49     int32x4_t vacc0_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa0)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
50     int32x4_t vacc1_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa1)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
51     int32x4_t vacc2_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa2)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
52     int32x4_t vacc3_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa3)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
58     vacc0_lo = vmlaq_s32(vacc0_lo, vmovl_s16(vget_low_s16(vxb0)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
59     vacc1_lo = vmlaq_s32(vacc1_lo, vmovl_s16(vget_low_s16(vxb1)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
60     vacc2_lo = vmlaq_s32(vacc2_lo, vmovl_s16(vget_low_s16(vxb2)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
61     vacc3_lo = vmlaq_s32(vacc3_lo, vmovl_s16(vget_low_s16(vxb3)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
115     int32x4_t vacc0_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa0)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
116     int32x4_t vacc1_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa1)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
117     int32x4_t vacc0_hi = vmulq_s32(vmovl_s16(vget_high_s16(vxa0)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
118     int32x4_t vacc1_hi = vmulq_s32(vmovl_s16(vget_high_s16(vxa1)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
123     vacc0_lo = vmlaq_s32(vacc0_lo, vmovl_s16(vget_low_s16(vxb0)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
124     vacc1_lo = vmlaq_s32(vacc1_lo, vmovl_s16(vget_low_s16(vxb1)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
125     vacc0_hi = vmlaq_s32(vacc0_hi, vmovl_s16(vget_high_s16(vxb0)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
126     vacc1_hi = vmlaq_s32(vacc1_hi, vmovl_s16(vget_high_s16(vxb1)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
159     int32x4_t vacc_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
163     int32x4_t vacc_hi = vmulq_s32(vmovl_s16(vget_high_s16(vxa)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
166     vacc_lo = vmlaq_s32(vacc_lo, vmovl_s16(vget_low_s16(vxb)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
170     vacc_hi = vmlaq_s32(vacc_hi, vmovl_s16(vget_high_s16(vxb)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
202     int32x4_t vacc_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
206     int32x4_t vacc_hi = vmulq_s32(vmovl_s16(vget_high_s16(vxa)), va_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
209     vacc_lo = vmlaq_s32(vacc_lo, vmovl_s16(vget_low_s16(vxb)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()
213     vacc_hi = vmlaq_s32(vacc_hi, vmovl_s16(vget_high_s16(vxb)), vb_multiplier);  in xnn_qu8_vadd_minmax_ukernel__neon()