[Intel-gfx] [PATCH] intel: Stop using sized buckets for the BO cache.

Chris Wilson chris at chris-wilson.co.uk
Mon Jun 7 15:15:53 CEST 2010


On Sun,  6 Jun 2010 23:52:59 -0700, Eric Anholt <eric at anholt.net> wrote:
> This is the patch I'm testing on my desktop currently to deal with the
> couple of reports we've had recently of our overallocation for the BO
> cache seriously hurting some interesting workloads (HD video playback
> and UrbanTerror on gen3).  I'd love to see some reports from more than
> just me, and especially some performance numbers on CPU-bound apps if
> anyone can come up with some, since I'm having a hard time getting
> stable numbers out of my system right now.

Some numbers from cairo-perf-trace on i945 and g45:

old: pot-buckets
new: nonpot-buckets
Speedups
========
   gl-g45              swfdec-giant-steps  2240.42 (2535.87 5.91%) -> 2049.88 (2541.73 10.49%):  1.09x speedup
[This seems to be within the margin of error for the measurement, so is
not significant.]

Slowdowns
=========
 xlib-g45              swfdec-giant-steps  2194.40 (2202.42 2.05%) -> 2305.49 (2322.09 2.36%):  1.05x slowdown
 xlib-i945           firefox-planet-gnome  45345.45 (45439.72 0.38%) -> 47691.74 (47702.83 0.13%):  1.05x slowdown
 xlib-g45               firefox-talos-svg  40301.31 (40428.13 0.57%) -> 42592.23 (42619.36 0.31%):  1.06x slowdown
   gl-g45                         poppler  10355.56 (10373.69 0.89%) -> 10989.65 (11188.48 1.27%):  1.06x slowdown
 xlib-i945             swfdec-giant-steps  8656.30 (8657.33 0.80%) -> 9443.07 (9471.56 0.41%):  1.09x slowdown
   gl-g45               xfce4-terminal-a1  4039.91 (4329.24 8.63%) -> 4417.84 (4599.02 5.40%):  1.09x slowdown
 xlib-g45              gnome-terminal-vim  28238.96 (28313.46 0.19%) -> 33882.79 (33913.10 0.23%):  1.20x slowdown
   gl-g45            firefox-planet-gnome  20323.68 (20538.31 2.18%) -> 24902.90 (25221.52 1.50%):  1.23x slowdown
 xlib-g45                         poppler  8865.63 (8889.82 0.52%) -> 10942.57 (10972.90 0.47%):  1.23x slowdown

Even with -fno-omit-frame-pointer, perf is not providing usuable insights
into the difference on the 64bit g45 system.
-ickle

-- 
Chris Wilson, Intel Open Source Technology Centre



More information about the Intel-gfx mailing list