[Intel-gfx] XRenderFillRectangles performance further degrading with 2.8.99

Clemens Eisserer linuxhippy at gmail.com
Tue Sep 22 10:25:13 CEST 2009


Hello,

On my i945GM with 2.8.99, XRenderFillRectangles performance has even
further degraded compared to 2.7 (and maybe 2.8?):

XRenderFillRectangles:
   2.2/XAA:      294664.411765 Ops/s; rects (!); 15x15   (100%)
   2.7/EXA:      155345.525653 Ops/s; rects (!); 15x15  (53%)
   2.7/UXA:        93812.748533 Ops/s; rects (!); 15x15  (32%)
   2.8.99/UXA:  52848.098667 Ops/s; rects (!); 15x15  (18%)

So we are now at ~1/6 of XAA performance for this test.

Also note the performance regression for XPutImage:
XPutImage:
   2.2/XAA:      168811.567164 Ops/s; put composition (!); 15x15 (100%)
   2.2/EXA:       62233.493810 Ops/s; put composition (!); 15x15 (36%)
   2.8.99/UXA:  14132.306553 Ops/s; put composition (!); 15x15 (8,3%)

The "rects" test simulates scanline-based rendering, as performed by
the java2d xrender backend. All those regressions are turning my
backend from something that worked really well into something you
don't really want to use.
Too bad its not based on cairo (Carl suggested that even before I
started working on it), so there won't unfourtunatly be any time to
investigate this. How sad.
The benchmark can be found at: http://93.83.133.214/RenderMark.html

- Clemens



More information about the Intel-gfx mailing list