[Mesa-users] llvmpipe os mesa with my own llvm

[Brian Paul]

On 06/14/2013 06:07 PM, burlen wrote:
> On 05/24/2013 04:24 PM, Burlen Loring wrote:
>> On 5/23/2013 1:09 PM, Burlen Loring wrote:
>>> On 5/23/2013 6:57 AM, Brian Paul wrote:
>>>> On 05/22/2013 04:40 PM, Burlen Loring wrote:
>>>>> On 5/22/2013 9:02 AM, Brian Paul wrote:
>>>>>> On 05/21/2013 04:43 PM, Burlen Loring wrote:
>>>> Not sure what to say about the performance difference.  If the app
>>>> is dominated by vertex transformation, then varying LP_NUM_THREADS
>>>> might not make a lot of difference.
>>
>> still looking at why threading doesn't seem to be working as expected.
>
> To close the loop on this, I found why threading wasn't working for me.
> It has nothing to do with Mesa. When running on the Cray XC30 the
> default CPU affinity used by the application launcher is "-cc cpu" where
> each process and any threads it creates are pinned to the same core. One
> has to either override this by passing the number of threads per process
> -d option, use numa node affinity "-cc numa_node" or disable the feature
> by passing the "-cc none" to the application launcher aprun.
>
> also I tried a couple of runs using a larger dataset(54M tris), as you
> said with a large number of vertices to process the threading in the
> fragment pipeline helps only so much. Hopefully the vertex pipeline
> could be threaded in a future release. Are there plans for that?

No specific plans, but we'd like to see that too.  Another area for 
improvement is to have some overlap between vertex transformation and 
rasterization.  They're totally serialized now.

-Brian