211     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
212     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
213     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
214     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
218     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
219     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
220     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
221     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
225     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
226     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
227     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
228     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
232     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
233     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
234     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
235     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
239     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
240     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
241     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
242     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
246     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
247     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
248     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
249     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));  in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()