92       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
93       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
94       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
95       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
96       const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
97       const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
98       const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
99       const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
146       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
147       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
148       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
149       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
236     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
237     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
238     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
239     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
250 …int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCD…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
251 …int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCD…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
252 …int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCD…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
253 …int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCD…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
255     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
256     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
289 …int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
290 …int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8…  in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()