"cudnn operators change to cudnn kernel"

paddle/framework/data_layout.h

@@ -14,7 +14,9 @@ limitations under the License. */
 
 #pragma once
 
-#include <iostream>
+#include <cctype>
+#include <ostream>
+
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -27,12 +29,19 @@ enum class DataLayout {
 };
 
 inline DataLayout StringToDataLayout(const std::string& str) {
-  if (str == "NHWC" || str == "nhwc") {
+  std::string s(str);
+  for (size_t i = 0; i < s.size(); ++i) {
+    s[i] = toupper(s[i]);
+  }
+
+  if (s == "NHWC") {
     return DataLayout::kNHWC;
-  } else if (str == "NCHW" || str == "nchw") {
+  } else if (s == "NCHW") {
     return DataLayout::kNCHW;
+  } else if (s == "ANYLAYOUT") {
+    return DataLayout::kAnyLayout;
   } else {
-    PADDLE_THROW("Unknown storage order string: %s", str);
+    PADDLE_THROW("Unknown storage order string: %s", s);
   }
 }
 
@@ -49,7 +58,7 @@ inline std::string DataLayoutToString(const DataLayout& data_layout) {
   }
 }
 
-inline std::ostream& operator<<(std::ostream& out, DataLayout l) {
+inline std::ostream& operator<<(std::ostream& out, const DataLayout& l) {
   out << DataLayoutToString(l);
   return out;
 }

paddle/framework/data_transform.cc

@@ -19,16 +19,19 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-Tensor* DataTransform(const OpKernelType& expected_kernel_type,
-                      const OpKernelType& kernel_type_for_var,
-                      const Tensor& input_tensor) {
-  Tensor* out = nullptr;
+void DataTransform(const OpKernelType& expected_kernel_type,
+                   const OpKernelType& kernel_type_for_var,
+                   const Tensor& input_tensor, Tensor* out) {
+  if (expected_kernel_type.library_type_ != kernel_type_for_var.library_type_) {
+    Copy(input_tensor, expected_kernel_type.place_, out);
+    return;
+  }
+
   if (!platform::is_same_place(kernel_type_for_var.place_,
                                expected_kernel_type.place_)) {
     out = DeviceTransform(input_tensor, expected_kernel_type.place_);
   }
   PADDLE_ENFORCE_NOT_NULL(out, "out should not be null");
-  return out;
 }
 
 void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,

paddle/framework/data_transform.h

@@ -30,9 +30,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-Tensor* DataTransform(const OpKernelType& expected_kernel_type,
-                      const OpKernelType& kernel_type_for_var,
-                      const Tensor& input_tensor);
+void DataTransform(const OpKernelType& expected_kernel_type,
+                   const OpKernelType& kernel_type_for_var,
+                   const Tensor& input_tensor, Tensor* out);
 
 void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
                             Variable& out_var);

paddle/framework/operator.cc

@@ -26,51 +26,6 @@ namespace framework {
 
 std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
 
-void UseCPU() {
-  kKernelPriority.clear();
-  /*Plain CPU*/
-  auto pair0 = std::make_tuple(platform::CPUPlace(), LibraryType::kPlain);
-  kKernelPriority.insert(kKernelPriority.begin(), pair0);
-}
-
-void UseMKLDNN() {
-  UseCPU();
-#if PADDLE_WITH_MKLML
-  {
-    /*MKLDNN Kernel*/
-    auto pair0 = std::make_tuple(platform::CPUPlace(), LibraryType::kMKLDNN);
-    kKernelPriority.insert(kKernelPriority.begin(), pair0);
-  }
-#endif
-}
-
-void UseCUDA() {
-  UseMKLDNN();
-#if PADDLE_WITH_CUDA
-  /*Plain GPU*/
-  auto pair0 = std::make_tuple(platform::CUDAPlace(0), LibraryType::kPlain);
-  kKernelPriority.insert(kKernelPriority.begin(), pair0);
-#endif
-}
-
-void UseCUDNN() {
-  UseCUDA();
-#if PADDLE_WITH_CUDA
-  if (platform::dynload::HasCUDNN()) {
-    /*CUDNN Kernel*/
-    auto pair0 = std::make_tuple(platform::CUDAPlace(0), LibraryType::kCUDNN);
-    kKernelPriority.insert(kKernelPriority.begin(), pair0);
-  }
-#endif
-}
-
-void UseALL() {
-  UseCPU();
-  UseMKLDNN();
-  UseCUDA();
-  UseCUDNN();
-}
-
 static DDim GetDims(const Scope& scope, const std::string& name) {
   Variable* var = scope.FindVar(name);
   if (var == nullptr) {
@@ -247,36 +202,33 @@ static bool VarIsTensor(const Variable* var) {
   return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
 }
 
-static const Tensor* GetTensorFromVar(const Variable* var) {
-  const Tensor* t = nullptr;
+static const Tensor* GetTensorFromVar(Variable* var) {
   if (var->IsType<LoDTensor>()) {
-    t = &(var->Get<LoDTensor>());
+    return var->GetMutable<LoDTensor>();
   } else if (var->IsType<SelectedRows>()) {
-    t = &(var->Get<SelectedRows>().value());
+    return var->GetMutable<SelectedRows>()->mutable_value();
   } else {
     PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                  var->Type().name());
   }
-  return t;
 }
 
 static Tensor* GetMutableTensorFromVar(Variable* var) {
-  Tensor* t = nullptr;
   if (var->IsType<LoDTensor>()) {
-    t = var->GetMutable<LoDTensor>();
+    return var->GetMutable<LoDTensor>();
   } else if (var->IsType<SelectedRows>()) {
-    t = var->GetMutable<SelectedRows>()->mutable_value();
+    return var->GetMutable<SelectedRows>()->mutable_value();
   } else {
     PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
                  var->Type().name());
   }
-  return t;
 }
 
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   auto* var = InputVar(name);
-  return var == nullptr ? nullptr : GetTensorFromVar(var);
+  return var == nullptr ? nullptr
+                        : GetTensorFromVar(const_cast<Variable*>(var));
 }
 
 template <>
@@ -319,6 +271,7 @@ bool OpSupportGPU(const std::string& op_type) {
   auto it = all_kernels.find(op_type);
   if (it == all_kernels.end()) {
     // All control operator must support GPU
+
     return true;
   }
   for (auto& kern_pair : it->second) {
@@ -492,58 +445,64 @@ void OperatorWithKernel::Run(const Scope& scope,
   }
 
   ExecutionContext ctx(*this, scope, *dev_ctx);
-  auto expected_kernel_key = this->GetExpectedKernelType(ctx);
 
   OpKernelMap& kernels = kernels_iter->second;
 
-  for (auto& candidate : kKernelPriority) {
-    auto candidate_key =
-        OpKernelType(expected_kernel_key.data_type_, std::get<0>(candidate),
-                     expected_kernel_key.data_layout_, std::get<1>(candidate));
+  // TODO(dzhwinter) : kernel fallback mechanism will be added when all the
+  // transform functions are ready.
 
-    if ((candidate_key == expected_kernel_key) ||
-        (kernels.count(candidate_key))) {
-      expected_kernel_key = candidate_key;
-      break;
-    }
-  }
+  // for (auto& candidate : kKernelPriority) {
+  //   Do selection
+  // }
+
+  auto expected_kernel_key = this->GetExpectedKernelType(ctx);
 
   VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
   Scope& new_scope = scope.NewScope();
 
-  for (auto& var_name_item : this->Inputs()) {
-    for (auto& var_name : var_name_item.second) {
-      auto* var = scope.FindVar(var_name);
-      if (var && VarIsTensor(var)) {
-        auto* tensor_in = GetTensorFromVar(var);
-        if (tensor_in->IsInitialized()) {
-          auto kernel_type_for_var = this->GetKernelTypeForVar(
-              var_name_item.first, *tensor_in, expected_kernel_key);
-          if (kernel_type_for_var != expected_kernel_key) {
-            auto out_var_names = OutputVars(true);
-            if (std::find(out_var_names.begin(), out_var_names.end(),
-                          var_name) != out_var_names.end()) {
-              PADDLE_THROW(
-                  "var %s is both input and output, "
-                  "does not support transform",
-                  var_name);
-            }
-            VLOG(3) << "need to do transform for var " << var_name;
-            auto* trans_var = new_scope.Var(var_name);
-            auto* out = DataTransform(expected_kernel_key, kernel_type_for_var,
-                                      *tensor_in);
-            CopyVariableWithTensor(*var, *out, *trans_var);
-          }
-        }
-      }
-    }
+  if (expected_kernel_key.library_type_ != LibraryType::kCUDNN) {
+    // for (auto& var_name_item : this->Inputs()) {
+    //   for (auto& var_name : var_name_item.second) {
+    //     auto* var = scope.FindVar(var_name);
+    //     if (var && VarIsTensor(var)) {
+    //       auto* tensor_in = GetTensorFromVar(var);
+    //       if (tensor_in->IsInitialized()) {
+    //         auto kernel_type_for_var = this->GetKernelTypeForVar(
+    //             var_name_item.first, *tensor_in, expected_kernel_key);
+    //         if (kernel_type_for_var != expected_kernel_key) {
+    //           auto out_var_names = OutputVars(true);
+    //           if (std::find(out_var_names.begin(), out_var_names.end(),
+    //                         var_name) != out_var_names.end()) {
+    //             PADDLE_THROW(
+    //                 "var %s is both input and output, "
+    //                 "does not support transform",
+    //                 var_name);
+    //           }
+    //           VLOG(3) << "need to do transform for var " << var_name;
+    //           auto* trans_var = new_scope.Var(var_name);
+    //           std::shared_ptr<Tensor> out(new Tensor);
+    //           DataTransform(expected_kernel_key, kernel_type_for_var,
+    //                         *tensor_in, out.get());
+    //           CopyVariableWithTensor(*var, *(out.get()), *trans_var);
+    //         }
+    //       }
+    //     }
+    //   }
+    // }
   }
 
   auto kernel_iter = kernels.find(expected_kernel_key);
 
-  kernel_iter->second->Compute(ExecutionContext(
-      *this, new_scope, *pool.Get(expected_kernel_key.place_)));
+  auto new_ctx =
+      ExecutionContext(*this, new_scope, *pool.Get(expected_kernel_key.place_));
+  VLOG(3) << "construct ExecutionContext ";
+  if (kernel_iter == kernels.end()) {
+    VLOG(3) << " Not such kernel";
+  }
+  auto& final_kernel = kernel_iter->second;
+  VLOG(3) << " before compute ";
+  final_kernel->Compute(new_ctx);
 }
 
 proto::DataType OperatorWithKernel::IndicateDataType(

paddle/framework/operator.h

@@ -54,33 +54,9 @@ constexpr char kGradVarSuffix[] = "@GRAD";
 constexpr char kZeroVarSuffix[] = "@ZERO";
 
 // define some kernel priority
+/* Define multiple kernel type fallback order*/
 extern std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
 
-/**
- * @brief Use cpu kernel only
- */
-void UseCPU();
-
-/**
- * @brief Perfer MKLDNN kernel than Plain CPU kernel
- */
-void UseMKLDNN();
-
-/**
- * @brief Perfer CUDA kernel than Plain CPU kernel
- */
-void UseCUDA();
-
-/**
- * @brief Perfer cudnn kernel than Plain CUDA kernel
- */
-void UseCUDNN();
-
-/**
- * @brief Use all available kernels
- */
-void UseALL();
-
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
 }

paddle/operators/CMakeLists.txt

@@ -136,8 +136,6 @@ op_library(sequence_softmax_op DEPS softmax)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)
 op_library(adagrad_op DEPS selected_rows_functor)
-op_library(conv_op DEPS vol2col)
-op_library(pool_op DEPS pooling)
 op_library(maxout_op DEPS maxouting)
 op_library(unpool_op DEPS unpooling)
 op_library(pool_with_index_op DEPS pooling)
@@ -148,12 +146,21 @@ op_library(max_sequence_len_op DEPS lod_rank_table)
 op_library(sequence_conv_op DEPS context_project)
 op_library(sequence_pool_op DEPS sequence_pooling)
 op_library(lstm_op DEPS sequence2batch lstm_compute)
-op_library(conv_transpose_op DEPS vol2col)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op DEPS executor)
 op_library(warpctc_op DEPS dynload_warpctc sequence_padding math_function)
 op_library(cos_sim_op DEPS cos_sim_functor)
 op_library(parallel_do_op DEPS executor)
+
+# Regist multiple Kernel to pybind
+op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS vol2col)
+op_library(pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling)
+op_library(conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
+  conv_transpose_cudnn_op.cu.cc DEPS vol2col)
+file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d, CUDNN);\n")
+file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(pool2d, CUDNN);\n")
+file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d_transpose, CUDNN);\n")
+
 # FIXME(typhoonzero): save/load depends lodtensor serialization functions
 op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)