`glm::max()` returns incorrect result in CUDA `__global__` or `__device__` functions

[WindVChen]

Hello,

Thank you for this great library.

I recently encountered unexpected behavior when using glm::max() within a CUDA __global__ or __device__ function in the latest source version of GLM.

Here’s a minimal example illustrating the issue:

__global__ void function2_CUDA(...)
{
    glm::vec2 h_0_ = glm::vec2(1e-4, 1e-4);
    glm::vec2 h_0_1 = glm::vec2(1e-5, 1e3);
    glm::vec2 h_0_2 = max(h_0_, h_0_1);
    
    printf("h_0_:   (%f, %f)\n", h_0_.x, h_0_.y);
    printf("h_0_1:  (%f, %f)\n", h_0_1.x, h_0_1.y);
    printf("h_0_2:  (%f, %f)\n", h_0_2.x, h_0_2.y);
}

void function1(...)
{
    function2_CUDA<<<256, 256>>>(...);
}

Expected output for h_0_2:

(1e-4, 1e3)

Actual output observed:

(1e-5, 1e3)

I verified that GLM version 1.0.1 produces the correct result. However, the latest source from the repository gives the incorrect output.

This leads me to suspect a compatibility issue in the current implementation of max() when used in cuda code. Alternatively, am I possibly missing something in its usage with CUDA?

Any insights or clarification would be appreciated!