81       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
82       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
83       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
84       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
85       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
86       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
87       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
88       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
93       vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
94       vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
95       vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
96       vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
97       vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
98       vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
99       vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
100       vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
201       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
202       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
203       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
204       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);  in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()