static void FUNC(transform_8x8_dc_add)(
    uint8_t *_dst, int16_t *coeffs, ptrdiff_t _stride) {
    int i, j;
    pixel    *dst    = (pixel *)_dst;
    int      stride  = _stride/sizeof(pixel);
    int      shift   = 14 - 8;
    int      add     = 1 << (shift - 1);
    int      coeff   = (((coeffs[0] + 1) >> 1) + add) >> shift;

    for (j = 0; j < 8; j++) {
        for (i = 0; i < 8; i++) {
            dst[i+j*stride] = av_clip_pixel(dst[i+j*stride] + coeff);
        }
    }
}


cglobal hevc_put_transform8x8_dc_add_8, 3, 4, 3, dst, coeffs, stride, temp
    xor            tempq, tempq
    mov            tempw, [coeffsq]
    add            tempd, 1
    shr            tempd, 1
    add            tempd, 32
    shr            tempd, 6
    movd              m0, tempd
    punpcklwd         m0, m0
    pshufd            m0, m0, 0
    pxor              m1, m1
    xor            tempq, tempq
    mov            tempd, 8
.loop
    movq              m2, [dstq]                                               ; load data from source
    punpcklbw         m2, m1
    paddw             m2, m0
    packuswb          m2, m2
    movq          [dstq], m2
    lea             dstq, [dstq+strideq]      ; dst += dststride
    dec            tempd
    jnz                 .loop                 ; height loop
    RET