Support for `sincos` ?

[ax3l]

Hi,

I know it is not standard in C/C++, but popular stdlibs (e.g. gnu, SYCL, CUDA, etc.) provide an efficient implementation to calculate the sin and cos for an argument at the same time: sincos(arg) and sincospi(arg).

Would it be sensible to have support for sincos(arg) math in vir-simd, too? I checked the code and I think when it refers to sincos it means sin and cos functions separately, did I understand that right?

At the same time, I read reports like this and wonder what you think about vectorized versions of sincos?

[mattkretz]

I did implement sincos for Vc (loong time ago): https://github.com/VcDevel/Vc/blob/8f862ab10d5ae99cc156ec625893fecb0870bf05/src/trigonometric.cpp#L335-L373

The way I implemented these functions, yes, sincos was able to reuse a lot of work of the individual calls and thus be more efficient. Adding this in vir-simd and calling into internals of libstdc++'s sin/cos implementations is possible (though fragile and unsupported, of course; as long as I'm the maintainer of libstdc++'s implementation and vir-simd that's not too big of a problem).

The libstdc++ implementation is derived from the Vc implementation, so it should be fairly easy.