Eric Anholt <e...@anholt.net> writes: > Ian Romanick <i...@freedesktop.org> writes: > >> On 05/21/2013 06:11 PM, Eric Anholt wrote: >>> We were expanding the live range too far, breaking register_coalesce_2() >>> and compute_to_mrf() on 16-wide shaders. Turning it back on improves >>> GLB2.7 performance by 0.239355% +/- 0.0850649% (n=398), though some >>> 16-wide shaders are lost. shader-db stats are: >>> >>> total instructions in shared programs: 1627211 -> 1609262 (-1.10%) >>> instructions in affected programs: 450351 -> 432402 (-3.99%) >>> >>> While 33 new 16-wide shaders are gained, 70 are lost. >> >> Dare I ask about the performance delta is one of the apps that has a >> shader that loses 16-wide? > > I've now had my IVB run for about 2 total days in probably 5 attempts > trying to get numbers out of tropics, which is what's impacted the most > according to shader-db (the other being savage2, for which we don't have > benchmarking, and the last being sanctuary which should be the same). > Tropics has failed every time where it gets stuck just continuing > rendering and never ending the test, except for last night where I > bumped the lame macbook power adapter and it ran out of battery. I give > up. > > The initial pair of results I got from a couple of runs showed minor > improvements, except that you have to account for GPU turbo warming up > so it's invalid. > > On HSW, I saw a small performance improvement long term, but it's not > rendering correctly on HSW so that's invalid.
OK, I lied. I didn't give up. I found that: 1) The HSW had stale /etc/drirc from debian, thus the bad rendering. 2) Tropics just stops bothering once DPMS comes on. So, added drirc and cranked up the screensaver, and I got some results. +.42% +/- .16% (n=7/8, obvious high-speed outlier from the first "before" run dropped.)
pgprxUPppFv3c.pgp
Description: PGP signature
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
