Questions about type aliases

[AlexeySachkov]

Hi folks,

I've been looking into intel/llvm SYCL implementation details after vector_aliases SYCL-CTS update and whilst looking into the SYCL spec I got a few questions about whether we want to add a few more type aliases into it or about how of the existing aliases are defined.

No more aliases for OpenCL vector types?

In SYCL 1.2.1 we have aliases like cl_int2 -> cl::sycl::vec<cl::sycl::cl_int, 2>. However, in SYCL 2020 we only have sycl::opencl::cl_int, but not sycl::opencl::cl_int2. I was wondering: is there a particular reason for not including them?

Potentially confusing alias long{n}

4.14.2.2. Aliases defines aliases long{n} and ulong{n}, which correspond to sycl::vec<int64_t, n> and sycl::vec<uint64_t, n>. However, long data type in C++ is not guaranteed to be 64 bits.

To me, this looks a bit weird to have sycl::vec<long, 2> and sycl::long2 which are potentially different types. Someone who is not familiar with exact details of vector aliases may think that they are the same, but that may not always be true. I suppose that the idea could be to better match OpenCL, but isn't it better to align with C++ instead and add OpenCL-related aliases specific to the corresponding backend?

Missing longlong{n} alias?

Also, there is no alias for long long type. For someone, who comes from C++ background, where only long long is guaranteed to be at least 64 bits, I think it maybe useful to have such an alias. Note that we had such an alias in SYCL 1.2.1 spec, but it didn't make it to SYCL 2020 by some reason.

[gmlueck]

To me, this looks a bit weird to have sycl::vec<long, 2> and sycl::long2 which are potentially different types.

I definitely agree that this is a problem, and I think this should be fixed in the SYCL 2020 spec. The problem exists also for the marray aliases, and it exists for other alias variants like sycl::short.

there is no alias for long long type

This seems like an oversight to me.

[fraggamuffin]

same as #353

[keryell]

We need to review the whole story again for the CTS.