
Marek Olšák on Friday merged another dozen commits with a new varying linking optimization pass, which for now is just enabled for the RadeonSI Gallium3D driver. The new optimizations carry out more dead input/output removal, output deduplication, backward inter-shader code motion, compaction, and more.
But for those not caring about all the low-level details, the exciting aspect of this low-level code improvement is:
"I also noticed that GLCTS finished 30% faster with this on Radeon 7600, probably because the pass moves a lot of code from FS to VS (including slow FP64 code) due to how the tests are written."
The OpenGL Conformance Test Suite (GL CTS) is running 30% faster with this latest merge request. The OpenGL CTS is commonly used for regression testing and ensuring the driver support is in good shape. Running 30% faster in 2024 is certainly a nice achievement.
It will be interesting to see if this merge request benefits other workloads significantly but even a 30% faster run-time for the OpenGL CTS is important for CI/CD and developer productivity on RadeonSI Gallium3D. This code is merged for Mesa 24.1.