INIT_XMM sse2
; void ff_hevc_idct16_dc_add_8_mmxext(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
cglobal hevc_idct16_dc_add_8, 3, 4, 0
    movsx             r3, word [r1]
    DC_ADD_INIT       r3, r2
    DC_ADD_OP       mova, r0, r2, r3
    lea               r0, [r0+r2*4]
    DC_ADD_OP       mova, r0, r2, r3
    lea               r0, [r0+r2*4]
    DC_ADD_OP       mova, r0, r2, r3
    lea               r0, [r0+r2*4]
    DC_ADD_OP       mova, r0, r2, r3
    RET

%if HAVE_AVX2_EXTERNAL
INIT_YMM avx2
; void ff_hevc_idct16_dc_add_8_mmxext(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
cglobal hevc_idct32_dc_add_8, 3, 4, dst, coeffs, stride
    movsx             r3, word [r1]
    DC_ADD_INIT_AVX2  r3, r2
    DC_ADD_OP       movu, r0, r2, r3,
 %rep 7
    lea               r0, [r0+r2*4]
    DC_ADD_OP       movu, r0, r2, r3
%endrep
    RET
%endif ;HAVE_AVX2_EXTERNAL