258     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
259     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
260     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
261     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
262     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
263     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
265     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
266     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
267     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
268     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
269     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
270     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
272     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
273     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
274     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
275     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
276     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
277     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
279     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
280     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
281     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
282     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
283     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
284     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
286     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
287     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
288     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
289     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
290     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
291     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
293     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
294     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
295     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
296     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
297     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
298     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
300     const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
301     const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
302     const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
303     const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
304     const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
305     const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
307     const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
308     const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
309     const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
310     const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
311     const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
312     const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);  in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()