120         psimd_f32 va2 = psimd_load_f32(a2);  in xnn_f32_igemm_ukernel_6x8s4__psimd()  local
135         vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c0);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
141         vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c0);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
149         va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
156         va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_igemm_ukernel_6x8s4__psimd()
167         vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c1);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
173         vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c1);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
181         va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
188         va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_igemm_ukernel_6x8s4__psimd()
199         vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c2);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
205         vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c2);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
213         va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
220         va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_igemm_ukernel_6x8s4__psimd()
231         vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c3);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
237         vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c3);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
256           const psimd_f32 va2 = psimd_load_splat_f32(a2);  in xnn_f32_igemm_ukernel_6x8s4__psimd()  local
269           vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);  in xnn_f32_igemm_ukernel_6x8s4__psimd()
270           vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);  in xnn_f32_igemm_ukernel_6x8s4__psimd()