Performance of the logarithm algorithm on bfloats



On 4/25/10 11:21 PM, Richard Fateman wrote:
> I wonder if you are noticing the time used to convert this number to
> base 10, and depending on what front end you are using,
> this could be done a smart way or not.
>
No, It's not the time to convert the number.  The original routine,
fplog, is still in the sources, and (fplog <test value>) takes a very
long time.  It's not surprising since the algorithm divides the test
value by %e until it's near one.  Since the test value is about
2b31455787102040073147561, it's no surprise this takes a long time.

Ray