I don't really want to discourage anyone from following their interests, but my point is that any reasonably performant implementation is sufficient for the current usage in Gallium and one that is 3x faster or slower isn't going to make any difference to our overall performance at this point. If there's a real problem with what we've got, sure lets fix it, but the real gains in graphics performance lie elsewhere.
Keith ________________________________________ From: Joakim Sindholt [b...@zhasha.com] Sent: Monday, November 30, 2009 8:55 AM To: Keith Whitwell Cc: mesa3d-dev@lists.sourceforge.net Subject: Re: [Mesa3d-dev] minor u_math.h speedup fun It sounds like you're ignoring the case where gallium is not built with gcc (i.e. doesn't have __builtin_popcount available). We still need an implementation to fall back on. On Mon, 2009-11-30 at 01:00 -0800, Keith Whitwell wrote: > Although popcount is a routine that has had huge optimization attention over > the years and yields itself to some very cool approaches, our usage of it in > Mesa is afaik never on a critical performance path. That may change one day, > eg a software rasterizer may find itself doing many popcounts per triangle > for some reason, but right now it's not on our performance path. >From that > point of view, pushing optimizations to gcc is definitely the preferred > option at this point. > > Keith > ________________________________________ > From: tom fogal [tfo...@alumni.unh.edu] > Sent: Sunday, November 29, 2009 7:12 PM > To: Matt Turner > Cc: mesa3d-dev@lists.sourceforge.net > Subject: Re: [Mesa3d-dev] minor u_math.h speedup fun > > Matt Turner <matts...@gmail.com> writes: > > On Sat, Nov 28, 2009 at 2:13 PM, Yang Zhao <y...@yangman.ca> wrote: > > > The speed-up is definitely there, but __builtin_popcount() > > > will still be drastically faster when architecture-specific > > > optimizations are enabled: > > > > I don't think this is the case (except for with SSE4's popcnt > > instruction, which your CFLAGS seem to be enabling.) > > > > Even when compiling with the Intel CC, which can undoubtedly > > can optimize code for Core 2 better than gcc, fast_bitcount is > > significantly faster. > > IMHO, the best long term solution is to use gcc's builtin, and ping > the gcc folks with the implementation and benchmark results of > fast_bitcount. If it's really better, they'll eventually adopt it on > the platforms where that makes sense. > > This would benefit a larger community, and would hopefully mean that > nobody needs to come across this again in <N> years when common CPUs > might be completely different -- Mesa would just benefit transparently > from someone in the gcc community noticing the need for a new tuning. > > -tom > > ------------------------------------------------------------------------------ > Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day > trial. Simplify your report design, integration and deployment - and focus on > what you do best, core application coding. Discover what's new with > Crystal Reports now. http://p.sf.net/sfu/bobj-july > _______________________________________________ > Mesa3d-dev mailing list > Mesa3d-dev@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/mesa3d-dev > > ------------------------------------------------------------------------------ > Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day > trial. Simplify your report design, integration and deployment - and focus on > what you do best, core application coding. Discover what's new with > Crystal Reports now. http://p.sf.net/sfu/bobj-july > _______________________________________________ > Mesa3d-dev mailing list > Mesa3d-dev@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/mesa3d-dev ------------------------------------------------------------------------------ Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day trial. Simplify your report design, integration and deployment - and focus on what you do best, core application coding. Discover what's new with Crystal Reports now. http://p.sf.net/sfu/bobj-july _______________________________________________ Mesa3d-dev mailing list Mesa3d-dev@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mesa3d-dev
