[Mesa-dev] Adjust MIMG writemask to only the used components

Michel Dänzer michel at daenzer.net
Wed Apr 10 01:32:54 PDT 2013


On Mit, 2013-04-10 at 10:26 +0200, Christian König wrote: 
> Am 03.04.2013 15:43, schrieb Michel Dänzer:
> > On Mit, 2013-04-03 at 15:38 +0200, Christian König wrote:
> >> Am 03.04.2013 15:19, schrieb Michel Dänzer:
> >>> On Mit, 2013-03-27 at 17:57 +0100, Christian König wrote:
> >>>> the attached patches dynamically adjust the writemask of MIMG
> >>>> instructions depending on the used components. Additional to that it
> >>>> also adjust the register class of MIMG instruction so that we also
> >>>> reduce register pressure.
> >>> There's mainly one thing that isn't clear to me: My understanding from
> >>> the ISA is that even if the instruction writemask operand has 'gaps',
> >>> e.g. 0x9, the enabled components will always be written to consecutive
> >>> VGPRs. So in this example, the instruction would write to two
> >>> consecutive VGPRs, one for component 0/'x' and one for component 3/'w'
> >>> of a four component vector. Is this handled correctly?
> >>>
> >> Yes your understanding is correct, and it is indeed handled correctly.
> >>
> >> I just analyze the used register components to figure out the needed
> >> writemask and while doing so remember which instruction uses which
> >> component. So after updating the writemask I just change the index of
> >> the used component as well.
> > Excellent. Sounds and looks good to me then.
> >
> 
> Sorry for the delay, I was busy with UVD for the last few days.
> 
> Does that count as a reviewed by from your side?

Yes.


-- 
Earthling Michel Dänzer           |                   http://www.amd.com
Libre software enthusiast         |          Debian, X and DRI developer




More information about the llvm-commits mailing list