auto l = mlir::dyn_cast<TypedValue<RankedTensorType>>(op.getX());
    auto r = mlir::dyn_cast<TypedValue<RankedTensorType>>(op.getY());
    if (!l || !r) return failure();

    auto element_type = getElementTypeOrSelf(l.getType());
    if (!element_type.isBF16()) return failure();

    auto out_type = op.getZ().getType();

    l = rewriter.create<ConvertOp>(op.getLoc(), l, rewriter.getF32Type());