6% performance drop in GpuTest v0.7 Volplosion with "remove GLSL IR optimisation loop" (-> fails to compile as SIMD16)
Submitted by Eero Tamminen
Assigned to Intel 3D Bugs Mailing List
Link to original bug (#100913)
Description
Patch series culminating in this commit changed performance in several tests:
commit ad55b1a7 Author: Timothy Arceri timothy.arceri@collabora.com AuthorDate: Wed Jan 18 10:28:22 2017 +1100 Commit: Timothy Arceri tarceri@itsqueeze.com CommitDate: Mon Apr 24 12:08:14 2017 +1000
i965: remove GLSL IR optimisation loop
IVB is running into some spilling issues in piglit with the
loop removed. However those tests are not really reflective
of a real world use case, also fp64 is brand new to IVB
so we leave the spilling issues to be resolved at a later
time.
On most of the platforms GpuTest v0.7 PixMark Volplosion test performance dropped by ~6%, on SKL GT4e, a bit more.
(There were few performance improvement around same time, but they were small enough that I couldn't bisect them to this commit, besides the SynMark shader compilation speed test which improved by ~9%.)
Volplosion test is fairly pure ALU test. While it has one copy shader, none of the other shaders do any texture accesses, including the heavy main fragment shader, i.e. texture access scheduling doesn't affect it.
Version: git