33   const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()  local
124       float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
125       float32x4_t vout4567n0 = vminq_f32(vacc4567n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
126       float32x4_t vout89ABn0 = vminq_f32(vacc89ABn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
127       float32x4_t voutCDEFn0 = vminq_f32(vaccCDEFn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
128       float32x4_t voutGHIJn0 = vminq_f32(vaccGHIJn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
129       float32x4_t voutKLMNn0 = vminq_f32(vaccKLMNn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
130       float32x4_t voutOPQRn0 = vminq_f32(vaccOPQRn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
131       float32x4_t voutSTUVn0 = vminq_f32(vaccSTUVn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
132       float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
133       float32x4_t vout4567n1 = vminq_f32(vacc4567n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
134       float32x4_t vout89ABn1 = vminq_f32(vacc89ABn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
135       float32x4_t voutCDEFn1 = vminq_f32(vaccCDEFn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
136       float32x4_t voutGHIJn1 = vminq_f32(vaccGHIJn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
137       float32x4_t voutKLMNn1 = vminq_f32(vaccKLMNn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
138       float32x4_t voutOPQRn1 = vminq_f32(vaccOPQRn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
139       float32x4_t voutSTUVn1 = vminq_f32(vaccSTUVn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
140       float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
141       float32x4_t vout4567n2 = vminq_f32(vacc4567n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
142       float32x4_t vout89ABn2 = vminq_f32(vacc89ABn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
143       float32x4_t voutCDEFn2 = vminq_f32(vaccCDEFn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
144       float32x4_t voutGHIJn2 = vminq_f32(vaccGHIJn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
145       float32x4_t voutKLMNn2 = vminq_f32(vaccKLMNn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
146       float32x4_t voutOPQRn2 = vminq_f32(vaccOPQRn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
147       float32x4_t voutSTUVn2 = vminq_f32(vaccSTUVn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
148       float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
149       float32x4_t vout4567n3 = vminq_f32(vacc4567n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
150       float32x4_t vout89ABn3 = vminq_f32(vacc89ABn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
151       float32x4_t voutCDEFn3 = vminq_f32(vaccCDEFn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
152       float32x4_t voutGHIJn3 = vminq_f32(vaccGHIJn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
153       float32x4_t voutKLMNn3 = vminq_f32(vaccKLMNn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
154       float32x4_t voutOPQRn3 = vminq_f32(vaccOPQRn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
155       float32x4_t voutSTUVn3 = vminq_f32(vaccSTUVn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
267         float32x4_t vout0123 = vminq_f32(vacc0123, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
268         float32x4_t vout4567 = vminq_f32(vacc4567, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
269         float32x4_t vout89AB = vminq_f32(vacc89AB, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
270         float32x4_t voutCDEF = vminq_f32(vaccCDEF, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
271         float32x4_t voutGHIJ = vminq_f32(vaccGHIJ, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
272         float32x4_t voutKLMN = vminq_f32(vaccKLMN, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
273         float32x4_t voutOPQR = vminq_f32(vaccOPQR, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
274         float32x4_t voutSTUV = vminq_f32(vaccSTUV, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
354         float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
355         float32x4_t vout4567n0 = vminq_f32(vacc4567n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
356         float32x4_t vout89ABn0 = vminq_f32(vacc89ABn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
357         float32x4_t voutCDEFn0 = vminq_f32(vaccCDEFn0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
358         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
359         float32x4_t vout4567n1 = vminq_f32(vacc4567n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
360         float32x4_t vout89ABn1 = vminq_f32(vacc89ABn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
361         float32x4_t voutCDEFn1 = vminq_f32(vaccCDEFn1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
362         float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
363         float32x4_t vout4567n2 = vminq_f32(vacc4567n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
364         float32x4_t vout89ABn2 = vminq_f32(vacc89ABn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
365         float32x4_t voutCDEFn2 = vminq_f32(vaccCDEFn2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
366         float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
367         float32x4_t vout4567n3 = vminq_f32(vacc4567n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
368         float32x4_t vout89ABn3 = vminq_f32(vacc89ABn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
369         float32x4_t voutCDEFn3 = vminq_f32(vaccCDEFn3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
434           float32x4_t vout0123 = vminq_f32(vacc0123, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
435           float32x4_t vout4567 = vminq_f32(vacc4567, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
436           float32x4_t vout89AB = vminq_f32(vacc89AB, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
437           float32x4_t voutCDEF = vminq_f32(vaccCDEF, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
489         float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
490         float32x4_t vout4567n0 = vminq_f32(vacc4567n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
491         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
492         float32x4_t vout4567n1 = vminq_f32(vacc4567n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
493         float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
494         float32x4_t vout4567n2 = vminq_f32(vacc4567n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
495         float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
496         float32x4_t vout4567n3 = vminq_f32(vacc4567n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
539           float32x4_t vout0123 = vminq_f32(vacc0123, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
540           float32x4_t vout4567 = vminq_f32(vacc4567, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
579         float32x4_t vout0123n0 = vminq_f32(vacc0123n0, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
580         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
581         float32x4_t vout0123n2 = vminq_f32(vacc0123n2, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
582         float32x4_t vout0123n3 = vminq_f32(vacc0123n3, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
614           float32x4_t vout0123 = vminq_f32(vacc0123, vmax);  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
651         float32x2_t vout01n0 = vmin_f32(vacc01n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
652         float32x2_t vout01n1 = vmin_f32(vacc01n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
653         float32x2_t vout01n2 = vmin_f32(vacc01n2, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
654         float32x2_t vout01n3 = vmin_f32(vacc01n3, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
686           float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
722         float32x2_t vout0n0 = vmin_f32(vacc0n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
723         float32x2_t vout0n1 = vmin_f32(vacc0n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
724         float32x2_t vout0n2 = vmin_f32(vacc0n2, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
725         float32x2_t vout0n3 = vmin_f32(vacc0n3, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
757           float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()