[Mesa-dev] [PATCH] r600g: Check for reduction instructions when merging instruction groups

[Christian König]

Am Mittwoch, den 05.10.2011, 14:48 -0400 schrieb Tom Stellard:
> On Wed, 2011-10-05 at 13:25 -0400, Christian König wrote:
> > Am Mittwoch, den 05.10.2011, 10:57 -0400 schrieb Tom Stellard:
> > > Reduction instructions can't share an instruction group with other
> > > instructions, so we need to check for them when trying to merge
> > > instruction groups. 
> > 
> > Hui? Where do you got that info?
> 
> pg. 60 r600isa.pdf

Oh, yes indeed, but on the other hand it also says the following:
>When any of the reduction instructions (DOT4, DOT4_IEEE, CUBE, and MAX4)
>is used, it must be executed on all four elements of a single vector.

> > I especially created that code to let a DOT4 and an ADD instruction
> > share the same instruction group in the first place, so I'm pretty sure
> > that it work.
> > 
> 
> If I understand the docs correctly, this is possible if the ADD
> instruction is in the ALU.Trans slot.  I overlooked this in the initial
> patch.  I'll try to revise it to account for this.

Correct.

> > It is indeed not possible to have a reduction instruction + something
> > else in the vector slots, but since a reduction instruction should
> > always take all four vector slots that shouldn't be a problem.
> > 
> > So what is the reason/problem behind it?
> > 
> 
> Good point.  I was thinking that one reduction slot plus three empty
> slots was acceptable, but maybe all slots do have to be exactly the
> same.  I'll try to play around with this some more and find out exactly
> how it works.

I only checked the r600isa while writing the code, so its possible that
later generations could actually cause a problem here by not using all
vector slots, but I don't think so.

Christian.