Trying Out The AMDGPU Kernel Driver's Experimental GPU Scheduler

Yesterday after finishing up the first open-source AMDGPU driver benchmarks of the Radeon R9 Fury, I decided to also give a go on the new AMDGPU driver scheduler that's an experimental feature for Linux 4.3.

In July AMD posted patches for a GPU scheduler for AMDGPU and it will be added to Linux 4.3. For this next kernel version, the GPU scheduler is disabled by default but can be enabled by setting the amdgpu.enable_scheduler=1 kernel command-line parameter for the AMDGPU module.

This scheduler is explained as, "This is more for OpenGL workloads (and compute going through kernel graphics driver I guess). The HSA stack uses a hardware-based scheduler, but the primary purpose of that scheduler is to provide support for an arbitrary number of user-space compute rings. The kernel graphics driver doesn't need that because it uses a single kernel ring and allows multiple processes to take turns submitting to it. The GPU scheduler Alex just posted basically maintains a number of kernel rings that are not connected directly to the hardware, and feeds work from those rings into that 'single kernel ring' which the HW pulls from."

I tried it out on the same system with DRM-Next, LLVM 3.8 SVN, and Mesa 11.1-dev Git as used in yesterday's article. Unfortunately, I didn't make it that far.

While the system boot up fine to the Xfce desktop, launching the Tesseract game yielded the system hanging on its loading screen. When trying to run Xonotic, the system didn't even make it as far as the loading screen before the system became unresponsive.

I'll give it another shot when it comes time for Linux 4.4.