[Mesa-dev] [PATCH 2/2] i965/skl: Disable SIMD16 when 3-source instructions are used

[Ian Romanick]

On 03/04/2015 06:46 AM, Neil Roberts wrote:
> Ilia Mirkin <imirkin at alum.mit.edu> writes:
> 
>> On Wed, Mar 4, 2015 at 9:33 AM, Neil Roberts <neil at linux.intel.com> wrote:
>>> Stepping C0 of Skylake fails when using SIMD16 with 3-source
>>> instructions (such as MAD). This patch just makes it disable SIMD16 in
>>> that case.
>>>
>>> This implements WaDisableSIMD16On3SrcInstr and fixes ~190 Piglit
>>> tests.
>>
>> Just curious -- is this faster than lowering MAD into MUL + ADD but
>> staying in SIMD16? (Should be fairly easy to implement as a post-RA
>> fixup I'd think, although not totally familiar with how the i965
>> compiler works.)
> 
> I haven't tested it but I think you're right that it would probably be
> faster to lower the instruction. However that sounds like more work and
> it's only relevant for a limited set of hardware so it doesn't seem
> worth doing. I think there are also other 3-source instructions and it
> might be more difficult to lower those (although I'm not sure).

It probably wouldn't be that bad if you did it the generator (instead of
the visitor)... but in this instance it probably isn't worth it.

> Regards,
> - Neil
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev