Re: wut do they mean by "fixed point" saying "
Hi Vahid,
Fixed Point arithmetic or representation basically says that you are limiting the representation of fractional part to a specific (fixed) number of bits. And hence the precision is fixed to a specific value and the decimal point are fixed to specific number of digits.
Now lets take an example say the maximum value of the data you have is 11.3567812 ( I consider the absolute value so the range is bounded by -11.**** to +11.******). Now your processor has a data width for storage of integer data thats say 32 bits. How many bits are required for your integer part in this case 11 to be represented in binary 4 bits right. Now the fixed point representation of 11.3567812 in a 32 bit processor, will be in Q27 format which will give you least error in fixed point computation.
Now let me define Q27, here you have used 1 for the sign bit and 4 bits for integer-part so numbers of bits left for fraction is 32-4-1 = 27 hence Q27 format.
Now you must note that the maximum representation possible in this Q27 notation is -16.0 to 15.999999992549419403076171875
NOTE: Fixed point representation will have an error of 1/2^(no.of bit in fraction +1) as error. So the precission is limited by the number of bit allocated to the fraction. You can as well restrict your self to storing data in just 16 bits but you will be increasing the error.
Now coming the the way "Fast Fixed point ICA Algorithm......" or what ever might be the case that you are referring to, you will definitely not see the author talking in these terms, because the author already assume that the float to fixed conversion is done and that the results are within tolerable error criteria. Now comes the problem of optimizing the fixed point code. Here author might talk about making the algorithm fast w.r.t a processor or making some fundamental computation changes in the way the fixed point code is executed so as that the code runs faster.
I hope I'm clear without creating confusion.
~Kalyan.