[SYCL][CUDA] Reduction ext unsupported

[bjoernknafla]

The reduction extension is not yet supported by CUDA. Mark LIT tests as
unsupported.

Signed-off-by: Bjoern Knafla bjoern@codeplay.com

[v-klochkov]

The reduction extension is not yet supported by CUDA. Mark LIT tests as
unsupported.

Signed-off-by: Bjoern Knafla bjoern@codeplay.com

I have one question and 1 comment:

1. When these 2 big patches were merged all tests reported passes.
   [SYCL] Implement basic reduction for parallel_for() accepting nd_range #1585
   [SYCL] Add faster reduction implementations using atomic or/and intel… #1615
   Do you think CUDA testing was turned OFF in CI in those days, or something changed in CUDA testing settings?

2. reduction feature only uses other existing sycl features. 3 basic features used by reduction:

• parallel_for accepting nd_range. (Is that supported on CUDA?)
• sycl::atomic class and operations like fetch_add(), fetch_min(), etc. (same question)
• intel::reduce() algorithm

In the lines:
// UNSUPPORTED: cuda
would you please be more specific. For example:
// UNSUPPORTED: cuda
// cuda does not yet support sycl::atomic, intel::reduce() features used by reduction implementation

[bader]

If these tests are not working due to the bugs rather than unsupported features than we should probably disable them using XFAIL instead of UNSUPPORTED.

[vladimirlaz]

The reduction extension is not yet supported by CUDA. Mark LIT tests as
unsupported.
Signed-off-by: Bjoern Knafla bjoern@codeplay.com

I have one question and 1 comment:

1. When these 2 big patches were merged all tests reported passes.
   [SYCL] Implement basic reduction for parallel_for() accepting nd_range #1585
   [SYCL] Add faster reduction implementations using atomic or/and intel… #1615
   Do you think CUDA testing was turned OFF in CI in those days, or something changed in CUDA testing settings?

• CUDA testing was turned OFF when [SYCL] Add faster reduction implementations using atomic or/and intel… #1615 was tested.

[Ruyk]

• intel::reduce() algorithm

Intel reduce algorithm is not a standard or "basic" SYCL feature. We are implementing SYCL 1.2.1 features first.

In the lines:
// UNSUPPORTED: cuda
would you please be more specific. For example:
// UNSUPPORTED: cuda
// cuda does not yet support sycl::atomic, intel::reduce() features used by reduction implementation

Atomic should be there once we finish the work on builtins.
Reduction is on the roadmap for later, but not implemented yet. Makes sense to clarify that on the comment.

[bjoernknafla]

Good call to document why the tests are unsupported. It is due to OpenCL 2.x alike work-group builtins which the CUDA implementation is not supporting yet. I will add the description to the tests.