New oapv_dc_removed_had8x8_neon() implementation.

[lordnn]

Binary equality but 30% faster.

[kpchoi]

The first pass runs in s16 with a gain of 8, so this kernel only matches the scalar oapv_dc_removed_had8x8() while max|sample| <= 4095; at 4096 they diverge. Same limit as the SSE and AVX kernels, so not a regression, but it is documented nowhere in the NEON file.

Please add a note for it in the same place oapv_dc_removed_had8x8_avx() has one — first thing in the function body, before int satd = 0;:

    /* first pass is register-wise on 128-bit s16 row vectors, so its values
       reach 8 * 4095 and only just fit in s16; the input is therefore limited
       to 12-bit samples. after a transpose the second pass runs register-wise
       on 128-bit s32 vectors, since its values can reach 64 * 4095 and do not
       fit in s16 */

[kpchoi]

VADDVQ_S32 is down to a single call site now, so the macro no longer earns its keep. Please drop the definition at the top of the file and inline it — int satd = 0; in the function header can go with it:

    int satd = vaddvq_s32(satv);
    return (satd + 2) >> 2;

That is the last thing from me.

[cpncf]

LGTM