Skip to content

Commit a89c5bd

Browse files
authored
Restrict redux max usage to sm_100 (#7520)
Looks like it is not supported on sm_120 :(
1 parent 620237e commit a89c5bd

File tree

1 file changed

+1
-1
lines changed

1 file changed

+1
-1
lines changed

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/TargetInfo.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -95,7 +95,7 @@ static std::optional<NVVM::ReduxKind> matchReduxKind(triton::ReduceOp op,
9595
Operation *reduceOp = op.getSingleCombiner();
9696
if (!reduceOp)
9797
return std::nullopt;
98-
if (computeCapability >= 100 && reduceOp->getResultTypes()[0].isF32()) {
98+
if (computeCapability == 100 && reduceOp->getResultTypes()[0].isF32()) {
9999
if (isa<arith::MinimumFOp, arith::MaximumFOp>(reduceOp))
100100
useNanQualifier = true;
101101
if (isa<arith::MaxNumFOp, arith::MaximumFOp>(reduceOp))

0 commit comments

Comments
 (0)