compiler: Fix parlang reductions over >= 4 loops #2417
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #2417 +/- ##
==========================================
- Coverage 86.74% 86.70% -0.05%
==========================================
Files 235 235
Lines 44701 44727 +26
Branches 8288 8291 +3
==========================================
+ Hits 38777 38779 +2
- Misses 5195 5221 +26
+ Partials 729 727 -2 ☔ View full report in Codecov by Sentry. |
| @@ -664,6 +664,16 @@ def max_mem_trans_size(self, dtype): | |||
| assert self.max_mem_trans_nbytes % np.dtype(dtype).itemsize == 0 | |||
| return int(self.max_mem_trans_nbytes / np.dtype(dtype).itemsize) | |||
|
|
|||
| def limits(self, compiler=None, language=None): | |||
There was a problem hiding this comment.
Do compiler and language get used anywhere? Why couldn't this be a property?
There was a problem hiding this comment.
they will in the future yeah, some limits may be compiler and/or language dependent
| return cluster | ||
|
|
||
| # If there are more parallel dimensions than the maximum allowed by the | ||
| # target platform, we must restrain the number of candidates | ||
| max_par_dims = platform.limits()['max-par-dims'] |
There was a problem hiding this comment.
why not just platform.max-par-dims as a property?
There was a problem hiding this comment.
because (see comment above) some of these limits may be e.g. programming-model dependent, besides being architecture-dependent
There was a problem hiding this comment.
The it's the same as having the compiler store it's complex/f16 dtype it doesn't belong there and should be part of a pass with if/dispatch
devito/arch/archinfo.py
Outdated
| language. | ||
| """ | ||
| return { | ||
| 'max-par-dims': np.inf, |
There was a problem hiding this comment.
I think np.inf might be problematic if used for indexing
There was a problem hiding this comment.
and indeed one test is failing because of this :D
norms of time functions with Buffer(1), where the time loop becomes parallel just like x/y/z because you don't have any to/t1/... -- see test |
FabioLuporini
force-pushed
the
restrain-device-blocking
branch
from
b22bb1e to
8c62ae9
Compare
No description provided.