127         const float32x4_t va0c0 = vdupq_lane_f32(vget_low_f32(va0), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
128         const float32x4_t va1c0 = vdupq_lane_f32(vget_low_f32(va1), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
129         const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
130         const float32x4_t va3c0 = vdupq_lane_f32(vget_low_f32(va3), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
131         const float32x4_t va4c0 = vdupq_lane_f32(vget_low_f32(va4), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
132         const float32x4_t va5c0 = vdupq_lane_f32(vget_low_f32(va5), 0);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
149         const float32x4_t va0c1 = vdupq_lane_f32(vget_low_f32(va0), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
150         const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
151         const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
152         const float32x4_t va3c1 = vdupq_lane_f32(vget_low_f32(va3), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
153         const float32x4_t va4c1 = vdupq_lane_f32(vget_low_f32(va4), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
154         const float32x4_t va5c1 = vdupq_lane_f32(vget_low_f32(va5), 1);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
310       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
311       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
312       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
313       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
314       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()
315       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);  in xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128()