Function count_nonzero_bytes_384

numpy/core/src/multiarray/item_selection.c:2360–2399

* count number of nonzero bytes in 48 byte block * w must be aligned to 8 bytes * * even though it uses 64 bit types its faster than the bytewise sum on 32 bit * but a 32 bit type version would make it even faster on these platforms */

 * but a 32 bit type version would make it even faster on these platforms
 */
static inline npy_intp
count_nonzero_bytes_384(const npy_uint64 * w)
{
    const npy_uint64 w1 = w[0];
    const npy_uint64 w2 = w[1];
    const npy_uint64 w3 = w[2];
    const npy_uint64 w4 = w[3];
    const npy_uint64 w5 = w[4];
    const npy_uint64 w6 = w[5];
    npy_intp r;

    /*
     * last part of sideways add popcount, first three bisections can be
     * skipped as we are dealing with bytes.
     * multiplication equivalent to (x + (x>>8) + (x>>16) + (x>>24)) & 0xFF
     * multiplication overflow well defined for unsigned types.
     * w1 + w2 guaranteed to not overflow as we only have 0 and 1 data.
     */
    r = ((w1 + w2 + w3 + w4 + w5 + w6) * 0x0101010101010101ULL) >> 56ULL;

    /*
     * bytes not exclusively 0 or 1, sum them individually.
     * should only happen if one does weird stuff with views or external
     * buffers.
     * Doing this after the optimistic computation allows saving registers and
     * better pipelining
     */
    if (NPY_UNLIKELY(
             ((w1 | w2 | w3 | w4 | w5 | w6) & 0xFEFEFEFEFEFEFEFEULL) != 0)) {
        /* reload from pointer to avoid a unnecessary stack spill with gcc */
        const char * c = (const char *)w;
        npy_uintp i, count = 0;
        for (i = 0; i < 48; i++) {
            count += (c[i] != 0);
        }
        return count;
    }

    return r;
}

#if NPY_SIMD
/* Count the zero bytes between `*d` and `end`, updating `*d` to point to where to keep counting from. */