Add support for kernel function interop

Author: aidan.belton

sycl/include/CL/sycl/accessor.hpp

@@ -2134,6 +2134,12 @@ class __SYCL_SPECIAL_CLASS accessor<DataT, Dimensions, AccessMode,
   bool operator!=(const accessor &Rhs) const { return !(*this == Rhs); }
 };
 
+// local_accessor
+// Same semantics and restrictions as accessor with target::local
+template <typename DataT, int Dimensions = 1>
+using local_accessor =
+    accessor<DataT, Dimensions, access_mode::read_write, access::target::local>;
+
 /// Image accessors.
 ///
 /// Available only when accessTarget == access::target::image.

sycl/include/CL/sycl/backend.hpp

@@ -115,15 +115,25 @@ inline backend_return_t<backend::opencl, event> get_native<
 }
 #endif
 
-// Native handle of an accessor should be accessed through interop_handler
+// Native accessor handle for kernel function interop
 template <backend BackendName, typename DataT, int Dimensions,
           access::mode AccessMode, access::target AccessTarget,
           access::placeholder IsPlaceholder>
 auto get_native(const accessor<DataT, Dimensions, AccessMode, AccessTarget,
                                IsPlaceholder> &Obj) ->
-    typename detail::interop<
-        BackendName, accessor<DataT, Dimensions, AccessMode, AccessTarget,
-                              IsPlaceholder>>::type = delete;
+    typename detail::interop<BackendName,
+                             accessor<DataT, Dimensions, AccessMode,
+                                      AccessTarget, IsPlaceholder>>::type {
+#ifdef __SYCL_DEVICE_ONLY__
+  return reinterpret_cast<typename detail::interop<
+      BackendName, accessor<DataT, Dimensions, AccessMode, AccessTarget,
+                            IsPlaceholder>>::type>(Obj.get_pointer().get());
+
+#else
+  throw runtime_error("Get native accessor is not support on host.",
+                      PI_INVALID_VALUE);
+#endif
+}
 
 namespace detail {
 // Forward declaration

sycl/include/sycl/ext/oneapi/experimental/backend/backend_traits_cuda.hpp

@@ -30,13 +30,6 @@ typedef struct CUstream_st *CUstream;
 typedef struct CUevent_st *CUevent;
 typedef struct CUmod_st *CUmodule;
 
-// As defined in the CUDA 10.1 header file. This requires CUDA version > 3.2
-#if defined(_WIN64) || defined(__LP64__)
-typedef unsigned long long CUdeviceptr;
-#else
-typedef unsigned int CUdeviceptr;
-#endif
-
 __SYCL_INLINE_NAMESPACE(cl) {
 namespace sycl {
 namespace detail {
@@ -72,33 +65,38 @@ template <> struct interop<backend::ext_oneapi_cuda, program> {
 };
 #endif
 
-// TODO the interops for accessor is used in the already deprecated class
-// interop_handler and can be removed after API cleanup.
 template <typename DataT, int Dimensions, access::mode AccessMode>
 struct interop<backend::ext_oneapi_cuda,
                accessor<DataT, Dimensions, AccessMode, access::target::device,
                         access::placeholder::false_t>> {
-  using type = CUdeviceptr;
+  using type = DataT *;
 };
 
 template <typename DataT, int Dimensions, access::mode AccessMode>
 struct interop<
     backend::ext_oneapi_cuda,
     accessor<DataT, Dimensions, AccessMode, access::target::constant_buffer,
              access::placeholder::false_t>> {
-  using type = CUdeviceptr;
+  using type = DataT *;
+};
+
+template <typename DataT, int Dimensions, access::mode AccessMode>
+struct interop<backend::ext_oneapi_cuda,
+               accessor<DataT, Dimensions, AccessMode, access::target::local,
+                        access::placeholder::false_t>> {
+  using type = DataT *;
 };
 
 template <typename DataT, int Dimensions, typename AllocatorT>
 struct BackendInput<backend::ext_oneapi_cuda,
                     buffer<DataT, Dimensions, AllocatorT>> {
-  using type = CUdeviceptr;
+  using type = DataT *;
 };
 
 template <typename DataT, int Dimensions, typename AllocatorT>
 struct BackendReturn<backend::ext_oneapi_cuda,
                      buffer<DataT, Dimensions, AllocatorT>> {
-  using type = void *;
+  using type = DataT *;
 };
 
 template <> struct BackendInput<backend::ext_oneapi_cuda, context> {

sycl/include/sycl/ext/oneapi/experimental/backend/cuda.hpp

@@ -21,12 +21,12 @@ namespace oneapi {
 namespace cuda {
 
 // Implementation of cuda::make<device>
-__SYCL_EXPORT device make_device(pi_native_handle NativeHandle) {
+inline __SYCL_EXPORT device make_device(pi_native_handle NativeHandle) {
   return detail::make_device(NativeHandle, backend::cuda);
 }
 
 // Implementation of cuda::make<platform>
-__SYCL_EXPORT platform make_platform(pi_native_handle NativeHandle) {
+inline __SYCL_EXPORT platform make_platform(pi_native_handle NativeHandle) {
   return detail::make_platform(NativeHandle, backend::cuda);
 }
 
@@ -36,7 +36,7 @@ __SYCL_EXPORT platform make_platform(pi_native_handle NativeHandle) {
 
 // CUDA context specialization
 template <>
-auto get_native<backend::ext_oneapi_cuda, context>(const context &C)
+inline auto get_native<backend::ext_oneapi_cuda, context>(const context &C)
     -> backend_return_t<backend::ext_oneapi_cuda, context> {
   // create a vector to be returned
   backend_return_t<backend::ext_oneapi_cuda, context> ret;
@@ -52,15 +52,15 @@ auto get_native<backend::ext_oneapi_cuda, context>(const context &C)
 
 // CUDA device specialization
 template <>
-device make_device<backend::ext_oneapi_cuda>(
+inline device make_device<backend::ext_oneapi_cuda>(
     const backend_input_t<backend::ext_oneapi_cuda, device> &BackendObject) {
   pi_native_handle NativeHandle = static_cast<pi_native_handle>(BackendObject);
   return ext::oneapi::cuda::make_device(NativeHandle);
 }
 
 // CUDA platform specialization
 template <>
-auto get_native<backend::ext_oneapi_cuda, platform>(const platform &C)
+inline auto get_native<backend::ext_oneapi_cuda, platform>(const platform &C)
     -> backend_return_t<backend::ext_oneapi_cuda, platform> {
   // get list of platform devices, and transfer to native platform type
   std::vector<device> platform_devices = C.get_devices();
@@ -75,7 +75,7 @@ auto get_native<backend::ext_oneapi_cuda, platform>(const platform &C)
 }
 
 template <>
-platform make_platform<backend::ext_oneapi_cuda>(
+inline platform make_platform<backend::ext_oneapi_cuda>(
     const backend_input_t<backend::ext_oneapi_cuda, platform> &BackendObject) {
   pi_native_handle NativeHandle =
       detail::pi::cast<pi_native_handle>(&BackendObject);
@@ -84,7 +84,7 @@ platform make_platform<backend::ext_oneapi_cuda>(
 
 // Specialisation of interop_handles get_native_context
 template <>
-backend_return_t<backend::ext_oneapi_cuda, context>
+inline backend_return_t<backend::ext_oneapi_cuda, context>
 interop_handle::get_native_context<backend::ext_oneapi_cuda>() const {
 #ifndef __SYCL_DEVICE_ONLY__
   return std::vector{reinterpret_cast<CUcontext>(getNativeContext())};