52           const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
54           vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
55           vacc4567c0 = vfmaq_lane_f32(vacc4567c0, va4567, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
56           vacc89ABc0 = vfmaq_lane_f32(vacc89ABc0, va89AB, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
57           vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
58           vacc4567c1 = vfmaq_lane_f32(vacc4567c1, va4567, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
59           vacc89ABc1 = vfmaq_lane_f32(vacc89ABc1, va89AB, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
100             const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
101             vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
102             vacc4567 = vfmaq_f32(vacc4567, va4567, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
103             vacc89AB = vfmaq_f32(vacc89AB, va89AB, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
144             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
146             vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
147             vacc4567c0 = vfmaq_lane_f32(vacc4567c0, va4567, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
148             vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
149             vacc4567c1 = vfmaq_lane_f32(vacc4567c1, va4567, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
182               const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
183               vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
184               vacc4567 = vfmaq_f32(vacc4567, va4567, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
217             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
219             vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
220             vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
245               const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
246               vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
276             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
278             vacc01c0 = vfma_lane_f32(vacc01c0, va01, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
279             vacc01c1 = vfma_lane_f32(vacc01c1, va01, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
304               const float32x2_t vb = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
305               vacc01 = vfma_f32(vacc01, va01, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()
334             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
336             vacc0c0 = vfma_lane_f32(vacc0c0, va0, vb, 0);  in xnn_f32_spmm_ukernel_12x2__neonfma()
337             vacc0c1 = vfma_lane_f32(vacc0c1, va0, vb, 1);  in xnn_f32_spmm_ukernel_12x2__neonfma()
362               const float32x2_t vb = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_12x2__neonfma()  local
363               vacc0 = vfma_f32(vacc0, va0, vb);  in xnn_f32_spmm_ukernel_12x2__neonfma()