The Generic SSE2 Loop

In response to a couple of comments on my post about find-first-bit-set in SSE2 registers, amounting to “what use is a routine that only does 16-byte bitvecs”, I thought I’d post the canonic, generic loop through memory using SSE2 ops. The basic principles are:

• valid memory segments in x86 always end on multiple of 16 bytes (or more)
• aligned loads of XMM registers (i.e. on 16-byte boundaries) are always faster than unaligned loads.

The generic loop over unaligned memory amounts to:

•  do an aligned load of the first part of a byte vector, possibly throwing away the low bytes of what’s loaded
• loop through aligned 16-byte chunks
• do an aligned load of the last part of a byte vector, possibly throwing away the high bytes of what’s loaded.

You’ll see this in my examples on SSE2-based strstr. Here’s “find-first-bit-set” for an arbitrary bit-vector:

 
// For a vector of 16 bytes, NZBYTES returns a 16-bit mask
//   of the non-zero bytes.
#define NZBYTES(p) (0xFFFF ^ _mm_movemask_epi8(\
                               _mm_cmpeq_epi8(*(__m128i const*)(p),\
                                              _mm_setzero_si128())))
int bitvec_ffs(uint8_t const*vec, int len)
{
    unsigned i = 15 & (uintptr_t)vec;
    uint8_t const *bp = vec - i;
    uint16_t m;
    if (i) {
        if ((m = NZBYTES(bp) >> i))
            return i = FFS(m), i*8 + FFS(vec[i]);
        len -= i, bp += 16;
    }

    for (; len >= 16; len -= 16, bp += 16)
        if ((m = NZBYTES(bp)))
            return i = FFS(m), (bp-vec+i)*8 + FFS(bp[i]);

    if (len > 0)
        if ((m = NZBYTES(bp) & ~(-1 << len)))
            return i = FFS(m), (bp-vec+i)*8 + FFS(bp[i]);

    return -1;
}

A more succinct way of solving the problem partitions it into the three parts implicitly, by careful arrangement of state changes. You be the judge of what’s more understandable 🙂

int bitvec_ffs(uint8_t const*vec, int len)
{
    int off = 15 & (intptr_t)vec;
    uint8_t const *bp = vec - off;  // back up to a 16b boundary
    uint16_t m = NZBYTES(bp);
    len += off;
    if (len < 16)
        m &= ~(-1 << len);  // mask off trailing bits (tiny vec[])
    for (m &= -1 <= 16; m = NZBYTES(bp), len -= 16)
        bp += 16;
    if (len < 16)
        m &= ~(-1 << len); // mask off trailing bits (tail of vec[])
    if (!m) return -1;
    off = ffs(m) - 1;
    return 8 * (off + bp - vec) + ffs(bp[off]) - 1;
}

Advertisements