46     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()  local
83       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
99       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
115       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
131       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
180       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
181       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
182       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
183       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
209       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
224         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
239           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
248     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
255     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
260     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
266 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), v…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
271 …const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0x…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()