convert c_float to ieee_11073_16bit_float

[alex_u]

Hello,

I need to convert a standard 32 bit floating point number to IEEE-11073 16-bit FLOAT with 12-bit mantissa and 4-bit exponent (source: https://infocenter.nordicsemi.com). Is there already an implementation in the sdk or an example somewhere? I tried it the following way, but the numbers in the nrf connect app where wrong.

#include "ble_bps.h"
ieee_float16_t c_float32_to_ieee_11073_float16(float input)
{
    ieee_float16_t ret;
    uint32_t input_raw = *((uint32_t*)&input);
    ret.mantissa = (input_raw >> 23) & 0xf;
    ret.exponent = (input_raw & 0xfff);
    return ret;
}

[haakonsh]

See similar post https://devzone.nordicsemi.com/f/nordic-q-a/28727/32-bit-ieee-11073-float-data-type-parser 

Otherwise, I'd check the ARM CMSIS DSP library, arm_math.h, https://github.com/ARM-software/CMSIS/blob/master/CMSIS/Include/arm_math.h

[alex_u]

the arm_math.h and the post your referenced are using 32 bit floating numbers. I already checked those. I am also confused about the 12 bit mantissa and 4 bit exponent format. The standard half precision floating point number seems to be 5 bit exponent and 11 bit mantissa (source: wikipedia). Could you give me a more detailed description of the IEEE-11073 16-bit FLOAT format? I cant find anything about exponent encoding and INF and NAN definition.

[alex_u]

I currently have no access to the specification. But I want to use your SDK to transmit bloodpressure measurments. It seems like this is only possible by using the 16 bit float defined in IEEE-11073. Or is there another solution?

[alex_u]

I managed to get it done.

Here is the final code:

ieee_float16_t c_float32_to_ieee_11073_float16(float input)
{
	ieee_float16_t ret;	
	ret.exponent = 0;
	while(input > (float)0x7ff){
		input /= 10;
		ret.exponent += 1;
	}
	while(input*10 < (float)0x7ff){
		input *= 10;
		ret.exponent -= 1;
	}
	ret.mantissa = (int16_t)input;
	return ret;
}

Thanks for your help.

Alex

[haakonsh]

That's not gonna work, for larger numbers the mantissa gets truncated. 
You need an algorithm that always returns a float with 12-bit precision. 

[alex_u]

Your right, thanks. I added an -INF and +INF limitation. For convinience, I set the max values a little bit less than the real limits of the 16 bit float. This should solve the issue.

 

ieee_float16_t c_float32_to_ieee_11073_float16(float input)
{
	ieee_float16_t ret;	
	if(input != input) //check for NAN
	{
		ret.exponent = 0;
		ret.mantissa = 0x07ff;
		return ret;
	}
	else if(input > (float)2E10) //check for +INF
	{
		ret.exponent = 0;
		ret.mantissa = 0x07FE;
		return ret;
	}
	else if(input < (float)-2E10) //check for -INF
	{
		ret.exponent = 0;
		ret.mantissa = 0x0802;
		return ret;
	}
	else
	{
		ret.exponent = 0;
		while(input > (float)0x7ff){
			input /= 10;
			ret.exponent += 1;
		}
		while(input*10 < (float)0x7ff){
			input *= 10;
			ret.exponent -= 1;
		}
		ret.mantissa = ((int16_t)input) & 0xfff;
		ret.exponent &= 0xf;
		return ret;
	}
}

[haakonsh]

Given that the largest mantissa is 2^11 - 1, and the largest exponent is 7,
the largest number you can represent will be (2^11 - 1) * 10^7 = 2.047 * 10^10. 

The precision of the mantissa will cause significant errors with very large or very small values, ie
the number 123456789: exp = 5, mantissa = 1235 --> 0.1235 * 10^5 = 123500000.

or 0.000001234567: exp = -5, mantissa = 1235 --> 0.1235 * 10^-5 =  0.000001235.

This is from my limited understanding, it might be wrong. 

There are also at least 5 cases that need to be handled differently:

1. input == zero.
2. input >= 1.0
3. input < 1.0
4. input >= -1.0
5. input < -1.0