200     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
201     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
202     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
203     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
207     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
208     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
209     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
210     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
214     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
215     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
216     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
217     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
221     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
222     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
223     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
224     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
228     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
229     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
230     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
231     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
235     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
236     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
237     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
238     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
242     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
243     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
244     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
245     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
249     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
250     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
251     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
252     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));  in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()