15% perf drop in GfxBench Manhattan 3.1 performance
Setup:
- HW: 18EU BXT, CML-H GT2, SKL GT4e (= all that I have working)
- OS: Ubuntu 20.04
- XDM / Desktop: lightdm / Unity
- kernel: drm-tip v5.8
- X server: git head (= dmabuf / E2E RBC enabled)
Between following Mesa commits:
- 2020-11-24 15:47:06, 7f223a23: spirv: Implement SpvOpConvertUToAccelerationStructureKHR
- 2020-11-25 16:46:16, ff9ea469: intel/disasm: Don't rely on FALLTHROUGHTs to print unsupported SFID
Iris GfxBench Manhattan 3.0 & 3.1 performance dropped on all platforms that I'm still tracking.
On BXT, (both onscreen & offscreen) performance regression is:
- 15% in GfxBench Manhattan 3.1
- 10% in GfxBench Manhattan 3.0
Notes:
- It probably affects all GENs as in your trends it's visible also on GEN11: http://otc-mesa-android.jf.intel.com/results/mesa_master/history#gfxbench5_manhattan_o
- No performance change with HadesCanyon (RadeonSI / Vega) i.e. regression is Intel specific
- GfxBench AztecRuins GL, GLES and Vulkan versions performance improved by 1-3% in the same commit range (that could be due to a different commit though)
- No significant performance changes in my other benchmarks, but in your trends I'm seeing a drop also in Team Fortress 2. Its previous result is so old though that it may not be related: http://otc-mesa-android.jf.intel.com/results/mesa_master/history#Team%20Fortress%202