Archives Discussions

[diapolo]

It seems like amd_bitalign(x, x, (uint)(32 - y)) and rotate(x, (uint)y) currently generate the same code (checked vor Tahiti and Cypress). Is this true for future releases, too, so that I can safely remove amd_bitalign() calls from a kernel. It was only used as a replacement for rotate() in my context.

Thanks,

Dia

[MicahVillmow · Accepted solution]

Dia,

It is safe to remove amd_bitalign in place of rotate as we now optimize the rotate pattern into bitalign.

[diapolo]

Thanks for that info. By the way, there are still no OpenCL changelog infos in the Catalyst release notes, what's up with that? It was promised quite a few weeks ago.

Dia

[vvelichkov]

Having OpenCL release notes in the monthly Catalyst updates is critically important! We all know that this is an overhead, but updating to a new driver release should not be "plug & pray" exercise for the users/devs