It is better explained in the RP2040 datasheet, see "2.8.3.2.1. Implementation Details" in the section "2.8. Bootrom", one very pertinent part I reproduce below:
Anyway the function results should be usable for most sane applications, but for these functions it would frequently not be possible to reproduce the numerical results exactly on a normal PC.The scientific functions are calculated using internal fixed-point representations so accuracy (as measured in ULP error rather than in absolute terms) is poorer in situations where converting the result back to floating point entails a large normalising shift. This occurs, for example, when calculating the sine of a value near a multiple of pi, the cosine of a value near an odd multiple of pi/2, or the logarithm of a value near 1.
Statistics: Posted by katak255 — Mon Jul 22, 2024 12:44 pm