119       const uint8x8_t vi0 = vld1_u8(i0); i0 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
120       const uint8x8_t vi1 = vld1_u8(i1); i1 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
121       const uint8x8_t vi2 = vld1_u8(i2); i2 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
122       const uint8x8_t vi3 = vld1_u8(i3); i3 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
123       const uint8x8_t vi4 = vld1_u8(i4); i4 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
124       const uint8x8_t vi5 = vld1_u8(i5); i5 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
125       const uint8x8_t vi6 = vld1_u8(i6); i6 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
126       const uint8x8_t vi7 = vld1_u8(i7); i7 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
127       const uint8x8_t vi8 = vld1_u8(i8); i8 += 8;  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
192       const uint8x8_t vi0 = vld1_u8(i0);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
193       const uint8x8_t vi1 = vld1_u8(i1);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
194       const uint8x8_t vi2 = vld1_u8(i2);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
195       const uint8x8_t vi3 = vld1_u8(i3);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
196       const uint8x8_t vi4 = vld1_u8(i4);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
197       const uint8x8_t vi5 = vld1_u8(i5);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
198       const uint8x8_t vi6 = vld1_u8(i6);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
199       const uint8x8_t vi7 = vld1_u8(i7);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()
200       const uint8x8_t vi8 = vld1_u8(i8);  in xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8()