Some questions regarding the bizarre (to me) half2 datatype:
- How much faster is arithmetic done on 1x half2 object vs. 2x half objects?
- Why are the intrinsics for half2 so limited and strange? Just one example (of which there are several):
__hbeq2 - Performs half2 vector if-equal comparison and returns boolean true iff both half results are true, boolean false otherwise.
Why would I want a single boolean result from say, __hbeq2(a, b)? I want to know both results, that is, is
a.x != b.x
AND is
a.y != b.y
Also why is there not an intrinsic for checking if
a.x != a.y
?
Ultimately, my questions about this type are driven by my assumption that 1x half2 operation is faster than 2x half operation.
Thanks!