AMD Linux Driver's LRU Bulk Moves Can Be A Big Help For Demanding Linux Games

[phoronix]

Phoronix: AMD Linux Driver's LRU Bulk Moves Can Be A Big Help For Demanding Linux Games

Sadly not currently queued as a fix for the Linux 5.4 kernel, re-enabling the LRU bulk moves functionality can be a significant boost for helping with the Radeon graphics driver performance for Linux gaming...

http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-LRU-Bulk-Moves-Valve

[xxmitsu]

Michael, you should subscribe to mesa gitlab ACO pending for inclusion: https://gitlab.freedesktop.org/mesa/..._requests/2009

[Venemo]

How do you enable the LRU bulk moves if you want to compile a kernel that has this feature?

[M@GOid]

Unfortunately, Padoka's PPA is not updated for about 2 months. Anyone knows another PPA that is not oibaf's, since he updates daily and that gets old fast.

[skeevy420]

How do you enable the LRU bulk moves if you want to compile a kernel that has this feature?

It patches into 5.2 cleanly....

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 48349e4f0701..fd3fbaa73fa3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -603,14 +603,12 @@ void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
     struct ttm_bo_global *glob = adev->mman.bdev.glob;
     struct amdgpu_vm_bo_base *bo_base;
 
-#if 0
     if (vm->bulk_moveable) {
         spin_lock(&glob->lru_lock);
         ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
         spin_unlock(&glob->lru_lock);
         return;
     }
-#endif
 
     memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
 
--
2.17.1

[AsuMagic]

To help putting things into perspective, to hit 60FPS you need to be below 16.6ms/frame.
I don't know how long the rest of the frametime in these tests was, but this may be a very significant improvement in this scenario.

[ms178]

bridgman When speaking of a big help for performance, could you please forward this to the concerned team which should get in touch with the glibc developers (Red Hat Engineer Partner Management or contact Florian Weimer directly) as it seems that glibc does enable a fast path only for capable Intel CPUs because they need more insights into AMD's architectures: https://sourceware.org/bugzilla/show_bug.cgi?id=24979

As this issue came up recently, here is another Mailing list entry of interest to justify the current behaviour: https://sourceware.org/ml/libc-alpha.../msg00155.html

[pal666]

To help putting things into perspective, to hit 60FPS you need to be below 16.6ms/frame.

they submit on separate thread, so it doesn't add to other per-frame work time if you have enough cpu cores. still, for 240 fps you need 4 ms and it was not possible without fix

[[TV]]

Unfortunately, Padoka's PPA is not updated for about 2 months. Anyone knows another PPA that is not oibaf's, since he updates daily and that gets old fast.

I recently switched from Padoka's PPA back to "Ubuntu-X" team PPA: team: https://launchpad.net/~ubuntu-x-swat...ubuntu/updates

[ernstp]

Unfortunately, Padoka's PPA is not updated for about 2 months. Anyone knows another PPA that is not oibaf's, since he updates daily and that gets old fast.

I made my own PPAs if anyones interested.

Release candidates:

https://launchpad.net/~ernstp/+archive/ubuntu/mesarc

Mesa ACO:

https://launchpad.net/~ernstp/+archive/ubuntu/mesaaco