Questions on FastNumericArrayConverter

[guohaoqiang]

It converts 4 uint8 to 4 half precision. I have no idea why it adds 0x66006600 here. Anyone could help with it please? Thanks

[hwu36]

there is a certain range in which a data d's float hex is the same as its int hex. then, we can add a bias to a value to push it to be within this range. we just need to subtract the bias after the conversion to restore its previous value.

the assumption here is that add/sub is faster than the real conversion.

[guohaoqiang]

Thank you. I'm trying to figure out if 0x6600 is unique, namely whether 0x6600 is replaceable or not.
I will post my code and findings later this week.

[hwu36]

it is not unique. any bias that can push the initial value to the range in which float and int hex are the same is okay.

[guohaoqiang]

How to get the magic number is a trick.

let's say:
int8_t a = -1; // binary 1111 1111
int16_t b = a + 0x6600; // b = 0110 0101 1111 1111

Now we have
b(hex) = 0x65ff
b(decimal) = 26111
but b(__half) = 1535 != 26111

https://evanw.github.io/float-toy/

[hwu36]

0x6600 is for uint8_t, not int8_t

[hwu36]

you can choose a random int value, e.g. 1. to go through these instructions to see how the hex change.

[hwu36]

this tool might be helpful for you to understand https://gregstoll.com/~gregstoll/floattohex/