176       int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()  local
195       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
213       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
231       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
249       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
267       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
285       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
303       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
321       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
339       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
357       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi9xOPQRSTUV), vget_high_s16(vk9xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
375       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi10xOPQRSTUV), vget_high_s16(vk10xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
393       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi11xOPQRSTUV), vget_high_s16(vk11xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
411       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
429       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
447       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi14xOPQRSTUV), vget_high_s16(vk14xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
465       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
483       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
501       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi17xOPQRSTUV), vget_high_s16(vk17xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
519       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi18xOPQRSTUV), vget_high_s16(vk18xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
537       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi19xOPQRSTUV), vget_high_s16(vk19xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
555       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi20xOPQRSTUV), vget_high_s16(vk20xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
573       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi21xOPQRSTUV), vget_high_s16(vk21xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
591       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi22xOPQRSTUV), vget_high_s16(vk22xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
609       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi23xOPQRSTUV), vget_high_s16(vk23xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
627       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi24xOPQRSTUV), vget_high_s16(vk24xOPQRSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
636       float32x4_t vfpaccSTUV = vcvtq_f32_s32(vaccSTUV);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
654       vaccSTUV = vcvtnq_s32_f32(vfpaccSTUV);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
660       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
673       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));  in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()