You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Add direct support for the LLVM `maximumnum` and `minimumnum`
intrinsics for NVPTX, rather than lowering them to a sequence of
compare + select instructions.
The `maximumnum` and `minimumnum` intrinsics map directly to PTX
`max`/`min` instructions.
In future, the LLVM `maxnum`/`minnum` intrinsics might need some fix-ups
for sNaN handling added, but currently both `llvm.maxnum` and
`llvm.maximumnum` will map directly to PTX `max` instructions.
0 commit comments