SIMD (AVX2) mask store and pack

Asked Aug 16 '16 at 23:07

Active Aug 16 '16 at 23:07

Viewed 508 times

I am trying to perform the following operation in AVX2 code (dest, data, and mask are int32 pointers):

int j=0;
for(i=0; i<N; ++i){
    if (mask[i]){
        dest[j++] = data[i];
    }
}

In AVX2, the closest instruction I can find is:

_mm256_maskstore_epi32(dest, (__m256*)mask, (__m256*)data);

But then I would need to perform the packing manually, which seems inefficient...

asked Aug 16 '16 at 23:07

nineties

1

exact duplicate: producing a filtered copy of an array with AVX2. – Peter Cordes Aug 16 '16 at 23:12
1

My googling skill isn't quite there yet, but :) – nineties Aug 16 '16 at 23:15
Feel free to upvote my answer there. I'm still pretty proud of the cool BMI2 trick using `pext`'s left-packing hardware to generate a shuffle mask. :) – Peter Cordes Aug 17 '16 at 01:02
done :). Thanks. I wish stackoverflow was more thanking friendly – nineties Aug 17 '16 at 20:38

0 Answers0