NEO/Win10 performance regression from .6025 onward including today's .6194

[allanmac]

Every NEO/Win10 driver post .6025 exhibits the same regression (huge spillage).

As noted here, .6025 works perfectly.

All kernels listed here are decorated with a reqd_subroup_size(8) and compiled with ioc64.

Let me know who I can send the kernels to.

[paigeale]

Hello Allan,

Thank you for making us aware of this potential regression. Please send your kernels to my email (listed in my profile) with a detailed description on how to reproduce. Our compiler team will take a look and see if we can reproduce this behavior and provide a fix.

-Alex

[allanmac]

Thanks, sent!

[paigeale]

The commit attached (1451db9) removes the spills you were seeing post .6025. Please verify and close.

[allanmac]

Thanks, I will confirm as soon as the next compute runtime release appears.

[paigeale]

Hello Allan,
Could you please verify that the latest compute runtime release solved your spill issues. This issue is fairly old and I would like to close this issue if it is no longer relevant.