insert dtype in std::map

Question

I want to do a map that takes a pair of pybind11::dtype and int and maps it into an OpenCV format:

static std::map<std::pair<pybind11::dtype, int>, int> ocv_types;

So I inserted all combinations but there seems to be a problem when adding int32_t and float_t:

    ocv_types.insert(std::make_pair(std::make_pair(pybind11::dtype::of<std::int32_t>() , 3), CV_32SC3));

    ocv_types.insert(std::make_pair(std::make_pair(pybind11::dtype::of<std::float_t>() , 3), CV_32FC3));

When I do this, only the CV_32SC3 is really inserted , my guess that somewhere the program "thinks" that both elements are equal and therefore is not going to insert the second one.

How can I actually add these 2?

P.S. I did this check just to "prove" that the types are not equal:

    if(pybind11::dtype::of<std::int32_t>() == pybind11::dtype::of<std::float_t>())
    {
        std::cout << "std::int32_t == std::float_t" << std::endl;
    }
    else
    {
        std::cout << "std::int32_t != std::float_t" << std::endl;
    }

... And of course they are not.

EDIT

I added the < function for dtype and used it in the compare function for the map, but not all elements are present in the map:

int getVal(pybind11::dtype type)
{
    if(type.is(pybind11::dtype::of<std::uint8_t>()))
        return 1;
    if(type.is(pybind11::dtype::of<std::uint16_t>()))
        return 2;
    if(type.is(pybind11::dtype::of<std::int16_t>()))
        return 3;
    if(type.is(pybind11::dtype::of<std::int32_t>()))
        return 4;
    if(type.is(pybind11::dtype::of<std::float_t>()))
        return 5;
    if(type.is(pybind11::dtype::of<std::double_t>()))
        return 6;
}

inline bool operator <(const pybind11::dtype a, const pybind11::dtype b) //friend claim has to be here
{
    return getVal(a) < getVal(b);
}

auto comp = [](const std::pair<pybind11::dtype, int> a, const std::pair<pybind11::dtype, int> b)
{
    return a < b;
};
static std::map<std::pair<pybind11::dtype, int>, int, decltype(comp)> ocv_types(comp);

Answer 1

As you noted pybind11::dtype do not have any particular order. So IMO best approach is to use std::unordered_map and provide respective hashes. pybind11 already has some hash function, so it is needed to adopt it for std::hash.

Here is test I've wrote (using Catch2) and it passes on my machine:

main.cpp:

#include "catch2/catch_all.hpp"
#include <pybind11/embed.h>
#include <pybind11/numpy.h>
#include <unordered_map>

template<>
struct std::hash<pybind11::dtype>
{
    size_t operator()(const pybind11::dtype &t) const
    {
        return pybind11::hash(t);
    }
};

template<>
struct std::hash<std::pair<pybind11::dtype, int>>
{
    size_t operator()(const std::pair<pybind11::dtype, int> &t) const
    {
        return std::hash<pybind11::dtype>{}(t.first) ^ static_cast<size_t>(t.second);
    }
};


TEST_CASE("map_with_dtype") {
    constexpr auto CV_32SC3 = 1;
    constexpr auto CV_32FC3 = 2;

    pybind11::scoped_interpreter guard{};

    std::unordered_map<std::pair<pybind11::dtype, int>, int> ocv_types;
    REQUIRE(ocv_types.empty());

    auto a = ocv_types.insert(std::make_pair(std::make_pair(pybind11::dtype::of<std::int32_t>() , 3), CV_32SC3));
    REQUIRE(a.second);

    auto b = ocv_types.insert(std::make_pair(std::make_pair(pybind11::dtype::of<std::float_t>() , 3), CV_32FC3));
    REQUIRE(b.second);
    CHECK(b.first->second == CV_32FC3);

    CHECK(ocv_types.size() == 2);
}

CMakeLists.txt:

cmake_minimum_required(VERSION 3.16)

# set the project name
project(MapOfPyBind11)

find_package(Catch2 REQUIRED)
find_package(pybind11 REQUIRED)

# add the executable
add_executable(MapOfPyBind11Test main.cpp)
target_link_libraries(MapOfPyBind11Test PRIVATE Catch2::Catch2 pybind11::module pybind11::embed)

include(CTest)
include(Catch)
catch_discover_tests(MapOfPyBind11Test)