76       psimd_f32 va2 = psimd_load_f32(a2);  in xnn_f32_gemm_ukernel_4x8s4__psimd()  local
87       vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c0);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
91       vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c0);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
97       va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
102       va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_4x8s4__psimd()
111       vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c1);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
115       vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c1);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
121       va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
126       va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_4x8s4__psimd()
135       vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c2);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
139       vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c2);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
145       va2 = __builtin_shufflevector(va2, va2, 1, 2, 3, 0);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
150       va2 = __builtin_shuffle(va2, va2, (psimd_s32) { 1, 2, 3, 0 });  in xnn_f32_gemm_ukernel_4x8s4__psimd()
159       vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123c3);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
163       vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567c3);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
176         const psimd_f32 va2 = psimd_load_splat_f32(a2);  in xnn_f32_gemm_ukernel_4x8s4__psimd()  local
187         vacc2x0123 = psimd_qfma_f32(vacc2x0123, va2, vb0123);  in xnn_f32_gemm_ukernel_4x8s4__psimd()
191         vacc2x4567 = psimd_qfma_f32(vacc2x4567, va2, vb4567);  in xnn_f32_gemm_ukernel_4x8s4__psimd()