99       const float32x4_t va0c0 = vdupq_lane_f32(vget_low_f32(va0), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
100       const float32x4_t va1c0 = vdupq_lane_f32(vget_low_f32(va1), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
101       const float32x4_t va2c0 = vdupq_lane_f32(vget_low_f32(va2), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
102       const float32x4_t va3c0 = vdupq_lane_f32(vget_low_f32(va3), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
103       const float32x4_t va4c0 = vdupq_lane_f32(vget_low_f32(va4), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
104       const float32x4_t va5c0 = vdupq_lane_f32(vget_low_f32(va5), 0);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
121       const float32x4_t va0c1 = vdupq_lane_f32(vget_low_f32(va0), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
122       const float32x4_t va1c1 = vdupq_lane_f32(vget_low_f32(va1), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
123       const float32x4_t va2c1 = vdupq_lane_f32(vget_low_f32(va2), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
124       const float32x4_t va3c1 = vdupq_lane_f32(vget_low_f32(va3), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
125       const float32x4_t va4c1 = vdupq_lane_f32(vget_low_f32(va4), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
126       const float32x4_t va5c1 = vdupq_lane_f32(vget_low_f32(va5), 1);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
285       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
286       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
287       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
288       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
289       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()
290       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);  in xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128()