Skip to content

Nearly 10% performance drop in SynMark CSCloth from "i965: Emit CS stall before MEDIA_VFE_STATE"

Submitted by Eero Tamminen

Assigned to Kenneth Graunke @kwg

Link to original bug (#104100)

Description

Following commit drops performance in SynMark CSCloth test, as expected:

commit 55a97db5 Author: Kenneth Graunke kenneth@whitecape.org AuthorDate: Tue Oct 31 09:57:54 2017 -0700 Commit: Matt Turner mattst88@gmail.com CommitDate: Mon Dec 4 10:02:46 2017 -0800

i965: Emit CS stall before MEDIA_VFE_STATE.

This fixes hangs on GFXBench 5's Aztec Ruins benchmark.

Unfortunately, it regresses OglCSCloth performance by about 10%. There
are some ideas for fixing that.

The Vulkan driver already emits this stall.

Reviewed-by: Matt Turner <mattst88@gmail.com>

The drop is:

  • 8-9% on SKL GT4e
  • 5-6% on KBL GT3e
  • 3-4% on SKL & KBL GT2
  • within (largish) variance on BDW GT2 & GT3

On HSW GT it would seem to actually improve perf marginally.

Latter is interesting because Kenneth's trial fix for the performance regression (verified to fix it): https://cgit.freedesktop.org/~kwg/mesa/commit/?h=vfe-stall-3&id=ab9bff0ac43f0d878106ba124e84edca9b359c8a

Is according to Kenneth causing problems on HSW. Maybe it could be applied only on newer GENs?

Version: git

To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information