Converting double values file to 2 byte per sample.

Hi,

I have a .txt file with 128 x 11900 vlaues. These values are doubles ranging from -0.00023462 to +1.0719. I need to convert them to 2 bytes per sample and store in a .bin file. I am not getting how to do it. I tried using the 'fread' function of matlab but all I am able to do is convert each sample to 8 bytes(double precision). Is there a way I can convert each sample into 2 bytes each?

Answer 1

[Walter Roberson]

two_bytes = uint16(floor(YourMatrix/1.636E-5)+15);

and fwrite() as uint16 .

reconstructed = (double(two_bytes) - 15) * 1.636E-5

the reconstruction error can be up to 1.636E-5 . You can reduce the average reconstruction error by using round() instead of floor(), but then you have to work on the boundary condition (the +15 or -15) (though there happens to be just enough slack room in the representation so that it is not a problem.)

[ratnesh pandey]

Yes, it solved my problem. But why 15?

[Walter Roberson]

Divide the difference between the max and min by the number of different 16 bit values, 65536. You get a number for the difference between adjacent quantified samples, but it is not a convenient number to put in code, so tweak it to be a slightly smaller "nice" number, delta.

Now for any one sample, the representation in 2^16th units is floor of (the sample divided by delta). Apply that to the lower bound of your data and you get -15. Your data is quantized between -15*delta and 65519 times delta.

At this point you could use a signed representation but it is a waste to effectively devote an entire bit for the matter of whether the quantized value is in the range -15 to -1 and you would need to use a delta twice as large. Easier to add 15 to the quantized value to shift to the range 0 to 65534 .

If you round instead of floor then the shifted quantized value becomes the range 1 to 65535 which is fine.

[ratnesh pandey]

"Now for any one sample, the representation in 2^16th units is floor of (the sample divided by delta). Apply that to the lower bound of your data and you get -15"

you mean -14? floor(-0.00023462/1.636E-5)

[Walter Roberson]

No, I mean -15

>> floor(-0.00023462/1.636E-5)

ans =

   -15

Answer 2

[James Tursa]

If you want a two byte floating point representation you can use half precision. E.g.,

https://www.mathworks.com/matlabcentral/fileexchange/23173-ieee-754r-half-precision-floating-point-converter

https://www.mathworks.com/help/fixedpoint/ref/half.html

This is one of the standard floating point formats, but will result in less precision than Walter's method because some of the bits are used for sign and exponent.