Intel® Integrated Performance Primitives
Deliberate problems developing high-performance vision, signal, security, and storage applications.

ippsSqrt_32f doesn't handle denormals correctly

Dan_R_
Beginner
598 Views

Hi,

It looks like both the ippsSqrt_32f and ippsSqrt_32f_I return +Inf for a positive denormal number.

See example in https://gist.github.com/yairchu/730ee5fa0368758c7fe8050447c43016 - the output is 

4.3665076719685603e-39 6.6079555878345896e-20
inf 0

This happens with both the latest 2018 Update 3 and in 2017 Update 4.

Looks like a bug unless I'm misunderstanding something.

Thanks,

Dan

0 Kudos
2 Replies
Igor_A_Intel
Employee
598 Views

Hi Dan,

Thank you for discovering this bug - I've managed to reproduce it.

We'll fix it in the one of the nearest releases, for the current moment I recommend you to use a workaround - ippsSqrt_32f_A24(src, dst, len) function from the ippVM domain (#include ippvm.h).

regards, Igor.

0 Kudos
Dan_R_
Beginner
598 Views

Great, thanks for the quick reply!

Dan

0 Kudos
Reply