cuda-backend-interop.asciidoc

Test plan for CUDA backend interoperability

This is a test plan for the supported backend interoperability APIs for the CUDA backend described in the CUDA backend specification.

1. Testing scope

1.1. Backend coverage

All of the tests described in this plan are performed only for SYCL implementations which support the CUDA backend.

1.2. Device coverage

All of the tests described in this plan are performed only for Nvidia GPU devices available under the CUDA backend.

1.3. Type coverage

The tests described in this plan are performed for all supported types for the type they are associated with.

When T is the data type for accessor or local_accessor it is tested for the following types:

• char

• signed char

• unsigned char

• short int

• unsigned short int

• int

• unsigned int

• long int

• unsigned long int

• long long int

• unsigned long long int

• float

• double

• bool

• std::byte

• std::int8_t

• std::int16_t

• std::int32_t

• std::int64_t

• std::uint8_t

• std::uint16_t

• std::uint32_t

• std::uint64_t

• std::size_t

• vec<T, int dim> where T is each of the scalar types listed above except for bool and dim is 1, 2, 3, 4, 8, and 16.

• marray<T, size_t dim> where T is each of the scalar types listed above and dim is 1, 2, 5, and 10.

• A user-defined struct with several scalar member variables, no constructor, destructor or member functions.

• A user-defined class with several scalar member variables and a user-defined default constructor.

• A user-defined class with several scalar member variables, a deleted default constructor, and a user-defined (non-default) constructor.

When T is the data type for unsampled_image_accessor or sampled_image_accessor it is tested for the following types:

• sycl::int4

• sycl::uint4

• sycl::float4

• sycl::half4

2. Tests

2.1. Backend availability

• Check that the macro SYCL_BACKEND_CUDA is defined.

• Check that the sycl::backend enum has the member cuda.

2.2. Application interoperability

All of the tests in this section are run within SYCL application scope.

2.2.1. Check availability of backend interop traits

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::device> is available and resolves to an alias of CUdevice.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::context> is available and resolves to an alias of CUcontext.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::queue> is available and resolves to an alias of CUstream.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::event> is available and resolves to an alias of CUevent.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::device> is available and resolves to an alias of CUdevice.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::context> is available and resolves to an alias of std::vector<CUcontext>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::queue> is available and resolves to an alias of CUstream.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::event> is available and resolves to an alias of CUevent.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::buffer> is available and resolves to an alias of void *.

2.2.2. Check availability of make_* interop functions

• Check that the sycl::make_device template function specialization for sycl::backend::cuda is defined and can take a valid CUdevice and returns a valid sycl::device.

• Check that the sycl::make_context template function specialization for sycl::backend::cuda is defined and can take a valid CUcontext and returns a valid sycl::context.

• Check that the sycl::make_queue template function specialization for sycl::backend::cuda is defined and can take a valid CUstream and returns a valid sycl::queue.

• Check that the sycl::make_event template function specialization for sycl::backend::cuda is defined and can take a valid CUevent and returns a valid sycl::event.

2.2.3. Check availability of get_native interop functions

• Check that the sycl::get_native template function specialization for sycl::backend::cuda and sycl::device is defined and can take a valid sycl::device and returns a valid CUdevice.

• Check that the sycl::get_native template function specialization for sycl::backend::cuda and sycl::context is defined and can take a valid sycl::context and returns a valid std::vector<CUcontext>.

• Check that the sycl::get_native template function specialization for sycl::backend::cuda and sycl::queue is defined and can take a valid sycl::queue and returns a valid CUstream.

• Check that the sycl::get_native template function specialization for sycl::backend::cuda and sycl::event is defined and can take a valid sycl::event and returns a valid CUevent.

2.2.4. Check availability of interop_handle get_native_* interop member functions

• Check that the sycl::interop_handle::get_native_device template member function specialization for sycl::backend::cuda is defined and returns a valid CUdevice.

• Check that the sycl::interop_handle::get_native_queue template member function specialization for sycl::backend::cuda is defined and returns a valid CUstream.

• Check that the sycl::interop_handle::get_native_context template member function specialization for sycl::backend::cuda is defined and returns a valid std::vector<CUcontext>.

• Check that the sycl::interop_handle::get_native_mem template member function specialization for sycl::backend::cuda and sycl::buffer is defined and returns a valid void *.

2.2.5. Check availability of CUDA backend-specific functions

• Check that the sycl::cuda::has_native_event(sycl::event) function is defined and returns a valid bool.

2.3. Kernel function interoperability

All of the tests in this section are run within SYCL kernel function scope.

2.3.1. Check availability of backend alias types

• Check that the sycl::cuda::texture template class is defined.

• Check that the sycl::cuda::surface template class is defined.

2.3.2. Check availability of backend interop traits

The following tests are not run for each of the types defined in Section 1.3.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::device>> is available and resolves to an alias of T *.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::constant_buffer>> is available and resolves to an alias of T *.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::local>> is available and resolves to an alias of T *.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sycl::local_accessor<T, Dims>> is available and resolves to an alias of T *.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sampled_image_accessor<T, 1, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 1>.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sampled_image_accessor<T, 2, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 2>.

• Check that the sycl::backend_input_t<sycl::backend::cuda, sampled_image_accessor<T, 3, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 3>.

• Check that the sycl::backend_input_t<sycl::backend::cuda, unsampled_image_accessor<T, 1, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 1>.

• Check that the sycl::backend_input_t<sycl::backend::cuda, unsampled_image_accessor<T, 2, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 2>.

• Check that the sycl::backend_input_t<sycl::backend::cuda, unsampled_image_accessor<T, 3, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 3>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::device>> is available and resolves to an alias of T *.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::constant_buffer>> is available and resolves to an alias of T *.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::accessor<T, Dims, Mode, target::local>> is available and resolves to an alias of T *.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sycl::local_accessor<T, Dims>> is available and resolves to an alias of T *.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sampled_image_accessor<T, 1, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 1>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sampled_image_accessor<T, 2, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 2>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, sampled_image_accessor<T, 3, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::texture<T, 3>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, unsampled_image_accessor<T, 1, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 1>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, unsampled_image_accessor<T, 2, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 2>.

• Check that the sycl::backend_return_t<sycl::backend::cuda, unsampled_image_accessor<T, 3, Mode, image_target::device>> is available and resolves to an alias of sycl::cuda::surface<T, 3>.

2.3.3. Check calling a CUDA function

The following tests are not run for each of the types defined in Section 1.3.

• Check that a CUDA function can be declared using SYCL_EXTERNAL.

• Check that a CUDA function can be called with the backend interop type T * converted from an accessor or local_accessor.

• Check that a CUDA function can be called with the backend interop type texture, converted from a sampled_image_accessor.

• Check that a CUDA function can be called with the backend interop type surface, converted from an unsampled_image_accessor.