39   const __m256 vtwo = _mm256_set1_ps(2.0f);  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()  local
212     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
213     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
214     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
215     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
216     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
217     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
218     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
219     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
220     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
221     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
222     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
223     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
224     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
225     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
226     vr7 = _mm256_mul_ps(vr7, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr7, vd7)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
227     vr7 = _mm256_mul_ps(vr7, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr7, vd7)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
228     vr8 = _mm256_mul_ps(vr8, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr8, vd8)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
229     vr8 = _mm256_mul_ps(vr8, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr8, vd8)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
299     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
300     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
338     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()
339     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72()