[go-nuts] Happiness is a compiled language

I created a Go implementation of the Parks–McClellan filter design
algorithm. C and Ruby implementations were compared. There's a Fortran
example too but I'm not going to try compiling code from 1972.

The source is here:
https://github.com/AE9RB/firpm

Times from a 3.0GHz Core 2 Duo:

          C: 2.07ms
         Go: 3.03ms
Ruby 2.0.0: 85.50ms
Ruby 1.9.2: 190.02ms


If you check that the computed results match you will find that they do not. Sin and Cos in Go uses a different algorithm than clang (and probably gcc). See for yourself:


package main

// #include <math.h>

import "C"

import "fmt"

import "math"

func main() {

  fmt.Println(C.cos(math.Pi / 2))

  fmt.Println(math.Cos(math.Pi / 2))

}


In practice, the firpm results are accurate enough that the filters will perform identically. However, writing tests against a C reference implementation is difficult. It's possible there's a bug introducing inaccuracy beyond Sin and Cos differences. I suppose I could swap in a new math library that made C calls for testing. But I won't yet because I just wanted a rough idea of Go performance for the kinds of problems I was hoping to use it for.


Looking at the Go source I see a lot of i386 assembly for math which uses FPU hardware but the amd64 uses Go code. I'm probably going to write a float32 math library so I'll have to learn more about why this is.


-david from OreGOn