127         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, vget_low_f32(va0), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
128         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, vget_low_f32(va1), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
129         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, vget_low_f32(va2), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
130         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, vget_low_f32(va3), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
131         vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, vget_low_f32(va4), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
132         vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c0, vget_low_f32(va5), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
133         vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, vget_low_f32(va0), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
134         vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, vget_low_f32(va1), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
135         vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, vget_low_f32(va2), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
136         vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, vget_low_f32(va3), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
137         vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, vget_low_f32(va4), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
138         vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c0, vget_low_f32(va5), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
143         vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, vget_low_f32(va0), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
144         vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, vget_low_f32(va1), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
145         vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, vget_low_f32(va2), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
146         vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, vget_low_f32(va3), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
147         vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, vget_low_f32(va4), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
148         vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c1, vget_low_f32(va5), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
149         vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c1, vget_low_f32(va0), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
150         vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c1, vget_low_f32(va1), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
151         vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c1, vget_low_f32(va2), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
152         vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c1, vget_low_f32(va3), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
153         vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c1, vget_low_f32(va4), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
154         vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c1, vget_low_f32(va5), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
286       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
287       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
288       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
289       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
290       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()
291       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128()