CODEWITHSUNDEEP
ssesimd
Michael IV
Michael IV

Reputation: 11506

SSE comparison returns vector of NANs

I am trying something like this:

__m128 cA = _mm_set_ps1(-2.0f);
__m128 cB = _mm_set_ps1(2.0f);
__m128 df = _mm_cmpgt_ps(cA, cB);

In this case df returns with zeros.

But if I do:

__m128 cA = _mm_set_ps1(2.0f);
__m128 cB = _mm_set_ps1(-2.0f);
__m128 df = _mm_cmpgt_ps(cA, cB);

It returns all -nan .Is is expected behaviour? If it is,how do I evaluate those nans?

Using Intel CPU,MS VisualStudio 2017

Upvotes: 1

Views: 579

Answers (1)

Peter Cordes
Peter Cordes

Reputation: 365667

SIMD compares produce a mask. All-one bits is the bit-pattern for -NaN. All-zero bits is the bit-pattern for +0.0

They're not intended to be interpreted as float. Use them with _mm_movemask_ps, blends, or things like _mm_and_ps.

e.g. _mm_and_ps( vec, cmp_result) zeros the elements where the compare was false. You can use this to do a conditional add by zeroing some input elements before an add.

For more about how to use SIMD, look up a guide/tutorial. https://stackoverflow.com/tags/sse/info

Upvotes: 4

Related Questions