[Mesa-dev] gallium scaled types

Dave Airlie airlied at gmail.com
Tue Nov 1 07:27:50 PDT 2011


On Tue, Nov 1, 2011 at 2:11 PM, Dave Airlie <airlied at gmail.com> wrote:
>>
>> After wasting so much time on this topic, and conceding on most points, I'm disappointed that the new pure integers formats were added without support for converting to/from floats (my 2nd request above) which ended up causing obscures bugs in llvmpipe (my 1st request).
>>
>> I regret having been so laissez-faire -- I should had made a fuss in the review request when my 2nd request was ignored --, but I learnt my lesson.
>
> I just added a fix to make sure llvmpipe didn't say it supported pure
> integer types so it should be unable to execute any of the problematic
> paths.
>
> The thing is converting pure integers to floats at any point in the
> pipeline is wrong and is a bug and should be treated as such. I left
> scope for translating ints to floats but I can find no reason in any
> API to enable this and unless someone has proof of such I don't feel
> we should, since you cannot represent a 32-bit integers in a 32-bit
> float without loss of precision.
>
> I'll do another llvmpipe run vs 7.11 and see what falls out later
> today and fix it, I did one before merge though I had to do a rebase
> since.
>
> But vlee has been fairly good at pointing me at piglit regressions so
> far and I've tried to fix them as they came in.
>
> Adding GLSL1.30 support to llvmpipe is going to involve doing it
> properly and not adding int->float conversions.

Ah now I see the fix I did for the last regression in the test code
broke stuff, so actually I should have just pushed the test code fix
and not the llvmpipe driver change.

Sorry, will piglit even over the seemingly "trivial" changes next time.

Dave.


More information about the mesa-dev mailing list