66     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
67     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
68     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
69     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
86     float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
87     float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
88     float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
89     float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
91     vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
92     vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
93     vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
94     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
97     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
98     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
99     float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
100     float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
102     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
103     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
104     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
105     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
107     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
108     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
109     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
110     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
112     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
113     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
114     vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
115     vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
126     float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
127     float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
128     float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
129     float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
169     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
180     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
181     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
184     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
185     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
186     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
187     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
194     float32x4_t vf = vmlaq_f32(vs, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
226     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
237     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
238     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
241     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
242     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
243     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
244     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
251     float32x4_t vf = vmlaq_f32(vs, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()