invariance across calls to geometric functions
Creating an issue based on discussion here: https://github.com/KhronosGroup/OpenCL-Docs/issues/1024#issuecomment-1841263676
In the previous discussion we noted that the rounding mode used by a given geometric function is irrelevant as long as the function meets the stated accuracy requirement for the function. However, we also noted that we currently do not have any invariance requirements for geometric functions (or any other functions, really), so it's unclear whether it is allowable to implement a function with one rounding mode one place in the program code, and with a different rounding mode (or any other floating-point configuration?) in a different place in the program code.
Should we have any invariance requirements for geometric functions, or for any other functions for that matter?