39   const __m256 vtwo = _mm256_set1_ps(2.0f);  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()  local
178     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
179     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
180     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
181     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
182     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
183     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
184     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
185     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
186     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
187     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
188     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
189     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
190     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
191     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
253     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
254     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
292     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()
293     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56()