Harden linear_to_gamma() function?

[hollasch]

We currently have the following implementation:

inline double linear_to_gamma(double linear_component)
{
    return sqrt(linear_component);
}

This will barf on negative values, returning NaN. Should we harden this like so?

inline double linear_to_gamma(double linear_component)
{
    if (linear_component < 0)
        return 0;

    return sqrt(linear_component);
}

or

inline double linear_to_gamma(double linear_component)
{
    return sqrt(fmax(0, linear_component));
}