CODEWITHSUNDEEP
armsimdneon
Lakshmi
Lakshmi

Reputation: 101

Perform a horizontal logical/bitwise AND operation across all lanes of uint8x8 Neon vector

I have a uint8x8 Neon vector which is a result of some operation. I need to perform a logical AND operation of the all the lanes to get my final result. Each element is either 0xff (TRUE) or 0x00 (FALSE). How do I perform it in Neon?

Upvotes: 3

Views: 1000

Answers (2)

Jake 'Alquimista' LEE
Jake 'Alquimista' LEE

Reputation: 6354

In that case, you can simply do a binary negation, and check if the 64bit result is 0.

vmvn d0, d0    
vpaddl.u32 d0, d0 // 64bit vceq isn't possible.
vceq.i32 d0, d0, #0

You now have the desired result in d0.

If you are working on aarch64, 64bit cmeq is possible

mvn v0.16b, v0.16b
cmeq v0.2d, v0.2d, #0

The best thing about this algorithm is that you don't need any other registers, because zero is the only immediate value accepted by the compare instructions.

Upvotes: 2

Paul R
Paul R

Reputation: 212959

Simple/obvious method (pseudo-code):

v = VAND(v, v >> 8)
v = VAND(v, v >> 16)
v = VAND(v, v >> 32)

3 x shifts and 3 x bitwise ANDs = 6 instructions.

Possibly more efficient method: do a horizontal sum of all elements, then return TRUE if sum == -8, otherwise return FALSE.
Possibly simpler method: just compare vector with a vector of all 1s.

return v == 0xffffffffffffffff;

Doing this efficiently is left as an exercise for the reader (may require 2 x 32 bit compares ?).

Upvotes: 1

Related Questions