103         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
104         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
105         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
106         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
107         vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
108         vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
109         vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
110         vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
115         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
116         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
117         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
118         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
119         vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
120         vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
121         vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
122         vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
222       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
223       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
224       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()
225       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);  in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128()