On 14/06/2024 10:53, Christian König wrote:
if (domain & abo->preferred_domains &
AMDGPU_GEM_DOMAIN_VRAM &&
- !(adev->flags & AMD_IS_APU))
- places[c].flags |= TTM_PL_FLAG_FALLBACK;
+ !(adev->flags & AMD_IS_APU)) {
+ /*
+ * When GTT is just an alternative to VRAM make sure
that we
+ * only use it as fallback and still try to fill up VRAM
first.
+ */
+ if (abo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT)
+ places[c].flags |= TTM_PL_FLAG_FALLBACK;
+
+ /*
+ * Enable GTT when the threshold of moved bytes is
+ * reached. This prevents any non essential buffer move
+ * when the links are already saturated.
+ */
+ places[c].flags |= TTM_PL_FLAG_MOVE_THRESHOLD;
+ }
For the APU case I *think* this works, but for discrete I am not sure
yet.
Agree, APUs are basically already fine as they are. VRAM is just used so
that it isn't wasted there.
Well yeah it works, but because re-validation is broken so it cannot hit
the broken migration budget. ;)
As a side note and disclaimer, the TTM "resource compatible" logic has
a half-life of about one week in my brain until I need to almost
re-figure it all out. I don't know if it just me, but I find it really
non-intuitive and almost like double, triple, or even quadruple
negation way of thinking about things.
Yeah I was also going back and forth between the different approaches
multiple times and just ended up in this implementation because it
seemed to do what I wanted to have.
It's certainly not very intuitive what's going on here.
It is not helping that with this proposal you set threshold on just
one of the possible object placements which further increases the
asymmetry. For me intuitive thing would be that thresholds apply to
the act of changing the current placement directly. Not indirectly via
playing with one of the placement flags dynamically.
Interesting idea, how would the handling then be? Currently we have only
the stages - 'don't evict' and 'evict'. Should we make it something more
like 'don't move', 'move', 'evict' ?
Intuitively I would think "don't move" aligns with the "out of migration
budget" concept.
Since in this patch you add move_threshold to ttm_operation_context
could it simply be used as the overall criteria if it is set?
In a way like:
1. If the current placement is from the list of userspace supplied
valid ones, and
2. Migration limit has been set, and
3. It is spent.
-> Then just don't migrate, return "all is good" from ttm_bo_validate.
Though I am not sure at the moment how that would interact with the
amdgpu_evict_flags and placements userspace did not specify.
Anyway, lets see.. So you set TTM_PL_FLAG_MOVE_THRESHOLD and
TTM_PL_FLAG_FALLBACK on the GTT placement, with the logic that it will
be considered compatible while under the migration budget?
(Side note, the fact both flags are set I also find very difficult to
mentally model.)
Say a buffer was evicted to GTT already. What then brings it back to
VRAM?
The first subsequent ttm_bo_validate pass (!evicting) says GTT is fine
(applicable) while ctx->bytes_moved < ctx->move_threshold, no? Isn't
that the opposite of what would be required and causes nothing to be
migrated back in? What am I missing?
The flag says that GTT is fine when ctx->bytes_moved >=
ctx->move_threshold. The logic is exactly inverted to what you described.
This way a BO will be moved back into VRAM as long as bytes moved
doesn't exceed the threshold.
I'm afraid I need to sketch it out... If buffer is currently in GTT and
placements are VRAM+GTT.
ttm_bo_validate(evicting=false)
1st iteration:
res=GTT != place=VRAM
continue
2nd iteration:
res=GTT == place=GTT+FALLBACK+THRESHOLD
ttm_place_applicable(GTT)
moved < threshold
return true
Buffer stays in GTT while under migration budget -> wrong, no? Or am I
still confused?
Regards,
Tvrtko
Setting both flags has the effect of saying: It's ok that the BO stays
in GTT when you either above the move threshold or would have to evict
something.
Regards,
Christian.
Regards,
Tvrtko