[BACKEND] Support dot op when the output is mma encoding and allowtf32 is true (#937)

This commit is contained in:
Keren Zhou
2022-12-03 11:14:12 -08:00
committed by GitHub
parent 8edfe813a5
commit f2fcaeabf3
5 changed files with 105 additions and 72 deletions

View File

@@ -81,9 +81,11 @@ void MembarAnalysis::transfer(Operation *op, RegionInfo *regionInfo,
return;
}
if (isa<triton::gpu::AsyncWaitOp>(op)) {
// If the current op is an async wait, we insert a barrier op and sync
// previous reads and writes.
if (isa<triton::gpu::AsyncWaitOp>(op) &&
!isa<gpu::BarrierOp>(op->getNextNode())) {
// If the current op is an async wait and the next op is not a barrier we
// insert a barrier op and sync
regionInfo->sync();
OpBuilder::InsertionGuard g(*builder);
builder->setInsertionPointAfter(op);
builder->create<gpu::BarrierOp>(op->getLoc());