[Mlir-commits] [mlir] [mlir][nvgpu] add `nvgpu.rcp` op (PR #100965)

[Guray Ozen]

================
@@ -802,4 +814,24 @@ def NVGPU_WarpgroupMmaInitAccumulatorOp : NVGPU_Op<"warpgroup.mma.init.accumulat
   let hasVerifier = 1;
 }
 
+def NVGPU_RcpOp : NVGPU_Op<"rcp", [Pure,
+                                   SameOperandsAndResultShape,
+                                   PredOpTrait<"in and out have same element type",
+                                            TCresVTEtIsSameAsOp<0, 0>>]> {
+  let summary = "F32 rcp calculation for vector types using nvvm.rcp.* OP";
+  let description = [{
+    F32 vector reciprocal calculation using `nvvm.rcp.*`. Currently only
+    `nvvm.approx.ftz.f` is supported.
+    The input and output are both F32 vector with same shape.
+  }];
----------------
grypp wrote:

```suggestion
  let description = [{
    Reciprocal calculation for `vector` types using `nvvm.rcp` OPs.

    Currently, only the `approx` rounding mode and `ftz` are supported, and only for the `f32` type.

    The input and output must be of the same vector type and shape.
  }];
```

https://github.com/llvm/llvm-project/pull/100965