int main(int argc, char* argv[]) { if (argc == 1) { std::cerr << argv[0] << " <matrix> [Num threads]" << std::endl; exit(1); } init_paralution(); if (argc > 2) { set_omp_threads_paralution(atoi(argv[2])); } info_paralution(); LocalVector<double> x; LocalVector<double> rhs; LocalMatrix<double> mat; mat.ReadFileMTX(std::string(argv[1])); mat.info(); x.Allocate("x", mat.get_nrow()); rhs.Allocate("rhs", mat.get_nrow()); x.info(); rhs.info(); rhs.Ones(); mat.Apply(rhs, &x); std::cout << "dot=" << x.Dot(rhs) << std::endl; mat.ConvertToELL(); mat.info(); mat.MoveToAccelerator(); x.MoveToAccelerator(); rhs.MoveToAccelerator(); mat.info(); rhs.Ones(); mat.Apply(rhs, &x); std::cout << "dot=" << x.Dot(rhs) << std::endl; stop_paralution(); return 0; }
int main(int argc, char* argv[]) { if (argc == 1) { std::cerr << argv[0] << " <matrix> [Num threads]" << std::endl; exit(1); } init_paralution(); if (argc > 2) { set_omp_threads_paralution(atoi(argv[2])); } info_paralution(); // int ii; LocalVector<double> x; LocalVector<double> rhs; LocalMatrix<double> mat; struct timeval ti1,ti2;//timer mat.ReadFileMTX(std::string(argv[1])); mat.info(); x.Allocate("x", mat.get_nrow()); rhs.Allocate("rhs", mat.get_nrow()); x.info(); rhs.info(); rhs.Ones(); gettimeofday(&ti1,NULL); /* read starttime in t1 */ mat.Apply(rhs, &x); gettimeofday(&ti2,NULL); /* read endtime in t2 */ fflush(stderr); fprintf(stderr, "\nTime cost host spmv code microseconds: %ld microseconds\n", ((ti2.tv_sec - ti1.tv_sec)*1000000L +ti2.tv_usec) - ti1.tv_usec ); std::cout << "\ndot=" << x.Dot(rhs) << std::endl; mat.ConvertToBCSR(); mat.info(); mat.MoveToAccelerator(); x.MoveToAccelerator(); rhs.MoveToAccelerator(); mat.info(); rhs.Ones(); // exit(1); gettimeofday(&ti1,NULL); /* read starttime in t1 */ mat.Apply(rhs, &x); gettimeofday(&ti2,NULL); /* read endtime in t2 */ fflush(stderr); fprintf(stderr, "\nTime cost for accelerator spmv microseconds: %ld microseconds\n", ((ti2.tv_sec - ti1.tv_sec)*1000000L +ti2.tv_usec) - ti1.tv_usec ); std::cout << "\ndot=" << x.Dot(rhs) << std::endl; stop_paralution(); return 0; }
int main(int argc, char* argv[]) { if (argc == 1) { std::cerr << argv[0] << " <matrix> [Num threads]" << std::endl; exit(1); } init_paralution(); if (argc > 2) { set_omp_threads_paralution(atoi(argv[2])); } info_paralution(); LocalVector<double> b, b_old, *b_k, *b_k1, *b_tmp; LocalMatrix<double> mat; mat.ReadFileMTX(std::string(argv[1])); // Gershgorin spectrum approximation double glambda_min, glambda_max; // Power method spectrum approximation double plambda_min, plambda_max; // Maximum number of iteration for the power method int iter_max = 10000; double tick, tack; // Gershgorin approximation of the eigenvalues mat.Gershgorin(glambda_min, glambda_max); std::cout << "Gershgorin : Lambda min = " << glambda_min << "; Lambda max = " << glambda_max << std::endl; mat.MoveToAccelerator(); b.MoveToAccelerator(); b_old.MoveToAccelerator(); b.Allocate("b_k+1", mat.get_nrow()); b_k1 = &b; b_old.Allocate("b_k", mat.get_nrow()); b_k = &b_old; b_k->Ones(); mat.info(); tick = paralution_time(); // compute lambda max for (int i=0; i<=iter_max; ++i) { mat.Apply(*b_k, b_k1); // std::cout << b_k1->Dot(*b_k) << std::endl; b_k1->Scale(double(1.0)/b_k1->Norm()); b_tmp = b_k1; b_k1 = b_k; b_k = b_tmp; } // get lambda max (Rayleigh quotient) mat.Apply(*b_k, b_k1); plambda_max = b_k1->Dot(*b_k) ; tack = paralution_time(); std::cout << "Power method (lambda max) execution:" << (tack-tick)/1000000 << " sec" << std::endl; mat.AddScalarDiagonal(double(-1.0)*plambda_max); b_k->Ones(); tick = paralution_time(); // compute lambda min for (int i=0; i<=iter_max; ++i) { mat.Apply(*b_k, b_k1); // std::cout << b_k1->Dot(*b_k) + plambda_max << std::endl; b_k1->Scale(double(1.0)/b_k1->Norm()); b_tmp = b_k1; b_k1 = b_k; b_k = b_tmp; } // get lambda min (Rayleigh quotient) mat.Apply(*b_k, b_k1); plambda_min = (b_k1->Dot(*b_k) + plambda_max); // back to the original matrix mat.AddScalarDiagonal(plambda_max); tack = paralution_time(); std::cout << "Power method (lambda min) execution:" << (tack-tick)/1000000 << " sec" << std::endl; std::cout << "Power method Lambda min = " << plambda_min << "; Lambda max = " << plambda_max << "; iter=2x" << iter_max << std::endl; LocalVector<double> x; LocalVector<double> rhs; x.CloneBackend(mat); rhs.CloneBackend(mat); x.Allocate("x", mat.get_nrow()); rhs.Allocate("rhs", mat.get_nrow()); // Chebyshev iteration Chebyshev<LocalMatrix<double>, LocalVector<double>, double > ls; rhs.Ones(); x.Zeros(); ls.SetOperator(mat); ls.Set(plambda_min, plambda_max); ls.Build(); tick = paralution_time(); ls.Solve(rhs, &x); tack = paralution_time(); std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl; // PCG + Chebyshev polynomial CG<LocalMatrix<double>, LocalVector<double>, double > cg; AIChebyshev<LocalMatrix<double>, LocalVector<double>, double > p; // damping factor plambda_min = plambda_max / 7; p.Set(3, plambda_min, plambda_max); rhs.Ones(); x.Zeros(); cg.SetOperator(mat); cg.SetPreconditioner(p); cg.Build(); tick = paralution_time(); cg.Solve(rhs, &x); tack = paralution_time(); std::cout << "Solver execution:" << (tack-tick)/1000000 << " sec" << std::endl; stop_paralution(); return 0; }
int main(int argc, char* argv[]) { if (argc == 1) { std::cerr << argv[0] << " <matrix> <initial_guess> <rhs> [Num threads]" << std::endl; exit(1); } init_paralution(); // if (argc > 4) { // set_omp_threads_paralution(atoi(argv[5])); // } set_omp_threads_paralution(8); info_paralution(); struct timeval now; double tick, tack, b=0.0f,s=0.0f, lprep=0.0f, sol_norm, diff_norm, ones_norm; double *phi_ptr=NULL; int *bubmap_ptr=NULL, phisize, maxbmap, setlssd, lvst_offst; int xdim, ydim, zdim, defvex_perdirec, defvex_perdirec_y, defvex_perdirec_z; DPCG<LocalMatrix<double>, LocalVector<double>, double > ls; #ifdef BUBFLO xdim=atoi(argv[5]); setlssd=atoi(argv[6]); defvex_perdirec=atoi(argv[7]); lvst_offst=atoi(argv[8]); phisize=(xdim+2*lvst_offst)*(ydim+2*lvst_offst)*(zdim+2*lvst_offst); #endif LocalVector<double> x; LocalVector<double>refsol; LocalVector<double>refones; LocalVector<double>chk_r; LocalVector<double> rhs; LocalMatrix<double> mat; LocalVector<double> Dinvhalf_min; LocalVector<double> Dinvhalf_plus; #ifdef GUUS LocalMatrix<double> Zin; #endif mat.ReadFileMTX(std::string(argv[1])); mat.info(); #ifdef GUUS Zin.ReadFileMTX(std::string(argv[2])); Zin.info(); #endif x.Allocate("x", mat.get_nrow()); refsol.Allocate("refsol", mat.get_nrow()); refones.Allocate("refones", mat.get_nrow()); rhs.Allocate("rhs", mat.get_nrow()); chk_r.Allocate("chk_r", mat.get_nrow()); #ifdef BUBFLO x.ReadFileASCII(std::string(argv[2])); #endif rhs.ReadFileASCII(std::string(argv[3])); #ifdef GUUS x.SetRandom(0.0,1.0,1000); refsol.ReadFileASCII(std::string(argv[4])); refones.Ones(); #endif //refsol.Ones(); // // // Uncomment for GPU #ifdef GPURUN mat.MoveToAccelerator(); x.MoveToAccelerator(); rhs.MoveToAccelerator(); chk_r.MoveToAccelerator(); Dinvhalf_min.MoveToAccelerator(); Dinvhalf_plus.MoveToAccelerator(); #endif gettimeofday(&now, NULL); tick = now.tv_sec*1000000.0+(now.tv_usec); #ifdef BUBFLO if(setlssd){ LocalVector<double> phi; LocalVector<int> bubmap; phi.Allocate("PHI", phisize); bubmap.Allocate("bubmap",mat.get_nrow()); phi.ReadFileASCII(std::string(argv[4])); bubmap.LeaveDataPtr(&bubmap_ptr); phi.LeaveDataPtr(&phi_ptr); bubmap_create(phi_ptr, bubmap_ptr, xdim, xdim, xdim, mat.get_nrow(), &maxbmap, lvst_offst); phi.Clear(); } ls.Setxdim(xdim); ls.SetNVectors_eachdirec(defvex_perdirec+1, defvex_perdirec+2, defvex_perdirec+3); ls.Set_alldims(xdim, xdim, xdim); ls.Setlvst_offst(lvst_offst); ls.SetNVectors(defvex_perdirec); ls.SetZlssd(setlssd); #endif gettimeofday(&now, NULL); tack = now.tv_sec*1000000.0+(now.tv_usec); lprep=(tack-tick)/1000000; std::cout << "levelset_prep" << lprep << " sec" << std::endl; // Linear Solver // return 0; gettimeofday(&now, NULL); tick = now.tv_sec*1000000.0+(now.tv_usec); #ifdef SCALIN mat.ExtractInverseDiagonal_sqrt(&Dinvhalf_min, -1); mat.ExtractInverseDiagonal_sqrt(&Dinvhalf_plus, 1); mat.DiagonalMatrixMult(Dinvhalf_min); mat.DiagonalMatrixMult_fromL(Dinvhalf_min); //x.PointWiseMult(Dinvhalf_plus); rhs.PointWiseMult(Dinvhalf_min); // rhs.Scale(0.3); #endif #ifdef GUUS ls.SetZ(Zin); #endif ls.SetOperator(mat); ls.Init(0.0, 1e-6, 1e8, 200000); // ls.RecordResidualHistory(); #ifdef BUBFLO ls.MakeZ_CSR(); // requires xdim_ and novecni_ and zlssd_ to be set if(setlssd) ls.MakeZLSSD(bubmap_ptr, maxbmap); // bubmap must be ready and maxbmap available #endif // // stop_paralution(); // return 0; ls.Build(); #ifdef MATDIA mat.ConvertToDIA(); #endif gettimeofday(&now, NULL); tack = now.tv_sec*1000000.0+(now.tv_usec); b=(tack-tick)/1000000; std::cout << "Building:" << b+lprep << " sec" << std::endl; // ls.Verbose(2); mat.info(); gettimeofday(&now, NULL); tick = now.tv_sec*1000000.0+(now.tv_usec); ls.Solve(rhs, &x); gettimeofday(&now, NULL); tack = now.tv_sec*1000000.0+(now.tv_usec); s=(tack-tick)/1000000; std::cout << "Solver execution:" << s << " sec" << std::endl; std::cout << "Total execution:" << s+b << " sec" << std::endl; #ifdef SCALIN x.PointWiseMult(Dinvhalf_min); #endif // #ifdef GUUS // x.WriteFileASCII("x_solution_shell_inv_neumann.rec"); //ls.RecordHistory("res__ongpu_tns.rec"); x.MoveToHost(); x.WriteFileASCII("x_neumann.rec"); x.MoveToAccelerator(); sol_norm=x.Norm(); mat.Apply(x, &chk_r); chk_r.ScaleAdd(double(-1.0), rhs); cout<<"\n Real Residual Norm is "<<chk_r.Norm(); cout<<"\n Norm of Solution is "<<sol_norm<<endl; cout<<"\n Norm of Reference Solution is "<<refsol.Norm()<<endl; cout<<"\n Norm of Ones is "<<refones.Norm()<<endl; x.MoveToHost(); refones.AddScale(x,(double)-1.0f); x.AddScale(refsol,(double)-1.0f); diff_norm=x.Norm(); ones_norm=refones.Norm(); cout<<"\n Relative Norm of Calculated Solution w.r.t. Reference is "<<((double)diff_norm/(double)sol_norm)<<endl; cout<<"\n Relative Norm of Calculated Solution w.r.t. Ones is "<<((double)ones_norm/(double)sol_norm)<<endl; #endif ls.Clear(); stop_paralution(); return 0; }