Don't emit atomics for thread-safe assignment patterns

Fixes tensor-compiler#316
Infinoid · Feb 10, 2021 · 5e5ef41 · 5e5ef41
1 parent 1114197
commit 5e5ef41
Showing 1 changed file with 8 additions and 2 deletions.
diff --git a/src/index_notation/transformations.cpp b/src/index_notation/transformations.cpp
@@ -1033,8 +1033,14 @@ IndexStmt scalarPromote(IndexStmt stmt, ProvenanceGraph provGraph,
  return;
  }
 
- stmt = forall(i, body, foralli.getParallelUnit(),
- foralli.getOutputRaceStrategy(), foralli.getUnrollFactor());
+ auto parallelunit = foralli.getParallelUnit();
+ auto racestrategy = foralli.getOutputRaceStrategy();
+ auto unrollfactor = foralli.getUnrollFactor();
+ // compilers reject openmp atomic pragmas for assignments from thread-local temporaries. See bug #316.
+ if(parallelunit == ParallelUnit::CPUThread && racestrategy == OutputRaceStrategy::Atomics) {
+ racestrategy = OutputRaceStrategy::IgnoreRaces;
+ }
+ stmt = forall(i, body, parallelunit, racestrategy, unrollfactor);
  for (const auto& consumer : consumers) {
  stmt = where(consumer, stmt);
  }