Lines Matching refs:__m128i
41 const __m128i tmp0 = _mm_avg_epu8(k, (in)); /* (k + in + 1) / 2 */ \
42 const __m128i tmp1 = _mm_and_si128((ij), st); /* (ij) & (s^t) */ \
43 const __m128i tmp2 = _mm_xor_si128(k, (in)); /* (k^in) */ \
44 const __m128i tmp3 = _mm_or_si128(tmp1, tmp2); /* ((ij) & (s^t)) | (k^in) */\
45 const __m128i tmp4 = _mm_and_si128(tmp3, one); /* & 1 -> lsb_correction */ \
51 const __m128i ta = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \
52 const __m128i tb = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \
53 const __m128i t1 = _mm_unpacklo_epi8(ta, tb); \
54 const __m128i t2 = _mm_unpackhi_epi8(ta, tb); \
55 _mm_store_si128(((__m128i*)(out)) + 0, t1); \
56 _mm_store_si128(((__m128i*)(out)) + 1, t2); \
61 const __m128i one = _mm_set1_epi8(1); \
62 const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]); \
63 const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]); \
64 const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]); \
65 const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]); \
67 const __m128i s = _mm_avg_epu8(a, d); /* s = (a + d + 1) / 2 */ \
68 const __m128i t = _mm_avg_epu8(b, c); /* t = (b + c + 1) / 2 */ \
69 const __m128i st = _mm_xor_si128(s, t); /* st = s^t */ \
71 const __m128i ad = _mm_xor_si128(a, d); /* ad = a^d */ \
72 const __m128i bc = _mm_xor_si128(b, c); /* bc = b^c */ \
74 const __m128i t1 = _mm_or_si128(ad, bc); /* (a^d) | (b^c) */ \
75 const __m128i t2 = _mm_or_si128(t1, st); /* (a^d) | (b^c) | (s^t) */ \
76 const __m128i t3 = _mm_and_si128(t2, one); /* (a^d) | (b^c) | (s^t) & 1 */ \
77 const __m128i t4 = _mm_avg_epu8(s, t); \
78 const __m128i k = _mm_sub_epi8(t4, t3); /* k = (a + b + c + d) / 4 */ \
79 __m128i diag1, diag2; \