[Mesa-dev] [PATCH 06/21] i965/fs: Implement SIMD16 64-bit integer multiplies on Gen 7.

[Matt Turner]

On 12/22/14, Ben Widawsky <benjamin.widawsky at intel.com> wrote:
> From: Matt Turner <mattst88 at gmail.com>
>
> v2 by Ben:
> Keep the Gen8 code:
>   Gen8 will be fixed later. The intent is to limit the platform impact of
> the
>   patch to fixing just Gen7. The original Gen7 code as written by Matt did
> not
>   work properly on Gen8 (I can't remember why). I haven't actually tested
> the code
>   on Gen8 that was modified by me. It's possible it does work.
>
> Fix the MUL/MACH:
>   The docs clearly say that the channel enables need to be used in the
>   mul/mach/mov sequence. I am not quite certain of the reason for this when
> doing
>   an 8 channel operation, however I presume it has to be with the weird way
> the
>   accumulator is used (gaining extra precision for the operation).

It's the same problem I described in the previous patch.

If the channel enable bits aren't the same across the first and second
groups of eight, we'll see problems. The existing piglit tests don't
seem to expose this problem.