[AIE2P] Fix modeling of resource conflicts in the scheduling model #313
Conversation
katerynamuts
requested review from
abhinay-anubola,
abnikant,
andcarminati,
F-Stuckmann,
gbossu,
khallouh,
konstantinschwarz,
martien-de-jong,
niwinanto,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
katerynamuts
force-pushed
the
kmuts.scheduling_model
branch
2 times, most recently
from
d06f251 to
ae601f7
Compare
|
I added a test for the known conflict that is fixed by this PR, namely, the conflict between |
| # | ||
| # (c) Copyright 2025 Advanced Micro Devices, Inc. or its affiliates | ||
| # RUN: llc -march=aie2p -run-pass=postmisched %s -o - | FileCheck %s | ||
|
|
There was a problem hiding this comment.
Perhaps add a comment about which resource we are checking and in which cycles they are used by the instruction.
Also see AIE/aie2/schedule/resource/*
Is this series of NOPs really caused by a resource conflict, or is there a data dependence?
There was a problem hiding this comment.
I changed the tests a little bit so that there is no data dependency.
| ; CHECK-NEXT: $lfl0 = VMOV_alu_mv_mv_x killed $x9 | ||
| ; CHECK-NEXT: NOP | ||
| $p1, $lf0, $r25 = VLDA_FILL_512 $p1, $lf0, $r25 | ||
| $lfl0 = VMOV_alu_mv_mv_x $x9 |
There was a problem hiding this comment.
maybe $lfl1 here to avoid any WAW dependency?
katerynamuts
force-pushed
the
kmuts.scheduling_model
branch
from
ae601f7 to
ba91f0b
Compare
katerynamuts
force-pushed
the
kmuts.scheduling_model
branch
from
ba91f0b to
c458323
Compare
| ; CHECK-NEXT: NOP | ||
| ; CHECK-NEXT: $lfl1 = VMOV_alu_mv_mv_x killed $x9 | ||
| ; CHECK-NEXT: NOP | ||
| ; CHECK-NEXT: NOP |
There was a problem hiding this comment.
Nit: maybe add the test in a pre-commit to show that the instruction is moved up by one cycle to avoid the conflict after the change.
I've run QoR tests with the updated scheduling model and the failed and passed tests remain the same.