[glob2-devel] Optimizing calculations

[Erik Søe Sørensen]

Kai Antweiler skrev:

> Looks like you use a small Taylor polynomials for cos and calculate it using the
> Horner scheme.  I don't think this can be made faster.
> Or could we use some discrete calculation for approximation?
>
> 2.
> Instead of:
>  float val, amp = 1.0, sum = 0.0;
>  for ( int i = 0;i < oct;i++, amp *= 0.5, tp *= 2.0 ) {
>    val = Noise3d( f_tp );
>    if ( hard ) val = fabs( val );
>    sum += amp * val;
>  }
>  return 0.5 + 0.5*( sum * ( ( float ) ( 1 << oct ) / ( float ) ( ( 1
> << ( oct + 1 ) ) - 1 ) ) );
>
> We could try:
>  float val, sum = 0.0;
>  for ( int i = 0;i < oct;i++, tp *= 2.0 ) {
>    val = Noise3d( f_tp );
>    if ( hard ) val = fabs( val );
>    sum += val / (float) (1<<i);
>  }
>  return 0.5 + 0.5*( sum * ( ( float ) ( 1 << oct ) / ( float ) ( ( 1
> << ( oct + 1 ) ) - 1 ) ) );
>
> But that is probably worse.
>  
It probably is - float->int and int->float conversions are expensive.
Keeping calculations in integers is usually way faster.

> Hey why isn't there a thing like float-shift?  After all this would
> only require to add/subtract 1 from the exponent!
>  
Well, there is such an instruction (FSCALE for x86); it doesn't seem 
like gcc uses it though... Oo

> 3.
> DynamicClouds.cpp:
> vpX += (viewPortX-vpX%64+96)%64-32;
> This might be faster using a mask:
> vpX += (viewPortX-vpX&63+96)&63-32;
>  
The compiler ought to figure this out for itself - at least for unsigned 
values (for signed it makes two shifts instead).

/Erik