
   }
-  ret = this->executor_->Prepare(this->nodes_);
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "Prepare CpuExecutor failed";
-    return ret;
+
+#ifdef SUPPORT_TRAIN
+  for (auto out_tensor : out_tensors_) {  // increase RefCount of output tensors, such that Run will not free them
+    out_tensor->set_ref_count(out_tensor->ref_count() + 1);
+  }
+#endif
+
+  for (auto *kernel : nodes_) {
+    MS_ASSERT(nullptr != kernel);
+    auto ret = kernel->PreProcess();
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "PreProcess kernel failed, name: " << kernel->name();
+      return ret;
+    }
+    ret = kernel->Run(before, after);
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "run kernel failed, name: " << kernel->name();
+      return ret;
+    }
+    ret = kernel->PostProcess();
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "PostProcess kernel failed, name: " << kernel->name();
+      return ret;
+    }
   }
   return RET_OK;
 }

mindspore/lite/src/sub_graph_kernel.h

@@ -94,7 +94,7 @@ class SubGraphKernel : public LiteKernel {
   // called before Run
   int PreProcess() override { return mindspore::lite::RET_OK; }
 
-  int Run() override;
+  int Run() override { return Run(nullptr, nullptr); }
 
   int Run(const KernelCallBack &before, const KernelCallBack &after) override;
   // called after Run
@@ -134,10 +134,8 @@ class CpuSubGraph : public SubGraphKernel {
   int Prepare() override;
   int Init() override { return SubGraphKernel::Init(); }
   int PreProcess() override { return SubGraphKernel::PreProcess(); }
-  int Run() override { return SubGraphKernel::Run(); }
-  int Run(const KernelCallBack &before, const KernelCallBack &after) override {
-    return SubGraphKernel::Run(before, after);
-  };
+  int Run() override { return Run(nullptr, nullptr); }
+  int Run(const KernelCallBack &before, const KernelCallBack &after) override;
   int PostProcess() override { return SubGraphKernel::PostProcess(); }
 };
 

mindspore/lite/src/train/train_session.cc

@@ -35,6 +35,7 @@
 #include "src/kernel_registry.h"
 #include "src/runtime/kernel/arm/fp32_grad/convolution.h"
 #include "src/runtime/kernel/arm/fp32/batchnorm_fp32.h"
+#include "src/common/tensor_util.h"
 
 namespace mindspore {
 namespace lite {
@@ -188,17 +189,33 @@ int TrainSession::RunGraph(const KernelCallBack &before, const KernelCallBack &a
     return lite::RET_NULL_PTR;
   }
   auto run_kernel = (train_mode_) ? train_kernels_ : inference_kernels_;
-  lite::CpuExecutor executor;
-  auto ret = RET_OK;
-  if (before == nullptr && after == nullptr) {
-    ret = executor.Run(this->inputs_, this->outputs_, run_kernel, this->context_->allocator.get());
-  } else {
-    ret = executor.Run(this->inputs_, this->outputs_, run_kernel, this->context_->allocator.get(), before, after);
-  }
+
+  auto ret = CheckTensorsInvalid(inputs_);
   if (ret != RET_OK) {
-    MS_LOG(ERROR) << "failed to run model";
+    MS_LOG(ERROR) << "CheckInputs failed";
     return ret;
   }
+  for (auto out_tensor : outputs_) {  // increase RefCount of output tensors, such that Run will not free them
+    out_tensor->set_ref_count(out_tensor->ref_count() + 1);
+  }
+  for (auto *kernel : run_kernel) {
+    MS_ASSERT(nullptr != kernel);
+    ret = kernel->PreProcess();
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "PreProcess kernel failed, name: " << kernel->name();
+      return ret;
+    }
+    ret = kernel->Run(before, after);
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "run kernel failed, name: " << kernel->name();
+      return ret;
+    }
+    ret = kernel->PostProcess();
+    if (RET_OK != ret) {
+      MS_LOG(ERROR) << "PostProcess kernel failed, name: " << kernel->name();
+      return ret;
+    }
+  }
 
   if (train_mode_ && virtual_batch_multiplier_) {
     virtual_batch_idx_++;