Forum: Programming
2020-01-19, 19:47
|
Replies: 43
Views: 25,062
|
Forum: Programming
2020-01-19, 02:49
|
Replies: 43
Views: 25,062
I forgot to add, C++ defines...
I forgot to add, C++ defines reinterpret_cast<pointer> (www.cplusplus.com/doc/tutorial/typecasting/) as part of the language standard, thus in a sense enshrining the necessary-ugliness of...
|
Forum: Programming
2020-01-18, 23:26
|
Replies: 43
Views: 25,062
@Paul: I consider type-punning aliasing a...
@Paul: I consider type-punning aliasing a necessary evil, e.g. my quad-float code relies on it. There's also the "when you need to be 100% sure that double-const you are using is translated...
|
Forum: Programming
2020-01-17, 19:24
|
Replies: 43
Views: 25,062
Bzipped code attached. Note in the meanwhile I...
Bzipped code attached. Note in the meanwhile I got a PM form DrSardonicus about mapping a target interval [0,pi/2] to [-pi/4,+pi/4]. The code George posted in #14 of the thread was for sin and cos,...
|
Forum: Programming
2020-01-16, 21:42
|
Replies: 43
Views: 25,062
I've been playing with the use of a...
I've been playing with the use of a Chebyshev-polynomial basis expansion for approximating e.g. sin and cos - this gives Pade-similar accuracy but is much more computationally convenient. Longer...
|
Forum: Programming
2020-01-12, 19:36
|
Replies: 43
Views: 25,062
One could, say, loop over a bunch of...
One could, say, loop over a bunch of evenly-spaced points in the interval of interest, use the various formulae to compute sincos in DP, compute errors using expensive extended-precision sincos as a...
|
Forum: Programming
2020-01-11, 19:36
|
Replies: 43
Views: 25,062
That is more suitable for non-FMA architectures...
That is more suitable for non-FMA architectures ... for FMA-supporting ones, George's code needs 8 FMA (2 of which are pure-MUL), compared to 10 separate operations (mix of ADD, MUL, and FMA) for the...
|
Forum: Programming
2020-01-10, 22:44
|
Replies: 43
Views: 25,062
George, presumably you need this functionality...
George, presumably you need this functionality for on-the-fly computation of FFT twiddles, i.e. for complex roots of unity of form exp(I*2*Pi/n), where n is some highly composite integer, likely of...
|