74     int32x4_t vacc3xCDEF = vacc0xCDEF;  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()  local
128 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
150 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
172 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
194 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
218 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
240 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
262 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
284 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
318 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
341 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
364 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
387 …      vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
410 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
433 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
456 …     vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
483     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
519     vacc3xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc3xCDEF, vmagic_bias));  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
537     vacc3xCDEF = vqsubq_s32(vacc3xCDEF, vmagic_bias_less_output_zero_point);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
547     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()
562     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));  in xnn_qu8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane()