Concepts for numeric_types

[osayamenja]

Hello,

In a nutshell, is there a more concise way to define a C++ concept for all cutlass numeric types?

I am currently writing some C++ code using cute numeric types and I would like to define a concise concept that captures all its numeric types.

For example, for some TensorType t using std::same_as<t, cute::half_t> || std::same_as<t, cute::float_e4m3_t> || ... || std::same_as<t, cute::complex<half_t>> || ... should be fine.

However, is there a fundamental type, that would perhaps allow for one shot check with e.g. std::common_with or std::derived_from?

I ask this question because in my code, I am invoking something along the lines of t_0 = TensorType(0.0), which I know all cute numeric types support. I am also very open to suggestions that may not involve using concepts at all.

Thanks.

[osayamenja]

Forgot to mention that is_integral and is_floating_point fails a static assertion for integer_subbyte and cute float structs, respectively; so, this would not apply, except I am doing something completely wrong here :)

Also, I am using concepts via libcu++

[thakkarV]

Unfortunately, we do not have plans to move to C++20, so we can't really start using concepts anytime soon. That said, I would expect is_integral to just work for integer subbyte too, and if they do not its likely an oversight that we would appreciate an MR for :)

[osayamenja]

Understood. OTOH, libcu++ backports concepts to C++ 14; but with limited functionality as the concepts or requires keywords are unavailable.

I decided on using the simple concept below to capture my use case as it applies both to cute numeric and, more generally, to language native types.

template<typename V>
    concept TensorValueType = requires(V v){
        {V(0)} -> std::same_as<V>;
    };

Regarding the MR, going off this reference, I think making this work would require redefining the alias types for integer_subbyte, that is these would become, for example, using uint1b_t = decltype(integer_subbyte<1, false>.Storage)

What do you think? @thakkarV

[thakkarV]

I do not think it would require new aliases -- instead you can just define new specializations for is_integral for the subbyte ints

struct cute::is_integral<cutlass::uint1b_t> : true_type {};

[osayamenja]

Side-note, how do we get the underlying value type of a tensor? I use below but is there a simpler way?

// For cute::Tensor input
using TensorValueType = typename cute::iterator_traits<decltype(input.data())>::value_type;

[thakkarV]

https://github.com/NVIDIA/cutlass/blob/main/include/cute/tensor.hpp#L110

should be what you want

[osayamenja]

For the above snippet, would this suggest replacing input.data() with input.engine().begin()? If not, could you kindly give an example?

[thakkarV]

simpler still

using ValueType = typename decltype(tensor)::value_type;

[osayamenja]

thanks!! @thakkarV