Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into mobilenet_gpu

cmake/generic.cmake

@@ -93,7 +93,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 if(NOT APPLE)
     find_package(Threads REQUIRED)
     link_libraries(${CMAKE_THREAD_LIBS_INIT})
-    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl")
+    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
 endif(NOT APPLE)
 
 function(merge_static_libs TARGET_NAME)

paddle/framework/CMakeLists.txt

@@ -12,7 +12,7 @@ cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
 cc_library(operator SRCS operator.cc DEPS op_desc protobuf)
-cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
+cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry place)
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto op_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry operator)
 py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)

paddle/framework/op_registry.h

@@ -147,13 +147,13 @@ class OpRegisterHelper {
   }
 };
 
-#define REGISTER_OP(__op_class, __op_maker_class, __op_type)         \
+#define REGISTER_OP(type, op_class, op_maker_class)                         \
-  class __op_class##Register {                                       \
+  class op_class##Register {                                                \
-   private:                                                          \
+   private:                                                                 \
-    const static OpRegisterHelper<__op_class, __op_maker_class> reg; \
+    const static OpRegisterHelper<op_class, op_maker_class> reg;            \
-  };                                                                 \
+  };                                                                        \
-  const OpRegisterHelper<__op_class, __op_maker_class>               \
+  const OpRegisterHelper<op_class, op_maker_class> op_class##Register::reg( \
-      __op_class##Register::reg(#__op_type);
+      #type)
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry_test.cc

@@ -26,7 +26,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-REGISTER_OP(CosineOp, CosineOpProtoAndCheckerMaker, cos_sim)
+REGISTER_OP(cos_sim, CosineOp, CosineOpProtoAndCheckerMaker);
 
 class MyTestOp : public OperatorBase {
  public:
@@ -53,7 +53,7 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-REGISTER_OP(MyTestOp, MyTestOpProtoAndCheckerMaker, my_test_op)
+REGISTER_OP(my_test_op, MyTestOp, MyTestOpProtoAndCheckerMaker);
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/operator_test.cc

@@ -45,7 +45,7 @@ class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-REGISTER_OP(OperatorTest, OperatorTestProtoAndCheckerMaker, test_operator)
+REGISTER_OP(test_operator, OperatorTest, OperatorTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
   OpDesc op_desc;
@@ -69,5 +69,55 @@ TEST(OperatorBase, all) {
   delete op;
 }
 
+class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+ public:
+  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of test op");
+    AddOutput("output", "output of test op");
+    AddAttr<float>("scale", "scale of cosine op")
+        .SetDefault(1.0)
+        .LargerThan(0.0);
+    AddType("test_operator");
+    AddComment("This is test op");
+  }
+};
+
+class OpWithKernelTest : public OperatorWithKernel {
+ public:
+  void InferShape(const std::shared_ptr<Scope>& scope) const override {}
+};
+
+class CPUKernelTest : public OpKernel {
+ public:
+  void Compute(const KernelContext& context) const {
+    float scale = context.op_.GetAttr<float>("scale");
+    ASSERT_NEAR(scale, 3.14, 1e-5);
+    std::cout << "this is cpu kernel" << std::endl;
+    std::cout << context.op_.DebugString() << std::endl;
+  }
+};
+
+REGISTER_OP(op_with_kernel, OpWithKernelTest, OpKernelTestProtoAndCheckerMaker);
+REGISTER_OP_KERNEL(op_with_kernel, platform::CPUPlace, CPUKernelTest);
+
+TEST(OpKernel, all) {
+  OpDesc op_desc;
+  op_desc.set_type("op_with_kernel");
+  *op_desc.mutable_inputs()->Add() = "IN1";
+  *op_desc.mutable_outputs()->Add() = "OUT1";
+  auto attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("scale");
+  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_f(3.14);
+
+  platform::CPUDeviceContext cpu_device_context;
+  auto scope = std::make_shared<Scope>();
+
+  OperatorBase* op = paddle::framework::OpRegistry::CreateOp(op_desc);
+  op->Run(scope, cpu_device_context);
+
+  delete op;
+}
 }  // namespace framework
 }  // namespace paddle

paddle/platform/CMakeLists.txt

@@ -1,11 +1,12 @@
 add_subdirectory(dynload)
 
-nv_test(cuda_test SRCS cuda_test.cu DEPS dyload_cuda)
+nv_test(cuda_test SRCS cuda_test.cu)
 
 cc_library(place SRCS place.cc)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
+
 IF(WITH_GPU)
-    set(GPU_CTX_DEPS dyload_cuda dynamic_loader )
+    set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
 ELSE()
     set(GPU_CTX_DEPS)
 ENDIF()

paddle/platform/device_context.cc

@@ -1,13 +1,30 @@
-#include <paddle/platform/device_context.h>
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+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 "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace platform {
-namespace dynload {
-namespace dummy {
-// Make DeviceContext A library.
-int DUMMY_VAR_FOR_DEV_CTX = 0;
 
-}  // namespace dummy
+template <>
-}  // namespace dynload
+Eigen::DefaultDevice* DeviceContext::get_eigen_device<Eigen::DefaultDevice>() {
+  return reinterpret_cast<CPUDeviceContext*>(this)->eigen_device();
+}
+
+#ifndef PADDLE_ONLY_CPU
+template <>
+Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() {
+  return reinterpret_cast<CUDADeviceContext*>(this)->eigen_device();
+}
+#endif
+
 }  // namespace platform
-}  // namespace paddle
+}  // namespace paddle

paddle/platform/device_context.h

@@ -1,11 +1,8 @@
 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
 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.
@@ -23,6 +20,7 @@ limitations under the License. */
 #define EIGEN_USE_GPU
 #endif
 #include <paddle/platform/place.h>
+#include <memory>
 #include <unsupported/Eigen/CXX11/Tensor>
 
 namespace paddle {
@@ -32,14 +30,27 @@ class DeviceContext {
  public:
   virtual ~DeviceContext() {}
   virtual Place GetPlace() const = 0;
+
+  template <typename DeviceType>
+  DeviceType* get_eigen_device();
 };
 
 class CPUDeviceContext : public DeviceContext {
  public:
+  Eigen::DefaultDevice* eigen_device() {
+    if (!eigen_device_) {
+      eigen_device_.reset(new Eigen::DefaultDevice());
+    }
+    return eigen_device_.get();
+  }
+
   Place GetPlace() const override {
     Place retv = CPUPlace();
     return retv;
   }
+
+ private:
+  std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
 };
 
 #ifndef PADDLE_ONLY_CPU
@@ -64,8 +75,8 @@ class CUDADeviceContext : public DeviceContext {
     GPUPlaceGuard guard(gpu_place_);
     paddle::platform::throw_on_error(cudaStreamCreate(&stream_),
                                      "cudaStreamCreate failed");
-    eigen_stream_ = new Eigen::CudaStreamDevice(&stream_);
+    eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
-    eigen_device_ = new Eigen::GpuDevice(eigen_stream_);
+    eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
   }
 
   Place GetPlace() const override {
@@ -80,7 +91,7 @@ class CUDADeviceContext : public DeviceContext {
 
   cudaStream_t stream() { return stream_; }
 
-  Eigen::GpuDevice eigen_device() { return *eigen_device_; }
+  Eigen::GpuDevice* eigen_device() { return eigen_device_.get(); }
 
   cublasHandle_t cublas_handle() {
     if (!blas_handle_) {
@@ -145,10 +156,8 @@ class CUDADeviceContext : public DeviceContext {
                          rand_generator_) == CURAND_STATUS_SUCCESS,
                      "curandDestroyGenerator failed");
     }
-
+    eigen_stream_.reset();
-    delete eigen_stream_;
+    eigen_device_.reset();
-    delete eigen_device_;
-
     paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
                                      "cudaStreamDestroy failed");
   }
@@ -157,8 +166,8 @@ class CUDADeviceContext : public DeviceContext {
   GPUPlace gpu_place_;
   cudaStream_t stream_;
 
-  Eigen::CudaStreamDevice* eigen_stream_;
+  std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
-  Eigen::GpuDevice* eigen_device_;
+  std::unique_ptr<Eigen::GpuDevice> eigen_device_;
 
   cublasHandle_t blas_handle_{nullptr};
 
@@ -167,6 +176,8 @@ class CUDADeviceContext : public DeviceContext {
   int random_seed_;
   curandGenerator_t rand_generator_{nullptr};
 };
+
 #endif
+
 }  // namespace platform
 }  // namespace paddle

paddle/platform/device_context_test.cc

@@ -15,13 +15,26 @@ limitations under the License. */
 #include "paddle/platform/device_context.h"
 #include "gtest/gtest.h"
 
-TEST(CUDADeviceContext, Init) {
+using DEVICE_GPU = Eigen::GpuDevice;
+TEST(Device, Init) {
+  int count = paddle::platform::GetDeviceCount();
+  for (int i = 0; i < count; i++) {
+    paddle::platform::DeviceContext* device_context =
+        new paddle::platform::CUDADeviceContext(i);
+    Eigen::GpuDevice* gpu_device =
+        device_context->template get_eigen_device<DEVICE_GPU>();
+    ASSERT_NE(nullptr, gpu_device);
+    delete device_context;
+  }
+}
+
+TEST(Device, CUDADeviceContext) {
   int count = paddle::platform::GetDeviceCount();
   for (int i = 0; i < count; i++) {
     paddle::platform::CUDADeviceContext* device_context =
         new paddle::platform::CUDADeviceContext(i);
-    Eigen::GpuDevice gpu_device = device_context->eigen_device();
+    Eigen::GpuDevice* gpu_device = device_context->eigen_device();
-    ASSERT_NE(nullptr, gpu_device.stream());
+    ASSERT_NE(nullptr, gpu_device);
     cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
     ASSERT_NE(nullptr, cudnn_handle);
     cublasHandle_t cublas_handle = device_context->cublas_handle();

paddle/platform/dynload/CMakeLists.txt

@@ -1,2 +1,2 @@
 cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags)
-nv_library(dyload_cuda SRCS cublas.cc cudnn.cc curand.cc)
+nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc)

paddle/pybind/pybind.cc

@@ -12,6 +12,7 @@ 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 <Python.h>
 #include <paddle/framework/scope.h>
 #include <pybind11/pybind11.h>
 
@@ -43,4 +44,4 @@ All parameter, weight, gradient are variables in Paddle.
            py::return_value_policy::reference);
 
   return m.ptr();
-}
+}