56           const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
58           vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
59           vacc4567c0 = vfmaq_lane_f32(vacc4567c0, va4567, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
60           vacc89ABc0 = vfmaq_lane_f32(vacc89ABc0, va89AB, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
61           vaccCDEFc0 = vfmaq_lane_f32(vaccCDEFc0, vaCDEF, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
62           vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
63           vacc4567c1 = vfmaq_lane_f32(vacc4567c1, va4567, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
64           vacc89ABc1 = vfmaq_lane_f32(vacc89ABc1, va89AB, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
65           vaccCDEFc1 = vfmaq_lane_f32(vaccCDEFc1, vaCDEF, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
114             const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
115             vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
116             vacc4567 = vfmaq_f32(vacc4567, va4567, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
117             vacc89AB = vfmaq_f32(vacc89AB, va89AB, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
118             vaccCDEF = vfmaq_f32(vaccCDEF, vaCDEF, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
162             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
164             vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
165             vacc4567c0 = vfmaq_lane_f32(vacc4567c0, va4567, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
166             vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
167             vacc4567c1 = vfmaq_lane_f32(vacc4567c1, va4567, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
200               const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
201               vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
202               vacc4567 = vfmaq_f32(vacc4567, va4567, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
235             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
237             vacc0123c0 = vfmaq_lane_f32(vacc0123c0, va0123, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
238             vacc0123c1 = vfmaq_lane_f32(vacc0123c1, va0123, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
263               const float32x4_t vb = vld1q_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
264               vacc0123 = vfmaq_f32(vacc0123, va0123, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
294             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
296             vacc01c0 = vfma_lane_f32(vacc01c0, va01, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
297             vacc01c1 = vfma_lane_f32(vacc01c1, va01, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
322               const float32x2_t vb = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
323               vacc01 = vfma_f32(vacc01, va01, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()
352             const float32x2_t vb = vld1_f32(w); w += 2;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
354             vacc0c0 = vfma_lane_f32(vacc0c0, va0, vb, 0);  in xnn_f32_spmm_ukernel_16x2__neonfma()
355             vacc0c1 = vfma_lane_f32(vacc0c1, va0, vb, 1);  in xnn_f32_spmm_ukernel_16x2__neonfma()
380               const float32x2_t vb = vld1_dup_f32(w); w += 1;  in xnn_f32_spmm_ukernel_16x2__neonfma()  local
381               vacc0 = vfma_f32(vacc0, va0, vb);  in xnn_f32_spmm_ukernel_16x2__neonfma()