64     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vx0123, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
65     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vx4567, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
66     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vx89AB, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
67     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vxCDEF, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
84     float32x4_t vt0123 = vmlaq_f32(vx0123, vn0123, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
85     float32x4_t vt4567 = vmlaq_f32(vx4567, vn4567, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
86     float32x4_t vt89AB = vmlaq_f32(vx89AB, vn89AB, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
87     float32x4_t vtCDEF = vmlaq_f32(vxCDEF, vnCDEF, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
89     vt0123 = vmlaq_f32(vt0123, vn0123, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
90     vt4567 = vmlaq_f32(vt4567, vn4567, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
91     vt89AB = vmlaq_f32(vt89AB, vn89AB, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
92     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
95     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
96     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
97     float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
98     float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
100     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
101     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
102     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
103     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
105     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
106     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
107     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
108     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
110     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
111     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
112     vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
113     vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
124     float32x4_t vf0123 = vmlaq_f32(vs0123, vp0123, vt0123);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
125     float32x4_t vf4567 = vmlaq_f32(vs4567, vp4567, vt4567);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
126     float32x4_t vf89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
127     float32x4_t vfCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
165     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
176     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
177     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
180     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
181     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
182     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
183     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
190     float32x4_t vf = vmlaq_f32(vs, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
222     float32x4_t vn = vmlaq_f32(vmagic_bias, vx, vlog2e);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
233     float32x4_t vt = vmlaq_f32(vx, vn, vminus_ln2_hi);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
234     vt = vmlaq_f32(vt, vn, vminus_ln2_lo);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
237     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
238     vp = vmlaq_f32(vc3, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
239     vp = vmlaq_f32(vc2, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
240     vp = vmlaq_f32(vc1, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
247     float32x4_t vf = vmlaq_f32(vs, vp, vt);  in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()