150         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()  local
179         vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
192         vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
205         vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
218         vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
231         vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
244         vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
257         vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
270         vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
283         vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
296         vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
309         vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
322         vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
335         vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
348         vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
361         vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
374         vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);  in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()