99       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
100       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
101       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
102       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
103       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
104       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
105       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
106       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
107       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
108       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
109       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
110       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__neon()
122       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
123       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
124       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
125       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
126       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
127       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
128       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
129       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
130       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
131       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
132       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
133       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__neon()
145       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
146       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
147       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
148       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
149       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
150       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
151       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
152       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
153       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
154       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
155       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
156       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__neon()
168       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
169       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
170       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
171       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
172       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
173       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
174       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
175       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
176       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
177       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
178       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
179       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__neon()
197         vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
198         vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
199         vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
200         vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
201         vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
202         vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__neon()
203         vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()
204         vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()
205         vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()
206         vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()
207         vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()
208         vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__neon()