577     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
578     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
584     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
585     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
591     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
592     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
598     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
599     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
605     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
606     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
612     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
613     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
619     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
620     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
626     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
627     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
633     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
634     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
640     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
641     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
647     const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
648     const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
654     const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
655     const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
661     const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
662     const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
668     const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
669     const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
675     const int32x2_t vsum3x89 = vpadd_s32(vpsum3x8, vpsum3x9);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
676     const int32x2_t vsum3xAB = vpadd_s32(vpsum3xA, vpsum3xB);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
682     const int32x2_t vsum3xCD = vpadd_s32(vpsum3xC, vpsum3xD);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
683     const int32x2_t vsum3xEF = vpadd_s32(vpsum3xE, vpsum3xF);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()