2009/9/24 Hugh Fisher <[email protected]> > Paul Brook wrote: > >> By your own numbers conventional 4-element SIMD gets <70% ALU utilization. >> That leaves quite a lot of scope for improvement, be it full MIMD or several >> ALU threads controlled by a single dispatch unit (ala. Larabee SIMD with >> software thread combining). >> > > I started this thread because the OGA2 designers were assuming > that GPU loads had become increasingly scalar. They're not, so > the design needs to be reconsidered. >
We did consider running a purely vector processor. We discussed of many issue between scalar and vector. We had, i don't know how many, discussion on what would be better and how we can squeeze more performance from what we have. We know that a good part of the load is vector. What we decided was a shader that would run vector operation as scalar. For simplicity and also for praticality: a vector processor is a lot harder to debug. > > I'm not opposed to MIMD full stop. If y'all can design a MIMD > processor that runs vector code fast, I'll be happy. > We didn't do a design who is MIMD, it actually closer to SPMD. http://en.wikipedia.org/wiki/SPMD We took that concept and extended it to run multiple program, one after the other so that each stage in the pipeline would process a different program. Each time we add an ALU that ALU follow the same code as the other ALU beside it. The data it's procesing come from a different thread(data set to process). If you want a name it could be called IPMD : Interleaved Process Multiple Data. Another way to view it is turn the data from a SIMD on the side. You change any lateral dependency of the data, you bring it back vertically. On a standard SIMD you would process different threads even with the same instruction one after the other. The shader we propose, we process all the same threads at once but we break down the instruction in more but simpler instruction.
_______________________________________________ Open-graphics mailing list [email protected] http://lists.duskglow.com/mailman/listinfo/open-graphics List service provided by Duskglow Consulting, LLC (www.duskglow.com)
