303     vacc0x0123 = vsraq_n_s32(vproduct0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
304     vacc0x4567 = vsraq_n_s32(vproduct0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
305     vacc0x89AB = vsraq_n_s32(vproduct0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
306     vacc0xCDEF = vsraq_n_s32(vproduct0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
307     vacc1x0123 = vsraq_n_s32(vproduct1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
308     vacc1x4567 = vsraq_n_s32(vproduct1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
309     vacc1x89AB = vsraq_n_s32(vproduct1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
310     vacc1xCDEF = vsraq_n_s32(vproduct1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
311     vacc2x0123 = vsraq_n_s32(vproduct2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
312     vacc2x4567 = vsraq_n_s32(vproduct2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
313     vacc2x89AB = vsraq_n_s32(vproduct2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
314     vacc2xCDEF = vsraq_n_s32(vproduct2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
315     vacc3x0123 = vsraq_n_s32(vproduct3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
316     vacc3x4567 = vsraq_n_s32(vproduct3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
317     vacc3x89AB = vsraq_n_s32(vproduct3x89AB, vbicq_s32(vacc3x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
318     vacc3xCDEF = vsraq_n_s32(vproduct3xCDEF, vbicq_s32(vacc3xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
319     vacc4x0123 = vsraq_n_s32(vproduct4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
320     vacc4x4567 = vsraq_n_s32(vproduct4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
321     vacc4x89AB = vsraq_n_s32(vproduct4x89AB, vbicq_s32(vacc4x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
322     vacc4xCDEF = vsraq_n_s32(vproduct4xCDEF, vbicq_s32(vacc4xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
323     vacc5x0123 = vsraq_n_s32(vproduct5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
324     vacc5x4567 = vsraq_n_s32(vproduct5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
325     vacc5x89AB = vsraq_n_s32(vproduct5x89AB, vbicq_s32(vacc5x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
326     vacc5xCDEF = vsraq_n_s32(vproduct5xCDEF, vbicq_s32(vacc5xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
327     vacc6x0123 = vsraq_n_s32(vproduct6x0123, vbicq_s32(vacc6x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
328     vacc6x4567 = vsraq_n_s32(vproduct6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
329     vacc6x89AB = vsraq_n_s32(vproduct6x89AB, vbicq_s32(vacc6x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
330     vacc6xCDEF = vsraq_n_s32(vproduct6xCDEF, vbicq_s32(vacc6xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
331     vacc7x0123 = vsraq_n_s32(vproduct7x0123, vbicq_s32(vacc7x0123, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
332     vacc7x4567 = vsraq_n_s32(vproduct7x4567, vbicq_s32(vacc7x4567, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
333     vacc7x89AB = vsraq_n_s32(vproduct7x89AB, vbicq_s32(vacc7x89AB, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
334     vacc7xCDEF = vsraq_n_s32(vproduct7xCDEF, vbicq_s32(vacc7xCDEF, vzero_shift_mask), 31);  in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()