Code: Difference between revisions

3,408 bytes added ,  27 November 2011
+ Fixed point toolbox
(+ Fixed point toolbox)
Line 85: Line 85:
*What does dicominfo do when a tag is not in its dictionary: skip it or give error? I was wondering about turning the tag into a variable name, something like Tag_3243_0010. (Matlab 6.5 (2002): Private__3243_0010)
*What does dicominfo do when a tag is not in its dictionary: skip it or give error? I was wondering about turning the tag into a variable name, something like Tag_3243_0010. (Matlab 6.5 (2002): Private__3243_0010)
*dicominfo: Items in sequences are not necessarily the same, so cannot be stored in arrays of structs. (Matlab 6.5 (2002): makes nested structs like dcm.RTDoseROISequence?.Item_1.DoseUnits?)
*dicominfo: Items in sequences are not necessarily the same, so cannot be stored in arrays of structs. (Matlab 6.5 (2002): makes nested structs like dcm.RTDoseROISequence?.Item_1.DoseUnits?)
== Fixed point toolbox ==
(initial announcement can be found [http://www.octave.org/octave-lists/archive/help-octave.2004/msg01274.html here], with the corresponding thread)
When implementing algorithms in hardware, it is common to reduce the accuracy of the representation of numbers to a smaller number of bits. This allows much lower complexity in the hardware, at the cost of accuracy and potential overflow problems. Such representations are known as fixed point.
OctaveForge now contains [http://octave.sourceforge.net/Fixedpoint/index.html a toolbox] to perform such fixed point calculations. This toolbox supplies a fixed point type that allows Octave to model the effects of such a reduction in accuracy of the representation of numbers. The major advantage of this toolbox is that with correctly written Octave scripts, the same code can be used to test both fixed and floating point representations of numbers.
What it does is create several new user types for fixed point scalar, complex scalars, matrices and complex matrices, and the corresponding operators on these types. As this code was first written against 2.1.50 there is no capabilities at this time for NDArray operations with this code, however I'm not sure this is a problem.
A typical use of the toolbox might be something like
n = 2;
a = rand (n, n);
b = rand (n, n);
## Create fixed-point version with 1 bit before decimal and 5 after.
af = fixed (1, 5, a);
bf = fixed (1, 5, b);
c = myfunc (a, b);
cf = myfunc (af, bf);
function y = myfunc (a, b)
    y = a * b;
endfunction
where as you can see the underlying function myfunc is unchanged, while it is called with either floating or fixed point types. The case above is for fixed-point values with 1-bit before the decimal point and 5 after, and for me gave a result of
octave:8> c
c =
  0.98105  0.94436
  0.82622  0.30831
octave:9> cf
cf =
  0.93750  0.90625
  0.78125  0.25000
which clearly shows the loss of precision of a fixed-point algorithm with only 6-bits of precision in a matrix multiply. One gotcha in this toolbox is the use of the concatenation operator "[ ]" which will implicitly reconvert fixed-point values back to floating-point, with Octave 2.1.57 or earlier. This is due to an internal limitation of octave that was removed in Octave 2.1.58. So for the best experience it is suggested you use this toolbox with octave 2.1.58 or later.
This package is only available with recent versions of OctaveForge (20040707 or later). With the package installed online help is available with the command
octave:1> fixedpoint info
As this package is relatively new, all feedback on its use would be most welcome.
Matlab also recently introduced a [http://www.mathworks.com/products/fixed Fixed Point Toolbox]. The Octave toolbox has been written independently of the Matlab toolbox and doesn't follow the same syntax. This might change in the future, if the author (DavidBateman) can be bothered to put the effort into making the required changes. Mathworks [http://www.mathworks.com/access/helpdesk/help/pdf_doc/fixedpoint/FPTUG.pdf documentation] for their fixed point toolbox also includes a generic discussion of fixed point numbers that might be a useful addition to the documentation supplied with Octave itself. Another short introduction to fixed point arithmetics is [http://home.earthlink.net/~yatescr/fp.pdf this].