void
CORE_cgetrf_reclap_init(void) {
int i;
for (i = 0; i < AMAX1BUF_SIZE; ++i) CORE_camax1buf[i] = -1.0;
sfmin = LAPACKE_slamch_work('S');
}
int check_factorization(int N, PLASMA_Complex32_t *A1, PLASMA_Complex32_t *A2, int LDA, int uplo)
{
float Anorm, Rnorm;
PLASMA_Complex32_t alpha;
int info_factorization;
int i,j;
float eps;
eps = LAPACKE_slamch_work('e');
PLASMA_Complex32_t *Residual = (PLASMA_Complex32_t *)malloc(N*N*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *L1 = (PLASMA_Complex32_t *)malloc(N*N*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *L2 = (PLASMA_Complex32_t *)malloc(N*N*sizeof(PLASMA_Complex32_t));
float *work = (float *)malloc(N*sizeof(float));
memset((void*)L1, 0, N*N*sizeof(PLASMA_Complex32_t));
memset((void*)L2, 0, N*N*sizeof(PLASMA_Complex32_t));
alpha= 1.0;
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,' ', N, N, A1, LDA, Residual, N);
/* Dealing with L'L or U'U */
if (uplo == PlasmaUpper){
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,'u', N, N, A2, LDA, L1, N);
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,'u', N, N, A2, LDA, L2, N);
cblas_ctrmm(CblasColMajor, CblasLeft, CblasUpper, CblasConjTrans, CblasNonUnit, N, N, CBLAS_SADDR(alpha), L1, N, L2, N);
}
else{
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,'l', N, N, A2, LDA, L1, N);
LAPACKE_clacpy_work(LAPACK_COL_MAJOR,'l', N, N, A2, LDA, L2, N);
cblas_ctrmm(CblasColMajor, CblasRight, CblasLower, CblasConjTrans, CblasNonUnit, N, N, CBLAS_SADDR(alpha), L1, N, L2, N);
}
/* Compute the Residual || A -L'L|| */
for (i = 0; i < N; i++)
for (j = 0; j < N; j++)
Residual[j*N+i] = L2[j*N+i] - Residual[j*N+i];
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, Residual, N, work);
Anorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, A1, LDA, work);
printf("============\n");
printf("Checking the Cholesky Factorization \n");
printf("-- ||L'L-A||_oo/(||A||_oo.N.eps) = %e \n",Rnorm/(Anorm*N*eps));
if ( isnan(Rnorm/(Anorm*N*eps)) || (Rnorm/(Anorm*N*eps) > 10.0) ){
printf("-- Factorization is suspicious ! \n");
info_factorization = 1;
}
else{
printf("-- Factorization is CORRECT ! \n");
info_factorization = 0;
}
free(Residual); free(L1); free(L2); free(work);
return info_factorization;
}
static int check_solution(PLASMA_enum uplo, PLASMA_enum trans, int N, int K,
PLASMA_Complex32_t alpha, PLASMA_Complex32_t *A, int LDA,
PLASMA_Complex32_t *B, int LDB,
float beta, PLASMA_Complex32_t *Cref, PLASMA_Complex32_t *Cplasma, int LDC)
{
int info_solution;
float Anorm, Bnorm, Cinitnorm, Cplasmanorm, Clapacknorm, Rnorm, result;
float eps;
PLASMA_Complex32_t beta_const;
float *work = (float *)malloc(max(N, K)* sizeof(float));
beta_const = -1.0;
Anorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm),
(trans == PlasmaNoTrans) ? N : K,
(trans == PlasmaNoTrans) ? K : N, A, LDA, work);
Bnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm),
(trans == PlasmaNoTrans) ? N : K,
(trans == PlasmaNoTrans) ? K : N, B, LDB, work);
Cinitnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, Cref, LDC, work);
Cplasmanorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, Cplasma, LDC, work);
cblas_cher2k(CblasColMajor, (CBLAS_UPLO)uplo, (CBLAS_TRANSPOSE)trans,
N, K, CBLAS_SADDR(alpha), A, LDA, B, LDB, (beta), Cref, LDC);
Clapacknorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, Cref, LDC, work);
cblas_caxpy(LDC*N, CBLAS_SADDR(beta_const), Cplasma, 1, Cref, 1);
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, Cref, LDC, work);
eps = LAPACKE_slamch_work('e');
printf("Rnorm %e, Anorm %e, Cinitnorm %e, Cplasmanorm %e, Clapacknorm %e\n",
Rnorm, Anorm, Cinitnorm, Cplasmanorm, Clapacknorm);
result = Rnorm / ((Anorm + Bnorm + Cinitnorm) * N * eps);
printf("============\n");
printf("Checking the norm of the difference against reference CHER2K \n");
printf("-- ||Cplasma - Clapack||_oo/((||A||_oo+||C||_oo).N.eps) = %e \n", result);
if ( isnan(Rnorm) || isinf(Rnorm) || isnan(result) || isinf(result) || (result > 10.0) ) {
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else {
printf("-- The solution is CORRECT ! \n");
info_solution= 0 ;
}
free(work);
return info_solution;
}
int check_solution(int M, int N, int NRHS, PLASMA_Complex32_t *A1, int LDA, PLASMA_Complex32_t *B1, PLASMA_Complex32_t *B2, int LDB)
{
int info_solution;
float Rnorm, Anorm, Xnorm, Bnorm;
PLASMA_Complex32_t alpha, beta;
float *work = (float *)malloc(max(M, N)* sizeof(float));
float eps;
eps = LAPACKE_slamch_work('e');
alpha = 1.0;
beta = -1.0;
Anorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, A1, LDA, work);
Xnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, NRHS, B2, LDB, work);
Bnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B1, LDB, work);
cblas_cgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, M, NRHS, N, CBLAS_SADDR(alpha), A1, LDA, B2, LDB, CBLAS_SADDR(beta), B1, LDB);
if (M >= N) {
PLASMA_Complex32_t *Residual = (PLASMA_Complex32_t *)malloc(M*NRHS*sizeof(PLASMA_Complex32_t));
memset((void*)Residual, 0, M*NRHS*sizeof(PLASMA_Complex32_t));
cblas_cgemm(CblasColMajor, CblasConjTrans, CblasNoTrans, N, NRHS, M, CBLAS_SADDR(alpha), A1, LDA, B1, LDB, CBLAS_SADDR(beta), Residual, M);
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, NRHS, Residual, M, work);
free(Residual);
}
else {
PLASMA_Complex32_t *Residual = (PLASMA_Complex32_t *)malloc(N*NRHS*sizeof(PLASMA_Complex32_t));
memset((void*)Residual, 0, N*NRHS*sizeof(PLASMA_Complex32_t));
cblas_cgemm(CblasColMajor, CblasConjTrans, CblasNoTrans, N, NRHS, M, CBLAS_SADDR(alpha), A1, LDA, B1, LDB, CBLAS_SADDR(beta), Residual, N);
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, Residual, N, work);
free(Residual);
}
printf("============\n");
printf("Checking the Residual of the solution \n");
printf("-- ||Ax-B||_oo/((||A||_oo||x||_oo+||B||)_oo.N.eps) = %e \n",Rnorm/((Anorm*Xnorm+Bnorm)*N*eps));
if (isnan(Rnorm / ((Anorm * Xnorm + Bnorm) * N * eps)) || (Rnorm / ((Anorm * Xnorm + Bnorm) * N * eps) > 10.0) ) {
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else {
printf("-- The solution is CORRECT ! \n");
info_solution= 0 ;
}
free(work);
return info_solution;
}
static int check_solution(PLASMA_enum side, PLASMA_enum uplo, int M, int N,
PLASMA_Complex32_t alpha, PLASMA_Complex32_t *A, int LDA,
PLASMA_Complex32_t *B, int LDB,
PLASMA_Complex32_t beta, PLASMA_Complex32_t *Cref, PLASMA_Complex32_t *Cplasma, int LDC)
{
int info_solution, NrowA;
float Anorm, Bnorm, Cinitnorm, Cplasmanorm, Clapacknorm, Rnorm;
float eps;
PLASMA_Complex32_t beta_const;
float result;
float *work = (float *)malloc(max(M, N)* sizeof(float));
beta_const = (PLASMA_Complex32_t)-1.0;
NrowA = (side == PlasmaLeft) ? M : N;
Anorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), NrowA, NrowA, A, LDA, work);
Bnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, B, LDB, work);
Cinitnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
Cplasmanorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cplasma, LDC, work);
cblas_csymm(CblasColMajor, (CBLAS_SIDE)side, (CBLAS_UPLO)uplo, M, N, CBLAS_SADDR(alpha), A, LDA, B, LDB, CBLAS_SADDR(beta), Cref, LDC);
Clapacknorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
cblas_caxpy(LDC * N, CBLAS_SADDR(beta_const), Cplasma, 1, Cref, 1);
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
eps = LAPACKE_slamch_work('e');
printf("Rnorm %e, Anorm %e, Bnorm %e, Cinitnorm %e, Cplasmanorm %e, Clapacknorm %e\n",Rnorm,Anorm,Bnorm,Cinitnorm,Cplasmanorm,Clapacknorm);
result = Rnorm / ((Anorm + Bnorm + Cinitnorm) * N * eps);
printf("============\n");
printf("Checking the norm of the difference against reference CSYMM \n");
printf("-- ||Cplasma - Clapack||_oo/((||A||_oo+||B||_oo+||C||_oo).N.eps) = %e \n", result );
if ( isinf(Clapacknorm) || isinf(Cplasmanorm) || isnan(result) || isinf(result) || (result > 10.0) ) {
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else {
printf("-- The solution is CORRECT ! \n");
info_solution= 0 ;
}
free(work);
return info_solution;
}
int check_solution(int N, int NRHS, PLASMA_Complex32_t *A1, int LDA, PLASMA_Complex32_t *B1, PLASMA_Complex32_t *B2, int LDB )
{
int info_solution;
float Rnorm, Anorm, Xnorm, Bnorm;
PLASMA_Complex32_t alpha, beta;
float *work = (float *)malloc(N*sizeof(float));
float eps;
eps = LAPACKE_slamch_work('e');
/* Initialize A1 and A2 for Symmetric Positive Matrix */
alpha = 1.0;
beta = -1.0;
Xnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B2, LDB, work);
Anorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, A1, LDA, work);
Bnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B1, LDB, work);
cblas_cgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, N, NRHS, N, CBLAS_SADDR(alpha), A1, LDA, B2, LDB, CBLAS_SADDR(beta), B1, LDB);
Rnorm = LAPACKE_clange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B1, LDB, work);
printf("============\n");
printf("Checking the Residual of the solution \n");
printf("-- ||Ax-B||_oo/((||A||_oo||x||_oo+||B||_oo).N.eps) = %e \n",Rnorm/((Anorm*Xnorm+Bnorm)*N*eps));
if (Rnorm/((Anorm*Xnorm+Bnorm)*N*eps) > 10.0){
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else{
printf("-- The solution is CORRECT ! \n");
info_solution = 0;
}
free(work);
return info_solution;
}
int check_solution(int N, int NRHS, float *A1, int LDA, float *B1, float *B2, int LDB)
{
int info_solution;
float Rnorm, Anorm, Xnorm, Bnorm;
float alpha, beta;
float *work = (float *)malloc(N*sizeof(float));
float eps;
eps = LAPACKE_slamch_work('e');
alpha = 1.0;
beta = -1.0;
Xnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B2, LDB, work);
Anorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, N, A1, LDA, work);
Bnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B1, LDB, work);
cblas_sgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, N, NRHS, N, (alpha), A1, LDA, B2, LDB, (beta), B1, LDB);
Rnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), N, NRHS, B1, LDB, work);
printf("============\n");
printf("Checking the Residual of the solution \n");
printf("-- ||Ax-B||_oo/((||A||_oo||x||_oo+||B||_oo).N.eps) = %e \n",Rnorm/((Anorm*Xnorm+Bnorm)*N*eps));
if ( isnan(Rnorm/((Anorm*Xnorm+Bnorm)*N*eps)) || (Rnorm/((Anorm*Xnorm+Bnorm)*N*eps) > 10.0) ){
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else{
printf("-- The solution is CORRECT ! \n");
info_solution = 0;
}
free(work);
return info_solution;
}
float LAPACKE_slamch( char cmach )
{
return LAPACKE_slamch_work( cmach );
}
int testing_cher2k(int argc, char **argv)
{
/* Check for number of arguments*/
if ( argc != 7 ){
USAGE("HER2K", "alpha beta M N LDA LDB LDC",
" - alpha : alpha coefficient\n"
" - beta : beta coefficient\n"
" - N : number of columns and rows of matrix C and number of row of matrix A and B\n"
" - K : number of columns of matrix A and B\n"
" - LDA : leading dimension of matrix A\n"
" - LDB : leading dimension of matrix B\n"
" - LDC : leading dimension of matrix C\n");
return -1;
}
PLASMA_Complex32_t alpha = (PLASMA_Complex32_t) atol(argv[0]);
float beta = (float) atol(argv[1]);
int N = atoi(argv[2]);
int K = atoi(argv[3]);
int LDA = atoi(argv[4]);
int LDB = atoi(argv[5]);
int LDC = atoi(argv[6]);
int NKmax = max(N, K);
int NminusOne = N - 1;
float eps;
int info_solution;
int info, u, t;
size_t LDAxK = LDA*NKmax;
size_t LDBxK = LDB*NKmax;
size_t LDCxN = LDC*N;
PLASMA_Complex32_t *A = (PLASMA_Complex32_t *)malloc(LDAxK*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *B = (PLASMA_Complex32_t *)malloc(LDBxK*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *C = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *Cinit = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *Cfinal = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *WORK = (PLASMA_Complex32_t *)malloc(2*LDC*sizeof(PLASMA_Complex32_t));
float *D = (float *) malloc(LDC *sizeof(float));
/* Check if unable to allocate memory */
if ( (!A) || (!B) || (!Cinit) || (!Cfinal) || (!D) ){
printf("Out of Memory \n ");
return -2;
}
eps = LAPACKE_slamch_work('e');
printf("\n");
printf("------ TESTS FOR PLASMA CHER2K ROUTINE ------- \n");
printf(" Size of the Matrix C %d by %d\n", N, K);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is to be %e \n",eps);
printf(" Computational tests pass if scaled residuals are less than 10.\n");
/*----------------------------------------------------------
* TESTING CHER2K
*/
/* Initialize A,B */
LAPACKE_clarnv_work(IONE, ISEED, LDAxK, A);
LAPACKE_clarnv_work(IONE, ISEED, LDBxK, B);
/* Initialize C */
LAPACKE_slarnv_work(IONE, ISEED, LDC, D);
claghe(&N, &NminusOne, D, C, &LDC, ISEED, WORK, &info);
free(D); free(WORK);
for (u=0; u<2; u++) {
for (t=0; t<3; t++) {
if (trans[t] == PlasmaTrans) continue;
memcpy(Cinit, C, LDCxN*sizeof(PLASMA_Complex32_t));
memcpy(Cfinal, C, LDCxN*sizeof(PLASMA_Complex32_t));
/* PLASMA CHER2K */
PLASMA_cher2k(uplo[u], trans[t], N, K, alpha, A, LDA, B, LDB, beta, Cfinal, LDC);
/* Check the solution */
info_solution = check_solution(uplo[u], trans[t], N, K,
alpha, A, LDA, B, LDB, beta, Cinit, Cfinal, LDC);
if (info_solution == 0) {
printf("***************************************************\n");
printf(" ---- TESTING CHER2K (%5s, %s) ........... PASSED !\n", uplostr[u], transstr[t]);
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING CHER2K (%5s, %s) ... FAILED !\n", uplostr[u], transstr[t]);
printf("************************************************\n");
}
}
}
free(A); free(B); free(C);
free(Cinit); free(Cfinal);
return 0;
}
int testing_strsm(int argc, char **argv)
{
/* Check for number of arguments*/
if ( argc != 5 ) {
USAGE("TRSM", "alpha M N LDA LDB",
" - alpha : alpha coefficient\n"
" - M : number of rows of matrices B\n"
" - N : number of columns of matrices B\n"
" - LDA : leading dimension of matrix A\n"
" - LDB : leading dimension of matrix B\n");
return -1;
}
float alpha = (float) atol(argv[0]);
int M = atoi(argv[1]);
int N = atoi(argv[2]);
int LDA = atoi(argv[3]);
int LDB = atoi(argv[4]);
float eps;
int info_solution;
int s, u, t, d, i;
int LDAxM = LDA*max(M,N);
int LDBxN = LDB*max(M,N);
float *A = (float *)malloc(LDAxM*sizeof(float));
float *B = (float *)malloc(LDBxN*sizeof(float));
float *Binit = (float *)malloc(LDBxN*sizeof(float));
float *Bfinal = (float *)malloc(LDBxN*sizeof(float));
/* Check if unable to allocate memory */
if ( (!A) || (!B) || (!Binit) || (!Bfinal)){
printf("Out of Memory \n ");
return -2;
}
eps = LAPACKE_slamch_work('e');
printf("\n");
printf("------ TESTS FOR PLASMA STRSM ROUTINE ------- \n");
printf(" Size of the Matrix B : %d by %d\n", M, N);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is to be %e \n",eps);
printf(" Computational tests pass if scaled residuals are less than 10.\n");
/*----------------------------------------------------------
* TESTING STRSM
*/
/* Initialize A, B, C */
LAPACKE_slarnv_work(IONE, ISEED, LDAxM, A);
LAPACKE_slarnv_work(IONE, ISEED, LDBxN, B);
for(i=0;i<max(M,N);i++)
A[LDA*i+i] = A[LDA*i+i] + 2.0;
for (s=0; s<2; s++) {
for (u=0; u<2; u++) {
#ifdef COMPLEX
for (t=0; t<3; t++) {
#else
for (t=0; t<2; t++) {
#endif
for (d=0; d<2; d++) {
memcpy(Binit, B, LDBxN*sizeof(float));
memcpy(Bfinal, B, LDBxN*sizeof(float));
/* PLASMA STRSM */
PLASMA_strsm(side[s], uplo[u], trans[t], diag[d],
M, N, alpha, A, LDA, Bfinal, LDB);
/* Check the solution */
info_solution = check_solution(side[s], uplo[u], trans[t], diag[d],
M, N, alpha, A, LDA, Binit, Bfinal, LDB);
printf("***************************************************\n");
if (info_solution == 0) {
printf(" ---- TESTING STRSM (%s, %s, %s, %s) ...... PASSED !\n",
sidestr[s], uplostr[u], transstr[t], diagstr[d]);
}
else {
printf(" ---- TESTING STRSM (%s, %s, %s, %s) ... FAILED !\n",
sidestr[s], uplostr[u], transstr[t], diagstr[d]);
}
printf("***************************************************\n");
}
}
}
}
free(A); free(B);
free(Binit); free(Bfinal);
return 0;
}
/*--------------------------------------------------------------
* Check the solution
*/
static int check_solution(PLASMA_enum side, PLASMA_enum uplo, PLASMA_enum trans, PLASMA_enum diag,
int M, int N, float alpha,
float *A, int LDA,
float *Bref, float *Bplasma, int LDB)
{
int info_solution;
float Anorm, Binitnorm, Bplasmanorm, Blapacknorm, Rnorm, result;
float eps;
float mzone = (float)-1.0;
float *work = (float *)malloc(max(M, N)* sizeof(float));
int Am, An;
if (side == PlasmaLeft) {
Am = M; An = M;
} else {
Am = N; An = N;
}
Anorm = LAPACKE_slantr_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), lapack_const(uplo), lapack_const(diag),
Am, An, A, LDA, work);
Binitnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Bref, LDB, work);
Bplasmanorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Bplasma, LDB, work);
cblas_strsm(CblasColMajor, (CBLAS_SIDE)side, (CBLAS_UPLO)uplo, (CBLAS_TRANSPOSE)trans,
(CBLAS_DIAG)diag, M, N, (alpha), A, LDA, Bref, LDB);
Blapacknorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Bref, LDB, work);
cblas_saxpy(LDB * N, (mzone), Bplasma, 1, Bref, 1);
Rnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Bref, LDB, work);
eps = LAPACKE_slamch_work('e');
printf("Rnorm %e, Anorm %e, Binitnorm %e, Bplasmanorm %e, Blapacknorm %e\n",
Rnorm, Anorm, Binitnorm, Bplasmanorm, Blapacknorm);
result = Rnorm / ((Anorm + Blapacknorm) * max(M,N) * eps);
printf("============\n");
printf("Checking the norm of the difference against reference STRSM \n");
printf("-- ||Cplasma - Clapack||_oo/((||A||_oo+||B||_oo).N.eps) = %e \n", result);
if ( isinf(Blapacknorm) || isinf(Bplasmanorm) || isnan(result) || isinf(result) || (result > 10.0) ) {
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else {
printf("-- The solution is CORRECT ! \n");
info_solution= 0 ;
}
free(work);
return info_solution;
}
int testing_csymm(int argc, char **argv)
{
/* Check for number of arguments*/
if ( argc != 7 ){
USAGE("SYMM", "alpha beta M N K LDA LDB LDC",
" - alpha : alpha coefficient \n"
" - beta : beta coefficient \n"
" - M : number of rows of matrices A and C \n"
" - N : number of columns of matrices B and C \n"
" - LDA : leading dimension of matrix A \n"
" - LDB : leading dimension of matrix B \n"
" - LDC : leading dimension of matrix C\n");
return -1;
}
PLASMA_Complex32_t alpha = (PLASMA_Complex32_t) atol(argv[0]);
PLASMA_Complex32_t beta = (PLASMA_Complex32_t) atol(argv[1]);
int M = atoi(argv[2]);
int N = atoi(argv[3]);
int LDA = atoi(argv[4]);
int LDB = atoi(argv[5]);
int LDC = atoi(argv[6]);
int MNmax = max(M, N);
int MminusOne = MNmax - 1;
float eps;
int info_solution;
int i, j, s, u, info;
int LDAxM = LDA*max(M, N);
int LDBxN = LDB*N;
int LDCxN = LDC*N;
PLASMA_Complex32_t *A = (PLASMA_Complex32_t *)malloc(LDAxM*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *B = (PLASMA_Complex32_t *)malloc(LDBxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *C = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *Cinit = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *Cfinal = (PLASMA_Complex32_t *)malloc(LDCxN*sizeof(PLASMA_Complex32_t));
PLASMA_Complex32_t *WORK = (PLASMA_Complex32_t *)malloc(2*LDC*sizeof(PLASMA_Complex32_t));
float *D = (float *) malloc(LDC *sizeof(float));
/* Check if unable to allocate memory */
if ((!A)||(!B)||(!Cinit)||(!Cfinal)){
printf("Out of Memory \n ");
return -2;
}
eps = LAPACKE_slamch_work('e');
printf("\n");
printf("------ TESTS FOR PLASMA CSYMM ROUTINE ------- \n");
printf(" Size of the Matrix %d by %d\n", M, N);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is to be %e \n",eps);
printf(" Computational tests pass if scaled residuals are less than 10.\n");
/*----------------------------------------------------------
* TESTING CSYMM
*/
/* Initialize A */
LAPACKE_slarnv_work(IONE, ISEED, LDC, D);
claghe(&MNmax, &MminusOne, D, A, &LDA, ISEED, WORK, &info);
free(D); free(WORK);
/* Initialize B */
LAPACKE_clarnv_work(IONE, ISEED, LDBxN, B);
/* Initialize C */
LAPACKE_clarnv_work(IONE, ISEED, LDCxN, C);
for (s=0; s<2; s++) {
for (u=0; u<2; u++) {
/* Initialize Cinit / Cfinal */
for ( i = 0; i < M; i++)
for ( j = 0; j < N; j++)
Cinit[LDC*j+i] = C[LDC*j+i];
for ( i = 0; i < M; i++)
for ( j = 0; j < N; j++)
Cfinal[LDC*j+i] = C[LDC*j+i];
/* PLASMA CSYMM */
PLASMA_csymm(side[s], uplo[u], M, N, alpha, A, LDA, B, LDB, beta, Cfinal, LDC);
/* Check the solution */
info_solution = check_solution(side[s], uplo[u], M, N, alpha, A, LDA, B, LDB, beta, Cinit, Cfinal, LDC);
if (info_solution == 0) {
printf("***************************************************\n");
printf(" ---- TESTING CSYMM (%5s, %5s) ....... PASSED !\n", sidestr[s], uplostr[u]);
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING CSYMM (%s, %s) ... FAILED !\n", sidestr[s], uplostr[u]);
printf("************************************************\n");
}
}
}
free(A); free(B); free(C);
free(Cinit); free(Cfinal);
return 0;
}
int testing_ssygst(int argc, char **argv)
{
/* Check for number of arguments*/
if (argc != 3) {
USAGE("HEGST", "N LDA LDB",
" - N : size of the matrices A and B\n"
" - LDA : leading dimension of the matrix A\n"
" - LDB : leading dimension of the matrix B\n");
return -1;
}
float eps = LAPACKE_slamch_work('e');
int N = atoi(argv[0]);
int LDA = atoi(argv[1]);
int LDB = atoi(argv[2]);
int info_transformation, info_factorization;
int i, u;
int LDAxN = LDA*N;
int LDBxN = LDB*N;
float *A1 = (float *)malloc(LDAxN*sizeof(float));
float *A2 = (float *)malloc(LDAxN*sizeof(float));
float *B1 = (float *)malloc(LDBxN*sizeof(float));
float *B2 = (float *)malloc(LDBxN*sizeof(float));
float *Ainit = (float *)malloc(LDAxN*sizeof(float));
float *Binit = (float *)malloc(LDBxN*sizeof(float));
/* Check if unable to allocate memory */
if ((!A1)||(!A2)||(!B1)||(!B2)||(!Ainit)||(!Binit)){
printf("Out of Memory \n ");
return -2;
}
/*----------------------------------------------------------
* TESTING SSYGST
*/
/* Initialize A1 and A2 */
PLASMA_splgsy(0., N, A1, LDA, 5198);
LAPACKE_slacpy_work(LAPACK_COL_MAJOR, 'A', N, N, A1, LDA, Ainit, LDA);
/* Initialize B1 and B2 */
PLASMA_splgsy((float)N, N, B1, LDB, 4231);
LAPACKE_slacpy_work(LAPACK_COL_MAJOR, 'A', N, N, B1, LDB, Binit, LDB);
printf("\n");
printf("------ TESTS FOR PLASMA SSYGST ROUTINE ------- \n");
printf(" Size of the Matrix %d by %d\n", N, N);
printf("\n");
printf(" The matrices A and B are randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is to be %e \n",eps);
printf(" Computational tests pass if scaled residuals are less than 60.\n");
/*----------------------------------------------------------
* TESTING SSYGST
*/
for (i=0; i<3; i++) {
for (u=0; u<2; u++) {
memcpy(A2, Ainit, LDAxN*sizeof(float));
memcpy(B2, Binit, LDBxN*sizeof(float));
PLASMA_spotrf(uplo[u], N, B2, LDB);
PLASMA_ssygst(itype[i], uplo[u], N, A2, LDA, B2, LDB);
/* Check the Cholesky factorization and the transformation */
info_factorization = check_factorization(N, B1, B2, LDB, uplo[u], eps);
info_transformation = check_transformation(itype[i], uplo[u], N, A1, A2, LDA, B2, LDB, eps);
if ( (info_transformation == 0) && (info_factorization == 0) ) {
printf("***************************************************\n");
printf(" ---- TESTING SSYGST (%s, %s) ....... PASSED !\n", itypestr[i], uplostr[u]);
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING SSYGST (%s, %s) ... FAILED !\n", itypestr[i], uplostr[u]);
printf("************************************************\n");
}
}
}
free(A1);
free(A2);
free(B1);
free(B2);
free(Ainit);
free(Binit);
return 0;
}
int testing_sgemm(int argc, char **argv)
{
/* Check for number of arguments*/
if ( argc != 8) {
USAGE("GEMM", "alpha beta M N K LDA LDB LDC",
" - alpha : alpha coefficient\n"
" - beta : beta coefficient\n"
" - M : number of rows of matrices A and C\n"
" - N : number of columns of matrices B and C\n"
" - K : number of columns of matrix A / number of rows of matrix B\n"
" - LDA : leading dimension of matrix A\n"
" - LDB : leading dimension of matrix B\n"
" - LDC : leading dimension of matrix C\n");
return -1;
}
float alpha = (float) atol(argv[0]);
float beta = (float) atol(argv[1]);
int M = atoi(argv[2]);
int N = atoi(argv[3]);
int K = atoi(argv[4]);
int LDA = atoi(argv[5]);
int LDB = atoi(argv[6]);
int LDC = atoi(argv[7]);
float eps;
int info_solution;
int i, j, ta, tb;
int LDAxK = LDA*max(M,K);
int LDBxN = LDB*max(K,N);
int LDCxN = LDC*N;
float *A = (float *)malloc(LDAxK*sizeof(float));
float *B = (float *)malloc(LDBxN*sizeof(float));
float *C = (float *)malloc(LDCxN*sizeof(float));
float *Cinit = (float *)malloc(LDCxN*sizeof(float));
float *Cfinal = (float *)malloc(LDCxN*sizeof(float));
/* Check if unable to allocate memory */
if ((!A)||(!B)||(!Cinit)||(!Cfinal)){
printf("Out of Memory \n ");
return -2;
}
eps = LAPACKE_slamch_work('e');
printf("\n");
printf("------ TESTS FOR PLASMA SGEMM ROUTINE ------- \n");
printf(" Size of the Matrix %d by %d\n", M, N);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is to be %e \n",eps);
printf(" Computational tests pass if scaled residuals are less than 10.\n");
/*----------------------------------------------------------
* TESTING SGEMM
*/
/* Initialize A, B, C */
LAPACKE_slarnv_work(IONE, ISEED, LDAxK, A);
LAPACKE_slarnv_work(IONE, ISEED, LDBxN, B);
LAPACKE_slarnv_work(IONE, ISEED, LDCxN, C);
#ifdef COMPLEX
for (ta=0; ta<3; ta++) {
for (tb=0; tb<3; tb++) {
#else
for (ta=0; ta<2; ta++) {
for (tb=0; tb<2; tb++) {
#endif
for ( i = 0; i < M; i++)
for ( j = 0; j < N; j++)
Cinit[LDC*j+i] = C[LDC*j+i];
for ( i = 0; i < M; i++)
for ( j = 0; j < N; j++)
Cfinal[LDC*j+i] = C[LDC*j+i];
/* PLASMA SGEMM */
PLASMA_sgemm(trans[ta], trans[tb], M, N, K, alpha, A, LDA, B, LDB, beta, Cfinal, LDC);
/* Check the solution */
info_solution = check_solution(trans[ta], trans[tb], M, N, K,
alpha, A, LDA, B, LDB, beta, Cinit, Cfinal, LDC);
if (info_solution == 0) {
printf("***************************************************\n");
printf(" ---- TESTING SGEMM (%s, %s) ............... PASSED !\n", transstr[ta], transstr[tb]);
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING SGEMM (%s, %s) ... FAILED !\n", transstr[ta], transstr[tb]);
printf("************************************************\n");
}
}
}
#ifdef _UNUSED_
}}
#endif
free(A); free(B); free(C);
free(Cinit); free(Cfinal);
return 0;
}
static int check_solution(PLASMA_enum transA, PLASMA_enum transB, int M, int N, int K,
float alpha, float *A, int LDA,
float *B, int LDB,
float beta, float *Cref, float *Cplasma, int LDC)
{
int info_solution;
float Anorm, Bnorm, Cinitnorm, Cplasmanorm, Clapacknorm, Rnorm, result;
float eps;
float beta_const;
float *work = (float *)malloc(max(K,max(M, N))* sizeof(float));
int Am, An, Bm, Bn;
beta_const = -1.0;
if (transA == PlasmaNoTrans) {
Am = M; An = K;
} else {
Am = K; An = M;
}
if (transB == PlasmaNoTrans) {
Bm = K; Bn = N;
} else {
Bm = N; Bn = K;
}
Anorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), Am, An, A, LDA, work);
Bnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), Bm, Bn, B, LDB, work);
Cinitnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
Cplasmanorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cplasma, LDC, work);
cblas_sgemm(CblasColMajor, (CBLAS_TRANSPOSE)transA, (CBLAS_TRANSPOSE)transB, M, N, K,
(alpha), A, LDA, B, LDB, (beta), Cref, LDC);
Clapacknorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
cblas_saxpy(LDC * N, (beta_const), Cplasma, 1, Cref, 1);
Rnorm = LAPACKE_slange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaInfNorm), M, N, Cref, LDC, work);
eps = LAPACKE_slamch_work('e');
printf("Rnorm %e, Anorm %e, Bnorm %e, Cinitnorm %e, Cplasmanorm %e, Clapacknorm %e\n",
Rnorm, Anorm, Bnorm, Cinitnorm, Cplasmanorm, Clapacknorm);
result = Rnorm / ((Anorm + Bnorm + Cinitnorm) * N * eps);
printf("============\n");
printf("Checking the norm of the difference against reference SGEMM \n");
printf("-- ||Cplasma - Clapack||_oo/((||A||_oo+||B||_oo+||C||_oo).N.eps) = %e \n",
result);
if ( isnan(Rnorm) || isinf(Rnorm) || isnan(result) || isinf(result) || (result > 10.0) ) {
printf("-- The solution is suspicious ! \n");
info_solution = 1;
}
else {
printf("-- The solution is CORRECT ! \n");
info_solution= 0 ;
}
free(work);
return info_solution;
}
