diff --git a/src/nouveau/compiler/nak/ir.rs b/src/nouveau/compiler/nak/ir.rs
index f8e096349c7..06bfd2d6d9d 100644
--- a/src/nouveau/compiler/nak/ir.rs
+++ b/src/nouveau/compiler/nak/ir.rs
@@ -3977,7 +3977,9 @@ pub struct OpF2F {
     pub dst_type: FloatType,
     pub rnd_mode: FRndMode,
     pub ftz: bool,
-    /// Place the result into the upper 16 bits of the destination register
+    /// For 16-bit up-conversions, take the high 16 bits of the source register.
+    /// For 16-bit down-conversions, place the result into the upper 16 bits of
+    /// the destination register
     pub high: bool,
     /// Round to the nearest integer rather than nearest float
     ///
diff --git a/src/nouveau/compiler/nak/sm50.rs b/src/nouveau/compiler/nak/sm50.rs
index 99097729e80..74a59d72755 100644
--- a/src/nouveau/compiler/nak/sm50.rs
+++ b/src/nouveau/compiler/nak/sm50.rs
@@ -1698,6 +1698,7 @@ impl SM50Op for OpF2F {
         e.set_field(10..12, (self.src_type.bits() / 8).ilog2());
 
         e.set_rnd_mode(39..41, self.rnd_mode);
+        e.set_bit(41, self.high);
         e.set_bit(42, self.integer_rnd);
         e.set_bit(44, self.ftz);
         e.set_bit(50, false); // saturate