How IPP functions handle data that contain special values?

[www_q_]

Do IPP functions  handle special values like quiet NaN/inf  strictly following the IEEE 754 way?

[www_q_]

I mean, such as: if x is a quiet NaN value, then do functions like ippsAddC_32f(x, yptr, length_of_y) returns strictly all NaNs despite whatever values stored in yptr? Also how NaN values will be ranked when sorting some array that contain NaNs?

[SergeyKostrov]

>>...Do IPP functions handle special values like quiet NaN... >>...if x is a quiet NaN value, then do functions like ippsAddC_32f(x, yptr, length_of_y) returns strictly all NaNs despite whatever >>values stored in yptr? An IPP function could fail with Access Violation exception. It could also process data without Access Violation exception in a future release of IPP and please take a look at a similar thread: http://software.intel.com/en-us/forums/showthread.php?t=106039&o=a&s=lr

[www_q_]

Thanks, so the IPP functions are not stable with special values... thats too bad.

[SergeyKostrov]

>>... the IPP functions are not stable with special values... ...It could also process data without Access Violation exception in a future release... and please take it into account. I wonder if somebody from IPP team could provide an update regarding the problem.

[Igor_A_Intel]

IPP libraries strictly follow IEEE 754 for floating point numbers. That means a bug in realization if you find different/undesirable behavior for any IPP function - so be so kind - submit a bug to QuAD or post a message here, please. Regarding ippiRotate - it has been already fixed - use 7.1. version, please. 2 Sergey Kostrov: I don't know any other IPP function that fails with AV because of special FP values in parameters - if you know - be so kind, share this info here, please. PS I think you know an anecdote about Japanese gasoline-powered saw... Regards, Igor