39   const __m256 vtwo = _mm256_set1_ps(2.0f);  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()  local
144     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
145     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
146     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
147     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
148     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
149     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
150     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
151     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
152     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
153     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
207     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
208     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
246     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()
247     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40()