AMDGPU Gang Submit Sent In For Linux 6.1
On Friday AMD submitted a feature pull request to DRM-Next of some last minute changes they would like to see as part of the upcoming Linux 6.1 kernel.
The feature pull mostly consists of fixes and changes for new IP blocks with upcoming GPUs, but there is one major feature included: gang submit. The gang submit code allows ensuring work is executed on multiple engines at the same time -- and that the work won't start until all of the work is ready for execution across the affected engines. Mesa's Radeon Vulkan driver "RADV" is currently one user awaiting the gang submission support for completing the Vulkan mesh shader support and exposing it by default. So assuming no issues with this late pull request hitting DRM-Next for Linux 6.1, this next kernel could be playing nicely for the Vulkan mesh shader support with RADV.
Some of the other last-minute changes as part of the Friday pull include adding more firmware types to the DebugFS/ioctl version queries, DCN 3.1.x / 3.2.x fixes, fixes for compiling under LLVM Clang, GPUVM locking fixes, and also a Vega 10 PWM fan speed fix. It's mostly fixes in this pull besides the gang submit functionality and the additional firmware type version queries support for user-space.
The full list of these AMDGPU patches can be found via the mailing list post.
The feature pull mostly consists of fixes and changes for new IP blocks with upcoming GPUs, but there is one major feature included: gang submit. The gang submit code allows ensuring work is executed on multiple engines at the same time -- and that the work won't start until all of the work is ready for execution across the affected engines. Mesa's Radeon Vulkan driver "RADV" is currently one user awaiting the gang submission support for completing the Vulkan mesh shader support and exposing it by default. So assuming no issues with this late pull request hitting DRM-Next for Linux 6.1, this next kernel could be playing nicely for the Vulkan mesh shader support with RADV.
Some of the other last-minute changes as part of the Friday pull include adding more firmware types to the DebugFS/ioctl version queries, DCN 3.1.x / 3.2.x fixes, fixes for compiling under LLVM Clang, GPUVM locking fixes, and also a Vega 10 PWM fan speed fix. It's mostly fixes in this pull besides the gang submit functionality and the additional firmware type version queries support for user-space.
The full list of these AMDGPU patches can be found via the mailing list post.
6 Comments