119         float32x2_t vacc01n0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
120         float32x2_t vacc01n1 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
121         float32x2_t vacc01n2 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
122         float32x2_t vacc01n3 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
126             const float32x2_t vi01 = vld1_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
136         float32x2_t vout01n0 = vmin_f32(vacc01n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
137         float32x2_t vout01n1 = vmin_f32(vacc01n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
138         float32x2_t vout01n2 = vmin_f32(vacc01n2, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
139         float32x2_t vout01n3 = vmin_f32(vacc01n3, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
161           float32x2_t vacc01 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
165               const float32x2_t vi01 = vld1_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
167               const float32x2_t vw = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
171           float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
190         float32x2_t vacc0n0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
191         float32x2_t vacc0n1 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
192         float32x2_t vacc0n2 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
193         float32x2_t vacc0n3 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
197             const float32x2_t vi0 = vld1_dup_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
207         float32x2_t vout0n0 = vmin_f32(vacc0n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
208         float32x2_t vout0n1 = vmin_f32(vacc0n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
209         float32x2_t vout0n2 = vmin_f32(vacc0n2, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
210         float32x2_t vout0n3 = vmin_f32(vacc0n3, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
232           float32x2_t vacc0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
236               const float32x2_t vi0 = vld1_dup_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
238               const float32x2_t vw = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
242           float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()