Rusticl lead developer Karol Herbst of Red Hat has merged a set of patches for this Rust-written open-source OpenCL driver to reduce the amount of memory allocations that happen and in turn helping reduce CPU overhead.
Long story short, the set of 13 patches now merged to the Mesa 25.3 development code help in further reducing the number of memory allocations in turn providing a small efficiency win. On an AMD graphics card Karol observed around a 2% improvement from this patch series.
Karol explained in the merge for this generic Gallium3D OpenCL driver implementation:
“This is a collection of a bunch of small perf improvements lowering CPU overhead. Nothing of this is substantial enough that it does much, but it does reduce the amount of allocation when launching kernels. And also cleans up the code a bit.
On my AMD card this seems to have an impact of around 2%, so that’s great 🙃”
Every little bit counts. In case you missed my recent Rusticl benchmarking, see Rusticl vs. AMD ROCm Performance On Ryzen AI Max+ “Strix Halo” and Rusticl vs. Intel Compute Runtime Performance For OpenCL On Battlemage.
We’ll see what more Rusticl optimizations come by the time of the Mesa 25.3 feature release around November.
