ENH: einsum_sumprod.c.src to cpp in numpy.multiarray see #29528 #30210
Conversation
523e2df to
bec2370
Compare
|
@inakleinbottle quick fyi ping to keep you apprised of my current status |
| * and SIMDF64 */ | ||
| template <> | ||
| struct SumSIMD<npy_float, npy_float> { | ||
| using SimdType = NpySIMDF32; |
There was a problem hiding this comment.
I think this, and a great deal of the other supporting structure will be useful elsewhere and should probably be moved to a header file that can just be included in all these replacement files as they are added. It might be worth having a discussion with the optimization team too for the SIMD stuff to avoid replicating work.
| /* Template where (npy_double, npy_double) will allow the SIMD | ||
| * capable version.*/ | ||
| template <typename T, typename Temptype, | ||
| typename std::enable_if<std::is_same<T, Temptype>::value && |
There was a problem hiding this comment.
The enable_if construct is a little odd here. First, since we have access to C++17, you might want to use the _t and _v variants which make this a bit more readable. However, here I think simple overloading would serve the same purpose. C++ always prefers concrete instantiations over templates if it is the best match, and enable_if constructions have a large cost for instantiation whereas overloading has a much smaller cost. I think though, that what is needed here is actually a partially specializable struct template that can be used as a customization point inside the sum_of_arr function. For instance:
template <typename T, typename TempType, typename SFINAE=void>
struct SumOfArr {
static TempType eval(T* daa, npy_intp size) noexcept(?) {}
};
template <typename T>
static inline TempTypeOf<T> sum_of_arr(T* data, npy_intp count) noexcept (?) {
using Helper = SumOfArray<T, TempTypeOf<T>>;
return Helper::eval(data, count);
}This is just an idea, but it would give you relatively low-overhead control the implementation based on the type and TempType (which I've shorthanded here as a trait, but could be left as a template argument if multiple choices are necessary).
|
Thanks @athurdekoos for tagging me in. I had a quick look over the code and made a couple of what I hope are helpful comments, or at least things to think about. To be clear, I don't think these are necessary changes, so this isn't a review of any kind. Both comments are "looking ahead" to the other c.src modules too, so establishing some standard patterns and reusable components might be a good idea. Very happy discuss or explain further. |
This PR is a continuation of #29528.
Description
This change ports einsum_sumprod from the generated einsum_sumprod.c.src file to a C++ template source file (einsum_sumprod.cpp). The goal is to improve readability and maintainability while preserving the existing behavior and performance characteristics.
There are no intended changes to the public API.
Summary of changes:
einsum_sumprod.c.srcwitheinsum_sumprod.cpp.Notes:
Considerations not implemented:
Benchmarks
Please let me know if you'd like me to adjust naming or structure.