/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
magma_d_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
precond_par.solver = Magma_JACOBI;
magma_dsolverinfo_init( &solver_par, &precond_par );
double one = MAGMA_D_MAKE(1.0, 0.0);
double zero = MAGMA_D_MAKE(0.0, 0.0);
magma_d_sparse_matrix A, B, B_d;
magma_d_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=10; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
double *diag_vals;
magma_dmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_D_MAKE( 4.0, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_dmscale( &A, Magma_UNITDIAG );
magma_d_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_mfree(&B_d);
magma_d_mfree(&B);
magma_d_mfree(&A);
magma_d_vfree(&x);
magma_d_vfree(&b);
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_dparse_opts( argc, argv, &zopts, &i, queue );
real_Double_t res;
magma_d_sparse_matrix Z, A, AT, A2, B, B_d;
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
magma_dm_5stencil( laplace_size, &Z, queue );
} else { // file-matrix test
magma_d_csr_mtx( &Z, argv[i], queue );
}
printf( "# matrix info: %d-by-%d with %d nonzeros\n",
(int) Z.num_rows,(int) Z.num_cols,(int) Z.nnz );
// scale matrix
magma_dmscale( &Z, zopts.scaling, queue );
// remove nonzeros in matrix
magma_dmcsrcompressor( &Z, queue );
// convert to be non-symmetric
magma_d_mconvert( Z, &A, Magma_CSR, Magma_CSRL, queue );
// transpose
magma_d_mtranspose( A, &AT, queue );
// convert, copy back and forth to check everything works
magma_d_mconvert( AT, &B, Magma_CSR, zopts.output_format, queue );
magma_d_mfree(&AT, queue );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
magma_d_mfree(&B, queue );
magma_dmcsrcompressor_gpu( &B_d, queue );
magma_d_mtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue );
magma_d_mfree(&B_d, queue );
magma_d_mconvert( B, &AT, zopts.output_format,Magma_CSR, queue );
magma_d_mfree(&B, queue );
// transpose back
magma_d_mtranspose( AT, &A2, queue );
magma_d_mfree(&AT, queue );
magma_dmdiff( A, A2, &res, queue);
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester: ok\n");
else
printf("# tester: failed\n");
magma_d_mfree(&A, queue );
magma_d_mfree(&A2, queue );
magma_d_mfree(&Z, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_dlobpcg
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
solver_par.epsilon = 1e-5;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 32;
solver_par.solver = Magma_LOBPCG;
magma_d_preconditioner precond_par;
precond_par.solver = Magma_JACOBI;
int precond = 0;
int format = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_d_sparse_matrix A, B, dA;
B.blocksize = 8;
B.alignment = 8;
B.storage_type = Magma_CSR;
int i;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--format", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( format ) {
case 0: B.storage_type = Magma_CSR; break;
case 1: B.storage_type = Magma_ELL; break;
case 2: B.storage_type = Magma_ELLRT; break;
case 3: B.storage_type = Magma_SELLP; break;
}
}else if ( strcmp("--mscale", argv[i]) == 0 ) {
scale = atoi( argv[++i] );
switch( scale ) {
case 0: scaling = Magma_NOSCALE; break;
case 1: scaling = Magma_UNITDIAG; break;
case 2: scaling = Magma_UNITROW; break;
}
}else if ( strcmp("--precond", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( precond ) {
case 0: precond_par.solver = Magma_JACOBI; break;
}
}else if ( strcmp("--blocksize", argv[i]) == 0 ) {
B.blocksize = atoi( argv[++i] );
}else if ( strcmp("--alignment", argv[i]) == 0 ) {
B.alignment = atoi( argv[++i] );
}else if ( strcmp("--verbose", argv[i]) == 0 ) {
solver_par.verbose = atoi( argv[++i] );
} else if ( strcmp("--maxiter", argv[i]) == 0 ) {
solver_par.maxiter = atoi( argv[++i] );
} else if ( strcmp("--tol", argv[i]) == 0 ) {
sscanf( argv[++i], "%lf", &solver_par.epsilon );
} else if ( strcmp("--eigenvalues", argv[i]) == 0 ) {
solver_par.num_eigenvalues = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_dlobpcg"
" [ --format %d (0=CSR, 1=ELL, 2=ELLRT, 4=SELLP)"
" [ --blocksize %d --alignment %d ]"
" --mscale %d (0=no, 1=unitdiag, 2=unitrownrm)"
" --verbose %d (0=summary, k=details every k iterations)"
" --maxiter %d --tol %.2e"
" --preconditioner %d (0=Jacobi) "
" --eigenvalues %d ]"
" matrices \n\n", format, (int) B.blocksize, (int) B.alignment,
(int) scale,
(int) solver_par.verbose,
(int) solver_par.maxiter, solver_par.epsilon, precond,
(int) solver_par.num_eigenvalues);
while( i < argc ){
magma_d_csr_mtx( &A, argv[i] );
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// scale initial guess
magma_dmscale( &A, scaling );
solver_par.ev_length = A.num_cols;
magma_d_sparse_matrix A2;
A2.storage_type = Magma_SELLC;
A2.blocksize = 8;
A2.alignment = 4;
magma_d_mconvert( A, &A2, Magma_CSR, A2.storage_type );
// copy matrix to GPU
magma_d_mtransfer( A2, &dA, Magma_CPU, Magma_DEV);
magma_dsolverinfo_init( &solver_par, &precond_par ); // inside the loop!
// as the matrix size has influence on the EV-length
real_Double_t gpu_time;
// Find the blockSize smallest eigenvalues and corresponding eigen-vectors
gpu_time = magma_wtime();
magma_dlobpcg( dA, &solver_par );
gpu_time = magma_wtime() - gpu_time;
printf("Time (sec) = %7.2f\n", gpu_time);
printf("solver runtime (sec) = %7.2f\n", solver_par.runtime );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_mfree( &dA );
magma_d_mfree( &A2 );
magma_d_mfree( &A );
i++;
}
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_d_matrix A={Magma_CSR}, AT={Magma_CSR}, A2={Magma_CSR},
B={Magma_CSR}, B_d={Magma_CSR};
int i=1;
real_Double_t start, end;
CHECK( magma_dparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_dm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_d_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
// scale matrix
CHECK( magma_dmscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_dmcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_dmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_dmconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_dmcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_dmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_dmfree(&B_d, queue );
CHECK( magma_dmconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_dmfree(&B, queue );
// transpose back
CHECK( magma_dmtranspose( AT, &A2, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmdiff( A, A2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix compressor: ok\n");
else
printf("%% tester matrix compressor: failed\n");
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
i++;
}
cleanup:
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_d_solver_par solver_par;
magma_d_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.restart = 30;
solver_par.num_eigenvalues = 0;
solver_par.ortho = Magma_CGS;
double one = MAGMA_D_MAKE(1.0, 0.0);
double zero = MAGMA_D_MAKE(0.0, 0.0);
magma_d_sparse_matrix A, B, B_d;
magma_d_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=100; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
double *diag_vals;
magma_dmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_D_MAKE( 4.1, 0.0 );
diag_vals[1] = MAGMA_D_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_D_MAKE( -1.0, 0.0 );
magma_dmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_dmscale( &A, Magma_UNITDIAG );
magma_d_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_d_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// test CG ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dcg_res( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test BICGSTAB ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dbicgstab( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PBICGSTAB Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
/*
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);printf("here\n");
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test GMRES ####################################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// solver
magma_dgmres( B_d, b, &x, &solver_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
// test PGMRES Jacobi ############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);*/
// test PGMRES ILU ###############################
// vectors and initial guess
magma_d_vinit( &b, Magma_DEV, A.num_cols, one );
magma_d_vinit( &x, Magma_DEV, A.num_cols, one );
magma_d_spmv( one, B_d, x, zero, b ); // b = A x
magma_d_vfree(&x);
magma_d_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_dsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_d_precondsetup( B_d, b, &precond_par );
// solver
magma_dpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_dsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_dsolverinfo_free( &solver_par, &precond_par );
magma_d_vfree(&x);
magma_d_vfree(&b);
printf("all tests passed.\n");
magma_d_mfree(&B_d);
magma_d_mfree(&B);
magma_d_mfree(&A);
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_dmscale(
magma_d_matrix *A,
magma_scale_t scaling,
magma_queue_t queue )
{
magma_int_t info = 0;
double *tmp=NULL;
magma_d_matrix hA={Magma_CSR}, CSRA={Magma_CSR};
if( A->num_rows != A->num_cols && scaling != Magma_NOSCALE ){
printf("%% warning: non-square matrix.\n");
printf("%% Fallback: no scaling.\n");
scaling = Magma_NOSCALE;
}
if ( A->memory_location == Magma_CPU && A->storage_type == Magma_CSRCOO ) {
if ( scaling == Magma_NOSCALE ) {
// no scale
;
}
else if( A->num_rows == A->num_cols ){
if ( scaling == Magma_UNITROW ) {
// scale to unit rownorm
CHECK( magma_dmalloc_cpu( &tmp, A->num_rows ));
for( magma_int_t z=0; z<A->num_rows; z++ ) {
double s = MAGMA_D_MAKE( 0.0, 0.0 );
for( magma_int_t f=A->row[z]; f<A->row[z+1]; f++ )
s+= MAGMA_D_REAL(A->val[f])*MAGMA_D_REAL(A->val[f]);
tmp[z] = MAGMA_D_MAKE( 1.0/sqrt( MAGMA_D_REAL( s ) ), 0.0 );
}
for( magma_int_t z=0; z<A->nnz; z++ ) {
A->val[z] = A->val[z] * tmp[A->col[z]] * tmp[A->rowidx[z]];
}
}
else if (scaling == Magma_UNITDIAG ) {
// scale to unit diagonal
CHECK( magma_dmalloc_cpu( &tmp, A->num_rows ));
for( magma_int_t z=0; z<A->num_rows; z++ ) {
double s = MAGMA_D_MAKE( 0.0, 0.0 );
for( magma_int_t f=A->row[z]; f<A->row[z+1]; f++ ) {
if ( A->col[f]== z ) {
// add some identity matrix
//A->val[f] = A->val[f] + MAGMA_D_MAKE( 100000.0, 0.0 );
s = A->val[f];
}
}
if ( s == MAGMA_D_MAKE( 0.0, 0.0 ) ){
printf("%%error: zero diagonal element.\n");
info = MAGMA_ERR;
}
tmp[z] = MAGMA_D_MAKE( 1.0/sqrt( MAGMA_D_REAL( s ) ), 0.0 );
}
for( magma_int_t z=0; z<A->nnz; z++ ) {
A->val[z] = A->val[z] * tmp[A->col[z]] * tmp[A->rowidx[z]];
}
}
else {
printf( "%%error: scaling not supported.\n" );
info = MAGMA_ERR_NOT_SUPPORTED;
}
}
else {
printf( "%%error: scaling not supported.\n" );
info = MAGMA_ERR_NOT_SUPPORTED;
}
}
else {
magma_storage_t A_storage = A->storage_type;
magma_location_t A_location = A->memory_location;
CHECK( magma_dmtransfer( *A, &hA, A->memory_location, Magma_CPU, queue ));
CHECK( magma_dmconvert( hA, &CSRA, hA.storage_type, Magma_CSRCOO, queue ));
CHECK( magma_dmscale( &CSRA, scaling, queue ));
magma_dmfree( &hA, queue );
magma_dmfree( A, queue );
CHECK( magma_dmconvert( CSRA, &hA, Magma_CSRCOO, A_storage, queue ));
CHECK( magma_dmtransfer( hA, A, Magma_CPU, A_location, queue ));
}
cleanup:
magma_free_cpu( tmp );
magma_dmfree( &hA, queue );
magma_dmfree( &CSRA, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_dopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
real_Double_t res;
magma_d_matrix Z={Magma_CSR}, A={Magma_CSR}, AT={Magma_CSR},
A2={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_index_t *comm_i=NULL;
double *comm_v=NULL;
magma_int_t start, end;
int i=1;
CHECK( magma_dparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_dm_5stencil( laplace_size, &Z, queue ));
} else { // file-matrix test
CHECK( magma_d_csr_mtx( &Z, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(Z.num_rows), int(Z.num_cols), int(Z.nnz) );
// slice matrix
CHECK( magma_index_malloc_cpu( &comm_i, Z.num_rows ) );
CHECK( magma_dmalloc_cpu( &comm_v, Z.num_rows ) );
CHECK( magma_dmslice( 1, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 0, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 1, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
CHECK( magma_dmslice( 9, 8, Z, &A2, &AT, &B, comm_i, comm_v, &start, &end, queue ) );
magma_dprint_matrix( A2, queue );
magma_dprint_matrix( AT, queue );
magma_dprint_matrix( B, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&AT, queue );
magma_dmfree(&B, queue );
// scale matrix
CHECK( magma_dmscale( &Z, zopts.scaling, queue ));
// remove nonzeros in matrix
CHECK( magma_dmcsrcompressor( &Z, queue ));
// convert to be non-symmetric
CHECK( magma_dmconvert( Z, &A, Magma_CSR, Magma_CSRL, queue ));
// transpose
CHECK( magma_dmtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_dmconvert( AT, &B, Magma_CSR, zopts.output_format, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_dmfree(&B, queue );
CHECK( magma_dmcsrcompressor_gpu( &B_d, queue ));
CHECK( magma_dmtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_dmfree(&B_d, queue );
CHECK( magma_dmconvert( B, &AT, zopts.output_format,Magma_CSR, queue ));
magma_dmfree(&B, queue );
// transpose back
CHECK( magma_dmtranspose( AT, &A2, queue ));
magma_dmfree(&AT, queue );
CHECK( magma_dmdiff( A, A2, &res, queue));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester: ok\n");
else
printf("%% tester: failed\n");
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
comm_i=NULL;
comm_v=NULL;
magma_dmfree(&A, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
i++;
}
cleanup:
magma_free_cpu( comm_i );
magma_free_cpu( comm_v );
magma_dmfree(&AT, queue );
magma_dmfree(&A, queue );
magma_dmfree(&B, queue );
magma_dmfree(&B_d, queue );
magma_dmfree(&A2, queue );
magma_dmfree(&Z, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}