#ifndef ASSEMBLER
blasint sgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xgetf2_k(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint sgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xgetrf_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint sgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xgetrf_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
int slaswp_plus (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, blasint *, BLASLONG);
int slaswp_minus(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, blasint *, BLASLONG);
int dlaswp_plus (BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, blasint *, BLASLONG);
int dlaswp_minus(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG, double *, BLASLONG, blasint *, BLASLONG);
int qlaswp_plus (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, blasint *, BLASLONG);
int qlaswp_minus(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, blasint *, BLASLONG);
int claswp_plus (BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, blasint *, BLASLONG);
int claswp_minus(BLASLONG, BLASLONG, BLASLONG, float, float, float *, BLASLONG, float *, BLASLONG, blasint *, BLASLONG);
int zlaswp_plus (BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, blasint *, BLASLONG);
int zlaswp_minus(BLASLONG, BLASLONG, BLASLONG, double, double, double *, BLASLONG, double *, BLASLONG, blasint *, BLASLONG);
int xlaswp_plus (BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, blasint *, BLASLONG);
int xlaswp_minus(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, xdouble *, BLASLONG, xdouble *, BLASLONG, blasint *, BLASLONG);
int slaswp_ncopy(BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
int dlaswp_ncopy(BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
int qlaswp_ncopy(BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
int claswp_ncopy(BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG, blasint *, float *);
int zlaswp_ncopy(BLASLONG, BLASLONG, BLASLONG, double *, BLASLONG, blasint *, double *);
int xlaswp_ncopy(BLASLONG, BLASLONG, BLASLONG, xdouble *, BLASLONG, blasint *, xdouble *);
blasint sgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint sgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_R_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_C_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_R_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_C_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xgetrs_N_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_T_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_R_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_C_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint sgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint sgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_R_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cgetrs_C_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_R_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zgetrs_C_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xgetrs_N_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_T_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_R_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xgetrs_C_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint spotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint spotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dpotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dpotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qpotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qpotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cpotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cpotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zpotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zpotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xpotf2_U(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xpotf2_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint spotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint spotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dpotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dpotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qpotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qpotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cpotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cpotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zpotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zpotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xpotrf_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xpotrf_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint spotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint spotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dpotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dpotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qpotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qpotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint cpotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint cpotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zpotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zpotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xpotrf_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xpotrf_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint slauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint slauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dlauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dlauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qlauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qlauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint clauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint clauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zlauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zlauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xlauu2_U(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xlauu2_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint slauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint slauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dlauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dlauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qlauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qlauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint clauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint clauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zlauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zlauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xlauum_U_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xlauum_L_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint slauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint slauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dlauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dlauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qlauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qlauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint clauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint clauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zlauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zlauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xlauum_U_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xlauum_L_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint strti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dtrti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qtrti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint ctrti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ztrti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xtrti2_UU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrti2_UN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrti2_LU(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrti2_LN(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint strtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dtrtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qtrtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint ctrtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ztrtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xtrtri_UU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_UN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_LU_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_LN_single(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint strtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint strtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dtrtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dtrtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qtrtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qtrtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint ctrtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ctrtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint ztrtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint ztrtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xtrtri_UU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_UN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_LU_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xtrtri_LN_parallel(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
int sneg_tcopy(BLASLONG, BLASLONG, float *, BLASLONG, float *);
int dneg_tcopy(BLASLONG, BLASLONG, double *, BLASLONG, double *);
int qneg_tcopy(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
int cneg_tcopy(BLASLONG, BLASLONG, float *, BLASLONG, float *);
int zneg_tcopy(BLASLONG, BLASLONG, double *, BLASLONG, double *);
int xneg_tcopy(BLASLONG, BLASLONG, xdouble *, BLASLONG, xdouble *);
blasint slarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint slarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint dlarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint dlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint qlarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint qlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint clarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint clarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, float *, float *, BLASLONG);
blasint zlarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint zlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, double *, double *, BLASLONG);
blasint xlarf_L(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
blasint xlarf_R(blas_arg_t *, BLASLONG *, BLASLONG *, xdouble *, xdouble *, BLASLONG);
#endif