compiler: Tweak par-tile unrolling and revamp MPI topology #2426
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #2426 +/- ##
==========================================
- Coverage 86.99% 86.98% -0.01%
==========================================
Files 236 236
Lines 44789 44811 +22
Branches 8351 8359 +8
==========================================
+ Hits 38962 38981 +19
- Misses 5099 5100 +1
- Partials 728 730 +2 ☔ View full report in Codecov by Sentry. |
mloubout
left a comment
mloubout
left a comment
There was a problem hiding this comment.
Minor comments, good to me
| 'x': ('*', 1, 1), | ||
| 'y': (1, '*', 1), | ||
| 'z': (1, 1, '*'), | ||
| 'xy': ('*', '*', 1), |
There was a problem hiding this comment.
Maybe add the xz, yz, xyz too to be generic
There was a problem hiding this comment.
I avoided them deliberately because I know they cannot provide better performance, but if you want, I can put them in
| if topology and len(topology) == len(self.shape): | ||
| self._topology = topology | ||
| else: | ||
| self._topology = None |
There was a problem hiding this comment.
I guess could to pad with 1 if only partial provided but probably better yeah. Maybe add a warning so user knows input been ignored
There was a problem hiding this comment.
Just thought that padding with 1 is worse because it might lead to crashes (by eg creating a 1,1,1 topology)
There was a problem hiding this comment.
Added a warning
devito/passes/clusters/blocking.py
Outdated
| self.umt_sparse = UnboundTuple(*par_tile[0], 1) | ||
| if par_tile.is_multi: | ||
| # The user has supplied one specific par-tile per blocked nest | ||
| self.umt = par_tile |
There was a problem hiding this comment.
self.umt = par_tile can be moved outside the if
FabioLuporini
force-pushed
the
tweak-tuner
branch
from
62533b8 to
1efe5d4
Compare
Admittedly two orthogonal things