45     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()  local
61       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
71       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
81       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
91       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
102       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
112       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
122       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
132       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
148       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
159         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
170 …       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
181 …       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
192 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
203 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
214 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
227     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
234     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
239     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
245 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
250 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()