RadeonSI Ups Its Compiler Threads To Let Shader-DB Run Faster On Modern Systems

The RadeonSI compiler queue can now run across more CPU cores/threads of modern systems though it appears this will primarily just benefit those running the shader-db shader test cases.

Marek Olšák of AMD has increased the number of compiler threads depending upon the CPU. The compiler queue in the RadeonSI Gallium3D driver was previously limited to three threads, which is a bottleneck for shader-db when running on modern systems commonly featuring eight or now sixteen threads.

This three thread queue has been replaced with some new logic where for systems with more than 12 threads it will use the number of threads times three divided by four, systems with six to 11 threads will use that count minus two, or systems with two to five threads use the thread count minus one.

That change is now in Git for Mesa 18.2. Even if you are not running the shader-db shader test case collection, this is good news in general as hopefully more people will be running it now or more frequently in looking for driver performance changes/regressions now that it should be able to run quicker on modern systems.