[Bf-committers] 3D rendering time regression

[Giuseppe Ghibò]

Brecht Van Lommel ha scritto:
> Hi,
>
> The first jump in that graph can be explained by the area light change, where it now does more shading calculations. The second one I'm not sure, it might be autothreads which was enabled by default around that time. Did you have the threads set to a number other than the number of cpu cores when running the benchmark?
>
> Brecht.
>   
No, I always used Nthreads = NCpus, even when the autothread patch to 
which I contributed was not yet merged in svn. So the number
of threads is always normalized to 2, for the dual core, 4 for the quad 
core, and 8 for the dual quad core, for any svn revision.

Note also that using a number of threads greater than number of CPUs 
doesn't have such steps (e.g. 9 threads
on a dual-quad core [but you need to patch the MAX_THREADS_NUMBER from 8 
to 16 to allow more than 8 threads),
or 5 threads on a quad core).

Bye
Giuseppe.