70     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
71     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
72     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
73     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
74     float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
75     float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
96     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
97     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
98     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
99     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
100     float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
101     float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
103     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
104     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
105     vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
106     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
107     vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
108     vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
111     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
112     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
113     float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
114     float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
115     float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
116     float32x4_t vpKLMN = vmlaq_f32(vc4, vc5, vtKLMN);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
118     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
119     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
120     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
121     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
122     vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
123     vpKLMN = vmlaq_f32(vc3, vpKLMN, vtKLMN);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
125     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
126     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
127     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
128     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
129     vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
130     vpKLMN = vmlaq_f32(vc2, vpKLMN, vtKLMN);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
132     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
133     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
134     vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
135     vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
136     vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
137     vpKLMN = vmlaq_f32(vc1, vpKLMN, vtKLMN);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
150     float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
151     float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
152     float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
153     float32x4_t veCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
154     float32x4_t veGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
155     float32x4_t veKLMN = vmlaq_f32(vsKLMN, vpKLMN, vtKLMN);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
246     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
257     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
258     vt = vmlaq_f32(vt, vn, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
261     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
262     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
263     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
264     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
271     float32x4_t ve = vmlaq_f32(vs, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
316     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
327     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
328     vt = vmlaq_f32(vt, vn, vln2_lo);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
331     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
332     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
333     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
334     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
341     float32x4_t ve = vmlaq_f32(vs, vp, vt);  in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()