26   v.val[0] = a;                     \
27   v.val[1] = b;                     \
31   v.val[0] = a;                        \
32   v.val[1] = b;                        \
33   v.val[2] = c;                        \
37   v.val[0] = a;                           \
38   v.val[1] = b;                           \
39   v.val[2] = c;                           \
40   v.val[3] = d;                           \
53   row01.val[0] = vreinterpretq_u64_s32(rows.val[0]);  in Transpose4x4()
54   row01.val[1] = vreinterpretq_u64_s32(rows.val[1]);  in Transpose4x4()
55   row23.val[0] = vreinterpretq_u64_s32(rows.val[2]);  in Transpose4x4()
56   row23.val[1] = vreinterpretq_u64_s32(rows.val[3]);  in Transpose4x4()
59     const uint64x1_t row0h = vget_high_u64(row01.val[0]);  in Transpose4x4()
60     const uint64x1_t row2l = vget_low_u64(row23.val[0]);  in Transpose4x4()
61     const uint64x1_t row1h = vget_high_u64(row01.val[1]);  in Transpose4x4()
62     const uint64x1_t row3l = vget_low_u64(row23.val[1]);  in Transpose4x4()
63     row01.val[0] = vcombine_u64(vget_low_u64(row01.val[0]), row2l);  in Transpose4x4()
64     row23.val[0] = vcombine_u64(row0h, vget_high_u64(row23.val[0]));  in Transpose4x4()
65     row01.val[1] = vcombine_u64(vget_low_u64(row01.val[1]), row3l);  in Transpose4x4()
66     row23.val[1] = vcombine_u64(row1h, vget_high_u64(row23.val[1]));  in Transpose4x4()
69     const int32x4x2_t out01 = vtrnq_s32(vreinterpretq_s32_u64(row01.val[0]),  in Transpose4x4()
70                                         vreinterpretq_s32_u64(row01.val[1]));  in Transpose4x4()
71     const int32x4x2_t out23 = vtrnq_s32(vreinterpretq_s32_u64(row23.val[0]),  in Transpose4x4()
72                                         vreinterpretq_s32_u64(row23.val[1]));  in Transpose4x4()
74     out.val[0] = out01.val[0];  in Transpose4x4()
75     out.val[1] = out01.val[1];  in Transpose4x4()
76     out.val[2] = out23.val[0];  in Transpose4x4()
77     out.val[3] = out23.val[1];  in Transpose4x4()