[Mesa-dev] [PATCH] i965/fs: Improve accuracy of dFdy() to match dFdx().

[Eric Anholt]

Paul Berry <stereotype441 at gmail.com> writes:

> Previously, we computed dFdy() using the following instruction:
>
>   add(8) dst<1>F src<4,4,0)F -src.2<4,4,0>F { align1 1Q }
>
> That had the disadvantage that it computed the same value for all 4
> pixels of a 2x2 subspan, which meant that it was less accurate than
> dFdx().  This patch changes it to the following instruction when
> c->key.high_quality_derivatives is set:
>
>   add(8) dst<1>F src<4,4,1>.xyxyF -src<4,4,1>.zwzwF { align16 1Q }
>
> This gives it comparable accuracy to dFdx().
>
> Unfortunately, for some reason the SIMD16 version of this instruction:
>
>   add(16) dst<1>F src<4,4,1>.xyxyF -src<4,4,1>.zwzwF { align16 1H }
>
> Doesn't seem to work reliably (presumably the hardware designers never
> validated the combination of align16 mode with compressed
> instructions), so we unroll it to:

From the gen4 PRM vol4, page 340:

    "A compressed instruction must be in Align1 access mode. Align16
     mode instructions cannot be compressed."

Other than updating the comment about compressed due to the PRM cite,
this is:

Reviewed-by: Eric Anholt <eric at anholt.net>

Thanks for figuring this out.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20131002/2f43cb5d/attachment.pgp>