Digital Signal Processing - IP and Algorithms

[jjhewitson]

Hi,

Has anyone experieneced the floating-point core not behaving as specified with regards to overflow?

I've created a standard single precision add/sub core which behaves as expected for some numbers but in other respects not at all.

For example adding 20+56=76 //41A00000+42600000=42980000 (hex) --perfect.

However, when I'm tring to test the overflow detection it only seems to work if I add large numbers.

Adding 1 to the theoretical largest representable number won't overflow or equal infinity!?!?!?  (7F7FFFFF+3F800000 = 7F7FFFFF) !?!

Adding 1/2 the largest value to the largest value overflows and equals infinty as you'd expect as does adding 1/4 or 1/8th of the largest to itself too, but anything small, even 100+largest, doesn't overflow...?!?

What's going on here?

Thanks!

(Using ise13.2, V5SX95, I'm talking about isim simulations)

[viviany]

Questions about the core itself (e.g. function, performance, parameters and etc.) should go to the correspongding board for that core. The board for floating-point core is "Digital Signal Processing - IP and Algorithms". I'll move your message to that board.

Questions about using Coregen tool, Core generating issue should be posted on the Design Entry board.

Vivian

[jjhewitson · Accepted solution]

Hi,

Appologies, I see what's going on here now.

In case anyone else makes the same mistake;

The largest representable number in single precision is;

7F7FFFFF = 340282346638528859811704183484516925440

Adding 1 = 340282346638528859811704183484516925441, which if you convert to single precision floating point is also 7F7FFFFF....  

Unless the value I add is sufficiently large enough that it would cause a rounding to a larger number the core won't detect an overflow as is correct according to the IEEE754 standard.