98       psimd_f32 va5 = psimd_load_f32(a5);  in xnn_f32_gemm_ukernel_6x8s4__psimd()  local
110       vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123c0);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
116       vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567c0);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
124       va5 = __builtin_shufflevector(va5, va5, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
131       va5 = __builtin_shuffle(va5, va5, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_6x8s4__psimd()
142       vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123c1);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
148       vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567c1);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
156       va5 = __builtin_shufflevector(va5, va5, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
163       va5 = __builtin_shuffle(va5, va5, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_6x8s4__psimd()
174       vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123c2);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
180       vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567c2);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
188       va5 = __builtin_shufflevector(va5, va5, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
195       va5 = __builtin_shuffle(va5, va5, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_6x8s4__psimd()
206       vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123c3);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
212       vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567c3);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
230         const psimd_f32 va5 = psimd_load_splat_f32(a5);  in xnn_f32_gemm_ukernel_6x8s4__psimd()  local
242         vacc5x0123 = psimd_qfma_f32(vacc5x0123, va5, vb0123);  in xnn_f32_gemm_ukernel_6x8s4__psimd()
248         vacc5x4567 = psimd_qfma_f32(vacc5x4567, va5, vb4567);  in xnn_f32_gemm_ukernel_6x8s4__psimd()