On Fri, 2010-12-17 at 00:49 -0800, Dave Airlie wrote: > On Tue, Apr 27, 2010 at 7:10 AM, Dave Airlie <[email protected]> wrote: > >>>> buffers. > >>> But you'll have more shader instructions for writing to all these > >>> outputs right? I think that could still make a difference, though it > >>> might be more theoretical rather than in practice. > >>> > >> > >> Right, I was talking specifically about memory bandwidth. But it would > >> take a few extra instructions as well. I think the major impact will > >> be the additional memory bandwidth rather than the added instructions. > >> > > > > > > I also thought about this a bit more, and I'm guessing GPUs use this > > to optimise cache usage. If they know they are rendering the same data > > to all 1..n render targets they only need to store one copy in the dst > > cache. > > Okay I'm finally getting back to resurrecting this, > > http://people.freedesktop.org/~airlied/scratch/0001-gallium-r300g-fix-frag-color-writing-attempt-2.patch > > Basically I've added a TGSI property that states whether single writes > are meant to go to single out or all of them. > > I've still got to fixup softpipe and there's lots of debugging in > there, but with that r300g passes the fbo-drawbuffers-fragcolor test.
Dave, I think this is subtle enough that it needs some documentation. I trawled the revived thread to figure out the detail of what's here (having forgotten all about it), and it looks fine but I think this really needs a couple of good paragraphs in the documentation and a from-first-principles description in the patch. Also, WRITE_ALL is a bit ambiguous -- could it be something like "BROADCAST_COLOR0_WRITES" or maybe "WRITE_ALL_CBUFS"? Keith _______________________________________________ mesa-dev mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/mesa-dev
