On Sat, Jun 24, 2017 at 08:07:15PM +0200, Christian König wrote: > Am 23.06.2017 um 19:39 schrieb John Brooks: > >This patch series is intended to improve performance when limited CPU-visible > >VRAM is under pressure. > > > >Moving BOs into visible VRAM is essentially a housekeeping task. It's faster to > >access them in VRAM than GTT, but it isn't a hard requirement for them to be in > >VRAM. As such, it is unnecessary to spend valuable time blocking on this in the > >page fault handler or during command submission. Doing so translates directly > >into a longer frame time (ergo stalls and stuttering). > > Sorry, but that strongly sounds like you are messing with things you don't > fully understand. > > Blocking in the page fault handler is mandatory to handle the page fault > correctly. So that is not something we can change easily. I do understand that. Indeed, the page fault handler must block until the memory is accessible. But the memory doesn't have to be in visible VRAM to be accessible; it could also be in GTT. So, what I meant was that it does not have to block for the excessively long time it takes to move BOs into visible VRAM when it is already full. It could spend less time blocking by just moving it to GTT in that situation. I apologize for the poor wording. John > > Regards, > Christian. > > > > >The problem worsens when attempting to move BOs into visible VRAM when it is > >full. This takes much longer than a simple move because other BOs have to be > >evicted, which involves finding and then moving potentially hundreds of other > >BOs, which is very time consuming. In the case of limited visible VRAM, it's > >important to do this sometime to keep the contents of visible VRAM fresh, but > >it does not need to be a blocking operation. If visible VRAM is full, the BO > >can be read from GTT in the meantime and the BO can be moved to VRAM later. > > > >Thus, I have made it so that neither the command submission code nor page fault > >handler spends time evicting BOs from visible VRAM, and instead this is > >deferred to a workqueue function that's queued when CS requests BOs flagged > >AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED. > > > >Speaking of CPU_ACCESS_REQUIRED, I've changed the handling of that flag so that > >the kernel driver can clear it later even if it was set by userspace. This is > >because the userspace graphics library can't know whether the application > >really needs it to be CPU_ACCESS_REQUIRED forever. The kernel driver can't know > >either, but it does know when page faults occur, and if a BO doesn't appear to > >have any page faults when it's moved somewhere inaccessible, the flag can be > >removed and it doesn't have to take up space in CPU-visible memory anymore. > >This change was based on IRC discussions with Michel. > > > >Patch 7 fixes a problem with BO moverate throttling that causes visible VRAM > >moves to not be throttled if total VRAM isn't full enough. > > > >I've also added a vis_vramlimit module parameter for debugging purposes. It's > >similar to the vramlimit parameter except it limits only visible VRAM. > > > >I have tested this patch set with the two games I know to be affected by > >visible VRAM pressure: DiRT Rally and Dying Light. It practically eliminates > >eviction-related stuttering in DiRT Rally as well as very low performance if > >visible VRAM is limited to 64MB. It also fixes severely low framerates that > >occurred in some areas of Dying Light. All my testing was done with an R9 290 > >with 4GB of visible VRAM with an Intel i7 4790. > > > >-- > >John Brooks (Frogging101) > > > >_______________________________________________ > >amd-gfx mailing list > >amd-gfx at lists.freedesktop.org > >https://lists.freedesktop.org/mailman/listinfo/amd-gfx > >
