51         unsigned ind = row * size_sse_row;  in shift_left2D()  local
52         unsigned x0 = X[ind].m128i_u32[0] & 1;  in shift_left2D()
54         X[ind] =  in shift_left2D()
55             _mm_or_si128(_mm_srli_epi16(X[ind], 1), _mm_slli_epi16(_mm_srli_si128(X[ind], 2), 15));  in shift_left2D()
57         unsigned x1 = X[ind + 1].m128i_u32[0] & 1;  in shift_left2D()
58         X[ind + 1] = _mm_or_si128(_mm_srli_epi16(X[ind + 1], 1),  in shift_left2D()
59                                   _mm_slli_epi16(_mm_srli_si128(X[ind + 1], 2), 15));  in shift_left2D()
60         X[ind].m128i_u32[3] |= x1 << 31;  in shift_left2D()
62         unsigned x2 = X[ind + 2].m128i_u32[0] & 1;  in shift_left2D()
63         X[ind + 2] = _mm_or_si128(_mm_srli_epi16(X[ind + 2], 1),  in shift_left2D()
64                                   _mm_slli_epi16(_mm_srli_si128(X[ind + 2], 2), 15));  in shift_left2D()
65         X[ind + 1].m128i_u32[3] |= x2 << 31;  in shift_left2D()
67         unsigned* dst = (unsigned*)&X[ind];  in shift_left2D()
74         unsigned ind = row * size_sse_row;  in shift_right2D()  local
76         unsigned x0 = X[ind].m128i_u32[3];  in shift_right2D()
78         X[ind] =  in shift_right2D()
79             _mm_or_si128(_mm_slli_epi16(X[ind], 1), _mm_srli_epi16(_mm_slli_si128(X[ind], 2), 15));  in shift_right2D()
81         unsigned x1 = X[ind + 1].m128i_u32[3];  in shift_right2D()
83         X[ind + 1] = _mm_or_si128(_mm_slli_epi16(X[ind + 1], 1),  in shift_right2D()
84                                   _mm_srli_epi16(_mm_slli_si128(X[ind + 1], 2), 15));  in shift_right2D()
85         X[ind + 1].m128i_u32[0] |= x0;  in shift_right2D()
87         unsigned* dst = (unsigned*)&X[ind];  in shift_right2D()
90         X[ind + 2] = _mm_or_si128(_mm_slli_epi16(X[ind + 2], 1),  in shift_right2D()
91                                   _mm_srli_epi16(_mm_slli_si128(X[ind + 2], 2), 15));  in shift_right2D()
92         X[ind + 2].m128i_u32[0] |= x1;  in shift_right2D()
93         X[ind].m128i_u32[0] |= x2;  in shift_right2D()