[SYCL] Deny pragma spelling for SYCL-specific attributes

[MrSidims]

Signed-off-by: Dmitry Sidorov [email protected]

[MrSidims]

TBH, I think we can also remove 'pragma spelling' for IntelReqdSubGroupSize as well. Yet I would prefer to leave ReqdWorkGroupSize attribute untouched. @erichkeane what do you think?

[erichkeane]

TBH, I think we can also remove 'pragma spelling' for IntelReqdSubGroupSize as well. Yet I would prefer to leave ReqdWorkGroupSize attribute untouched. @erichkeane what do you think?

I don't have a strong opinion, but I don't know how heavy the pragma spelling for thoseis used for OpenCL. Do we know?

[MrSidims]

TBH, I think we can also remove 'pragma spelling' for IntelReqdSubGroupSize as well. Yet I would prefer to leave ReqdWorkGroupSize attribute untouched. @erichkeane what do you think?

I don't have a strong opinion, but I don't know how heavy the pragma spelling for thoseis used for OpenCL. Do we know?

Except of loop pragmas (unroll, ivdep etc), I have never encountered any pragmas in OpenCL.

[erichkeane]

TBH, I think we can also remove 'pragma spelling' for IntelReqdSubGroupSize as well. Yet I would prefer to leave ReqdWorkGroupSize attribute untouched. @erichkeane what do you think?

I don't have a strong opinion, but I don't know how heavy the pragma spelling for thoseis used for OpenCL. Do we know?

Except of loop pragmas (unroll, ivdep etc), I have never encountered any pragmas in OpenCL.

Then I think we can remove the pragma spelling for IntelReqdGroupSize, and if there's a regression revert that part.

Whats the logic for wanting to leave ReqWorkGroupSize untouched?

[MrSidims]

Whats the logic for wanting to leave ReqWorkGroupSize untouched?

It's not Intel-specific attribute coming from OpenCL core spec itself.
At the same time quick googling showed me this:

No results found for "pragma reqd_work_group_size"

So it might be safe to remove the spelling for it as well.

[MrSidims]

Whats the logic for wanting to leave ReqWorkGroupSize untouched?

It's not Intel-specific attribute coming from OpenCL core spec itself.
At the same time quick googling showed me this:

No results found for "pragma reqd_work_group_size"

So it might be safe to remove the spelling for it as well.

@keryell may be you know if any of your customers use 'pragma' spelling?

[erichkeane]

I'm happy with this. @keryell stop this if you think its necessary, otherwise I think we can just wait to see if it breaks anything.

[Fznamznon]

I don't have OpenCL background, so I cannot say is correct this change or not, but this particular attribute ReqdWorkGroupSize is OpenCL specific and it is added not by us.
Suggest you go through community with change for this attribute first.
I think you will get answers on your questions regarding this attribute.

[MrSidims]

Ok, lets stick with leaving reqd_work_group_size attribute alone.

[keryell]

@keryell may be you know if any of your customers use 'pragma' spelling?

Traditional Xilinx tools use a lot of #pragma but not the SYCL path.

While I am an advocate of no user-visible #pragma or attribute in SYCL to use only C++ decorative functions (think Python @decorator...) or classes instead, #pragma play even less nicely with C++ than attributes. :-( So at the end I do not have issues with having only attributes, as a less evil solution. :-)

[erichkeane]

Thanks for the discussion guys, I think we've bottomed out, so I'm approving.