[BXT/GLK] Up to 5% perf drop in fullscreen onscreen benchmarks

Setup:

• HW1: BXT/APL J4205 (18EU)
• HW2: GLK (J4005)
• OS: Ubuntu 20.04
• XDM / Desktop: lightdm / Unity
• X server & Mesa: git head within past few months (= dmabuf / E2E RBC enabled)
• kernel: drm-tip git

Between following kernel commits:

• b058a44fd6 drm-tip: 2020y-09m-26d-21h-35m-55s UTC integration manifest
• a8a96de162 drm-tip: 2020y-09m-28d-15h-19m-34s UTC integration manifest

Kernel performance dropped on BXT in onscreen fullscreen tests:

• 5% GpuTest Triangle
• 2-3% GfxBench Manhattan 3.0 & T-Rex, SynMark HdrBloom
• 1-2% GfxBench Manhattan 3.1, SynMark TerrainFly*
• 0.5-1% GfxBench CarChase, several of SynMark Batch*, Geometry* and VS* tests

(Because it's several weeks since the regression, one can see clearly from trends which benchmarks have consistently lower perf even with <1% perf regressions.)

Exact perf regression is harder to get for GLK because of the #1205 (closed) functionality regression having larger impact on it, but it's approximately same as for BXT.

No idea whether this is in any way related to earlier (still open) BXT perf regression #2180.

Of note:

commit 55ea1cb178ef85677b35126685fbf5d8d649b62e
Author: Karthik B S <karthik.b.s@intel.com>
Date:   Mon Sep 21 16:32:10 2020 +0530

    drm/i915: Enable async flips in i915
    
    Enable asynchronous flips in i915 for gen9+ platforms.

but all i915 commits within that range are display related.

added UX feature: display/Other labels

There's another BXT, with ClearLinux (i.e. performance governor instead of powersave/ondemand used by Ubuntu) and Git version of Weston. On that device, I don't see performance regression with X11 version of GfxBench (i.e. when it's running on top of Xwayland), nor with Wayland version of GfxBench.

So it seems to be specific to X behaviour, and/or power management, as BXT devices are TDP limited.

Looking at the simplest, fullscreen triangle case, CPU uses in this case much less power than GPU (uncore). However, that small CPU power usage grew by 20%, and as result GPU was run at few percent slower speed, which could explain the performance drop.

Hi Reporter, can you try to reproduce this issue on latest drm-tip.

https://cgit.freedesktop.org/drm-tip

Please post your observations on it and any steps followed to reproduce the issue.

all i915 commits within that range are display related.

As that regression is visible only in onscreen, not offscreen versions of the benchmarks, and only in fullscreen ones, not windowed ones, it could be display related. It should be also noted that 3D benchmarks always run with Vsync disabled (for obvious reasons).

can you try to reproduce this issue on latest drm-tip.

Performance is still at the regressed level with the last evening drm-tip.

At end of January there was about same sized additional 3D perf drop, but that seems to have been fixed in late March => Currently kernel perf is still at the original regressed lower level.

mentioned in issue #2180

GfxBench tests are not any more regressed. Those ones were actually more than fixed by the March kernel change. For rest of the regressed cases, only the additional regression was taken care of.

X server change in mid June improved things marginally.

The X server change in September that both:

• removed the extra copy done for non-vsynched fullscreen updates (= perf improvement for mem-BW limited cases)
• disabled support for modifiers i.e. support for E2E RBC (= perf drop for mem-BW limited cases with well compressing screen content)

More than compensated for kernel's regression in GpuTest fullscreen case.

As to remaining SynMark cases, Mesa regression in mid-June made things worse for them, which the September X server change mostly compensated for, but in total their perf is still regressed.

Because there have been so many perf changes for these benchmarks from multiple components, and nobody seems to be looking at this >1 year old kernel regression, I'm closing this bug.

closed