[Liboil] complex types?

[Stephane Fillod]

On Sun, Nov 13, 2005 at 04:32:51PM -0800, David Schleef wrote:
> On Mon, Nov 14, 2005 at 01:07:52AM +0100, Stephane Fillod wrote:
> > Some new liboil functions I'd like to introduce need (math) complex types.
> > Those complex types may be based on 32 bit floats, double, 16 bit signed
> > short ints, 8 bit unsigned ints, etc. So we need to introduce new
> > type names in function names. Should it be _cf32, _cs16, _cu8, etc. ?
> > Dave, what do you advise, how the new types should be added to and handled
> > by the automatical function testing (ie. wrt OIL_DEFINE_CLASS) ?
> > 
> > What should be the function argument type? In the case of _cf32, should
> > it be "float *", or "struct { float data[2]; } *" ?
> > Rem: I prefer the first one.
> 
> Eu, gross.
> 
> Is there any reason to have complex integer types?  I can understand
> the need for complex float/double.

Some applications do complex fixed-point (integer) computations.
This is a lot faster on FPU-less systems.

> Do you mean 'float *' as in having two pointers, one for real and
> one for imaginary parts?  Or 'float *' as in 'float *d_2xn' that
> has n interleaved real and imag float values?  It's not clear which
> would be better from a speed standpoint.

I mean 'float *' as in 'float *d_2xn' that has n interleaved real and
imag float values. Speed can be one reason, existing data model of 
applications and/or hardware devices is another one.

> I'm inclined to avoid the struct idea.

Me too.

> Also, would it be possible to define functions without inventing
> new types?  I.e., oil_sincos_f64() could be thought of as a complex
> function.  I think I would prefer to explore that route, since
> types are exceedingly difficult to remove from the rolling ABI concept
> (if even possible).

oil_sincos_f64()-like prototype does not fit my need, because "complex" 
args are not interleaved. BTW, I don't understand why offset and interval
are passed by address and not by value in oil_sincos_f64.

So, shall we take the 'float *d_2xn' route?
NB: some args may be complex, some others may be real only.
As an example, what would be the name of a sincos outputing interleaved
complex doubles?  oil_sincos_cif64, or better name? sincos is not a good 
example, but we need to distinguish between complex or not, interleaved
or not, for functions like liboil_*mult_ because the maths are different.
What the OIL_DEFINE_CLASS should look like?

-- 
Stephane