On 10/14/2013 10:44 AM, Eric Anholt wrote:
> Chris Forbes <[email protected]> writes:
> 
>> Previously this was float, which caused the copy from g0 to mangle
>> everything.
> 
> If we face a choice of types for a raw mov, we should choose float --
> it's higher performance on IVB (they can get dispatched twice as fast,
> when there's a hyperthread waiting to dispatch a float op)
> 
> Patch 1/2 is
> 
> Reviewed-by: Eric Anholt <[email protected]>

Using float for raw MOVs is unsafe for things exposed by
ARB_fragment_program, ARB_vertex_program, or fixed-function vertex
processing.  We still use ALT mode there so we get proper 0^0=1 handling
for POW.

We should really use IEEE everywhere and emit special code for POW in
ARB_vp/fp.  But that's above and beyond fixing a regression.

_______________________________________________
mesa-dev mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to