50           const float32x2_t vw = vld1_f32(w); w += 2;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
107         float32x2_t vacc01n0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
108         float32x2_t vacc01n1 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
112             const float32x2_t vi01 = vld1_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
114             const float32x2_t vw = vld1_f32(w); w += 2;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
120         float32x2_t vout01n0 = vmin_f32(vacc01n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
121         float32x2_t vout01n1 = vmin_f32(vacc01n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
137           float32x2_t vacc01 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
141               const float32x2_t vi01 = vld1_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
143               const float32x2_t vw = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
147           float32x2_t vout01 = vmin_f32(vacc01, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
166         float32x2_t vacc0n0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
167         float32x2_t vacc0n1 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
171             const float32x2_t vi0 = vld1_dup_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
173             const float32x2_t vw = vld1_f32(w); w += 2;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
179         float32x2_t vout0n0 = vmin_f32(vacc0n0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
180         float32x2_t vout0n1 = vmin_f32(vacc0n1, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
196           float32x2_t vacc0 = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
200               const float32x2_t vi0 = vld1_dup_f32(input);  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
202               const float32x2_t vw = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
206           float32x2_t vout0 = vmin_f32(vacc0, vget_low_f32(vmax));  in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()