133     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
134     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
135     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
136     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
140     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
141     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
142     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
143     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
147     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
148     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
149     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
150     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
154     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
155     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
156     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
157     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));  in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()