Unexpected compMin behavior.

[MaxSac]

I tried to calculate the compMin from a vector which includes a NaN.
The result seems to be dependent on the position of the NaN in the vector.
It may be a different behavior than the c++ standard suggests [IEEE floating-point arithmetic (IEC 60559)] (and my naive intuition 😆 ).

#include <glm/glm.hpp>
#include <glm/gtx/component_wise.hpp>
#include <iostream>
#include <limits>

int main(int argc, char* argv[])
{
    // Option I.
    std::cout << glm::compMin(glm::vec2(std::numeric_limits<float>::quiet_NaN(), 1.f)) << std::endl;
    // >>> nan
    std::cout << glm::compMin(glm::vec2(1.f, std::numeric_limits<float>::quiet_NaN())) << std::endl;
    // >>> 1

     // Option II.
    std::cout << std::fmin(std::numeric_limits<float>::quiet_NaN(), 1.f) << std::endl;
    // >>> 1
    std::cout << std::fmin(1.f, std::numeric_limits<float>::quiet_NaN()) << std::endl;
    // >>> 1

    return 0;
}

Is this kind of functionality intentional?

[gottfriedleibniz]

GLSL defines min as:

Returns y if y < x; otherwise it returns x

which should match std::min's behaviour (not std::fmin). Unfortunately, the inconsistency in special-value behaviour between IEEE-operators vs ternary-logic will emerge elsewhere too.

[christophe-lunarg]

Erm, the comment, the example don't match. Without clarification, I will just close this issue. Maybe that's a feature request for a fcompMin function. unclear.

[MaxSac]

Sorry for the incorrect minimal working example. I wrote it yesterday in a hurry and have now corrected it.

To give a little bit of background.
I tried to calculate the interaction distance from a ray to a box, as sketched in the picture.

I wrote the following code and was quite surprised that a) and b) gave different results.

#include <array>
#include <glm/glm.hpp>
#include <glm/gtx/component_wise.hpp>
#include <iostream>
#include <limits>

std::array<float, 2> intersection(glm::vec2 corner1, glm::vec2 corner2,
    glm::vec2 origin, glm::vec2 invDirection)
{
    auto tMin = (corner1 - origin) * invDirection;
    auto tMax = (corner2 - origin) * invDirection;

    auto tEnter = glm::min(tMin, tMax);
    auto tExit = glm::max(tMin, tMax);

    auto enterMax = glm::compMax(tEnter); 
    auto exitMin = glm::compMin(tExit);  

    return std::array<float, 2> { enterMax, exitMin };
}

int main(int argc, char* argv[])
{
    auto c1 = glm::vec2(1, 1);
    auto c2 = glm::vec2(2, 2);

    // a)
    auto origin1 = glm::vec2(1, 0);
    auto invDirection1 = glm::vec2(std::numeric_limits<float>::quiet_NaN(), 1.f);
    auto intersection_points1 = intersection(c1, c2, origin1, invDirection1);
    std::cout << "entry: " << intersection_points1[0] << ", exit: " << intersection_points1[1] << std::endl;
    // >>> entry: nan, exit: nan


    // b)
    auto origin2 = glm::vec2(0, 1);
    auto invDirection2 = glm::vec2(1.f, std::numeric_limits<float>::quiet_NaN());
    auto intersection_points2 = intersection(c1, c2, origin2, invDirection2);
    std::cout << "entry: " << intersection_points2[0] << ", exit: " << intersection_points2[1] << std::endl;
    // >>> entry: 1, exit: 2

    return 0;
}

Technically I can understand what happens, but it was against my naive expectation.
I guess using something like fcompMax would help in this case and would be a nice feature.