[Bf-committers] Cycles CUDA Toolkit 8 transition

[Sergey Sharybin]

Hey again,

Spent majority of the day trying to solve the regression, without much
success. Even simplest kernel needed for BMW scene is about 10% slower.
This is mainly coming from bump nodes. Enabling all other features makes
things even worse performance wise.

I did some tweaks again to make sure all functions are inlined in the same
manner by CUDA 8.0 as they used to be before. So now PTax output shows
exactly same function, but for some reason spills are just higher with new
toolkit and at the same time stack usage is reasonably slower. Not sure yet
what's going on here and think we'd better leave this alone for until
official toolkit is released.

For the time being i've switched buildbots to use more complicated setup,
using CUDA 7.5 for all kernels except sm_60 and sm_61 (new generation
cards) and using new toolkit only for new kernels.

So hopefully now all maxwell and lower crads have same performance as
before. And yet users of new cards can have some degree of GPU rendering.


On Mon, Aug 1, 2016 at 6:15 PM, Sergey Sharybin <sergey.vfx at gmail.com>
wrote:

> Hey everyone,
>
> In order to make Cycles officially support new Pascal cards (GTX 10x0) it
> is crucial to use CUDA Toolkit 8. While it is still in the RC stage we
> started looking into making it official one for Cycles.
>
> There is already some initial work done:
>
> - Tweaks are done to Cycles kernel to have same level of inlined functions
> (new toolkit barely used inlined functions, which resulted in poor
> performance on all cards)
> - New toolkit was installed onto all release environments and buildbots.
>
> So on a positive side, next buildbot's builds will support Pascal cards,
> yay! :)
>
> But unfortunately, while i've managed to keep Kepler cards (GTX760) same
> performance as before, we had a report from Jens in IRC about Maxwell cards
> having a poor performance yet again (in his case, up to 40% worse).
>
> This is really unacceptable, we'll keep working on solution for this
> issue. It might take some time, so just be aware and please be patient :)
>
> I will keep posted here with our progress.
>
> --
> With best regards, Sergey Sharybin
>



-- 
With best regards, Sergey Sharybin