# Downsampling a 23 bit number to 8 bits

I'm trying to build a system that recieves coefficients and performs a Fourier approximation, I need to write the system in VHDL so I'm using tables to simulate sin functions, they output an 8 bit number which is the sin of the input, then I need to multiply each sin with it's given coefficient and to prevent overflow I output a 16 bit number, next I want to sum over all the sins I have and I've determined that I need to output a 23 bit number to prevent overflow, now I need to output the result to a display which expects an 8 bit number and I'm stuck at this step as I'm not sure how to sample my 23 bit number so I get the correct shape of the function without distortions.

Related Electrical Engineering News on Phys.org
Baluncore
2019 Award
Stricktly speaking, down-sampling involves changing the sample rate, not the bit depth.
You need to keep the signal in range on the screen. Can you display more than 8 bit data?
If your output is to be in dB then you might take the log (to the base two) of the 23 bit number.
If your output must be linear you will need to search for the maximum and scale the dataset.
display(i) = 255 * value(i) / maximum.
The process gain of the transform will increase the dynamic range of the output spectrum by the square root of the number of input samples.

Stricktly speaking, down-sampling involves changing the sample rate, not the bit depth.
You need to keep the signal in range on the screen. Can you display more than 8 bit data?
If your output is to be in dB then you might take the log (to the base two) of the 23 bit number.
If your output must be linear you will need to search for the maximum and scale the dataset.
display(i) = 255 * value(i) / maximum.
The process gain of the transform will increase the dynamic range of the output spectrum by the square root of the number of input samples.
Thanks for the reply! I was going for something similar, thing is my dataset is dynamic which means my maximum will change in real time, I suppose the easy solution is to always output the last 8 bits, am I correct?

Baluncore
2019 Award
I suppose the easy solution is to always output the last 8 bits, am I correct?
You could just display the most significant 8 bits but you will be throwing away much of your data.

If you do not have a fast multiply function then, rather than tracking the maximum, consider doing a 23 bit OR of all the data values into a common register that started out zero. After the accumulate, the first set bit will tell you how many bits to shift the output to show the most useful 8 bits.

sophiecentaur
Gold Member
Talk about a quart into a pint pot!
For really drastic bit reduction is could be worth using floating point arithmetic.
so I get the correct shape of the function without distortions.
Hardly 'without distortions' but what level of distortion would be acceptable?
You could always low pass filter the high res data, which would reduce apparent distortion.
There are so many possibilities but the best one would depend upon your actual requirement and what complexity (and delay) you can accept for the processing.

marcusl
Gold Member
Sampling with finite resolution also introduces noise into even ideal signals (like a sine wave read from a table). It can be shown that the noise is additive white Gaussian. Whether the reduction in SNR is significant depends on your particular experiment.

Baluncore
2019 Award
I'm trying to build a system that recieves coefficients and performs a Fourier approximation,
This appears to be a Fourier Synthesis exercise, computed in hardware using integers only.
Unless there is some intelligent control of the integer input coefficients, the output amplitude will need to be scaled.

sophiecentaur
Gold Member
This appears to be a Fourier Synthesis exercise, computed in hardware using integers only.
Unless there is some intelligent control of the integer input coefficients, the output amplitude will need to be scaled.
Raised Cos analysis is what's done for video bit reduction. I think it's chosen for speed and for the subjective qualities of the result (relevant here). There is a load of stuff about MPEG coding and its history (and Jpeg, too). The questions of scaling, blocking and interpolation have all been addressed since digital transmission of sound and vision have been implemented. No need to reinvent a wheel, I think.

Baluncore
2019 Award
No need to reinvent a wheel, I think.
I thought that this was an introductory exercise in VHDL. There was no OP mention of fast transform algorithms or data compression.

sophiecentaur