[llvm] [NVPTX] Add support for f32x2 mixed-precision add/sub (PR #221957)

[Srinivasa Ravi via llvm-commits]

================
@@ -2508,6 +2537,24 @@ foreach rnd = FPRoundingModes in {
              (f32 (fpextend type:$a)),
              f32:$b, rnd_imm))]>,
         Requires<[SM100]>;
+
+  let Predicates = [hasRubinFamilySupport] in {
+    foreach t = [F16X2RT, BF16X2RT] in {
+      def INT_NVVM_MIXED_ADD_ # rnd # _f32x2_ # t.PtxType :
+        BasicNVPTXInst<(outs B64:$dst), (ins B32:$a, B64:$b),
+          "add." # rnd # ".f32x2." # t.PtxType # ".f32x2",
+          [(set v2f32:$dst,
+            (int_nvvm_fadd (fpextend_v2f32_from_packed t.Ty:$a), v2f32:$b,
+                          rnd_imm))]>;
+
+      def : Pat<(v2f32 (int_nvvm_fadd
+                        (fpextend_v2f32_from_lanes t.Ty.ElementType:$a0,
+                                                    t.Ty.ElementType:$a1),
+                        v2f32:$b, rnd_imm)),
+                (!cast<Instruction>("INT_NVVM_MIXED_ADD_" # rnd # "_f32x2_" #
----------------
Wolfram70 wrote:

Makes sense, done!

https://github.com/llvm/llvm-project/pull/221957