Intel Just Released A Crazy Fast Acceleration Architecture
Unfortunately, the timing of the push isn't too great just before a busy weekend and some Bavarian colleagues being in town for a few days. That means there are no benchmarks to provide right now or even today. By early next week, however, I hope to have some test results available for Sandy Bridge and earlier generations of Intel hardware too. By then I'll also have had time to further analyze this new SNA method. Unfortunately, the automated Intel per-commit driver tracker isn't setup quite yet, but will be soon as well.
Until the more extensive analysis, here's the message from the commit introducing this massive new Intel acceleration architecture.
The premise is that switching between rings (i.e. the BLT and RENDER rings) on SandyBridge imposes a large latency overhead whilst rendering. The cause is that in order to switch rings, we need to split the batch earlier than is desired and to add serialisation between the rings. Both of which incur large overhead.
By switching to using a pure 3D blit engine (ok, not so pure as the BLT engine still has uses for the core drawing model which can not be easily represented without a combinatorial explosion of shaders) we can take advantage of additional efficiencies, such as relative relocations, that have been incorporated into recent hardware advances. However, even older hardware performs better from avoiding the implicit context switches and from the batching efficiency of the 3D pipeline...
But this is X, and PolyGlyphBlt still exists and remains in use. So for the operations that are not worth accelerating in hardware, we introduce a shadow buffer mechanism through out and reintroduce pixmap migration. Doing this efficiently is the cornerstone of ensuring that we do exploit the increased potential of recent hardware for running old applications and environments (i.e. so that the latest and greatest chip is actually faster than gen2!)
For the curious, sna is SandyBridge's New Acceleration. If you are running older chipsets and welcome the performance increase offered by this patch, then you may choose to call it Snazzy instead.
Speedups
========
gen3 firefox-fishtank 1203584.56 (1203842.75 0.01%) -> 85561.71 (125146.44 14.87%): 14.07x speedup
gen5 grads-heat-map 3385.42 (3489.73 1.44%) -> 350.29 (350.75 0.18%): 9.66x speedup
gen3 xfce4-terminal-a1 4179.02 (4180.09 0.06%) -> 503.90 (531.88 4.48%): 8.29x speedup
gen4 grads-heat-map 2458.66 (2826.34 4.64%) -> 348.82 (349.20 0.29%): 7.05x speedup
gen3 grads-heat-map 1443.33 (1445.32 0.09%) -> 298.55 (298.76 0.05%): 4.83x speedup
gen3 swfdec-youtube 3836.14 (3894.14 0.95%) -> 889.84 (979.56 5.99%): 4.31x speedup
gen6 grads-heat-map 742.11 (744.44 0.15%) -> 172.51 (172.93 0.20%): 4.30x speedup
gen3 firefox-talos-svg 71740.44 (72370.13 0.59%) -> 21959.29 (21995.09 0.68%): 3.27x speedup
gen5 gvim 8045.51 (8071.47 0.17%) -> 2589.38 (3246.78 10.74%): 3.11x speedup
gen6 poppler 3800.78 (3817.92 0.24%) -> 1227.36 (1230.12 0.30%): 3.10x speedup
gen6 gnome-terminal-vim 9106.84 (9111.56 0.03%) -> 3459.49 (3478.52 0.25%): 2.63x speedup
gen5 midori-zoomed 9564.53 (9586.58 0.17%) -> 3677.73 (3837.02 2.02%): 2.60x speedup
gen5 gnome-terminal-vim 38167.25 (38215.82 0.08%) -> 14901.09 (14902.28 0.01%): 2.56x speedup
gen5 poppler 13575.66 (13605.04 0.16%) -> 5554.27 (5555.84 0.01%): 2.44x speedup
gen5 swfdec-giant-steps 8941.61 (8988.72 0.52%) -> 3851.98 (3871.01 0.93%): 2.32x speedup
gen5 xfce4-terminal-a1 18956.60 (18986.90 0.07%) -> 8362.75 (8365.70 0.01%): 2.27x speedup
gen5 firefox-fishtank 88750.31 (88858.23 0.14%) -> 39164.57 (39835.54 0.80%): 2.27x speedup
gen3 midori-zoomed 2392.13 (2397.82 0.14%) -> 1109.96 (1303.10 30.35%): 2.16x speedup
gen6 gvim 2510.34 (2513.34 0.20%) -> 1200.76 (1204.30 0.22%): 2.09x speedup
gen5 firefox-planet-gnome 40478.16 (40565.68 0.09%) -> 19606.22 (19648.79 0.16%): 2.06x speedup
gen5 gnome-system-monitor 10344.47 (10385.62 0.29%) -> 5136.69 (5256.85 1.15%): 2.01x speedup
gen3 poppler 2595.23 (2603.10 0.17%) -> 1297.56 (1302.42 0.61%): 2.00x speedup
gen6 firefox-talos-gfx 7184.03 (7194.97 0.13%) -> 3806.31 (3811.66 0.06%): 1.89x speedup
gen5 evolution 8739.25 (8766.12 0.27%) -> 4817.54 (5050.96 1.54%): 1.81x speedup
gen3 evolution 1684.06 (1696.88 0.35%) -> 1004.99 (1008.55 0.85%): 1.68x speedup
gen3 gnome-terminal-vim 4285.13 (4287.68 0.04%) -> 2715.97 (3202.17 13.52%): 1.58x speedup
gen5 swfdec-youtube 5843.94 (5951.07 0.91%) -> 3810.86 (3826.04 1.32%): 1.53x speedup
gen4 poppler 7496.72 (7558.83 0.58%) -> 5125.08 (5247.65 1.44%): 1.46x speedup
gen4 gnome-terminal-vim 21126.24 (21292.08 0.85%) -> 14590.25 (15066.33 1.80%): 1.45x speedup
gen5 firefox-talos-svg 99873.69 (100300.95 0.37%) -> 70745.66 (70818.86 0.05%): 1.41x speedup
gen4 firefox-planet-gnome 28205.10 (28304.45 0.27%) -> 19996.11 (20081.44 0.56%): 1.41x speedup
gen5 firefox-talos-gfx 93070.85 (93194.72 0.10%) -> 67687.93 (70374.37 1.30%): 1.37x speedup
gen4 evolution 6696.25 (6854.14 0.85%) -> 4958.62 (5027.73 0.85%): 1.35x speedup
gen3 swfdec-giant-steps 2538.03 (2539.30 0.04%) -> 1895.71 (2050.62 62.43%): 1.34x speedup
gen4 gvim 4356.18 (4422.78 0.70%) -> 3276.31 (3281.69 0.13%): 1.33x speedup
gen6 evolution 1242.13 (1245.44 0.72%) -> 953.76 (954.54 0.07%): 1.30x speedup
gen6 firefox-planet-gnome 4554.23 (4560.69 0.08%) -> 3758.76 (3768.97 0.28%): 1.21x speedup
gen3 firefox-talos-gfx 6264.13 (6284.65 0.30%) -> 5261.56 (5370.87 1.28%): 1.19x speedup
gen4 midori-zoomed 4771.13 (4809.90 0.73%) -> 4037.03 (4118.93 0.85%): 1.18x speedup
gen6 swfdec-giant-steps 1557.06 (1560.13 0.12%) -> 1336.34 (1341.29 0.32%): 1.17x speedup
gen4 firefox-talos-gfx 80767.28 (80986.31 0.17%) -> 69629.08 (69721.71 0.06%): 1.16x speedup
gen6 midori-zoomed 1463.70 (1463.76 0.08%) -> 1331.45 (1336.56 0.22%): 1.10x speedup
Slowdowns
=========
gen6 xfce4-terminal-a1 2030.25 (2036.23 0.25%) -> 2144.60 (2240.31 4.29%): 1.06x slowdown
gen4 swfdec-youtube 3580.00 (3597.23 3.92%) -> 3826.90 (3862.24 0.91%): 1.07x slowdown
gen4 firefox-talos-svg 66112.25 (66256.51 0.11%) -> 71433.40 (71584.31 0.14%): 1.08x slowdown
gen4 gnome-system-monitor 5691.60 (5724.03 0.56%) -> 6707.56 (6747.83 0.33%): 1.18x slowdown
gen3 ocitysmap 3494.05 (3502.44 0.20%) -> 4321.99 (4524.42 2.78%): 1.24x slowdown
gen4 ocitysmap 3628.42 (3641.66 9.37%) -> 5177.16 (5828.74 8.38%): 1.43x slowdown
gen5 ocitysmap 4027.77 (4068.11 0.80%) -> 5748.26 (6282.25 7.38%): 1.43x slowdown
gen6 ocitysmap 1401.61 (1402.24 0.40%) -> 2365.74 (2379.14 4.12%): 1.69x slowdown
[Note the performance regression for ocitysmap comes from that we now attempt to support rendering to and (more importantly) from large surfaces. By enabling such operations is the only way to one day be faster than purely using the CPU, in the meantime we suffer regression due to the increased migration and aperture thrashing. The other couple of regressions will be eliminated with improved span and shader support, now that the framework for such is in place.]
The performance increase for Cairo completely overlooks the other critical aspects of the architecture:
World of Padman:
gen3 (800x600): 57.5 -> 96.2
gen4 (800x600): 47.8 -> 74.6
gen6 (1366x768): 100.4 -> 140.3 [F15]
144.3 -> 146.4 [drm-intel-next]
x11perf (gen6);
aa10text: 3.47 -> 14.3 Mglyphs/s [unthrottled!]
copywinwin10: 1.66 -> 1.99 Mops/s
copywinpix10: 2.28 -> 2.98 Mops/s
And we do not have a good measure for how much improvement the reworking of the fallback paths give, except that xterm is now over 4x faster...
PS: This depends upon the Xorg patchset "Remove the cacheing of the last scratch PixmapRec" for correct invalidations of scratch Pixmaps (used by the dix to implement SHM operations, used by chromium and gtk+ pixbufs.
PPS: ./configure --enable-sna
65 Comments
