[Mlir-commits] [mlir] [MLIR][Math][GPU] Add lowering of absi and fpowi to libdevice (PR #123422)

[Oleksandr Alex Zinenko]

================
@@ -16,37 +16,11 @@
 
 namespace mlir {
 
-/// Rewriting that replace SourceOp with a CallOp to `f32Func` or `f64Func` or
-/// `f32ApproxFunc` or `f16Func` depending on the element type and the
-/// fastMathFlag of that Op. The function declaration is added in case it was
-/// not added before.
-///
-/// If the input values are of bf16 type (or f16 type if f16Func is empty), the
-/// value is first casted to f32, the function called and then the result casted
-/// back.
-///
-/// Example with NVVM:
-///   %exp_f32 = math.exp %arg_f32 : f32
-///
-/// will be transformed into
-///   llvm.call @__nv_expf(%arg_f32) : (f32) -> f32
-///
-/// If the fastMathFlag attribute of SourceOp is `afn` or `fast`, this Op lowers
-/// to the approximate calculation function.
-///
-/// Also example with NVVM:
-///   %exp_f32 = math.exp %arg_f32 fastmath<afn> : f32
-///
-/// will be transformed into
-///   llvm.call @__nv_fast_expf(%arg_f32) : (f32) -> f32
-template <typename SourceOp>
-struct OpToFuncCallLowering : public ConvertOpToLLVMPattern<SourceOp> {
+template <typename SourceOp, typename DerivedTy>
----------------
ftynse wrote:

Please add some documentation. I also think the implicit convention for CRTP is to have the derived class the as the leading template argument.

https://github.com/llvm/llvm-project/pull/123422