Intel Xe Driver In Linux 7.1 Enabling THP For Device Pages As A Big SVM Win
Sent out today was a new batch of "drm-xe-next" material of Intel Xe kernel graphics driver improvements ready for the upcoming Linux 7.1 kernel cycle. Standing out in this pull is enabling Transparent Hugepages (THP) support for drm_pagemap as a big win for those making use of Shared Virtual Memory (SVM) for GPU compute and the like.
Intel engineers have been working in recent months on enabling Transparent Hugepages for GPU memory migration. THP for device pages during migration avoids splitting folios and needing to loop multiple times on all the pages to perform the necessary operations on a page-level basis.
Intel driver engineer Francois Dugast commented on the drastic benefit this THP support for the device pages has with the Intel Xe driver:
Wow! This THP support has quite a big impact and will surely be beneficial for Intel GPU compute, especially with the newly-announced Arc Pro B70 and the like.
This week's Intel Xe pull also introduces the new xe_vm_get_property_ioctl interface that allows user-space to request a list of observed failed page faults from a specified VM.
There is also improved VF FLR synchronization for enhancing Xe VFIO support, an L2 cache flush optimization for Xe3P_LPG, page reclamation fixes, and a variety of other low-level improvements.
The full list of these latest Intel Xe graphics driver changes ready for the Linux 7.1 kernel cycle can be found via this pull request now pending against DRM-Next.
Intel engineers have been working in recent months on enabling Transparent Hugepages for GPU memory migration. THP for device pages during migration avoids splitting folios and needing to loop multiple times on all the pages to perform the necessary operations on a page-level basis.
Intel driver engineer Francois Dugast commented on the drastic benefit this THP support for the device pages has with the Intel Xe driver:
"In Xe, this drastically improves performance when using SVM. The GT stats below collected after a 2MB page fault show overall servicing is more than 7 times faster, and thanks to reduced CPU overhead the time spent on the actual copy goes from 23% without THP to 80% with THP:"
Wow! This THP support has quite a big impact and will surely be beneficial for Intel GPU compute, especially with the newly-announced Arc Pro B70 and the like.
This week's Intel Xe pull also introduces the new xe_vm_get_property_ioctl interface that allows user-space to request a list of observed failed page faults from a specified VM.
There is also improved VF FLR synchronization for enhancing Xe VFIO support, an L2 cache flush optimization for Xe3P_LPG, page reclamation fixes, and a variety of other low-level improvements.
The full list of these latest Intel Xe graphics driver changes ready for the Linux 7.1 kernel cycle can be found via this pull request now pending against DRM-Next.