71         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()  local
85         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
92         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
99         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
103         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
110         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
117         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
121         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
128         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
135         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
139         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
146         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
157         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()  local
169         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
173         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
177         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
179         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
183         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
187         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
189         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
193         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
197         va0x0 = vext_s8(va0x0, va0x0, 2);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
199         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()
203         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);  in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal()