Comment #12 on issue 3006 by [email protected]: Inaccurate sin/cos values
http://code.google.com/p/v8/issues/detail?id=3006
I respectfully disagree. As the person making the change, the burden is on
you to demonstrate that the max relative error over a reasonably large
range is not significantly worse than the existing implementation. It may
be accurate enough for you (and that's fine), but there are millions of
applications out there and you can't arbitrarily change their results
because you've reduced the accuracy of the trig functions.
I 'm not asking you to mimic one approx with another. I would like accurate
values for sin and cos.
cos(2^120) does not need to be 0.477, because, well the correct answer is
-0.92. I concede that this is a corner case that requires accurate pi
reduction. This is a quality of implementation issue. But there are
differences for values as small as 2200*pi. The value of
Math.tan(Math.PI/2) is incorrect. Since tan(x) is +infinity from the left
and -infinity from the right, I don't see returning +Infinity is correct
either.
I do not see the ECMAscript spec very vague. It explicitly gives some
leeway, but the intent in 15.8.2 is clear when it says that the same
library can be used that is available to the C programmer. It also
recommended using the same algorithms as given in Sun's fdlibm. A library
written by an expert in the field. I'm pretty sure this library says
cos(2^120) = -0.92.
I would love to have something twice as fast, but it needs to be accurate
too.
--
You received this message because this project is configured to send all
issue notifications to this address.
You may adjust your notification preferences at:
https://code.google.com/hosting/settings
--
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.
