Enable Big/Super Pages for v3d by mairacanal · Pull Request #7285 · raspberrypi/linux

mairacanal · 2026-03-23T21:33:41Z

Considering that the Raspberry Pi is an embedded device with limited memory, memory fragmentation is an important aspect for performance. Using Big/Super Pages has clear benefits when it comes to reducing TLB misses, but also has an impact on memory fragmentation as we need to allocate aligned contiguous memory, increasing compaction pressure and memory waste for small BOs.

As Big/Super Pages only have benefits for larger BOs, create a minimum BO size to use the THP partition. After testing different thresholds, 512KB provides the most balanced results with clear improvements and no significant regressions. This means that Big/Super Pages will only be used for BOs of at least 512KB.

Here are some benchmark results. Each trace has been run twice to gather the results.

popcornmix · 2026-05-18T14:21:05Z

I have run various 3d related benchmarks to check the effect of these settings. 5 runs of each.
THP=1 is TRANSPARENT_HUGEPAGE_MADVISE=y
THP=2 is TRANSPARENT_HUGEPAGE_ALWAYS=y
BIG_V3D=1 is Only use Big/Super Pages if BO >= 512KB commit

The percentages are change from default. An ! means statistically significant.
I'm struggling to find a statistically significant benefit.

Any suggestions for a test that shows a clear benefit?

aqua5000: https://webglsamples.org/aquarium/aquarium.html with 5000 fish
aqua5000: https://webglsamples.org/aquarium/aquarium.html with 500 fish
motionmark: https://browserbench.org/MotionMark1.3.1/
speedometer: https://browserbench.org/Speedometer3.1/
youtube: https://www.youtube.com/watch?v=LXb3EKWsInQ&autoplay=1 and extract DrawAndSwap
volumeshader: https://volumeshaderbm.com/start/
glmark2-es2: glmark2-es2-drm from apt
vkmark: vkmark from apt
openarena: openarena timedemo from apt

=== aqua5000 ===
272.0 : Default [288, 255, 254, 280, 283]
272.2 : THP=2 / BIG_V3D=1 [224, 286, 276, 285, 290] +0.1%
274.2 : BIG_V3D=1 [222, 295, 283, 286, 285] +0.8%
276.6 : THP=1 [305, 277, 269, 276, 256] +1.7%
279.4 : THP=2 [279, 279, 283, 278, 278] +2.7%
281.2 : THP=1 / BIG_V3D=1 [264, 303, 286, 264, 289] +3.4%
=== aqua500 ===
700.0 : THP=1 [712, 693, 707, 694, 694] -2.3%!
701.0 : THP=2 [710, 696, 695, 697, 707] -2.2%
703.6 : THP=1 / BIG_V3D=1 [706, 709, 692, 705, 706] -1.8%
710.4 : THP=2 / BIG_V3D=1 [712, 708, 716, 708, 708] -0.9%
711.0 : BIG_V3D=1 [708, 703, 713, 720, 711] -0.8%
716.8 : Default [705, 718, 721, 704, 736]
=== glmark2-es2 ===
665.3 : THP=1 [666, 667, 667, 666, 667, 665, 666, 662, 665, 662] -1.9%!
665.4 : THP=2 [666, 667, 663, 666, 663, 666, 666, 665, 667, 665] -1.9%!
673.0 : THP=2 / BIG_V3D=1 [671, 674, 674, 673, 673] -0.8%!
673.4 : THP=1 / BIG_V3D=1 [673, 674, 673, 673, 674] -0.8%!
675.4 : BIG_V3D=1 [675, 675, 675, 676, 676] -0.5%!
678.5 : Default [679, 679, 679, 679, 678, 679, 678, 678, 678, 678]
=== motionmark ===
24022.0 : THP=2 / BIG_V3D=1 [24163, 23706, 24797, 24051, 23393] -11.9%!
25599.5 : THP=1 [24445, 25943, 24328, 27235, 24412, 23262, 24022, 28012, 28355, 25981] -6.1%
25655.4 : BIG_V3D=1 [29270, 27546, 23687, 23459, 24315] -5.9%
26255.5 : THP=2 [27648, 26508, 23661, 30018, 28348, 26604, 25809, 23730, 26815, 23414] -3.7%
26260.2 : THP=1 / BIG_V3D=1 [22966, 28149, 26818, 27127, 26241] -3.7%
27261.8 : Default [23926, 28065, 26559, 28385, 23741, 30594, 32586, 24177, 24391, 30194]
=== openarena ===
9792.0 : THP=1 [9780, 9760, 9750, 9790, 9810, 9830, 9800, 9790, 9800, 9810] -2.8%!
9801.0 : THP=2 [9820, 9830, 9770, 9780, 9800, 9810, 9780, 9790, 9790, 9840] -2.8%!
10012.0 : THP=2 / BIG_V3D=1 [10040, 10040, 9980, 10010, 9990] -0.7%!
10022.0 : THP=1 / BIG_V3D=1 [10010, 10050, 10030, 10000, 10020] -0.6%!
10046.0 : BIG_V3D=1 [9970, 10070, 10070, 10060, 10060] -0.3%
10079.0 : Default [10080, 10040, 10080, 10080, 10070, 10090, 10100, 10050, 10100, 10100]
=== speedometer ===
501.4 : Default [503, 497, 502, 507, 498]
501.6 : BIG_V3D=1 [495, 503, 503, 502, 505] +0.0%
502.0 : THP=1 [502, 499, 503, 502, 504] +0.1%
502.0 : THP=2 / BIG_V3D=1 [506, 499, 507, 497, 501] +0.1%
503.6 : THP=2 [501, 503, 508, 509, 497] +0.4%
506.8 : THP=1 / BIG_V3D=1 [500, 508, 507, 510, 509] +1.1%
=== vkmark ===
578.8 : THP=1 [579, 579, 579, 579, 578, 578, 579, 579, 579, 579] -4.4%!
578.8 : THP=2 [579, 578, 578, 578, 579, 579, 580, 579, 579, 579] -4.4%!
602.2 : THP=2 / BIG_V3D=1 [602, 602, 603, 602, 602] -0.6%!
602.8 : THP=1 / BIG_V3D=1 [603, 602, 603, 603, 603] -0.5%!
604.0 : BIG_V3D=1 [604, 604, 604, 604, 604] -0.3%!
605.7 : Default [605, 606, 606, 606, 606, 605, 606, 605, 606, 606]
=== volumeshader ===
630.0 : Default [630, 630, 630, 630, 630]
630.0 : BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=1 / BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=2 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=2 / BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
=== youtube ===
1718.2 : THP=1 / BIG_V3D=1 [1692, 1745, 1767, 1701, 1686] -2.2%
1747.0 : THP=2 / BIG_V3D=1 [1745, 1686, 1748, 1764, 1792] -0.5%
1756.0 : Default [1770, 1792, 1767, 1721, 1730]
1763.0 : BIG_V3D=1 [1600, 1795, 1783, 1848, 1789] +0.4%
1774.0 : THP=2 [1815, 1783, 1751, 1773, 1748] +1.0%
1776.8 : THP=1 [1712, 1859, 1742, 1776, 1795] +1.2%

mairacanal · 2026-05-25T14:10:40Z

I have run various 3d related benchmarks to check the effect of these settings. 5 runs of each. THP=1 is TRANSPARENT_HUGEPAGE_MADVISE=y THP=2 is TRANSPARENT_HUGEPAGE_ALWAYS=y BIG_V3D=1 is Only use Big/Super Pages if BO >= 512KB commit
The percentages are change from default. An ! means statistically significant. I'm struggling to find a statistically significant benefit.
Any suggestions for a test that shows a clear benefit?

Thanks for sharing these results. I'm under the impression that using traces might be masking other side-effects that using THP has on the system. I'll investigate it further and run benchmarks without traces.

popcornmix · 2026-05-26T12:40:50Z

I'm testing with RpiOS trixie mesa - are you testing with something closer to master?
Might there be a difference in results observed?

mairacanal · 2026-05-26T17:05:43Z

Yes, for the results I previously shared, I was using Mesa mainline (commit 3ea293a9d1d).

popcornmix · 2026-06-03T16:06:47Z

I've added GIT=1 to test matrix, which is using mesa from head of git (bb15c00).
There are some huge wins just from the mesa switch.
And, I think the THP/BIG_V3D are visible when combined with GIT=1 for some tests.

Any view on the safety if bumping mesa closer to git master in our trixie repo?
Would you expect any incompatibilities with trixie packages that use mesa?
Are the gains from a few commits that could be cherry-picked onto trixie mesa source?

=== aqua5000 ===
193.0 : Default [199, 182, 198, 188, 198]
193.2 : BIG_V3D=1 [190, 204, 184, 189, 199] +0.1%
194.6 : THP=1 / BIG_V3D=1 [202, 200, 201, 184, 186] +0.8%
199.3 : THP=2 [200, 199, 199] +3.3%
203.0 : THP=1 [196, 199, 207, 202, 211] +5.2%
207.8 : THP=2 / BIG_V3D=1 [222, 222, 193, 201, 201] +7.7%
218.2 : GIT=1 [220, 201, 232, 249, 189] +13.1%
289.6 : THP=2 / GIT=1 / BIG_V3D=1 [254, 294, 319, 326, 255] +50.1%!
299.5 : GIT=1 / BIG_V3D=1 [307, 303, 294, 294] +55.2%!
299.5 : THP=1 / GIT=1 [301, 300, 298, 299] +55.2%!
302.5 : THP=2 / GIT=1 [300, 302, 304, 304] +56.7%!
308.4 : THP=1 / GIT=1 / BIG_V3D=1 [289, 301, 297, 334, 321] +59.8%!
=== aqua500 ===
547.5 : BIG_V3D=1 [546, 546, 548, 550] -1.0%
548.4 : THP=2 / BIG_V3D=1 [538, 550, 545, 560, 549] -0.9%
551.0 : THP=1 / BIG_V3D=1 [549, 546, 557, 552] -0.4%
551.2 : THP=2 [531, 546, 551, 558, 570] -0.4%
553.2 : Default [547, 546, 556, 555, 562]
558.2 : THP=1 [546, 548, 562, 579, 556] +0.9%
657.8 : GIT=1 [657, 656, 660, 658] +18.9%!
758.8 : THP=1 / GIT=1 [756, 766, 764, 749] +37.2%!
762.2 : GIT=1 / BIG_V3D=1 [745, 781, 757, 770, 758] +37.8%!
763.6 : THP=2 / GIT=1 [748, 760, 778, 756, 776] +38.0%!
768.0 : THP=1 / GIT=1 / BIG_V3D=1 [766, 755, 788, 760, 771] +38.8%!
780.4 : THP=2 / GIT=1 / BIG_V3D=1 [751, 799, 779, 804, 769] +41.1%!
=== basemark ===
23720.0 : BIG_V3D=1 [23500, 23600, 24200, 24000, 23300] -0.8%
23920.0 : Default [24600, 23600, 23300, 24100, 24000]
23933.3 : THP=1 [23900, 23900, 24000] +0.1%
24320.0 : THP=2 / BIG_V3D=1 [24300, 24700, 24900, 24100, 23600] +1.7%
24460.0 : THP=1 / BIG_V3D=1 [24800, 24300, 24900, 24300, 24000] +2.3%
24650.0 : THP=2 [24800, 25000, 24600, 24200] +3.1%!
24920.0 : GIT=1 [24900, 24400, 25300, 25200, 24800] +4.2%!
29840.0 : THP=1 / GIT=1 / BIG_V3D=1 [32500, 28300, 29200, 28500, 30700] +24.7%!
30740.0 : THP=2 / GIT=1 [31600, 31300, 31800, 29600, 29400] +28.5%!
30860.0 : THP=2 / GIT=1 / BIG_V3D=1 [31700, 30300, 29600, 31300, 31400] +29.0%!
31325.0 : THP=1 / GIT=1 [31300, 31500, 30700, 31800] +31.0%!
31600.0 : GIT=1 / BIG_V3D=1 [32000, 31400, 31400, 31600] +32.1%!
=== glmark2-es2 ===
708.0 : THP=1 [708, 708, 708, 708, 708] -0.4%!
708.2 : THP=2 [708, 708, 708, 709, 708] -0.4%!
708.2 : THP=2 / BIG_V3D=1 [709, 708, 708, 708, 708] -0.4%!
708.4 : THP=1 / BIG_V3D=1 [709, 708, 709, 708, 708] -0.4%!
710.7 : GIT=1 [710, 710, 711, 711, 711, 711] -0.0%
711.0 : Default [711, 711, 712, 711, 711, 710]
711.0 : BIG_V3D=1 [711, 711, 711, 711, 711] +0.0%
717.4 : THP=2 / GIT=1 / BIG_V3D=1 [716, 718, 717, 718, 718] +0.9%!
717.6 : THP=2 / GIT=1 [718, 717, 717, 718, 718] +0.9%!
718.0 : THP=1 / GIT=1 / BIG_V3D=1 [716, 719, 718, 718, 719] +1.0%!
718.4 : THP=1 / GIT=1 [720, 718, 718, 718, 718] +1.0%!
718.8 : GIT=1 / BIG_V3D=1 [719, 719, 720, 717, 719] +1.1%!
=== jetstream ===
12979.0 : THP=2 [12962, 12965, 12998, 12991] -0.7%!
13032.2 : THP=2 / BIG_V3D=1 [13055, 12956, 13100, 13056, 12994] -0.2%
13048.4 : THP=1 [13050, 13098, 13034, 13075, 12985] -0.1%
13054.8 : THP=1 / BIG_V3D=1 [13025, 13097, 13069, 13087, 12996] -0.1%
13062.2 : GIT=1 [13093, 13061, 13098, 13051, 13008] -0.0%
13063.4 : BIG_V3D=1 [13124, 13016, 13041, 13072, 13064] -0.0%
13064.0 : Default [13047, 13081, 13107, 13044, 13041]
19516.0 : THP=2 / GIT=1 [19865, 19017, 19086, 19414, 20198] +49.4%!
20004.0 : THP=2 / GIT=1 / BIG_V3D=1 [20014, 19985, 20013] +53.1%!
20196.8 : GIT=1 / BIG_V3D=1 [20305, 20158, 20051, 20287, 20183] +54.6%!
20215.8 : THP=1 / GIT=1 [20220, 20238, 20169, 20259, 20193] +54.7%!
20235.8 : THP=1 / GIT=1 / BIG_V3D=1 [20255, 20235, 20215, 20238] +54.9%!
=== jetstream2 ===
8429.6 : THP=1 / GIT=1 [8406, 8357, 8483, 8432, 8470] -0.5%
8430.8 : THP=2 / GIT=1 / BIG_V3D=1 [8353, 8466, 8427, 8429, 8479] -0.5%
8447.4 : THP=2 / GIT=1 [8447, 8467, 8397, 8403, 8523] -0.3%
8462.5 : GIT=1 / BIG_V3D=1 [8448, 8452, 8465, 8485] -0.1%
8465.2 : THP=1 / GIT=1 / BIG_V3D=1 [8489, 8447, 8426, 8479, 8485] -0.1%
8467.0 : THP=2 [8507, 8413, 8469, 8489, 8457] -0.1%
8467.2 : THP=1 / BIG_V3D=1 [8408, 8487, 8418, 8531, 8492] -0.1%
8467.8 : GIT=1 [8575, 8387, 8398, 8501, 8478] -0.1%
8470.6 : BIG_V3D=1 [8489, 8538, 8422, 8544, 8360] -0.0%
8474.0 : Default [8484, 8545, 8432, 8462, 8447]
8478.2 : THP=1 [8484, 8477, 8468, 8484] +0.1%
8479.0 : THP=2 / BIG_V3D=1 [8491, 8467, 8472, 8486] +0.1%
=== motionmark ===
23532.0 : THP=2 / BIG_V3D=1 [20566, 24335, 23606, 26178, 22975] -1.8%
23968.8 : Default [23892, 23929, 23029, 24646, 24348]
24368.4 : BIG_V3D=1 [24717, 23221, 23324, 24770, 25810] +1.7%
24611.8 : THP=2 [24378, 24819, 24352, 24898] +2.7%
25068.4 : GIT=1 [24177, 23702, 25994, 25378, 26091] +4.6%
25158.0 : THP=1 / BIG_V3D=1 [25274, 25135, 25065] +5.0%!
25254.8 : THP=1 [25005, 25626, 25216, 25172] +5.4%!
28886.2 : THP=2 / GIT=1 / BIG_V3D=1 [29531, 26346, 29635, 27540, 31379] +20.5%!
28987.8 : GIT=1 / BIG_V3D=1 [32612, 26444, 29952, 28610, 27321] +20.9%!
29082.5 : THP=2 / GIT=1 [28172, 29061, 29430, 29667] +21.3%!
29214.0 : THP=1 / GIT=1 [29616, 29744, 28049, 28637, 30024] +21.9%!
30751.8 : THP=1 / GIT=1 / BIG_V3D=1 [33157, 29991, 31263, 28282, 31066] +28.3%!
=== octane ===
15724.2 : THP=1 / BIG_V3D=1 [15604, 15635, 15628, 15857, 15897] -1.2%
15741.0 : GIT=1 [16185, 15314, 15417, 15902, 15887] -1.0%
15907.4 : Default [15780, 15878, 16146, 15735, 15998]
15941.6 : THP=1 [16133, 16207, 15803, 16035, 15530] +0.2%
16029.8 : THP=2 / BIG_V3D=1 [15895, 16263, 15810, 16077, 16104] +0.8%
16035.0 : THP=2 [16181, 16026, 16090, 15977, 15901] +0.8%
16069.5 : BIG_V3D=1 [16095, 16012, 15999, 16172] +1.0%
25032.8 : THP=1 / GIT=1 [25007, 24918, 24647, 25201, 25391] +57.4%!
25046.4 : THP=1 / GIT=1 / BIG_V3D=1 [25425, 24571, 24825, 25320, 25091] +57.5%!
25084.6 : THP=2 / GIT=1 [25172, 25170, 24976, 25450, 24655] +57.7%!
25100.4 : GIT=1 / BIG_V3D=1 [24913, 25011, 25395, 24809, 25374] +57.8%!
25299.0 : THP=2 / GIT=1 / BIG_V3D=1 [25235, 25455, 25101, 25342, 25362] +59.0%!
=== openarena ===
10150.0 : THP=2 [10160, 10130, 10150, 10160] -0.6%!
10160.0 : THP=1 [10160, 10150, 10180, 10150] -0.5%!
10164.0 : THP=1 / BIG_V3D=1 [10230, 10120, 10170, 10160, 10140] -0.5%
10166.0 : THP=2 / BIG_V3D=1 [10250, 10150, 10150, 10130, 10150] -0.4%
10210.0 : BIG_V3D=1 [10280, 10170, 10200, 10200, 10200] -0.0%
10211.4 : Default [10290, 10290, 10170, 10170, 10180, 10180, 10200]
10441.4 : GIT=1 [10490, 10500, 10420, 10410, 10430, 10420, 10420] +2.3%!
10457.5 : THP=1 / GIT=1 / BIG_V3D=1 [10450, 10460, 10450, 10470] +2.4%!
10458.0 : THP=1 / GIT=1 [10450, 10450, 10480, 10440, 10470] +2.4%!
10460.0 : THP=2 / GIT=1 / BIG_V3D=1 [10450, 10460, 10480, 10450, 10460] +2.4%!
10470.0 : THP=2 / GIT=1 [10450, 10490, 10470, 10470, 10470] +2.5%!
10496.0 : GIT=1 / BIG_V3D=1 [10510, 10500, 10500, 10490, 10480] +2.8%!
=== speedometer ===
369.2 : GIT=1 [370, 365, 373, 370, 368] -1.2%!
370.6 : THP=2 / BIG_V3D=1 [368, 369, 372, 374, 370] -0.8%!
371.6 : BIG_V3D=1 [372, 370, 372, 375, 369] -0.6%
372.2 : THP=1 [367, 375, 375, 375, 369] -0.4%
373.0 : THP=2 [378, 374, 373, 371, 369] -0.2%
373.4 : THP=1 / BIG_V3D=1 [375, 375, 373, 369, 375] -0.1%
373.8 : Default [375, 373, 374, 373]
530.6 : THP=1 / GIT=1 [530, 534, 533, 528, 528] +42.0%!
532.2 : THP=2 / GIT=1 [532, 533, 533, 531, 532] +42.4%!
532.4 : THP=2 / GIT=1 / BIG_V3D=1 [526, 534, 534, 534, 534] +42.4%!
535.8 : THP=1 / GIT=1 / BIG_V3D=1 [538, 534, 533, 539, 535] +43.4%!
536.6 : GIT=1 / BIG_V3D=1 [542, 535, 538, 539, 529] +43.6%!
=== vkmark ===
617.8 : THP=1 / BIG_V3D=1 [618, 617, 618, 618, 618] -0.6%!
617.8 : THP=2 [618, 618, 618, 617, 618] -0.6%!
618.0 : THP=1 [618, 618, 618, 618, 618] -0.6%!
618.0 : THP=2 / BIG_V3D=1 [618, 619, 618, 618, 617] -0.6%!
619.8 : GIT=1 [619, 620, 620, 620, 620, 620] -0.3%!
621.8 : Default [622, 621, 622, 622, 622, 622]
622.6 : BIG_V3D=1 [623, 623, 623, 622, 622] +0.1%!
627.2 : THP=1 / GIT=1 / BIG_V3D=1 [627, 627, 627, 627, 628] +0.9%!
627.6 : THP=2 / GIT=1 / BIG_V3D=1 [628, 627, 628, 628, 627] +0.9%!
627.8 : THP=1 / GIT=1 [628, 628, 628, 627, 628] +1.0%!
628.2 : GIT=1 / BIG_V3D=1 [629, 628, 628, 628, 628] +1.0%!
628.2 : THP=2 / GIT=1 [629, 628, 628, 628, 628] +1.0%!
=== volumeshader ===
630.0 : Default [630, 630, 630, 630, 630]
630.0 : BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=1 / BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=2 [630, 630, 630, 630, 630] +0.0%
630.0 : THP=2 / BIG_V3D=1 [630, 630, 630, 630, 630] +0.0%
650.0 : GIT=1 [650, 650, 650, 650, 650] +3.2%!
650.0 : GIT=1 / BIG_V3D=1 [650, 650, 650, 650, 650] +3.2%!
650.0 : THP=1 / GIT=1 [650, 650, 650, 650, 650] +3.2%!
650.0 : THP=1 / GIT=1 / BIG_V3D=1 [650, 650, 650, 650, 650] +3.2%!
650.0 : THP=2 / GIT=1 [650, 650, 650, 650, 650] +3.2%!
650.0 : THP=2 / GIT=1 / BIG_V3D=1 [650, 650, 650, 650, 650] +3.2%!
=== youtube ===
1845.0 : THP=2 / BIG_V3D=1 [1845, 1842, 1848] -1.5%!
1863.3 : THP=1 / BIG_V3D=1 [1855, 1866, 1869] -0.5%
1868.8 : THP=2 [1912, 1832, 1855, 1876, 1869] -0.2%
1872.7 : Default [1880, 1862, 1876]
1873.0 : THP=1 [1894, 1842, 1852, 1901, 1876] +0.0%
1877.8 : BIG_V3D=1 [1862, 1842, 1883, 1894, 1908] +0.3%
1934.4 : GIT=1 [1965, 1938, 1949, 1908, 1912] +3.3%!
2183.5 : THP=1 / GIT=1 [2174, 2169, 2198, 2193] +16.6%!
2214.4 : THP=2 / GIT=1 [2242, 2203, 2198, 2217, 2212] +18.2%!
2233.2 : THP=2 / GIT=1 / BIG_V3D=1 [2257, 2242, 2188, 2252, 2227] +19.3%!
2236.4 : THP=1 / GIT=1 / BIG_V3D=1 [2252, 2242, 2217, 2273, 2198] +19.4%!
2300.0 : GIT=1 / BIG_V3D=1 [2299, 2299, 2294, 2309, 2299] +22.8%!

mairacanal · 2026-06-04T19:05:55Z

Any view on the safety if bumping mesa closer to git master in our trixie repo?
Would you expect any incompatibilities with trixie packages that use mesa?
Are the gains from a few commits that could be cherry-picked onto trixie mesa source?

I have the impression that bumping to 26.1 should be safe, but I’ll check with the Mesa experts to confirm. I’ll also see whether there is a small set of commits we can cherry-pick, and then I’ll get back to you with a solution.

mairacanal · 2026-06-05T10:54:16Z

@popcornmix, in order to make sure our results are similar and compatible, would you mind sharing your benchmarking scripts?

popcornmix · 2026-06-05T12:08:46Z

The browser benchmarks use CDP to launch benchmarks and grab results. I've attached the version I last used
browser-bench.sh

I'd be interested in your scores from any of those (especially the ones I report as having bigger performance gains).

Note: I convert a meaure from output to a score (bigger is better) with:

benchmark	filter
aquarium	`awk '/Avg FPS:/ {printf "%.0f\n", $3 * 10}'`
jetstream	`awk '/Overall:/ {v=$2} END {printf "%.0f\n", v * 100}'`
jetstream2	`awk '/Overall:/ {v=$2} END {printf "%.0f\n", v * 100}'`
motionmark	`awk '/Score:/ {v=$2} END {printf "%.0f\n", v * 100}'`
speedometer	`awk '/Score:/ {v=$2} END {printf "%.0f\n", v * 100}'`
basemark	`awk '/Score:/ {v=$2} END {printf "%.0f\n", v * 100}'`
volumeshader	`awk '/Avg FPS:/ {printf "%.0f\n", $3 * 1000}'`
youtube	`awk '/V3D durations/ && match($0, /gl_total=([0-9.]+)/, m) {printf "%.0f\n", (m[1]>0)?1000.0/m[1]:0}'`

I (with claude) extended this to also support Firefox (and BiDi) and the last run didn't show the significant boosts with latest mesa (even on chromium), so I am now a bit less confident with the gains observed. They were consistent enough across multiple runs and multiple benchmarks that it was measuring something, but there may be some other variable in play.

The framework that launches them has a lot of dependencies (e.g. switching branches and building/installing kernel and bootloader on our build server, building mesa, adjusting config.txt/cmdline.txt/eeprom-config) so might be quite a job to get running somewhere else.

mairacanal · 2026-06-05T12:31:41Z

Thanks for sharing the script! I'll use it to evaluate the possible gains across Mesa commits. About the performance gains, I'd also be a bit skeptical about +20% performance improvements. From our collection traces, there was a consistent measurable impact on the performance, but as you can see it averaged around +2%, with the biggest improvement being Google Maps with +10%.

Having said that, I'll run the benchmark locally and bring the results back to you.

Considering that the Raspberry Pi is an embedded device with limited memory, memory fragmentation is an important aspect for performance. Using Big/Super Pages has clear benefits when it comes to reducing TLB misses, but also has an impact on memory fragmentation as we need to allocate aligned contiguous memory, increasing compaction pressure and memory waste for small BOs. As Big/Super Pages only have benefits for larger BOs, create a minimum BO size to use the THP partition. After testing different thresholds, 512KB provides the most balanced results with clear improvements and no significant regressions. This means that Big/Super Pages will only be used for BOs of at least 512KB. Signed-off-by: Maíra Canal <mcanal@igalia.com>

Signed-off-by: Maíra Canal <mcanal@igalia.com>

popcornmix force-pushed the rpi-6.18.y branch 2 times, most recently from 2e9acf3 to 6ad963aCompare March 25, 2026 14:01

mairacanal force-pushed the v3d/downstream/super-pages-512k branch from c60b428 to 8298673Compare April 6, 2026 00:00

popcornmix force-pushed the rpi-6.18.y branch from 21da81b to 5558a45Compare April 8, 2026 13:43

mairacanal force-pushed the v3d/downstream/super-pages-512k branch from 8298673 to 48996f0Compare April 13, 2026 12:34

popcornmix force-pushed the rpi-6.18.y branch from 8fa1bed to 5970220Compare April 13, 2026 13:06

popcornmix force-pushed the rpi-6.18.y branch from e63eeee to e538839Compare April 21, 2026 13:45

mairacanal mentioned this pull request May 16, 2026
configs: Enable TRANSPARENT_HUGEPAGE_MADVISE #6390
Draft

mairacanal force-pushed the v3d/downstream/super-pages-512k branch from 48996f0 to b2316a1Compare June 5, 2026 10:50

mairacanal added 2 commits June 6, 2026 14:34

configs: enable CONFIG_TRANSPARENT_HUGEPAGE=y
ba81860
Signed-off-by: Maíra Canal <mcanal@igalia.com>

mairacanal force-pushed the v3d/downstream/super-pages-512k branch from b2316a1 to ba81860Compare June 6, 2026 17:34

Navigation Menu

Uh oh!

Enable Big/Super Pages for v3d - #7285

Enable Big/Super Pages for v3d#7285
mairacanal wants to merge 2 commits into
raspberrypi:rpi-6.18.yfrom
mairacanal:v3d/downstream/super-pages-512k

mairacanal commented Mar 23, 2026

Uh oh!

popcornmix commented May 18, 2026

Uh oh!

mairacanal commented May 25, 2026

Uh oh!

popcornmix commented May 26, 2026

Uh oh!

mairacanal commented May 26, 2026•
edited
Loading

Uh oh!

popcornmix commented Jun 3, 2026

Uh oh!

mairacanal commented Jun 4, 2026

Uh oh!

mairacanal commented Jun 5, 2026

Uh oh!

popcornmix commented Jun 5, 2026•
edited
Loading

Uh oh!

mairacanal commented Jun 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

mairacanal commented Mar 23, 2026

Uh oh!

popcornmix commented May 18, 2026

Uh oh!

mairacanal commented May 25, 2026

Uh oh!

popcornmix commented May 26, 2026

Uh oh!

mairacanal commented May 26, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

popcornmix commented Jun 3, 2026

Uh oh!

mairacanal commented Jun 4, 2026

Uh oh!

mairacanal commented Jun 5, 2026

Uh oh!

popcornmix commented Jun 5, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mairacanal commented Jun 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mairacanal commented May 26, 2026•
edited
Loading

popcornmix commented Jun 5, 2026•
edited
Loading