On HSW GT2, performance dropped in following tests:
- 10-20% in 4 SynMark2 v7 Terrain tests
- 4% in GfxBench Manhattan 3.0
- 2% in GfxBench T-Rex
- 1-2% in GfxBench CarChase
GfxBench tests perf drop both for onscreen & offscreen versions.
Similar, but much smaller drops are visible also on BYT. I don't see drops in these cases on SNB or GEN8+ platforms, so it seems GEN7 specific.
This drop happens between following commits:
- Good: 7f3eab03fe041bc28900a02f7ae503920753ea6b 2017-10-02 freedreno/a5xx: fix missing restore state
- Bad: df6b320a83c89b6401fde888375529b3fc66f4fa 2017-10-03 etnaviv: Set up unknown GC3000 states
Looking at the commits between this range, one of following changes could be the cause for the drop:
- intel: Always set Cube Face Enables for all surfaces.
- refactor for: i965: Implement ARB_indirect_parameters
- glsl refactor done by Ian
Hopefully somebody has time to bisect this further.
I don't have a Haswell GT2, but I tried to reproduce this on a Haswell GT3e.
I tried both OglTerrainFlyTess and Gl41TerrainFlyInst on both commits, and the framerates were basically identical. So, I guess I cannot reproduce this.
I can reproduce earlier good results, so the drop wasn't due to setup change on our HSW.
During past ~3 months, there haven't been significant changes in terrain tests results before this on HSW GT2 (I don't have data for GT3e). So, kernel and X versions shouldn't matter that much (only change was modesetting dropping perf on BYT). Our Mesa test setup is currently using kernel 4.13 and X server from same time frame (mid September), with rest being Ubuntu 16.04.
Maybe the issue requires GPU to be slower than yours compared to the CPU, to be visible? Our HSW is i7-4790, 1200Mhz GT2. If you don't have anything similar at your side, one option would be to try disabling one slice / eMMC or just lowering GPU max speed enough to make the tests fully GPU bound.
During the same commit range there's also clear 15% drop in SynMark PSPom performance on BYT (with 1-channel memory). On HSW GT2, PSPom drop is 2%.
Because GfxBench drops happen also with offscreen versions, I don't think it's required that Mesa would be used by the whole system (by X modesetting driver) like in our system.