Mapping a template type to a ScalarType

C++
1

I’m looking for a function - for a constexpr - that’ll map a type T to a ScalarType. I imagine it’d be called at::ScalarType<T>() or the like, but grepping through the PyTorch codebase I can’t find anything promising.

Concretely, I want to write something like

template <typename T> 
at::Tensor empty() {
    return at::empty({1}, at::dtype(at::ScalarType<T>());
}

so that a call like empty<float>() gives me back a at::kFloat tensor.

I understand this is not usually the way the conversion happens, and that control flow should generally be generic -> specific. I know that I could write this as empty(at::ScalarType) and then use the DISPATCH macros to do the templating. For the particular chunk of code I’m working on though, it’d be useful and it feels like it should exist, I just can’t find it.

(More generally, if anyone has advice on searching for a specific signature in a large codebase, that’d be great too. Best I can find is CppDepend, which is $$$)

2

I tripped over caffe2::TypeMeta::Make<T>() in the PyTorch source, so I’ve ended up using

template<typename T>
at::ScalarType dtype() { return at::typeMetaToScalarType(caffe2::TypeMeta::Make<T>()); }
1 Like
3

found another way: torch::CppTypeToScalarType<long>()

and there is ScalarTypeToCPPType in the same src file: c10/core/ScalarType.h

2 Likes