66     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
67     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
68     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
69     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
86     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
87     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
88     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
89     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
91     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
92     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
93     vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
94     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
97     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
98     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
99     float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
100     float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
102     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
103     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
104     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
105     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
107     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
108     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
109     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
110     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
112     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
113     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
114     vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
115     vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
126     float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
127     float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
128     float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
129     float32x4_t veCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
202     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
213     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
214     vt = vmlaq_f32(vt, vn, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
217     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
218     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
219     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
220     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
227     float32x4_t ve = vmlaq_f32(vs, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
272     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
283     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
284     vt = vmlaq_f32(vt, vn, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
287     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
288     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
289     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
290     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
297     float32x4_t ve = vmlaq_f32(vs, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()