CODEWITHSUNDEEP
x86ssesimd
Fan Zhang
Fan Zhang

Reputation: 629

SSE instruction need the data aligned

Must the data be 16-byte aligned so that it can be processed by the SSE instruction without segmentation fault? The compiler I tried is gcc with option -msse2. I want use _mm_cmpgt_epi32 to compare a large int array. I found that it can not be executed at any location of the array except the position with subscript of the multiples of 4.

Upvotes: 1

Views: 1168

Answers (1)

Paul R
Paul R

Reputation: 212969

Yes, when you load and store data to/from SSE registers it needs to be 16 byte aligned, unless you use the misaligned versions of the load/store instructions, e.g. _mm_loadu_si128/_mm_storeu_si128. There is typically a performance penalty for using these misaligned load/store instructions however, so one would normally try to ensure correct data alignment at all times and only use misaligned loads/stores as a last resort.

Upvotes: 6

Related Questions