[llvm] [NVPTX] Add float to tf32 conversion intrinsic (PR #121507)

[Artem Belevich via llvm-commits]

================
@@ -1466,6 +1466,15 @@ let TargetPrefix = "nvvm" in {
   def int_nvvm_e5m2x2_to_f16x2_rn_relu : ClangBuiltin<"__nvvm_e5m2x2_to_f16x2_rn_relu">,
       Intrinsic<[llvm_v2f16_ty], [llvm_i16_ty], [IntrNoMem, IntrNoCallback]>;
 
+// Convert Float to TF32
+def int_nvvm_cvt_float_to_tf32 : Intrinsic<[llvm_i32_ty],
+    [llvm_float_ty, // Input float
+     llvm_i8_ty,    // Flag for Rounding Modes
----------------
Artem-B wrote:

Naming, the hardest problem in computer science...

My usual position is that I prefer to avoid parameters as instruction selectors, unless it is absolutely necessary, and it does not seem to be the case here.
The main reason is that there's no good way to handle invalid selector values. If someone gets intrinsic name wrong, they do get a sensible diagnostic that the name is wong. With the wrong values, we'll either crash somewhere in LLVM because we can't lower the intrinsic, or will produce something invalid.
In many cases, the set of allowed variants is not a complete cartesian product of all parameters. There are usually invalid combinations. Exact enumeration goes a long way to avoid accidental use of those. It also makes it very obvious which combinations are supported/valid.

In other words, I would prefer to keep things simple, add a handful of new conversion intrinsics mapping 1:1 to the instructions, and be done with that.


https://github.com/llvm/llvm-project/pull/121507