67   MaskHelper::extractShuffle(mask, b01, b23, b45, b67);  in mm256ShufflePs()
95   MaskHelper::extractPermute(mask, b03, b47);  in mm256Permute2f128Ps()
128   Value s0 = mm256ShufflePs(ib, t0, t2, MaskHelper::shuffle<1, 0, 1, 0>());  in transpose4x8xf32()
129   Value s1 = mm256ShufflePs(ib, t0, t2, MaskHelper::shuffle<3, 2, 3, 2>());  in transpose4x8xf32()
130   Value s2 = mm256ShufflePs(ib, t1, t3, MaskHelper::shuffle<1, 0, 1, 0>());  in transpose4x8xf32()
131   Value s3 = mm256ShufflePs(ib, t1, t3, MaskHelper::shuffle<3, 2, 3, 2>());  in transpose4x8xf32()
132   vs[0] = mm256Permute2f128Ps(ib, s0, s1, MaskHelper::permute<2, 0>());  in transpose4x8xf32()
133   vs[1] = mm256Permute2f128Ps(ib, s2, s3, MaskHelper::permute<2, 0>());  in transpose4x8xf32()
134   vs[2] = mm256Permute2f128Ps(ib, s0, s1, MaskHelper::permute<3, 1>());  in transpose4x8xf32()
135   vs[3] = mm256Permute2f128Ps(ib, s2, s3, MaskHelper::permute<3, 1>());  in transpose4x8xf32()
158   Value sh0 = mm256ShufflePs(ib, t0, t2, MaskHelper::shuffle<1, 0, 3, 2>());  in transpose8x8xf32()
159   Value sh2 = mm256ShufflePs(ib, t1, t3, MaskHelper::shuffle<1, 0, 3, 2>());  in transpose8x8xf32()
160   Value sh4 = mm256ShufflePs(ib, t4, t6, MaskHelper::shuffle<1, 0, 3, 2>());  in transpose8x8xf32()
161   Value sh6 = mm256ShufflePs(ib, t5, t7, MaskHelper::shuffle<1, 0, 3, 2>());  in transpose8x8xf32()
164       mm256BlendPsAsm(ib, t0, sh0, MaskHelper::blend<0, 0, 1, 1, 0, 0, 1, 1>());  in transpose8x8xf32()
166       mm256BlendPsAsm(ib, t2, sh0, MaskHelper::blend<1, 1, 0, 0, 1, 1, 0, 0>());  in transpose8x8xf32()
168       mm256BlendPsAsm(ib, t1, sh2, MaskHelper::blend<0, 0, 1, 1, 0, 0, 1, 1>());  in transpose8x8xf32()
170       mm256BlendPsAsm(ib, t3, sh2, MaskHelper::blend<1, 1, 0, 0, 1, 1, 0, 0>());  in transpose8x8xf32()
172       mm256BlendPsAsm(ib, t4, sh4, MaskHelper::blend<0, 0, 1, 1, 0, 0, 1, 1>());  in transpose8x8xf32()
174       mm256BlendPsAsm(ib, t6, sh4, MaskHelper::blend<1, 1, 0, 0, 1, 1, 0, 0>());  in transpose8x8xf32()
176       mm256BlendPsAsm(ib, t5, sh6, MaskHelper::blend<0, 0, 1, 1, 0, 0, 1, 1>());  in transpose8x8xf32()
178       mm256BlendPsAsm(ib, t7, sh6, MaskHelper::blend<1, 1, 0, 0, 1, 1, 0, 0>());  in transpose8x8xf32()
180   vs[0] = mm256Permute2f128Ps(ib, s0, s4, MaskHelper::permute<2, 0>());  in transpose8x8xf32()
181   vs[1] = mm256Permute2f128Ps(ib, s1, s5, MaskHelper::permute<2, 0>());  in transpose8x8xf32()
182   vs[2] = mm256Permute2f128Ps(ib, s2, s6, MaskHelper::permute<2, 0>());  in transpose8x8xf32()
183   vs[3] = mm256Permute2f128Ps(ib, s3, s7, MaskHelper::permute<2, 0>());  in transpose8x8xf32()
184   vs[4] = mm256Permute2f128Ps(ib, s0, s4, MaskHelper::permute<3, 1>());  in transpose8x8xf32()
185   vs[5] = mm256Permute2f128Ps(ib, s1, s5, MaskHelper::permute<3, 1>());  in transpose8x8xf32()
186   vs[6] = mm256Permute2f128Ps(ib, s2, s6, MaskHelper::permute<3, 1>());  in transpose8x8xf32()
187   vs[7] = mm256Permute2f128Ps(ib, s3, s7, MaskHelper::permute<3, 1>());  in transpose8x8xf32()