fix bug in register gpu OpKernel

paddle/framework/op_registry.h

@@ -403,15 +403,16 @@ class GradOpRegisterHelper {
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
       __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
       "REGISTER_OP_KERNEL must be in global namespace");                  \
-  struct __op_kernel_register__##type##__ {                               \
-    __op_kernel_register__##type##__() {                                  \
+  struct __op_kernel_register__##type##__##DEVICE_TYPE##__ {              \
+    __op_kernel_register__##type##__##DEVICE_TYPE##__() {                 \
       ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
       key.place_ = PlaceType();                                           \
       ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
           .reset(new __VA_ARGS__());                                      \
     }                                                                     \
   };                                                                      \
-  static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
+  static __op_kernel_register__##type##__##DEVICE_TYPE##__                \
+      __reg_kernel_##type##__##DEVICE_TYPE##__;                           \
   int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
 
 // (type, KernelType)

paddle/framework/operator.h

@@ -199,7 +199,11 @@ class OperatorWithKernel : public OperatorBase {
       place_ = dev_ctx.GetPlace();
     }
 
-    bool operator==(const OpKernelKey& o) const { return place_ == o.place_; }
+    // bool operator==(const OpKernelKey& o) const { return place_ == o.place_;
+    // }
+    bool operator==(const OpKernelKey& o) const {
+      return platform::places_are_same_class(place_, o.place_);
+    }
   };
 
   struct OpKernelHash {

paddle/pybind/pybind.cc

@@ -80,9 +80,11 @@ PYBIND11_PLUGIN(core) {
              self.mutable_data<int>(place);
            })
       .def("set", paddle::pybind::PyCPUTensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyCUDATensorSetFromArray<float>)
       .def("set", paddle::pybind::PyCPUTensorSetFromArray<int>)
+#ifndef PADDLE_ONLY_CPU
+      .def("set", paddle::pybind::PyCUDATensorSetFromArray<float>)
       .def("set", paddle::pybind::PyCUDATensorSetFromArray<int>)
+#endif
       .def("shape",
            [](pd::Tensor& self) { return pd::vectorize(self.dims()); });
 

paddle/pybind/tensor_bind.h

@@ -42,9 +42,6 @@ template <size_t I, typename... ARGS>
 struct CastToPyBufferImpl<true, I, ARGS...> {
   using CUR_TYPE = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
   py::buffer_info operator()(framework::Tensor &tensor) {
-    PADDLE_ENFORCE(paddle::platform::is_cpu_place(tensor.holder_->place()),
-                   "Only CPU tensor can cast to numpy array");
-
     if (std::type_index(typeid(CUR_TYPE)) == tensor.holder_->type()) {
       auto dim_vec = framework::vectorize(tensor.dims());
       std::vector<size_t> dims_outside;
@@ -99,6 +96,7 @@ void PyCPUTensorSetFromArray(
   std::memcpy(dst, array.data(), sizeof(T) * array.size());
 }
 
+#ifndef PADDLE_ONLY_CPU
 template <typename T>
 void PyCUDATensorSetFromArray(
     framework::Tensor &self,
@@ -112,10 +110,10 @@ void PyCUDATensorSetFromArray(
 
   self.Resize(framework::make_ddim(dims));
   auto *dst = self.mutable_data<T>(place);
-  std::memcpy(dst, array.data(), sizeof(T) * array.size());
   paddle::platform::GpuMemcpySync(
       dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice);
 }
+#endif
 
 }  // namespace pybind
 }  // namespace paddle