void LAPACKE_sge_trans(int matrix_layout, lapack_int m, lapack_int n, const float *in, lapack_int ldin, float *out, lapack_int ldout)
void LAPACKE_sgg_trans(int matrix_layout, lapack_int m, lapack_int n, const float *in, lapack_int ldin, float *out, lapack_int ldout)