the first bad commit is:

b6eeb7a1f7efa591504070b606be655e27e6e9c2

Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Nov 5 13:03:41 2014 +0000

    Disable DRI3 by default

    The external libraries, both in git, and especially shipping already
    enabled in distributions, are buggy and lead to server crashes and
    lockups. Caveat emptor.

    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

DRI2 has more overhead (copy per frame update) than DRI3, so the performance regression from disabling DRI3 is expected, especially on memory bandwidth bound high FPS cases.

Only byt? On ivb I see:

DRI2 - FPS: 361.101 (MIN: 360.858, MAX: 394.804, STDEV [%]: 1.30636, S: 36)

DRI3 - FPS: 348.131 (MIN: 328.46, MAX: 370.083, STDEV [%]: 3.63791, S: 33)

Tests are run in FullHD.  BYT has (a lot) less memory bandwidth available than Core machines, compared to the memory bandwidth taken by FullHD frames.  On IVB/HSW, Batch0 test isn't memory bandwidth bound (at least as much as on BYT).

Test with below cases:


SynMark2_v6.0.0_OglBatch2
SynMark2_v6.0.0_OglBatch3
SynMark2_v6.0.0_OglBatch4
SynMark2_v6.0.0_OglGeomPoint
SynMark2_v6.0.0_OglGeomTriList
SynMark2_v6.0.0_OglGeomTriStrip

On byt N2820, 1080p:

dri2: FPS: 99.9493 (MIN: 89.2223, MAX: 123.483, CV: 7.00455%, S: 10)
dri3: FPS: 58.1767 (MIN: 42.9242, MAX: 66.0809, CV: 8.41218%, S: 10)

I am not sure if dri3 is being locked to screen refresh or not (max of 66fps suggests otherwise, but it is suspiciously close to refresh).

Tweaking DRI2 to use async flipping [http://cgit.freedesktop.org/~ickle/xserver/commit/?h=async-20141230&id=13372bcf995599b4fe1f427923635d02d2f4c71f]:

FPS: 138.5 (MIN: 67.0551, MAX: 185.931, CV: 19.661%, S: 10)

which does support that it is the swap ellision that may be the difference, but fwiw I cannot reproduce the reported DRI3 improvements.

(In reply to Chris Wilson from comment #6)
> On byt N2820, 1080p:
> 
> dri2: FPS: 99.9493 (MIN: 89.2223, MAX: 123.483, CV: 7.00455%, S: 10)
> dri3: FPS: 58.1767 (MIN: 42.9242, MAX: 66.0809, CV: 8.41218%, S: 10)
> 
> I am not sure if dri3 is being locked to screen refresh or not (max of 66fps
> suggests otherwise, but it is suspiciously close to refresh).

Based on bug 79715 debugging, swaps may be locked to refresh only part of the time.  Maybe it can get locked also to multiple of 60 FPS.


> Tweaking DRI2 to use async flipping
> [http://cgit.freedesktop.org/~ickle/xserver/commit/?h=async-
> 20141230&id=13372bcf995599b4fe1f427923635d02d2f4c71f]:
> 
> FPS: 138.5 (MIN: 67.0551, MAX: 185.931, CV: 19.661%, S: 10)
> 
> which does support that it is the swap ellision that may be the difference,
> but fwiw I cannot reproduce the reported DRI3 improvements.

Improvements are visible only in memory bound tests and only when they don't get wrongly refresh locked. Based on bug 79715, on what tests / when DRI3 locking happens, is pretty random. :-)

To recap on my byt and OglBatch1 and latest mesa + async patches:

dri3: 147.969 (MIN: 123.535, MAX: 164.291, CV: 5.35204%, S: 14)
dri2: 153.901 (MIN: 153.454, MAX: 165.578, CV: 3.32846%, S: 14)

This is now invalid since the boost came from DRI3 swap elision which is no longer used.