56     int32x4_t vacc1xCDEF = vacc0xCDEF;  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()  local
78       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
92       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
106       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
120       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
135       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
149       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
163       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
177       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
199       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
214 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
229 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
244 …      vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
259 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
274 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
289 …     vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
306     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
317     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
326     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
333 …const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
341 …x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_…  in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()