scalar implementation without code duplication

[actual-daniel]

I'm currently a bit confused about the purpose of xsimd::generic. My expectation was that its possible to use this architecture as a fallback or for development purposes. For this I would assume that xsimd::batch<T, xsimd::generic> just contains a single scalar value. But this is not the case - Instead the build fails because xsimd::has_simd_register is false.

I could specialize load_* and broadcast calls and try to separate it from the rest of the implementation which then hopefully works with scalar and simd types. But it feels like this could be abstracted through implementation of the generic architecture.

Is there a reason why this is not done? Are there examples how to write simd and scalar implementations without duplicating code?

[serge-sans-paille]

yeah, generic means "generic vector architecture". It would work with an hypothetical vector of one register, but that architecture fallback doesn't exist.

We do provide overloads for scalar values though, i.e. xsimd::cos fallsback to std::cos where it makes sense.

[actual-daniel]

@serge-sans-paille Thanks for answering! Is there a reason why it is not implemented? I guess the implementation would not be that complicated - Is there a chance that this functionality is merged if contributed?

If merging upstream is not feasible I guess the best way to resolve this is to maintain a custom "scalar" backend which is derived from generic?

[serge-sans-paille]

Instead of the "scalar" backend, which would conflict in essence with the scalar overload we provide for every operation, we could have an emulated<n> backed that would emulate a batch of n elements using scalar operations, relying on the generic implementation etc.

Is that what you have in mind? I'd happily implement that.

[actual-daniel]

@serge-sans-paille Yes this is exactly what I was looking for! Using a generic n as you suggested would actually be even better than what my initial request was.