50       float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()  local
51       float32x4_t vacc4567n1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
52       float32x4_t vacc89ABn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
53       float32x4_t vaccCDEFn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
54       float32x4_t vaccGHIJn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
55       float32x4_t vaccKLMNn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
56       float32x4_t vaccOPQRn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
57       float32x4_t vaccSTUVn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
82           vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
100       float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
232         float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()  local
233         float32x4_t vacc4567n1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
234         float32x4_t vacc89ABn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
235         float32x4_t vaccCDEFn1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
250             vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
260         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
341         float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()  local
342         float32x4_t vacc4567n1 = vacc0123n1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
353             vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
359         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
417         float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()  local
426             vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
430         float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax);  in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()