cache memory in local scope (#7058)

* add KernelTypeToString interface

* cache memory in local scope

* fix typo

* refine trans logic

paddle/framework/data_transform.h

@@ -27,7 +27,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-using DataTransformFN =
+using DataTransformFn =
     std::function<void(const std::vector<platform::DeviceContext*> ctx,
                        const Variable& in, Variable* out)>;
 using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
@@ -47,7 +47,7 @@ struct KernelTypePairHash {
 };
 
 using DataTransformMap =
-    std::unordered_map<KernelTypePair, DataTransformFN, KernelTypePairHash>;
+    std::unordered_map<KernelTypePair, DataTransformFn, KernelTypePairHash>;
 
 class DataTransformFnMap {
  public:
@@ -58,25 +58,25 @@ class DataTransformFnMap {
   }
 
   void Insert(const OpKernelType& left, const OpKernelType& right,
-              const DataTransformFN& data_tranform_fn) {
+              const DataTransformFn& data_tranform_fn) {
     Insert(std::make_pair(left, right), data_tranform_fn);
   }
 
   void Insert(const KernelTypePair& kernel_type_pair,
-              const DataTransformFN& data_tranform_fn) {
+              const DataTransformFn& data_tranform_fn) {
     PADDLE_ENFORCE(!Has(kernel_type_pair),
                    "KernelTypePair %s has been registered", "");
     map_.insert({kernel_type_pair, data_tranform_fn});
   }
 
-  const DataTransformFN& Get(const KernelTypePair& key_pair) const {
+  const DataTransformFn& Get(const KernelTypePair& key_pair) const {
     auto data_transformer = GetNullable(key_pair);
     PADDLE_ENFORCE_NOT_NULL(data_transformer,
-                            "DataTransformFN should not be NULL");
+                            "DataTransformFn should not be NULL");
     return *data_transformer;
   }
 
-  const DataTransformFN* GetNullable(const KernelTypePair& key_pair) const {
+  const DataTransformFn* GetNullable(const KernelTypePair& key_pair) const {
     auto it = map_.find(key_pair);
     if (it == map_.end()) {
       return nullptr;

paddle/framework/op_kernel_type.h

@@ -68,6 +68,8 @@ struct OpKernelType {
            data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
            library_type_ == o.library_type_;
   }
+
+  bool operator!=(const OpKernelType& o) const { return !(*this == o); }
 };
 
 inline std::ostream& operator<<(std::ostream& os,
@@ -78,5 +80,11 @@ inline std::ostream& operator<<(std::ostream& os,
   return os;
 }
 
+inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
+  std::ostringstream stream;
+  stream << kernel_key;
+  return stream.str();
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_kernel_type_test.cc

@@ -26,10 +26,8 @@ TEST(OpKernelType, ToString) {
   OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
                               LibraryType::kCUDNN);
 
-  std::ostringstream stream;
-  stream << op_kernel_type;
   ASSERT_EQ(
-      stream.str(),
+      paddle::framework::KernelTypeToString(op_kernel_type),
       "data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
 }
 

paddle/framework/operator.cc

@@ -413,37 +413,51 @@ void OperatorWithKernel::Run(const Scope& scope,
   }
 
   if (actual_kernel_key == expected_kernel_key) {
-    kernel_iter->second->Compute(ctx);
+    PADDLE_ENFORCE_EQ(actual_kernel_key.place_, expected_kernel_key.place_,
+                      "Currently, model parallelism is only supported between "
+                      "CPU and other devices. For example, multi-GPU model "
+                      "parallelism will failed.");
   } else {
-    Scope& op_scope = scope.NewScope();
+    const DataTransformFn* trans_fun =
+        DataTransformFnMap::Instance().GetNullable(
+            std::make_pair(actual_kernel_key, expected_kernel_key));
+    if (trans_fun) {
       auto input_vars = this->InputVars();
+      // TODO(qijun) filter the input vars that do not need to be transformed
+
+      // filter vars that has been transformed
+      std::vector<std::string> need_trans;
       for (auto var_name : input_vars) {
-      op_scope.Var(var_name);
+        auto var_name_trans =
+            var_name + framework::KernelTypeToString(expected_kernel_key);
+        if (!scope.FindVar(var_name_trans)) {
+          const_cast<Scope&>(scope).Var(var_name_trans);
+          need_trans.push_back(var_name);
+        }
       }
 
+      if (!need_trans.empty()) {
         // TODO(qijun) get appropriate DeviceContext from DeviceContext pool
         platform::DeviceContext* trans_dev_ctx = nullptr;
         std::vector<platform::DeviceContext*> trans_dev_ctx_vec{trans_dev_ctx};
 
-    // TODO(qijun) get appropriate DataTransformFN from global map
-    framework::DataTransformFN trans_fun = nullptr;
-
         // Wait for transform starting
         dev_ctx->Wait();
 
-    for (auto var_name : input_vars) {
-      trans_fun(trans_dev_ctx_vec, *(scope.FindVar(var_name)),
-                op_scope.FindVar(var_name));
+        for (auto var_name : need_trans) {
+          (*trans_fun)(trans_dev_ctx_vec, *(scope.FindVar(var_name)),
+                       scope.FindVar(var_name + framework::KernelTypeToString(
+                                                    expected_kernel_key)));
         }
         // Wait for data transform finishing
         for (auto ctx : trans_dev_ctx_vec) {
           ctx->Wait();
         }
-
-    // Create a new ExecutionContext
-    ExecutionContext op_ctx(*this, op_scope, *dev_ctx);
-    kernel_iter->second->Compute(op_ctx);
       }
+    }
+  }
+
+  kernel_iter->second->Compute(ctx);
 }
 
 OpKernelType OperatorWithKernel::GetActualKernelType(