[SYCL] Implement SYCL 2020 vec aliases#7889
Conversation
This commit added aliases for `sycl::vec` class as described in section 4.14.2.2. Aliases of SYCL 2020 specification, revision 6. Aliases which use OpenCL type, such as `cl_int2` are still preserved, because they are used internally, see intel#7888.
According to 4.14.2.2. Aliases of SYCL 2020 specification, revision 6,
`longlong{N}` aliases do not exists.
There is corresponding PR to remove them from the implementation:
intel/llvm#7889
|
/verify with intel/llvm-test-suite#1477 |
According to 4.14.2.2. Aliases of SYCL 2020 specification, revision 6,
`longlong{N}` aliases do not exists.
There is corresponding PR to remove them from the implementation:
intel/llvm#7889
…ement-sycl-2020-vec-aliases
| using half = sycl::detail::half_impl::half; | ||
|
|
||
| // FIXME: SYCL 2020 spec says that cl_* aliases should reside in sycl::opencl | ||
| // namespace. |
There was a problem hiding this comment.
What is blocking us from applying these FIXME comments right away?
There was a problem hiding this comment.
Apparently, we still use some of these aliases: #7888. I don't know exactly about cl_* scalar aliases, but we for sure use cl_* vector aliases.
I would prefer an incremental approach and get this merged first as a fix for vector aliases and then submit a separate patch for scalar aliases
There was a problem hiding this comment.
I took another look at this: vector cl_* aliases depend on scalar cl_* aliases and those in turn are still used (#7888)
There was a problem hiding this comment.
Apparently, we still use some of these aliases: #7888. I don't know exactly about
cl_*scalar aliases, but we for sure usecl_*vector aliases.I would prefer an incremental approach and get this merged first as a fix for vector aliases and then submit a separate patch for scalar aliases
I also prefer incremental approach, which should be remove cl_* alias usages from DPC++ code base first and remove alias definitions after that. This way there will be no technical debt.
This is a follow-up from intel#7889. Adjusted LIT tests to acount for the new definition of long{n} alias, which may not always be equal to sycl::vec<long, n> now.
|
Post-commit failures are being addressed in #7896 |
|
I've just realized that some of the failures are SYCL-CTS and with the current spec wording, that is a bug in CTS and not in the implementation. Submitted KhronosGroup/SYCL-CTS#446. We should probably temporarily disable some CTS tests to have green post-commit results |
This is a follow-up from #7889. Adjusted LIT tests to account for the new definition of `long{n}` alias, which may not always be equal to `sycl::vec<long, n>` now. Removed some dead code (type traits) from the runtime.
…suite#1477) According to 4.14.2.2. Aliases of SYCL 2020 specification, revision 6, `longlong{N}` aliases do not exists. There is corresponding PR to remove them from the implementation: intel#7889
AlexeySachkov
deleted the
private/asachkov/implement-sycl-2020-vec-aliases
branch
3 participants
This commit added aliases for
sycl::vecclass as described in section 4.14.2.2. Aliases of SYCL 2020 specification, revision 6.Aliases which use OpenCL type, such as
cl_int2are still preserved, because they are used internally, see #7888.