262     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
263     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
269     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
270     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
276     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
277     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
283     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
284     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
290     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
291     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
297     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
298     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
304     const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
305     const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
311     const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
312     const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);  in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()