By: rscheidegger (rscheidegger.delete@this.gmx.ch), August 9, 2011 5:03 pm
Room: Moderated Discussions
David Kanter (dkanter@realworldtech.com) on 8/9/11 wrote:
>Honestly, it's hard to tell. The ironlake documentation is basically the worst
>I've ever seen in my life (e.g. it lists the wrong number of registers, MAD support, etc.).
btw what makes you think that MAD isn't there in IL and it works in Sandy Bridge? I think the reason it isn't used in the linux driver is really because it's in vector-arithmetic group, which implies align16 mode I think and hence only usable in simd4 or simd4x2 mode (AOS), not simd8 or simd16 mode (SOA). So far I haven't seen anything in the docs implying this has changed for SNB. I always thought there's probably register fetch restrictions involved.
>Honestly, it's hard to tell. The ironlake documentation is basically the worst
>I've ever seen in my life (e.g. it lists the wrong number of registers, MAD support, etc.).
btw what makes you think that MAD isn't there in IL and it works in Sandy Bridge? I think the reason it isn't used in the linux driver is really because it's in vector-arithmetic group, which implies align16 mode I think and hence only usable in simd4 or simd4x2 mode (AOS), not simd8 or simd16 mode (SOA). So far I haven't seen anything in the docs implying this has changed for SNB. I always thought there's probably register fetch restrictions involved.
