285     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
286     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
287     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
288     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
292     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
293     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
294     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
295     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
299     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
300     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
301     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
302     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
306     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
307     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
308     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
309     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
313     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
314     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
315     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
316     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
320     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
321     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
322     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
323     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
327     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
328     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
329     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
330     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
334     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
335     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
336     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
337     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
341     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
342     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
343     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
344     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
348     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
349     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
350     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
351     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
355     const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
356     const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
357     const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
358     const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
362     const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
363     const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
364     const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
365     const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15));  in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()