Provide a constexpr size function about nsimd

[gquintin]

This is because of SVE basically where the size of SIMD vectors is not known at compile time but only at runtime. That's why none of the NSIMD constructs are constexpr. If you look at the headers you will see the use NSIMD_STRUCT so that when compiling for SVE this macro will be __sizeless_struct which is currently supported by Armclang and the Fujitsu compiler.

May I ask in which construct you need the size at compile time?

[eschnett]

I was defining a std::array where the size is rounded up to the next vector size, as in:

constexpr size_t vsize = pack<double>().size();
constexpr size_t N = (2*order+1 + vsize - 1) / vsize * vsize;
array<T,N> tmp;

I am currently using sizeof(pack<double>) / size(double) as a work-around.

[gquintin]

Ah ok, in this case maybe nsimd::max_len will be of use to you (from nsimd.h)

#if NSIMD_CXX >= 2014
template <NSIMD_CONCEPT_VALUE_TYPE T>
constexpr int max_len = max_len_t<T>::value;
#endif

The downside is that its value is huge (namely (2048 bits) / bits(T)) for two reasons:

• it is the maximum size of an SVE vector according to the SVE standard and
• we wanted a common value for all architectures in case, by some matters of facts, this value is (most probably indirectly) used as a parameter to code that is executed on different architectures (eg. serialization of data, computations on heterogeneous environments via MPI...)

In any case be careful with sizeof(pack<T>) as this code won't compile when you will target an SVE architecture. I don't know about Arm's plans but my guess is that you will find SVE everywhere in the near future including Apple's computers. (https://www.arm.com/blogs/blueprint/armv9)

[eschnett]

Thanks. I'm using the value only for a temporary array, so I don't have a need for an architecture-independent value.