[Glamor] Trapezoid performance regression with the shader enabled.

[He Junyan]

Hi Zhigang:

The trapezoid generating speed of the shader is relatively slow when the 
trapezoid area is big. We need to fallback to pixman
when the trapezoid's width and height is bigger than 300 pixels to avoid 
the regression of the benchmark you referred.

I think the shader’s advantage is that it can calculate multi-pixels 
concurrently and avoid uploading.
Pixman’s manner is based on scan line, the edge of the trapezoid need to 
consider anti-alias and pixman’s rasterization need to do
some jobs here, but inside the trapezoid, it use memset to assign value 
very effectively.
If use shader manner, it still needs to expand the vertex shader and use 
the fragment shader to caculate every pixel. So for
large area of trapezoid, the pixman's speed seems faster.
The benchmarks which have regression use the big trapezoid, we can add a 
workaround here to fallback to CPU when
the trapezoid is big to avoid the regression. The result on my IVB after 
workaround is:

grads-heat-map    1.201  ------------>  0.698
firefox-fishbowl   35.027  ------------>  20.308
firefox-asteroids    6.379  ------------>  5.464
firefox-fishtank    28.373  ------------>  21.747
firefox-planet-gnome 7.043  ------------>  6.694
evolution           1.910  ------------>  2.047
firefox-canvas-alpha 26.458  ------------>  22.258
firefox-canvas      32.026  ------------>  27.788
firefox-talos-svg    11.758  ------------>  11.130
gnome-system-monitor 1.387 ------------>  1.761
swfdec-giant-steps   1.626  ------------>  1.579
firefox-paintball     14.353  ------------>  15.477



> Hi Junyan,
>
> After workaround the glsl-like problem, I can get correct rendering with
> trapezoid shader enabled.
>
> Then I measured the performance and found that it does accelerate some
> cases significantly, but
>
> also bring some major performance regressions, here is a summary, please
> check it out.
>
> old: cairo-trace-glamor
>
> new: cairo-trace-glamor-trapezoid
>
> Speedups
>
> ========
>
> xlib             grads-heat-map  1241.46 (1243.12 0.82%) -> 593.86
> (593.89 12.62%):  2.09x speedup
>
> xlib           firefox-fishbowl  35292.75 (35378.87 0.53%) -> 25675.22
> (25710.04 0.80%):  1.37x speedup
>
> xlib          firefox-asteroids  6902.64 (7127.80 1.73%) -> 5518.59
> (5843.95 2.74%):  1.25x speedup
>
> xlib           firefox-fishtank  33430.00 (33504.70 0.15%) -> 28612.69
> (28622.39 0.05%):  1.17x speedup
>
> xlib       firefox-planet-gnome  7187.17 (7188.44 0.46%) -> 6544.22
> (6555.78 0.53%):  1.10x speedup
>
> Slowdowns
>
> =========
>
> xlib                  evolution  1769.50 (1774.23 0.13%) -> 1877.39
> (1891.87 4.37%):  1.06x slowdown
>
> xlib       firefox-canvas-alpha  29717.59 (29761.29 0.19%) -> 32037.14
> (32108.83 0.42%):  1.08x slowdown
>
> xlib             firefox-canvas  42736.86 (43217.24 1.21%) -> 46611.94
> (46761.75 0.43%):  1.09x slowdown
>
> xlib          firefox-talos-svg  13468.38 (13653.09 0.65%) -> 15275.94
> (15280.97 0.11%):  1.13x slowdown
>
> xlib       gnome-system-monitor  1408.31 (1417.22 0.87%) -> 1839.00
> (1863.20 0.69%):  1.31x slowdown
>
> xlib         swfdec-giant-steps  1733.63 (1751.49 0.84%) -> 2280.65
> (2284.43 9.88%):  1.32x slowdown
>
> xlib          firefox-paintball  18133.06 (18139.63 0.06%) -> 32112.83
> (32125.89 0.04%):  1.77x slowdown
>
> As we discussed this afternoon, I’d like to release the version 0.5 by
> the end of next week. I hope we can
>
> fix the regression before that time.
>
> Thanks.
>
>
>
> _______________________________________________
> Glamor mailing list
> Glamor at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/glamor
>