115         const int32x4_t vacc_hi = vaddw_s16(vbias, vreinterpret_s16_u16(vget_high_u16(vsum)));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()  local
118         vst1q_s32(b, vacc_hi); b += 4;  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
176         int32x4_t vacc_hi = vld1q_s32(b + 4);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()  local
188         vacc_hi = vaddw_s16(vacc_hi, vreinterpret_s16_u16(vget_high_u16(vsum)));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
191         vst1q_s32(b, vacc_hi); b += 4;  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
271         int32x4_t vacc_hi = vld1q_s32(b); b += 4;  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()  local
283         vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
286         const int32x4_t vneg_mask_hi = vreinterpretq_s32_u32(vcltq_s32(vacc_hi, vmovq_n_s32(0)));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
291         const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vget_low_s32(vmultiplier));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
292         const int64x2_t vproduct67 = vmull_high_s32(vacc_hi, vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
301         const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
302         const int64x2_t vproduct67 = vmull_s32(vget_high_s32(vacc_hi), vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
317 …  vacc_hi = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc45), vreinterpretq_s32_s64(vscaled_acc67));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
319 …const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_poin…  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
322         vacc_hi = vcombine_s32(vmovn_s64(vscaled_acc45), vmovn_s64(vscaled_acc67));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
324 …const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), voutput_…  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
345         int32x4_t vacc_hi = vld1q_s32(b);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()  local
357         vacc_hi = vaddw_s16(vacc_hi, vget_high_s16(vsum));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
360         const int32x4_t vneg_mask_hi = vreinterpretq_s32_u32(vcltq_s32(vacc_hi, vmovq_n_s32(0)));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
365         const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vget_low_s32(vmultiplier));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
366         const int64x2_t vproduct67 = vmull_high_s32(vacc_hi, vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
375         const int64x2_t vproduct45 = vmull_s32(vget_low_s32(vacc_hi), vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
376         const int64x2_t vproduct67 = vmull_s32(vget_high_s32(vacc_hi), vmultiplier);  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
391 …  vacc_hi = vuzp1q_s32(vreinterpretq_s32_s64(vscaled_acc45), vreinterpretq_s32_s64(vscaled_acc67));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
393 …const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), voutput_zero_poin…  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
396         vacc_hi = vcombine_s32(vmovn_s64(vscaled_acc45), vmovn_s64(vscaled_acc67));  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
398 …const int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), voutput_…  in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()