Is there a way to read 16bit float(mantissa 8bit)bin file?

I've found a function CustomFloat but I can't find a way to read the file...

it goes like this

00 00 07 00 F8 FF FD FF 01 00 01 00 01 00 FF FF

[dpb]

What's the encoding definition?

[Won Geun Shin]

I don't understand the definition of encoding definition.

But, I think it is little endian and it uses 16bit for floating point(1 sign bit, 7 exponent bits and 8bit for mantissa)

[James Tursa]

I think what dpb means, is the encoding "half precision" or "bfloat16" or something else? Can you provide what floating point numbers you would expect from your example hex codes? Then we can verify encoding.

[Won Geun Shin]

At the explanation, it says that it is fixed point number of 7 bits of precision. So it's not half precision or so. I don't know the expected number from that. Sorry....

[Walter Roberson]

Fixed point numbers do not use exponent fields.

If you have a floating point representation then you need to know the offset on the exponent. The exponent is not a signed field so the representation is

(1 - Signbit * 2) times 2^(exponent - Base) times 8 bit integer

but we do not know what the Base offset is.

Also we need to know whether the representation uses the "hidden bit" algorithm, as is common. If so then instead of times an 8 bit integer, would be times (256 plus 8 bit integer)

[Won Geun Shin]

sorry for the late response.

I had to communicate with other team for it.

So, apparently i knew it wrong..

MSB 8bit is the integer part and LSB 8bit is the fraction part(only the MSB bits is Affected by the sign bit) and it is 2's complement number!

so example goes like this

ex]

01000110 11000000 = 70.75 (70+0.75(0.5+0.25))

10111001 01000000 = -70.75(-71+0.25)

[Won Geun Shin]

so I guess I can use bin2num function to covert that..

Thanks a lot

[James Tursa]

@Walter Roberson Would also need to know whether the significand is normalized to 1.0 binary or 0.1 binary. This ties in with the exponent bias, of course.

Answer 1

[Walter Roberson]

str = "00 00 07 00 F8 FF FD FF 01 00 01 00 01 00 FF FF 46 C0 B9 40"

str = "00 00 07 00 F8 FF FD FF 01 00 01 00 01 00 FF FF 46 C0 B9 40"

A = sscanf(str, "%2x", [1 inf]);
if mod(numel(A), 2) == 1; A(end+1) = 0; end   %pad if odd length
MSB = A(1:2:end);
mask = MSB >= 128;
MSB(mask) = MSB(mask) - 256;
LSB = A(2:2:end);
out = MSB + LSB/256;

format long g
out.'

ans = 10×1

0 7 -7.00390625 -2.00390625 1 1 1 -0.00390625 70.75 -70.75

Answer 2

[Chunru]

str = "00 00 07 00 F8 FF FD FF 01 00 01 00 01 00 FF FF"

str = "00 00 07 00 F8 FF FD FF 01 00 01 00 01 00 FF FF"

a = sscanf(str, "%s%s", inf);
n = floor(length(a)/4)

n = 8

afi = fi(0, 1, 16, 7);  % signed, 16-bit word, 7-bit fraction
for i=1:n
    h = a((i-1)*4 + (1:4));       % the hex from data input    
    afi.hex = h;                  % assign the hex to fi
    d = afi.double;               % convert to double 
    fprintf("%s : %f\n", afi.hex, afi.double)
end

0000 : 0.000000
0700 : 14.000000
f8ff : -14.007812
fdff : -4.007812
0100 : 2.000000
0100 : 2.000000
0100 : 2.000000
ffff : -0.007812