Testing The LLVM SI Machine Instruction Scheduler

Landing last month in the LLVM SVN/Git code-base was the SI machine scheduler for the AMDGPU LLVM back-end. This scheduler has the potential to improve the performance for some hardware/workloads, but not by the wide margins originally reported by some early testers.

While the SI machine scheduler has been in the LLVM back-end, landing in Mesa Git a few days ago was an option for easily enabling it.

Assuming you are building against the latest LLVM 3.8 or LLVM 3.9 SVN code, the R600_DEBUG=sisched option on Mesa Git can be used for turning on the LLVM SI machine scheduler.

While it's on my TODO list for running some of these SI machine instruction scheduler tests, a Phoronix reader has already run a few tests on his R9 Fury in conjunction with the Linux 4.5 development kernel.


These SI machine scheduler tests have already been discussed within our forums, but if you missed these preliminary independent results, you can find them via this OpenBenchmarking.org result file. However, at least for this user's system, the sisched option doesn't make a huge difference. I'll run some of my own tests on multiple GPUs when time allows.