39   const __m256 vtwo = _mm256_set1_ps(2.0f);  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()  local
195     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
196     vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
197     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
198     vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
199     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
200     vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
201     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
202     vr3 = _mm256_mul_ps(vr3, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr3, vd3)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
203     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
204     vr4 = _mm256_mul_ps(vr4, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr4, vd4)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
205     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
206     vr5 = _mm256_mul_ps(vr5, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr5, vd5)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
207     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
208     vr6 = _mm256_mul_ps(vr6, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr6, vd6)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
209     vr7 = _mm256_mul_ps(vr7, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr7, vd7)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
210     vr7 = _mm256_mul_ps(vr7, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr7, vd7)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
276     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
277     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
315     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()
316     vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));  in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64()