a=randn(5*10^7)
b=randn(5*10^7)
f(a::Float64,b::Float64)::Float64 = sin(a) + cos(b)
@time @tturbo @. sin(a) + cos(b) # .05 seconds
@time @tturbo @. f(a,b) # 1 second
Is it possible to get the @tturbo speedup using the function f() version?
@tturbo, as a macro, doesn’t see the definition of the method of f, which is only available after the macro is expanded. So I’d say the answer is no.
You may want to have a look at:
Try f(a, b) = sin(a) + cos(b) instead.
@tturbo does not support being restricted Float64 inputs/outputs, as it transforms the loop to use SIMD types. If the f doesn’t work with these SIMD types, it’ll execute the original code without applying transforms.
Feel free to make a PR adding this to the README or docs.
@tturbo used to error out with a MethodError, which was more helpful in diagnosing this than silently falling back to scalar code.
It’s a tradeoff between being informative and explicit, vs making it easier to write generic code.
It does see f, but the only thing it sees is that it isn’t valid to call it with SIMD types.
Otherwise, if it were callable with SIMD types, it’d treat f as an expensive black box function, which is a reasonable way to treat it if f is calling something like sin and/or cos. It wouldn’t be reasonable if f were just cheap functions like +.