[AIEX] Scheduler improvements #147
Conversation
gbossu
requested review from
SagarMaheshwari99,
abhinay-anubola,
abnikant,
andcarminati,
khallouh,
konstantinschwarz,
martien-de-jong and
stephenneuendorffer
as code owners
There was a problem hiding this comment.
nit: I think we first do MaxExpensiveIterations, then fall back to the global safety margins. Reverse the two terms here, and perhaps make the comment more precise?
There was a problem hiding this comment.
nit: I think the comment should be on the outside of this if. And then perhaps also order the branches accordingly.
There was a problem hiding this comment.
Could we have a [presched, RA] example that actually demonstrates reduced spilling?
There was a problem hiding this comment.
I really tried to manipulate that example, but can't get it to spill without writing absolutely ugly code. I'd leave it like this if that's fine to you. In a follow-up PR where I change the MachinePipeliner, i'll add an end-to-end IR test that shows Add2D getting nicely SW pipelined.
There was a problem hiding this comment.
CHECK: isTrackingPressure() is trivially true here.
There was a problem hiding this comment.
perhaps assert that we don't exceed MaxLatency. (or ConflictHorizon)
There was a problem hiding this comment.
Note: I'm reducing the limit here, but then this get multiplied by the number of pressure units required by the reg class. E.g. the number of free units we try to maintain for W is 2, for X it is 4, and for Y it is 8.
We want to increase the safety margin for one instruciton at a time here, instead of doing it for all instructions at once.
There was a problem hiding this comment.
Is there a rationale behind this number?
There was a problem hiding this comment.
Yes and no. I feel anything over 50 is too much, and anything below 10 is not enough if we need to move a couple of instructions up by 2-3 cycles. So 25 felt like a good compromise. And this works well for loops with an II between 5 and 10 cycles, which is the territory of the PreRA pipeliner for us.
There was a problem hiding this comment.
Considering your change, does this error become more common without this increase?
There was a problem hiding this comment.
Oh it never triggered, I just changed the condition to account for the extra iterations, otherwise we would fail thinking we are in an infinite loop.
There was a problem hiding this comment.
nit: const unsigned MaxPressure
In a follow-up commit, the premisched will re-order the instructions to reduce the pressure and avoid spills during RA.
gbossu
dismissed stale reviews from andcarminati and martien-de-jong
via
8d997c4
gbossu
force-pushed
the
gaetan.improve.scheds
branch
from
9a38355 to
8d997c4
Compare
- Reserve a certain number of registers, not regunits - Be extra careful when the region max pressure exceeds limits
gbossu
force-pushed
the
gaetan.improve.scheds
branch
from
8d997c4 to
71614b9
Compare
4 participants
Pre-RA: More conservative scheduling when under high register pressure. This is very helpful for SW pipelining. I'll have another PR which makes the MachinePipeliner find more schedules, and this PR here helps us not spill
Post-RA: Change the loop-aware scheduling to have an "expensive convergence" mode, when we increase the latency safety margin per instruciton, instead of for all instructions.
QoR results below. Overall it's good. There are some regressions, but we'll get rid of them with less unrolling and more SWP.