update

paddle/fluid/lite/CMakeLists.txt

-cc_library(executor_lite SRCS executor.cc)
-cc_library(op_lite SRCS op_lite.cc)
-cc_library(memory_lite SRCS memory.cc)
-cc_library(tensor_lite SRCS tensor.cc DEPS memory_lite)
-cc_library(variable_lite SRCS variable.cc)
-cc_library(op_registry_lite SRCS op_registry.cc)
-cc_library(scope_lite SRCS scope.cc)
-
+add_subdirectory(core)
 add_subdirectory(x86)
 add_subdirectory(cuda)
 add_subdirectory(operators)
 add_subdirectory(kernels)
 add_subdirectory(model_parser)
-
-
-# tests
-cc_test(test_scope_lite SRCS scope_test.cc DEPS scope_lite)

paddle/fluid/lite/core/CMakeLists.txt

+cc_library(executor_lite SRCS executor.cc)
+cc_library(op_lite SRCS op_lite.cc)
+cc_library(memory_lite SRCS memory.cc)
+cc_library(tensor_lite SRCS tensor.cc DEPS memory_lite)
+cc_library(variable_lite SRCS variable.cc)
+cc_library(op_registry_lite SRCS op_registry.cc)
+cc_library(scope_lite SRCS scope.cc)
+
+cc_test(test_scope_lite SRCS scope_test.cc DEPS scope_lite)

paddle/fluid/lite/context.h → paddle/fluid/lite/core/context.h

@@ -25,6 +25,7 @@ class Context {
  public:
   using target_wrapper_t = TargetWrapper<Target>;
   using stream_t = typename TargetWrapper<Target>::stream_t;
+  using event_t = typename TargetWrapper<Target>::event_t;
 
   Context() = default;
   Context(int device_id, stream_t compute_stream, stream_t data_stream)
@@ -35,15 +36,20 @@ class Context {
   void SetDeviceId(int device_id) { device_id_ = device_id; }
   void SetComputeStream(stream_t x) { compute_stream_ = x; }
   void SetDataStream(stream_t x) { data_stream_ = x; }
+  void SetDependEvents(const std::vector<event_t>& events) {
+    depend_events_ = events;
+  }
 
   int device_id() const { return device_id_; }
   stream_t compute_stream() const { return compute_stream_; }
   stream_t data_stream() const { return data_stream_; }
+  const std::vector<event_t>& depend_events() const { return depend_events_; }
 
  private:
-  int device_id_;
+  int device_id_{0};
   stream_t compute_stream_;
   stream_t data_stream_;
+  std::vector<event_t> depend_events_;
 };
 
 class OpContext final {

paddle/fluid/lite/op_kernel.h → paddle/fluid/lite/core/kernel.h

@@ -19,8 +19,8 @@
 #include <map>
 #include <string>
 #include "paddle/fluid/framework/op_desc.h"
-#include "paddle/fluid/lite/context.h"
-#include "paddle/fluid/lite/target_wrapper.h"
+#include "context.h"
+#include "target_wrapper.h"
 #include "paddle/fluid/lite/utils/all.h"
 
 namespace paddle {

paddle/fluid/lite/memory.cc → paddle/fluid/lite/core/memory.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/lite/memory.h"
+#include "memory.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/lite/op_lite.h → paddle/fluid/lite/core/op_lite.h

@@ -19,7 +19,7 @@
 #include <map>
 #include <string>
 #include "context.h"
-#include "op_kernel.h"
+#include "kernel.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/variable.h"
@@ -40,11 +40,13 @@ struct Registry {
  * to eliminate overhead of some operations in current framework.
  *
  * The Operator are designed as follows:
- * - it can has some members to hold the argument addresses,
- * - it should act just like a function call, no more logic should included.
+ * - it can has some members to hold the argument and some other computation
+ * resources,
+ * - it should act like a function call, no more logic included.
  */
 class OpLite : public Registry {
  public:
+  // The strategies to pick a kernel from candidates.
   enum class KernelStrategy {
     // Return the user specified one.
     kStatic = 0,
@@ -57,18 +59,25 @@ class OpLite : public Registry {
   OpLite() {}
   OpLite(std::unique_ptr<OpContext> &&x) : op_context_(std::move(x)) {}
 
+  // Check the shape.
   virtual bool CheckShape() const { return true; }
+  // Inference the outputs' shape.
   virtual bool InferShape() const { return true; }
+  // Run this operator.
   virtual bool Run() = 0;
+  // Build the operator, attach it with the runtime environment.
   virtual bool Build(const framework::OpDesc &opdesc,
                      framework::Scope *scope) = 0;
+  // Human-readable information.
   virtual std::string DebugString() const = 0;
 
+ protected:
+  // Specify the kernel to run by default.
   virtual void StaticPickKernel(
       const std::vector<TargetType> &valid_targets) = 0;
 
-  void PickBestKernel(const std::vector<TargetType> &valid_places,
-                      KernelStrategy kernel_strategy = KernelStrategy::kStatic);
+  void PickKernel(const std::vector<TargetType> &valid_places,
+                  KernelStrategy kernel_strategy = KernelStrategy::kStatic);
 
   // Create all the kernels for the valid targets.
   void CreateKernels();

paddle/fluid/lite/op_registry.h → paddle/fluid/lite/core/op_registry.h

@@ -15,9 +15,9 @@
 #include <memory>
 #include <string>
 #include <unordered_map>
-#include "paddle/fluid/lite/op_kernel.h"
-#include "paddle/fluid/lite/op_lite.h"
-#include "paddle/fluid/lite/target_wrapper.h"
+#include "kernel.h"
+#include "op_lite.h"
+#include "target_wrapper.h"
 #include "paddle/fluid/lite/utils/all.h"
 
 namespace paddle {

paddle/fluid/lite/scope.cc → paddle/fluid/lite/core/scope.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/lite/scope.h"
+#include "scope.h"
 #include "scope.h"
 
 namespace paddle {

paddle/fluid/lite/scope.h → paddle/fluid/lite/core/scope.h

@@ -19,7 +19,7 @@
 #include <unordered_map>
 #include <utility>
 #include <vector>
-#include "paddle/fluid/lite/variable.h"
+#include "variable.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/scope_test.cc → paddle/fluid/lite/core/scope_test.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/lite/scope.h"
+#include "scope.h"
 #include <gtest/gtest.h>
 
 namespace paddle {

paddle/fluid/lite/target_wrapper.h → paddle/fluid/lite/core/target_wrapper.h

@@ -19,6 +19,7 @@ namespace paddle {
 namespace lite {
 
 enum class TargetType { kHost = 0, kX86, kCUDA, kARM, kLastAsPlaceHolder };
+// Some helper macro to get a specific TargetType.
 #define TARGET(item__) paddle::lite::TargetType::item__
 #define TARGET_VAL(item__) static_cast<int>(TARGET(item__))
 
@@ -36,20 +37,22 @@ using ARM = Target<TargetType::kARM>;
 
 enum class PrecisionType { kFloat = 0, kInt8, kLastAsPlaceHolder };
 
+// Some helper macro to get a specific PrecisionType.
 #define PRECISION(item__) paddle::lite::PrecisionType::item__
 #define PRECISION_VAL(item__) static_cast<int>(PRECISION(item__))
 constexpr int kNumPrecisions =
     PRECISION_VAL(kLastAsPlaceHolder) - PRECISION_VAL(kFloat);
 
 // Event sync for multi-stream devices like CUDA and OpenCL.
+// For the devices without support of stream, leave it empty.
 template <TargetType Target>
 class Event {};
 
 // Memory copy directions.
 enum class IoDirection {
-  HtoH = 0,
-  HtoD,
-  DtoH,
+  HtoH = 0,  // Host to host
+  HtoD,      // Host to device
+  DtoH,      // Device to host
 };
 
 // This interface should be specified by each kind of target.

paddle/fluid/lite/variable.cc → paddle/fluid/lite/core/variable.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/lite/variable.h"
+#include "variable.h"
 
 namespace paddle {
 namespace lite {}  // namespace lite

paddle/fluid/lite/variable.h → paddle/fluid/lite/core/variable.h

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #pragma once
-#include "paddle/fluid/lite/tensor.h"
+#include "paddle/fluid/lite/core/tensor.h"
 #include "paddle/fluid/lite/utils/all.h"
 
 namespace paddle {

paddle/fluid/lite/kernels/host/fc_compute.cc

@@ -14,7 +14,7 @@
 
 #include "paddle/fluid/lite/kernels/host/fc_compute.h"
 #include <Eigen/Core>
-#include "paddle/fluid/lite/op_registry.h"
+#include "paddle/fluid/lite/core/op_registry.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/kernels/host/fc_compute.h

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #pragma once
-#include "paddle/fluid/lite/op_kernel.h"
+#include "paddle/fluid/lite/core/kernel.h"
 #include "paddle/fluid/lite/operators/fc_op.h"
 
 namespace paddle {

paddle/fluid/lite/kernels/host/relu_compute.cc

@@ -12,4 +12,4 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/lite/kernels/relu_compute.h"
+#include "paddle/fluid/lite/kernels/host/relu_compute.h"

paddle/fluid/lite/kernels/host/relu_compute.h

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #pragma once
-#include "paddle/fluid/lite/op_kernel.h"
+#include "paddle/fluid/lite/core/kernel.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/model_parser/model_parser.cc

@@ -14,9 +14,9 @@
 
 #include "paddle/fluid/lite/model_parser/model_parser.h"
 #include <fstream>
-#include "paddle/fluid/lite/scope.h"
-#include "paddle/fluid/lite/tensor.h"
-#include "paddle/fluid/lite/variable.h"
+#include "paddle/fluid/lite/core/scope.h"
+#include "paddle/fluid/lite/core/tensor.h"
+#include "paddle/fluid/lite/core/variable.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/operators/fc_op.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "fc_op.h"
-#include "paddle/fluid/lite/op_registry.h"
+#include "paddle/fluid/lite/core/op_registry.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/operators/fc_op.h

@@ -14,8 +14,8 @@
 
 #include <string>
 #include <vector>
-#include "paddle/fluid/lite/op_lite.h"
-#include "paddle/fluid/lite/tensor.h"
+#include "paddle/fluid/lite/core/op_lite.h"
+#include "paddle/fluid/lite/core/tensor.h"
 #include "paddle/fluid/lite/utils/all.h"
 
 namespace paddle {

paddle/fluid/lite/operators/relu_op.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/lite/operators/relu_op.h"
-#include "paddle/fluid/lite/op_registry.h"
+#include "paddle/fluid/lite/core/op_registry.h"
 
 namespace paddle {
 namespace lite {

paddle/fluid/lite/operators/relu_op.h

@@ -14,8 +14,8 @@
 
 #include <string>
 #include <vector>
-#include "paddle/fluid/lite/op_lite.h"
-#include "paddle/fluid/lite/tensor.h"
+#include "paddle/fluid/lite/core/op_lite.h"
+#include "paddle/fluid/lite/core/tensor.h"
 #include "paddle/fluid/lite/utils/all.h"
 
 namespace paddle {

paddle/fluid/lite/x86/target_wrapper.h

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #pragma once
-#include "paddle/fluid/lite/target_wrapper.h"
+#include "paddle/fluid/lite/core/target_wrapper.h"
 
 namespace paddle {
 namespace framework {