efficient way to convert scatter indices into gather indices?

Question

I\'m trying to write a stream compaction (take an array and get rid of empty elements) with SIMD intrinsics. Each iteration of the loop processes 8 elements at a time (SIMD

Answer 1

If you want to emulate _mm_movemask_epi8 and you just need an 8 bit scalar mask from 8 byte elements then you can do something like this using AltiVec:

#include <stdio.h>

int main(void)
{
    const vector unsigned char vShift = { 0, 1, 2, 3, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0 };
                                            // constant shift vector

    vector unsigned char isValid = { 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
                                            // sample input

    vector unsigned char v1 = vec_sl(isValid, vShift);
                                            // shift input values
    vector unsigned int v2 = vec_sum4s(v1, (vector unsigned int)(0));
    vector signed int v3 = vec_sum2s((vector signed int)v2, (vector signed int)(0));
                                            // sum shifted values
    vector signed int v4 = vec_splat(v3, 1);
    unsigned int mask __attribute__ ((aligned(16)));
    vec_ste((vector unsigned int)v4, 0, &mask);
                                            // store sum in scalar

    printf("v1 = %vu\n", v1);
    printf("v2 = %#vlx\n", v2);
    printf("v3 = %#vlx\n", v3);
    printf("v4 = %#vlx\n", v4);
    printf("mask = %#x\n", mask);

    return 0;
}

This is 5 AltiVec instructions versus 1 in SSE. You might be able to lose the vec_splat and get it down to 4.