[PATCH v2] Visible VRAM Management Improvements

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Changes in this version:
- Dropped former patch 1 ("Separate placements and busy placements") as it was
  no longer necessary
- Dropped former patch 4 ("Don't force BOs into visible VRAM if they can go to
  GTT instead") as it was unnecessary and had too many side effects (Thanks
  Christian)
- Dropped former patches 8 and 9 ("Asynchronously move BOs to visible VRAM" and
  an associated optimization), as there may be better ways to address the root
  of the problem
- Added a u64 cast to patch 1 (Thanks Christian)
- Optimized patch 2 in the case where visible VRAM is not smaller than total
  VRAM (Thanks Michel)

The series as it is now fixes Dying Light and improves DiRT Rally frame times.
Without the asynchronous moves, there is still occasional stalling in DiRT
Rally.


The "Set/clear CPU_ACCESS_REQUIRED flag on page fault and CS" patch received
objections, but I have not removed it yet as there seems to be some support for
the idea. The precise details of when to set and clear the flag may need further
discussion.

I note that if the patch is removed (but retaining the preventative measures
against repeated moves back and forth between GTT and invisible VRAM), there
are very few evictions but the average framerate drops from ~89 to ~82.
Presumably we'll see either higher memory presure or lower framerates depending
on whether userspace over- or under-estimates when to set this flag. What we do
in the kernel may depend on which side of this tradeoff we want to end up on
for now.

--
John Brooks (Frogging101)



[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux