56     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()  local
60     int32x4_t vacc1x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
64     int32x4_t vacc2x89AB = vacc0x89AB;  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
104         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
122         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
140         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
158         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
177 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
195 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
213 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
231 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
255         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
274 …       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
293 …       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
312 …       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
331 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
350 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
369 …      vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
392     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
407     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
420     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
434 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
445 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()