69   mm1 = mm7; \
71   mm1 = _mm_slli_si64(mm1, 2 * BYTE_BIT);     /* mm1=( - 0 1 2) */ \
76   mm1 = _mm_or_si64(mm1, wk[r]);              /* mm1=(-1 0 1 2) */ \
83   mm1 = _mm_add_pi16(mm1, PW_EIGHT); \
88   mm1 = _mm_add_pi16(mm1, mm7); \
90   mm1 = _mm_srli_pi16(mm1, 4);                /* mm1=OutrLE=( 0  2  4  6) */ \
99   mm1 = _mm_or_si64(mm1, mm0);     /* mm1=OutrL=( 0  1  2  3  4  5  6  7) */ \
102   _mm_store_si64((__m64 *)outptr##r, mm1); \
114   __m64 mm0, mm1, mm2, mm3 = 0.0, mm4, mm5, mm6, mm7 = 0.0;  in jsimd_h2v2_fancy_upsample_mmi()  local
148     mm1 = _mm_load_si64((__m64 *)inptr_1);    /* mm1 = row[-1][0] */  in jsimd_h2v2_fancy_upsample_mmi()
155     mm5 = mm1;  in jsimd_h2v2_fancy_upsample_mmi()
156     mm1 = _mm_unpacklo_pi8(mm1, mm3);         /* mm1 = row[-1][0]( 0 1 2 3) */  in jsimd_h2v2_fancy_upsample_mmi()
168     mm1 = _mm_add_pi16(mm1, mm0);             /* mm1=Int0L=( 0 1 2 3) */  in jsimd_h2v2_fancy_upsample_mmi()
173     _mm_store_si64((__m64 *)outptr0, mm1);      /* temporarily save */  in jsimd_h2v2_fancy_upsample_mmi()
178     mm1 = _mm_and_si64(mm1, mm7);             /* mm1=( 0 - - -) */  in jsimd_h2v2_fancy_upsample_mmi()
181     wk[0] = mm1;  in jsimd_h2v2_fancy_upsample_mmi()
191         mm1 = _mm_load_si64((__m64 *)inptr_1 + 1);  /* mm1 = row[-1][1] */  in jsimd_h2v2_fancy_upsample_mmi()
198         mm5 = mm1;  in jsimd_h2v2_fancy_upsample_mmi()
199         mm1 = _mm_unpacklo_pi8(mm1, mm3);     /* mm1 = row[-1][1]( 0 1 2 3) */  in jsimd_h2v2_fancy_upsample_mmi()
208         mm1 = _mm_add_pi16(mm1, mm0);         /* mm1 = Int0L = ( 0 1 2 3) */  in jsimd_h2v2_fancy_upsample_mmi()
213         _mm_store_si64((__m64 *)outptr0 + 2, mm1);  /* temporarily save */  in jsimd_h2v2_fancy_upsample_mmi()
218         mm1 = _mm_slli_si64(mm1, (SIZEOF_MMWORD - 2) * BYTE_BIT); /* mm1=( - - - 0) */  in jsimd_h2v2_fancy_upsample_mmi()
221         wk[2] = mm1;  in jsimd_h2v2_fancy_upsample_mmi()
225         mm1 = _mm_cmpeq_pi8(mm1, mm1);  in jsimd_h2v2_fancy_upsample_mmi()
226         mm1 = _mm_slli_si64(mm1, (SIZEOF_MMWORD - 2) * BYTE_BIT);  in jsimd_h2v2_fancy_upsample_mmi()
227         mm2 = mm1;  in jsimd_h2v2_fancy_upsample_mmi()
230         mm1 = _mm_and_si64(mm1, mm_tmp);      /* mm1=( - - - 7) */  in jsimd_h2v2_fancy_upsample_mmi()
234         wk[2] = mm1;  in jsimd_h2v2_fancy_upsample_mmi()