CODEWITHSUNDEEP
c#signal-processingoctave
Dmitry Kh
Dmitry Kh

Reputation: 136

Octave precision for floats: Is it 128 bit under the hood?

I'm making a C# version of the Burg algorithm using an approach presented by Koen Vos in "A Fast Implementation of Burg’s Method".

I used the GNU Octave arburg function to compare the results with. The test results are almost the same when I use decimal for internal variables in my C# code (accuracy is 0.0000000000001) but quite different when I use double (accuracy is 0.01).

As I know GNU Octave uses 64-bit precision for floats, not 128-bit. Am I wrong?

/* Coefficients for comparision are taken from GNU Octave arburg() 
* t = [0:2000];
* x = sin( 2 * pi() * t / (512 / 5.2));
* output_precision(16)
* [a, v, k] = arburg(x(1:512), 4)
*/

The C# code is more than 300 lines so I think it's better to not put it here.

I think either GNU Octave uses 128-bit precision under the hood or I have a mistake in my C# code and increasing the precision of calculations mitigates this mistake somehow.

The question is it possible that float data in GNU Octave (or Matlab) are 128-bit inside?

Upvotes: 3

Views: 3159

Answers (1)

Cris Luengo
Cris Luengo

Reputation: 60660

Octave uses 64-bit floats by default, there is no way to force it to use a higher precision. It knows only double (64-bit floats) and single (32-bit floats).

Intel (and compatible) processors can do computation with 80-bit floats (in C this is long double), but they don't support 128-bit floats. Some software might emulate 128-bit float for improved precision, but Octave is not one of them (nor is MATLAB).

Upvotes: 5

Related Questions