void magmaf_dprint_gpu( magma_int_t *m, magma_int_t *n, devptr_t *dA, magma_int_t *ldda ) { magma_dprint_gpu( *m, *n, magma_ddevptr(dA), *ldda ); }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dprint */ int main( int argc, char** argv) { TESTING_INIT(); double *hA; magmaDouble_ptr dA; //magma_int_t ione = 1; //magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t M, N, lda, ldda; //size magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; lda = M; ldda = magma_roundup( M, opts.align ); // multiple of 32 by default //size = lda*N; /* Allocate host memory for the matrix */ TESTING_MALLOC_CPU( hA, double, lda *N ); TESTING_MALLOC_DEV( dA, double, ldda*N ); //lapackf77_dlarnv( &ione, ISEED, &size, hA ); for( int j = 0; j < N; ++j ) { for( int i = 0; i < M; ++i ) { hA[i + j*lda] = MAGMA_D_MAKE( i + j*0.01, 0. ); } } magma_dsetmatrix( M, N, hA, lda, dA, ldda ); printf( "A=" ); magma_dprint( M, N, hA, lda ); printf( "dA=" ); magma_dprint_gpu( M, N, dA, ldda ); TESTING_FREE_CPU( hA ); TESTING_FREE_DEV( dA ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dgetrf */ int main( int argc, char** argv) { TESTING_INIT(); magma_setdevice( 0 ); double *hA, *dA; /* Matrix size */ magma_int_t m = 5; magma_int_t n = 10; //magma_int_t ione = 1; //magma_int_t ISEED[4] = {0,0,0,1}; //magma_int_t size; magma_int_t lda, ldda; lda = ((m + 31)/32)*32; ldda = ((m + 31)/32)*32; /* Allocate host memory for the matrix */ TESTING_MALLOC_CPU( hA, double, lda *n ); TESTING_MALLOC_DEV( dA, double, ldda*n ); //size = lda*n; //lapackf77_dlarnv( &ione, ISEED, &size, hA ); for( int j = 0; j < n; ++j ) { for( int i = 0; i < m; ++i ) { hA[i + j*lda] = MAGMA_D_MAKE( i + j*0.01, 0. ); } } magma_dsetmatrix( m, n, hA, lda, dA, ldda ); printf( "A=" ); magma_dprint( m, n, hA, lda ); printf( "dA=" ); magma_dprint_gpu( m, n, dA, ldda ); //printf( "dA=" ); //magma_dprint( m, n, dA, ldda ); //printf( "A=" ); //magma_dprint_gpu( m, n, hA, lda ); /* Memory clean up */ TESTING_FREE_CPU( hA ); TESTING_FREE_DEV( dA ); /* Shutdown */ TESTING_FINALIZE(); }
extern "C" magma_int_t magma_dlobpcg( magma_d_sparse_matrix A, magma_d_solver_par *solver_par ){ #define residualNorms(i,iter) ( residualNorms + (i) + (iter)*n ) #define magmablas_swap(x, y) { pointer = x; x = y; y = pointer; } #define hresidualNorms(i,iter) (hresidualNorms + (i) + (iter)*n ) #define gramA( m, n) (gramA + (m) + (n)*ldgram) #define gramB( m, n) (gramB + (m) + (n)*ldgram) #define gevectors(m, n) (gevectors + (m) + (n)*ldgram) #define h_gramB( m, n) (h_gramB + (m) + (n)*ldgram) #define magma_d_bspmv_tuned(m, n, alpha, A, X, beta, AX) { \ magmablas_dtranspose( m, n, X, m, blockW, n ); \ magma_d_vector x, ax; \ x.memory_location = Magma_DEV; x.num_rows = m*n; x.nnz = m*n; x.val = blockW; \ ax.memory_location= Magma_DEV; ax.num_rows = m*n; ax.nnz = m*n; ax.val = AX; \ magma_d_spmv(alpha, A, x, beta, ax ); \ magmablas_dtranspose( n, m, blockW, n, X, m ); \ } //************************************************************** // Memory allocation for the eigenvectors, eigenvalues, and workspace solver_par->solver = Magma_LOBPCG; magma_int_t m = A.num_rows; magma_int_t n =(solver_par->num_eigenvalues); double *blockX = solver_par->eigenvectors; double *evalues = solver_par->eigenvalues; double *dwork, *hwork; double *blockP, *blockAP, *blockR, *blockAR, *blockAX, *blockW; double *gramA, *gramB, *gramM; double *gevectors, *h_gramB; double *pointer, *origX = blockX; double *eval_gpu; magma_int_t lwork = max( 2*n+n*magma_get_dsytrd_nb(n), 1 + 6*3*n + 2* 3*n* 3*n); magma_dmalloc_pinned( &hwork , lwork ); magma_dmalloc( &blockAX , m*n ); magma_dmalloc( &blockAR , m*n ); magma_dmalloc( &blockAP , m*n ); magma_dmalloc( &blockR , m*n ); magma_dmalloc( &blockP , m*n ); magma_dmalloc( &blockW , m*n ); magma_dmalloc( &dwork , m*n ); magma_dmalloc( &eval_gpu , 3*n ); //**********************************************************+ magma_int_t verbosity = 1; magma_int_t *iwork, liwork = 15*n+9; // === Set solver parameters === double residualTolerance = solver_par->epsilon; magma_int_t maxIterations = solver_par->maxiter; // === Set some constants & defaults === double c_one = MAGMA_D_ONE, c_zero = MAGMA_D_ZERO; double *residualNorms, *condestGhistory, condestG; double *gevalues; magma_int_t *activeMask; // === Check some parameters for possible quick exit === solver_par->info = 0; if (m < 2) solver_par->info = -1; else if (n > m) solver_par->info = -2; if (solver_par->info != 0) { magma_xerbla( __func__, -(solver_par->info) ); return solver_par->info; } magma_int_t *info = &(solver_par->info); // local info variable; // === Allocate GPU memory for the residual norms' history === magma_dmalloc(&residualNorms, (maxIterations+1) * n); magma_malloc( (void **)&activeMask, (n+1) * sizeof(magma_int_t) ); // === Allocate CPU work space === magma_dmalloc_cpu(&condestGhistory, maxIterations+1); magma_dmalloc_cpu(&gevalues, 3 * n); magma_malloc_cpu((void **)&iwork, liwork * sizeof(magma_int_t)); double *hW; magma_dmalloc_pinned(&hW, n*n); magma_dmalloc_pinned(&gevectors, 9*n*n); magma_dmalloc_pinned(&h_gramB , 9*n*n); // === Allocate GPU workspace === magma_dmalloc(&gramM, n * n); magma_dmalloc(&gramA, 9 * n * n); magma_dmalloc(&gramB, 9 * n * n); #if defined(PRECISION_z) || defined(PRECISION_c) double *rwork; magma_int_t lrwork = 1 + 5*(3*n) + 2*(3*n)*(3*n); magma_dmalloc_cpu(&rwork, lrwork); #endif // === Set activemask to one === for(int k =0; k<n; k++) iwork[k]=1; magma_setmatrix(n, 1, sizeof(magma_int_t), iwork, n ,activeMask, n); magma_int_t gramDim, ldgram = 3*n, ikind = 4; // === Make the initial vectors orthonormal === magma_dgegqr_gpu(ikind, m, n, blockX, m, dwork, hwork, info ); //magma_dorthomgs( m, n, blockX ); magma_d_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX ); // === Compute the Gram matrix = (X, AX) & its eigenstates === magma_dgemm(MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramM, n); magma_dsyevd_gpu( MagmaVec, MagmaUpper, n, gramM, n, evalues, hW, n, hwork, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, info ); // === Update X = X * evectors === magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockX, m, gramM, n, c_zero, blockW, m); magmablas_swap(blockW, blockX); // === Update AX = AX * evectors === magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockAX, m, gramM, n, c_zero, blockW, m); magmablas_swap(blockW, blockAX); condestGhistory[1] = 7.82; magma_int_t iterationNumber, cBlockSize, restart = 1, iter; //Chronometry real_Double_t tempo1, tempo2; magma_device_sync(); tempo1=magma_wtime(); // === Main LOBPCG loop ============================================================ for(iterationNumber = 1; iterationNumber < maxIterations; iterationNumber++) { // === compute the residuals (R = Ax - x evalues ) magmablas_dlacpy( MagmaUpperLower, m, n, blockAX, m, blockR, m); /* for(int i=0; i<n; i++){ magma_daxpy(m, MAGMA_D_MAKE(-evalues[i],0), blockX+i*m, 1, blockR+i*m, 1); } */ #if defined(PRECISION_z) || defined(PRECISION_d) magma_dsetmatrix( 3*n, 1, evalues, 3*n, eval_gpu, 3*n ); #else magma_ssetmatrix( 3*n, 1, evalues, 3*n, eval_gpu, 3*n ); #endif magma_dlobpcg_res( m, n, eval_gpu, blockX, blockR, eval_gpu); magmablas_dnrm2_cols(m, n, blockR, m, residualNorms(0, iterationNumber)); // === remove the residuals corresponding to already converged evectors magma_dcompact(m, n, blockR, m, residualNorms(0, iterationNumber), residualTolerance, activeMask, &cBlockSize); if (cBlockSize == 0) break; // === apply a preconditioner P to the active residulas: R_new = P R_old // === for now set P to be identity (no preconditioner => nothing to be done ) // magmablas_dlacpy( MagmaUpperLower, m, cBlockSize, blockR, m, blockW, m); /* // === make the preconditioned residuals orthogonal to X magma_dgemm(MagmaConjTrans, MagmaNoTrans, n, cBlockSize, m, c_one, blockX, m, blockR, m, c_zero, gramB(0,0), ldgram); magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n, c_mone, blockX, m, gramB(0,0), ldgram, c_one, blockR, m); */ // === make the active preconditioned residuals orthonormal magma_dgegqr_gpu(ikind, m, cBlockSize, blockR, m, dwork, hwork, info ); //magma_dorthomgs( m, cBlockSize, blockR ); // === compute AR magma_d_bspmv_tuned(m, cBlockSize, c_one, A, blockR, c_zero, blockAR ); if (!restart) { // === compact P & AP as well magma_dcompactActive(m, n, blockP, m, activeMask); magma_dcompactActive(m, n, blockAP, m, activeMask); /* // === make P orthogonal to X ? magma_dgemm(MagmaConjTrans, MagmaNoTrans, n, cBlockSize, m, c_one, blockX, m, blockP, m, c_zero, gramB(0,0), ldgram); magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n, c_mone, blockX, m, gramB(0,0), ldgram, c_one, blockP, m); // === make P orthogonal to R ? magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockR, m, blockP, m, c_zero, gramB(0,0), ldgram); magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, cBlockSize, c_mone, blockR, m, gramB(0,0), ldgram, c_one, blockP, m); */ // === Make P orthonormal & properly change AP (without multiplication by A) magma_dgegqr_gpu(ikind, m, cBlockSize, blockP, m, dwork, hwork, info ); //magma_dorthomgs( m, cBlockSize, blockP ); //magma_d_bspmv_tuned(m, cBlockSize, c_one, A, blockP, c_zero, blockAP ); magma_dsetmatrix( cBlockSize, cBlockSize, hwork, cBlockSize, dwork, cBlockSize); // magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, // m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m); // replacement according to Stan #if defined(PRECISION_s) || defined(PRECISION_d) magmablas_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m); #else magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m); #endif } iter = max(1,iterationNumber-10- (int)(log(1.*cBlockSize))); double condestGmean = 0.; for(int i = 0; i<iterationNumber-iter+1; i++) condestGmean += condestGhistory[i]; condestGmean = condestGmean / (iterationNumber-iter+1); if (restart) gramDim = n+cBlockSize; else gramDim = n+2*cBlockSize; /* --- The Raileight-Ritz method for [X R P] ----------------------- [ X R P ]' [AX AR AP] y = evalues [ X R P ]' [ X R P ], i.e., GramA GramB / X'AX X'AR X'AP \ / X'X X'R X'P \ | R'AX R'AR R'AP | y = evalues | R'X R'R R'P | \ P'AX P'AR P'AP / \ P'X P'R P'P / ----------------------------------------------------------------- */ // === assemble GramB; first, set it to I magmablas_dlaset(MagmaFull, ldgram, ldgram, c_zero, c_one, gramB, ldgram); // identity if (!restart) { magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockP, m, blockX, m, c_zero, gramB(n+cBlockSize,0), ldgram); magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockR, m, c_zero, gramB(n+cBlockSize,n), ldgram); } magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockR, m, blockX, m, c_zero, gramB(n,0), ldgram); // === get GramB from the GPU to the CPU and compute its eigenvalues only magma_dgetmatrix(gramDim, gramDim, gramB, ldgram, h_gramB, ldgram); lapackf77_dsyev("N", "L", &gramDim, h_gramB, &ldgram, gevalues, hwork, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, #endif info); // === check stability criteria if we need to restart condestG = log10( gevalues[gramDim-1]/gevalues[0] ) + 1.; if ((condestG/condestGmean>2 && condestG>2) || condestG>8) { // Steepest descent restart for stability restart=1; printf("restart at step #%d\n", (int) iterationNumber); } // === assemble GramA; first, set it to I magmablas_dlaset(MagmaFull, ldgram, ldgram, c_zero, c_one, gramA, ldgram); // identity magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockR, m, blockAX, m, c_zero, gramA(n,0), ldgram); magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockR, m, blockAR, m, c_zero, gramA(n,n), ldgram); if (!restart) { magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockP, m, blockAX, m, c_zero, gramA(n+cBlockSize,0), ldgram); magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockAR, m, c_zero, gramA(n+cBlockSize,n), ldgram); magma_dgemm(MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockAP, m, c_zero, gramA(n+cBlockSize,n+cBlockSize), ldgram); } /* // === Compute X' AX or just use the eigenvalues below ? magma_dgemm(MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramA(0,0), ldgram); */ if (restart==0) { magma_dgetmatrix(gramDim, gramDim, gramA, ldgram, gevectors, ldgram); } else { gramDim = n+cBlockSize; magma_dgetmatrix(gramDim, gramDim, gramA, ldgram, gevectors, ldgram); } for(int k=0; k<n; k++) *gevectors(k,k) = MAGMA_D_MAKE(evalues[k], 0); // === the previous eigensolver destroyed what is in h_gramB => must copy it again magma_dgetmatrix(gramDim, gramDim, gramB, ldgram, h_gramB, ldgram); magma_int_t itype = 1; lapackf77_dsygvd(&itype, "V", "L", &gramDim, gevectors, &ldgram, h_gramB, &ldgram, gevalues, hwork, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); for(int k =0; k<n; k++) evalues[k] = gevalues[k]; // === copy back the result to gramA on the GPU and use it for the updates magma_dsetmatrix(gramDim, gramDim, gevectors, ldgram, gramA, ldgram); if (restart == 0) { // === contribution from P to the new X (in new search direction P) magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m); magmablas_swap(dwork, blockP); // === contribution from R to the new X (in new search direction P) magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockR, m, gramA(n,0), ldgram, c_one, blockP, m); // === corresponding contribution from AP to the new AX (in AP) magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockAP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m); magmablas_swap(dwork, blockAP); // === corresponding contribution from AR to the new AX (in AP) magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockAR, m, gramA(n,0), ldgram, c_one, blockAP, m); } else { // === contribution from R (only) to the new X magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockR, m, gramA(n,0), ldgram, c_zero, blockP, m); // === corresponding contribution from AR (only) to the new AX magma_dgemm(MagmaNoTrans, MagmaNoTrans,m, n, cBlockSize, c_one, blockAR, m, gramA(n,0), ldgram, c_zero, blockAP, m); } // === contribution from old X to the new X + the new search direction P magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockX, m, gramA, ldgram, c_zero, dwork, m); magmablas_swap(dwork, blockX); //magma_daxpy(m*n, c_one, blockP, 1, blockX, 1); magma_dlobpcg_maxpy( m, n, blockP, blockX ); // === corresponding contribution from old AX to new AX + AP magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockAX, m, gramA, ldgram, c_zero, dwork, m); magmablas_swap(dwork, blockAX); //magma_daxpy(m*n, c_one, blockAP, 1, blockAX, 1); magma_dlobpcg_maxpy( m, n, blockAP, blockAX ); condestGhistory[iterationNumber+1]=condestG; if (verbosity==1) { // double res; // magma_dgetmatrix(1, 1, // (double*)residualNorms(0, iterationNumber), 1, // (double*)&res, 1); // // printf("Iteration %4d, CBS %4d, Residual: %10.7f\n", // iterationNumber, cBlockSize, res); printf("%4d-%2d ", (int) iterationNumber, (int) cBlockSize); magma_dprint_gpu(1, n, residualNorms(0, iterationNumber), 1); } restart = 0; } // === end for iterationNumber = 1,maxIterations ======================= // fill solver info magma_device_sync(); tempo2=magma_wtime(); solver_par->runtime = (real_Double_t) tempo2-tempo1; solver_par->numiter = iterationNumber; if( solver_par->numiter < solver_par->maxiter){ solver_par->info = 0; }else if( solver_par->init_res > solver_par->final_res ) solver_par->info = -2; else solver_par->info = -1; // ============================================================================= // === postprocessing; // ============================================================================= // === compute the real AX and corresponding eigenvalues magma_d_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX ); magma_dgemm(MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramM, n); magma_dsyevd_gpu( MagmaVec, MagmaUpper, n, gramM, n, gevalues, dwork, n, hwork, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, info ); for(int k =0; k<n; k++) evalues[k] = gevalues[k]; // === update X = X * evectors magmablas_swap(blockX, dwork); magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, dwork, m, gramM, n, c_zero, blockX, m); // === update AX = AX * evectors to compute the final residual magmablas_swap(blockAX, dwork); magma_dgemm(MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, dwork, m, gramM, n, c_zero, blockAX, m); // === compute R = AX - evalues X magmablas_dlacpy( MagmaUpperLower, m, n, blockAX, m, blockR, m); for(int i=0; i<n; i++) magma_daxpy(m, MAGMA_D_MAKE(-evalues[i], 0), blockX+i*m, 1, blockR+i*m, 1); // === residualNorms[iterationNumber] = || R || magmablas_dnrm2_cols(m, n, blockR, m, residualNorms(0, iterationNumber)); // === restore blockX if needed if (blockX != origX) magmablas_dlacpy( MagmaUpperLower, m, n, blockX, m, origX, m); printf("Eigenvalues:\n"); for(int i =0; i<n; i++) printf("%e ", evalues[i]); printf("\n\n"); printf("Final residuals:\n"); magma_dprint_gpu(1, n, residualNorms(0, iterationNumber), 1); printf("\n\n"); //=== Print residual history in a file for plotting ==== double *hresidualNorms; magma_dmalloc_cpu(&hresidualNorms, (iterationNumber+1) * n); magma_dgetmatrix(n, iterationNumber, (double*)residualNorms, n, (double*)hresidualNorms, n); printf("Residuals are stored in file residualNorms\n"); printf("Plot the residuals using: myplot \n"); FILE *residuals_file; residuals_file = fopen("residualNorms", "w"); for(int i =1; i<iterationNumber; i++) { for(int j = 0; j<n; j++) fprintf(residuals_file, "%f ", *hresidualNorms(j,i)); fprintf(residuals_file, "\n"); } fclose(residuals_file); magma_free_cpu(hresidualNorms); // === free work space magma_free( residualNorms ); magma_free_cpu( condestGhistory ); magma_free_cpu( gevalues ); magma_free_cpu( iwork ); magma_free_pinned( hW ); magma_free_pinned( gevectors ); magma_free_pinned( h_gramB ); magma_free( gramM ); magma_free( gramA ); magma_free( gramB ); magma_free( activeMask ); magma_free( blockAX ); magma_free( blockAR ); magma_free( blockAP ); magma_free( blockR ); magma_free( blockP ); magma_free( blockW ); magma_free( dwork ); magma_free( eval_gpu ); magma_free_pinned( hwork ); #if defined(PRECISION_z) || defined(PRECISION_c) magma_free_cpu( rwork ); #endif return MAGMA_SUCCESS; }
extern "C" magma_int_t magma_zlobpcg( magma_z_matrix A, magma_z_solver_par *solver_par, magma_z_preconditioner *precond_par, magma_queue_t queue ) { magma_int_t info = 0; #define residualNorms(i,iter) ( residualNorms + (i) + (iter)*n ) #define SWAP(x, y) { pointer = x; x = y; y = pointer; } #define hresidualNorms(i,iter) (hresidualNorms + (i) + (iter)*n ) #define gramA( m, n) (gramA + (m) + (n)*ldgram) #define gramB( m, n) (gramB + (m) + (n)*ldgram) #define gevectors(m, n) (gevectors + (m) + (n)*ldgram) #define h_gramB( m, n) (h_gramB + (m) + (n)*ldgram) #define magma_z_bspmv_tuned(m, n, alpha, A, X, beta, AX, queue) { \ magma_z_matrix x={Magma_CSR}, ax={Magma_CSR}; \ x.memory_location = Magma_DEV; x.num_rows = m; x.num_cols = n; x.major = MagmaColMajor; x.nnz = m*n; x.dval = X; x.storage_type = Magma_DENSE; \ ax.memory_location= Magma_DEV; ax.num_rows = m; ax.num_cols = n; ax.major = MagmaColMajor; ax.nnz = m*n; ax.dval = AX; ax.storage_type = Magma_DENSE; \ CHECK( magma_z_spmv(alpha, A, x, beta, ax, queue )); \ } //************************************************************** // Memory allocation for the eigenvectors, eigenvalues, and workspace solver_par->solver = Magma_LOBPCG; magma_int_t m = A.num_rows; magma_int_t n = (solver_par->num_eigenvalues); magmaDoubleComplex *blockX = solver_par->eigenvectors; double *evalues = solver_par->eigenvalues; solver_par->numiter = 0; solver_par->spmv_count = 0; magmaDoubleComplex *dwork=NULL, *hwork=NULL; magmaDoubleComplex *blockP=NULL, *blockAP=NULL, *blockR=NULL, *blockAR=NULL, *blockAX=NULL, *blockW=NULL; magmaDoubleComplex *gramA=NULL, *gramB=NULL, *gramM=NULL; magmaDoubleComplex *gevectors=NULL, *h_gramB=NULL; dwork = NULL; hwork = NULL; blockP = NULL; blockR = NULL; blockAP = NULL; blockAR = NULL; blockAX = NULL; blockW = NULL; gramA = NULL; gramB = NULL; gramM = NULL; gevectors = NULL; h_gramB = NULL; magmaDoubleComplex *pointer, *origX = blockX; double *eval_gpu=NULL; magma_int_t iterationNumber, cBlockSize, restart = 1, iter; //Chronometry real_Double_t tempo1, tempo2, tempop1, tempop2; magma_int_t lwork = max( 2*n+n*magma_get_dsytrd_nb(n), 1 + 6*3*n + 2* 3*n* 3*n); magma_int_t *iwork={0}, liwork = 15*n+9; magma_int_t gramDim, ldgram = 3*n, ikind = 3; magmaDoubleComplex *hW={0}; // === Set solver parameters === double residualTolerance = solver_par->rtol; magma_int_t maxIterations = solver_par->maxiter; double tmp; double r0=0; // set in 1st iteration // === Set some constants & defaults === magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; double *residualNorms={0}, *condestGhistory={0}, condestG={0}; double *gevalues={0}; magma_int_t *activeMask={0}; double *hresidualNorms={0}; #ifdef COMPLEX double *rwork={0}; magma_int_t lrwork = 1 + 5*(3*n) + 2*(3*n)*(3*n); CHECK( magma_dmalloc_cpu(&rwork, lrwork)); #endif CHECK( magma_zmalloc_pinned( &hwork , lwork )); CHECK( magma_zmalloc( &blockAX , m*n )); CHECK( magma_zmalloc( &blockAR , m*n )); CHECK( magma_zmalloc( &blockAP , m*n )); CHECK( magma_zmalloc( &blockR , m*n )); CHECK( magma_zmalloc( &blockP , m*n )); CHECK( magma_zmalloc( &blockW , m*n )); CHECK( magma_zmalloc( &dwork , m*n )); CHECK( magma_dmalloc( &eval_gpu , 3*n )); //**********************************************************+ // === Check some parameters for possible quick exit === solver_par->info = MAGMA_SUCCESS; if (m < 2) info = MAGMA_DIVERGENCE; else if (n > m) info = MAGMA_SLOW_CONVERGENCE; if (solver_par->info != 0) { magma_xerbla( __func__, -(info) ); goto cleanup; } solver_par->info = info; // local info variable; // === Allocate GPU memory for the residual norms' history === CHECK( magma_dmalloc(&residualNorms, (maxIterations+1) * n)); CHECK( magma_malloc( (void **)&activeMask, (n+1) * sizeof(magma_int_t) )); // === Allocate CPU work space === CHECK( magma_dmalloc_cpu(&condestGhistory, maxIterations+1)); CHECK( magma_dmalloc_cpu(&gevalues, 3 * n)); CHECK( magma_malloc_cpu((void **)&iwork, liwork * sizeof(magma_int_t))); CHECK( magma_zmalloc_pinned(&hW, n*n)); CHECK( magma_zmalloc_pinned(&gevectors, 9*n*n)); CHECK( magma_zmalloc_pinned(&h_gramB , 9*n*n)); // === Allocate GPU workspace === CHECK( magma_zmalloc(&gramM, n * n)); CHECK( magma_zmalloc(&gramA, 9 * n * n)); CHECK( magma_zmalloc(&gramB, 9 * n * n)); // === Set activemask to one === for(magma_int_t k =0; k<n; k++){ iwork[k]=1; } magma_setmatrix(n, 1, sizeof(magma_int_t), iwork, n , activeMask, n, queue); #if defined(PRECISION_s) ikind = 3; #endif // === Make the initial vectors orthonormal === magma_zgegqr_gpu(ikind, m, n, blockX, m, dwork, hwork, &info ); //magma_zorthomgs( m, n, blockX, queue ); magma_z_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX, queue ); solver_par->spmv_count++; // === Compute the Gram matrix = (X, AX) & its eigenstates === magma_zgemm( MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramM, n, queue ); magma_zheevd_gpu( MagmaVec, MagmaUpper, n, gramM, n, evalues, hW, n, hwork, lwork, #ifdef COMPLEX rwork, lrwork, #endif iwork, liwork, &info ); // === Update X = X * evectors === magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockX, m, gramM, n, c_zero, blockW, m, queue ); SWAP(blockW, blockX); // === Update AX = AX * evectors === magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockAX, m, gramM, n, c_zero, blockW, m, queue ); SWAP(blockW, blockAX); condestGhistory[1] = 7.82; tempo1 = magma_sync_wtime( queue ); // === Main LOBPCG loop ============================================================ for(iterationNumber = 1; iterationNumber < maxIterations; iterationNumber++) { // === compute the residuals (R = Ax - x evalues ) magmablas_zlacpy( MagmaFull, m, n, blockAX, m, blockR, m, queue ); /* for(magma_int_t i=0; i<n; i++) { magma_zaxpy( m, MAGMA_Z_MAKE(-evalues[i],0), blockX+i*m, 1, blockR+i*m, 1, queue ); } */ magma_dsetmatrix( 3*n, 1, evalues, 3*n, eval_gpu, 3*n, queue ); CHECK( magma_zlobpcg_res( m, n, eval_gpu, blockX, blockR, eval_gpu, queue )); magmablas_dznrm2_cols( m, n, blockR, m, residualNorms(0, iterationNumber), queue ); // === remove the residuals corresponding to already converged evectors CHECK( magma_zcompact(m, n, blockR, m, residualNorms(0, iterationNumber), residualTolerance, activeMask, &cBlockSize, queue )); if (cBlockSize == 0) break; // === apply a preconditioner P to the active residulas: R_new = P R_old // === for now set P to be identity (no preconditioner => nothing to be done ) //magmablas_zlacpy( MagmaFull, m, cBlockSize, blockR, m, blockW, m, queue ); //SWAP(blockW, blockR); // preconditioner magma_z_matrix bWv={Magma_CSR}, bRv={Magma_CSR}; bWv.memory_location = Magma_DEV; bWv.num_rows = m; bWv.num_cols = cBlockSize; bWv.major = MagmaColMajor; bWv.nnz = m*cBlockSize; bWv.dval = blockW; bRv.memory_location = Magma_DEV; bRv.num_rows = m; bRv.num_cols = cBlockSize; bRv.major = MagmaColMajor; bRv.nnz = m*cBlockSize; bRv.dval = blockR; tempop1 = magma_sync_wtime( queue ); CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, bRv, &bWv, precond_par, queue )); CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, bWv, &bRv, precond_par, queue )); tempop2 = magma_sync_wtime( queue ); precond_par->runtime += tempop2-tempop1; // === make the preconditioned residuals orthogonal to X if( precond_par->solver != Magma_NONE){ magma_zgemm( MagmaConjTrans, MagmaNoTrans, n, cBlockSize, m, c_one, blockX, m, blockR, m, c_zero, gramB(0,0), ldgram, queue ); magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n, c_neg_one, blockX, m, gramB(0,0), ldgram, c_one, blockR, m, queue ); } // === make the active preconditioned residuals orthonormal magma_zgegqr_gpu(ikind, m, cBlockSize, blockR, m, dwork, hwork, &info ); #if defined(PRECISION_s) // re-orthogonalization SWAP(blockX, dwork); magma_zgegqr_gpu(ikind, m, cBlockSize, blockR, m, dwork, hwork, &info ); #endif //magma_zorthomgs( m, cBlockSize, blockR, queue ); // === compute AR magma_z_bspmv_tuned(m, cBlockSize, c_one, A, blockR, c_zero, blockAR, queue ); solver_par->spmv_count++; if (!restart) { // === compact P & AP as well CHECK( magma_zcompactActive(m, n, blockP, m, activeMask, queue )); CHECK( magma_zcompactActive(m, n, blockAP, m, activeMask, queue )); /* // === make P orthogonal to X ? magma_zgemm( MagmaConjTrans, MagmaNoTrans, n, cBlockSize, m, c_one, blockX, m, blockP, m, c_zero, gramB(0,0), ldgram, queue ); magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n, c_neg_one, blockX, m, gramB(0,0), ldgram, c_one, blockP, m, queue ); // === make P orthogonal to R ? magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockR, m, blockP, m, c_zero, gramB(0,0), ldgram, queue ); magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, cBlockSize, cBlockSize, c_neg_one, blockR, m, gramB(0,0), ldgram, c_one, blockP, m, queue ); */ // === Make P orthonormal & properly change AP (without multiplication by A) magma_zgegqr_gpu(ikind, m, cBlockSize, blockP, m, dwork, hwork, &info ); #if defined(PRECISION_s) // re-orthogonalization SWAP(blockX, dwork); magma_zgegqr_gpu(ikind, m, cBlockSize, blockP, m, dwork, hwork, &info ); #endif //magma_zorthomgs( m, cBlockSize, blockP, queue ); //magma_z_bspmv_tuned(m, cBlockSize, c_one, A, blockP, c_zero, blockAP, queue ); magma_zsetmatrix( cBlockSize, cBlockSize, hwork, cBlockSize, dwork, cBlockSize, queue ); // replacement according to Stan #if defined(PRECISION_s) || defined(PRECISION_d) magmablas_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m, queue ); #else magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m, queue ); #endif } iter = max( 1, iterationNumber - 10 - int(log(1.*cBlockSize)) ); double condestGmean = 0.; for(magma_int_t i = 0; i<iterationNumber-iter+1; i++){ condestGmean += condestGhistory[i]; } condestGmean = condestGmean / (iterationNumber-iter+1); if (restart) gramDim = n+cBlockSize; else gramDim = n+2*cBlockSize; /* --- The Raileight-Ritz method for [X R P] ----------------------- [ X R P ]' [AX AR AP] y = evalues [ X R P ]' [ X R P ], i.e., GramA GramB / X'AX X'AR X'AP \ / X'X X'R X'P \ | R'AX R'AR R'AP | y = evalues | R'X R'R R'P | \ P'AX P'AR P'AP / \ P'X P'R P'P / ----------------------------------------------------------------- */ // === assemble GramB; first, set it to I magmablas_zlaset( MagmaFull, ldgram, ldgram, c_zero, c_one, gramB, ldgram, queue ); // identity if (!restart) { magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockP, m, blockX, m, c_zero, gramB(n+cBlockSize,0), ldgram, queue ); magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockR, m, c_zero, gramB(n+cBlockSize,n), ldgram, queue ); } magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockR, m, blockX, m, c_zero, gramB(n,0), ldgram, queue ); // === get GramB from the GPU to the CPU and compute its eigenvalues only magma_zgetmatrix( gramDim, gramDim, gramB, ldgram, h_gramB, ldgram, queue ); lapackf77_zheev("N", "L", &gramDim, h_gramB, &ldgram, gevalues, hwork, &lwork, #ifdef COMPLEX rwork, #endif &info); // === check stability criteria if we need to restart condestG = log10( gevalues[gramDim-1]/gevalues[0] ) + 1.; if ((condestG/condestGmean>2 && condestG>2) || condestG>8) { // Steepest descent restart for stability restart=1; printf("restart at step #%d\n", int(iterationNumber)); } // === assemble GramA; first, set it to I magmablas_zlaset( MagmaFull, ldgram, ldgram, c_zero, c_one, gramA, ldgram, queue ); // identity magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockR, m, blockAX, m, c_zero, gramA(n,0), ldgram, queue ); magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockR, m, blockAR, m, c_zero, gramA(n,n), ldgram, queue ); if (!restart) { magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, n, m, c_one, blockP, m, blockAX, m, c_zero, gramA(n+cBlockSize,0), ldgram, queue ); magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockAR, m, c_zero, gramA(n+cBlockSize,n), ldgram, queue ); magma_zgemm( MagmaConjTrans, MagmaNoTrans, cBlockSize, cBlockSize, m, c_one, blockP, m, blockAP, m, c_zero, gramA(n+cBlockSize,n+cBlockSize), ldgram, queue ); } /* // === Compute X' AX or just use the eigenvalues below ? magma_zgemm( MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramA(0,0), ldgram, queue ); */ if (restart==0) { magma_zgetmatrix( gramDim, gramDim, gramA, ldgram, gevectors, ldgram, queue ); } else { gramDim = n+cBlockSize; magma_zgetmatrix( gramDim, gramDim, gramA, ldgram, gevectors, ldgram, queue ); } for(magma_int_t k=0; k<n; k++) *gevectors(k,k) = MAGMA_Z_MAKE(evalues[k], 0); // === the previous eigensolver destroyed what is in h_gramB => must copy it again magma_zgetmatrix( gramDim, gramDim, gramB, ldgram, h_gramB, ldgram, queue ); magma_int_t itype = 1; lapackf77_zhegvd(&itype, "V", "L", &gramDim, gevectors, &ldgram, h_gramB, &ldgram, gevalues, hwork, &lwork, #ifdef COMPLEX rwork, &lrwork, #endif iwork, &liwork, &info); for(magma_int_t k =0; k<n; k++) evalues[k] = gevalues[k]; // === copy back the result to gramA on the GPU and use it for the updates magma_zsetmatrix( gramDim, gramDim, gevectors, ldgram, gramA, ldgram, queue ); if (restart == 0) { // === contribution from P to the new X (in new search direction P) magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m, queue ); SWAP(dwork, blockP); // === contribution from R to the new X (in new search direction P) magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockR, m, gramA(n,0), ldgram, c_one, blockP, m, queue ); // === corresponding contribution from AP to the new AX (in AP) magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockAP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m, queue ); SWAP(dwork, blockAP); // === corresponding contribution from AR to the new AX (in AP) magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockAR, m, gramA(n,0), ldgram, c_one, blockAP, m, queue ); } else { // === contribution from R (only) to the new X magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize, c_one, blockR, m, gramA(n,0), ldgram, c_zero, blockP, m, queue ); // === corresponding contribution from AR (only) to the new AX magma_zgemm( MagmaNoTrans, MagmaNoTrans,m, n, cBlockSize, c_one, blockAR, m, gramA(n,0), ldgram, c_zero, blockAP, m, queue ); } // === contribution from old X to the new X + the new search direction P magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockX, m, gramA, ldgram, c_zero, dwork, m, queue ); SWAP(dwork, blockX); //magma_zaxpy( m*n, c_one, blockP, 1, blockX, 1, queue ); CHECK( magma_zlobpcg_maxpy( m, n, blockP, blockX, queue )); // === corresponding contribution from old AX to new AX + AP magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, blockAX, m, gramA, ldgram, c_zero, dwork, m, queue ); SWAP(dwork, blockAX); //magma_zaxpy( m*n, c_one, blockAP, 1, blockAX, 1, queue ); CHECK( magma_zlobpcg_maxpy( m, n, blockAP, blockAX, queue )); condestGhistory[iterationNumber+1]=condestG; magma_dgetmatrix( 1, 1, residualNorms(0, iterationNumber), 1, &tmp, 1, queue ); if ( iterationNumber == 1 ) { solver_par->init_res = tmp; r0 = tmp * solver_par->rtol; if ( r0 < ATOLERANCE ) r0 = ATOLERANCE; } solver_par->final_res = tmp; if ( tmp < r0 ) { break; } if (cBlockSize == 0) { break; } if ( solver_par->verbose!=0 ) { if ( iterationNumber%solver_par->verbose == 0 ) { // double res; // magma_zgetmatrix( 1, 1, // (magmaDoubleComplex*)residualNorms(0, iterationNumber), 1, // (magmaDoubleComplex*)&res, 1, queue ); // // printf("Iteration %4d, CBS %4d, Residual: %10.7f\n", // iterationNumber, cBlockSize, res); printf("%4d-%2d ", int(iterationNumber), int(cBlockSize)); magma_dprint_gpu(1, n, residualNorms(0, iterationNumber), 1); } } restart = 0; } // === end for iterationNumber = 1,maxIterations ======================= // fill solver info tempo2 = magma_sync_wtime( queue ); solver_par->runtime = (real_Double_t) tempo2-tempo1; solver_par->numiter = iterationNumber; if ( solver_par->numiter < solver_par->maxiter) { info = MAGMA_SUCCESS; } else if ( solver_par->init_res > solver_par->final_res ) info = MAGMA_SLOW_CONVERGENCE; else info = MAGMA_DIVERGENCE; // ============================================================================= // === postprocessing; // ============================================================================= // === compute the real AX and corresponding eigenvalues magma_z_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX, queue ); magma_zgemm( MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, blockX, m, blockAX, m, c_zero, gramM, n, queue ); magma_zheevd_gpu( MagmaVec, MagmaUpper, n, gramM, n, gevalues, dwork, n, hwork, lwork, #ifdef COMPLEX rwork, lrwork, #endif iwork, liwork, &info ); for(magma_int_t k =0; k<n; k++) evalues[k] = gevalues[k]; // === update X = X * evectors SWAP(blockX, dwork); magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, dwork, m, gramM, n, c_zero, blockX, m, queue ); // === update AX = AX * evectors to compute the final residual SWAP(blockAX, dwork); magma_zgemm( MagmaNoTrans, MagmaNoTrans, m, n, n, c_one, dwork, m, gramM, n, c_zero, blockAX, m, queue ); // === compute R = AX - evalues X magmablas_zlacpy( MagmaFull, m, n, blockAX, m, blockR, m, queue ); for(magma_int_t i=0; i<n; i++) magma_zaxpy( m, MAGMA_Z_MAKE(-evalues[i], 0), blockX+i*m, 1, blockR+i*m, 1, queue ); // === residualNorms[iterationNumber] = || R || magmablas_dznrm2_cols( m, n, blockR, m, residualNorms(0, iterationNumber), queue ); // === restore blockX if needed if (blockX != origX) magmablas_zlacpy( MagmaFull, m, n, blockX, m, origX, m, queue ); printf("Eigenvalues:\n"); for(magma_int_t i =0; i<n; i++) printf("%e ", evalues[i]); printf("\n\n"); printf("Final residuals:\n"); magma_dprint_gpu(1, n, residualNorms(0, iterationNumber), 1); printf("\n\n"); //=== Prmagma_int_t residual history in a file for plotting ==== CHECK( magma_dmalloc_cpu(&hresidualNorms, (iterationNumber+1) * n)); magma_dgetmatrix( n, iterationNumber, residualNorms, n, hresidualNorms, n, queue ); solver_par->iter_res = *hresidualNorms(0, iterationNumber-1); printf("Residuals are stored in file residualNorms\n"); printf("Plot the residuals using: myplot \n"); FILE *residuals_file; residuals_file = fopen("residualNorms", "w"); for(magma_int_t i =1; i<iterationNumber; i++) { for(magma_int_t j = 0; j<n; j++) fprintf(residuals_file, "%f ", *hresidualNorms(j,i)); fprintf(residuals_file, "\n"); } fclose(residuals_file); cleanup: magma_free_cpu(hresidualNorms); // === free work space magma_free( residualNorms ); magma_free_cpu( condestGhistory ); magma_free_cpu( gevalues ); magma_free_cpu( iwork ); magma_free_pinned( hW ); magma_free_pinned( gevectors ); magma_free_pinned( h_gramB ); magma_free( gramM ); magma_free( gramA ); magma_free( gramB ); magma_free( activeMask ); if (blockX != (solver_par->eigenvectors)) magma_free( blockX ); if (blockAX != (solver_par->eigenvectors)) magma_free( blockAX ); if (blockAR != (solver_par->eigenvectors)) magma_free( blockAR ); if (blockAP != (solver_par->eigenvectors)) magma_free( blockAP ); if (blockR != (solver_par->eigenvectors)) magma_free( blockR ); if (blockP != (solver_par->eigenvectors)) magma_free( blockP ); if (blockW != (solver_par->eigenvectors)) magma_free( blockW ); if (dwork != (solver_par->eigenvectors)) magma_free( dwork ); magma_free( eval_gpu ); magma_free_pinned( hwork ); #ifdef COMPLEX magma_free_cpu( rwork ); rwork = NULL; #endif return info; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dlacpy */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; double c_neg_one = MAGMA_D_NEG_ONE; double *h_A, *h_B, *h_R; magmaDouble_ptr d_A, d_B; magma_int_t M, N, size, lda, ldb, ldda, lddb; magma_int_t ione = 1; magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull }; printf("%% uplo M N CPU GByte/s (ms) GPU GByte/s (ms) check\n"); printf("%%================================================================\n"); for( int iuplo = 0; iuplo < 3; ++iuplo ) { for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; lda = M; ldb = lda; ldda = magma_roundup( M, opts.align ); // multiple of 32 by default lddb = ldda; size = lda*N; if ( uplo[iuplo] == MagmaLower ) { // load & save lower trapezoid (with diagonal) if ( M > N ) { gbytes = 2. * sizeof(double) * (1.*M*N - 0.5*N*(N-1)) / 1e9; } else { gbytes = 2. * sizeof(double) * 0.5*M*(M+1) / 1e9; } } else if ( uplo[iuplo] == MagmaUpper ) { // load & save upper trapezoid (with diagonal) if ( N > M ) { gbytes = 2. * sizeof(double) * (1.*M*N - 0.5*M*(M-1)) / 1e9; } else { gbytes = 2. * sizeof(double) * 0.5*N*(N+1) / 1e9; } } else { // load & save entire matrix gbytes = 2. * sizeof(double) * 1.*M*N / 1e9; } TESTING_MALLOC_CPU( h_A, double, size ); TESTING_MALLOC_CPU( h_B, double, size ); TESTING_MALLOC_CPU( h_R, double, size ); TESTING_MALLOC_DEV( d_A, double, ldda*N ); TESTING_MALLOC_DEV( d_B, double, lddb*N ); /* Initialize the matrix */ for( int j = 0; j < N; ++j ) { for( int i = 0; i < M; ++i ) { h_A[i + j*lda] = MAGMA_D_MAKE( i + j/10000., j ); h_B[i + j*ldb] = MAGMA_D_MAKE( i - j/10000. + 10000., j ); } } /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_dsetmatrix( M, N, h_A, lda, d_A, ldda, opts.queue ); magma_dsetmatrix( M, N, h_B, ldb, d_B, lddb, opts.queue ); gpu_time = magma_sync_wtime( opts.queue ); //magmablas_dlacpy( uplo[iuplo], M-2, N-2, d_A+1+ldda, ldda, d_B+1+lddb, lddb, opts.queue ); // inset by 1 row & col magmablas_dlacpy( uplo[iuplo], M, N, d_A, ldda, d_B, lddb, opts.queue ); gpu_time = magma_sync_wtime( opts.queue ) - gpu_time; gpu_perf = gbytes / gpu_time; /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); //magma_int_t M2 = M-2; // inset by 1 row & col //magma_int_t N2 = N-2; //lapackf77_dlacpy( lapack_uplo_const(uplo[iuplo]), &M2, &N2, h_A+1+lda, &lda, h_B+1+ldb, &ldb ); lapackf77_dlacpy( lapack_uplo_const(uplo[iuplo]), &M, &N, h_A, &lda, h_B, &ldb ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gbytes / cpu_time; if ( opts.verbose ) { printf( "A= " ); magma_dprint( M, N, h_A, lda ); printf( "B= " ); magma_dprint( M, N, h_B, ldb ); printf( "dA=" ); magma_dprint_gpu( M, N, d_A, ldda ); printf( "dB=" ); magma_dprint_gpu( M, N, d_B, lddb ); } /* ===================================================================== Check the result =================================================================== */ magma_dgetmatrix( M, N, d_B, lddb, h_R, lda, opts.queue ); blasf77_daxpy(&size, &c_neg_one, h_B, &ione, h_R, &ione); error = lapackf77_dlange("f", &M, &N, h_R, &lda, work); printf("%5s %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n", lapack_uplo_const(uplo[iuplo]), (int) M, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., (error == 0. ? "ok" : "failed") ); status += ! (error == 0.); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( h_R ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } printf( "\n" ); } opts.cleanup(); TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dlaset Code is very similar to testing_dlacpy.cpp */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; double c_neg_one = MAGMA_D_NEG_ONE; double *h_A, *h_R; magmaDouble_ptr d_A; double offdiag = MAGMA_D_MAKE( 1.2000, 6.7000 ); double diag = MAGMA_D_MAKE( 3.1415, 2.7183 ); magma_int_t M, N, size, lda, ldda; magma_int_t ione = 1; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull }; printf("uplo M N CPU GByte/s (ms) GPU GByte/s (ms) check\n"); printf("=================================================================\n"); for( int iuplo = 0; iuplo < 3; ++iuplo ) { for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; //M += 2; // space for insets //N += 2; lda = M; ldda = ((M+31)/32)*32; size = lda*N; if ( uplo[iuplo] == MagmaLower || uplo[iuplo] == MagmaUpper ) { // save triangle (with diagonal) // TODO wrong for trapezoid gbytes = sizeof(double) * 0.5*N*(N+1) / 1e9; } else { // save entire matrix gbytes = sizeof(double) * 1.*M*N / 1e9; } TESTING_MALLOC_CPU( h_A, double, size ); TESTING_MALLOC_CPU( h_R, double, size ); TESTING_MALLOC_DEV( d_A, double, ldda*N ); /* Initialize the matrix */ for( int j = 0; j < N; ++j ) { for( int i = 0; i < M; ++i ) { h_A[i + j*lda] = MAGMA_D_MAKE( i + j/10000., j ); } } /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_dsetmatrix( M, N, h_A, lda, d_A, 0, ldda, opts.queue ); gpu_time = magma_sync_wtime( 0 ); //magmablas_dlaset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, 0, ldda, opts.queue ); // inset by 1 row & col magmablas_dlaset( uplo[iuplo], M, N, offdiag, diag, d_A, 0, ldda, opts.queue ); gpu_time = magma_sync_wtime( 0 ) - gpu_time; gpu_perf = gbytes / gpu_time; /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); //magma_int_t M2 = M-2; // inset by 1 row & col //magma_int_t N2 = N-2; //lapackf77_dlaset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda ); lapackf77_dlaset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gbytes / cpu_time; if ( opts.verbose ) { printf( "A= " ); magma_dprint( M, N, h_A, lda ); printf( "dA=" ); magma_dprint_gpu( M, N, d_A, 0, ldda, opts.queue ); } /* ===================================================================== Check the result =================================================================== */ magma_dgetmatrix( M, N, d_A, 0, ldda, h_R, lda, opts.queue ); blasf77_daxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_dlange("f", &M, &N, h_R, &lda, work); printf("%5s %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n", lapack_uplo_const( uplo[iuplo] ), (int) M, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., (error == 0. ? "ok" : "failed") ); status += ! (error == 0.); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_R ); TESTING_FREE_DEV( d_A ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } printf( "\n" ); } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dlat2s and slat2d */ int main( int argc, char** argv ) { #define A(i_,j_) ( A + (i_) + (j_)*lda) #define SA(i_,j_) (SA + (i_) + (j_)*lda) TESTING_INIT(); real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; float serror, swork[1]; double c_neg_one = MAGMA_D_NEG_ONE; float s_neg_one = MAGMA_S_NEG_ONE; magma_int_t ione = 1; magma_int_t n, lda, ldda, size, info; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; float *SA, *SR; double *A, *R; float *dSA; double *dA; magma_opts opts; parse_opts( argc, argv, &opts ); magma_uplo_t uplo[] = { MagmaLower, MagmaUpper }; printf("func uplo N CPU GB/s (ms) GPU GB/s (ms) ||R||_F\n"); printf("=====================================================================\n"); for( int iuplo = 0; iuplo < 2; ++iuplo ) { for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { n = opts.nsize[itest]; lda = n; ldda = ((n+31)/32)*32; // 0.5*(n+1)*n double-real loads and 0.5*(n+1)*n single-real stores (and vice-versa for slat2d) gbytes = (real_Double_t) 0.5*(n+1)*n * (sizeof(double) + sizeof(float)) / 1e9; size = ldda*n; // ldda >= lda TESTING_MALLOC_CPU( SA, float, size ); TESTING_MALLOC_CPU( A, double, size ); TESTING_MALLOC_CPU( SR, float, size ); TESTING_MALLOC_CPU( R, double, size ); TESTING_MALLOC_DEV( dSA, float, size ); TESTING_MALLOC_DEV( dA, double, size ); lapackf77_dlarnv( &ione, ISEED, &size, A ); lapackf77_slarnv( &ione, ISEED, &size, SA ); magma_dsetmatrix( n, n, A, lda, dA, ldda ); magma_ssetmatrix( n, n, SA, lda, dSA, ldda ); /* ===================================================================== Performs operation using LAPACK dlat2s =================================================================== */ info = 0; cpu_time = magma_wtime(); lapackf77_dlat2s( lapack_uplo_const(uplo[iuplo]), &n, A, &lda, SA, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gbytes / cpu_time; if (info != 0) printf("lapackf77_dlat2s returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ==================================================================== Performs operation using MAGMA dlat2s =================================================================== */ gpu_time = magma_sync_wtime(0); magmablas_dlat2s( uplo[iuplo], n, dA, ldda, dSA, ldda, &info ); gpu_time = magma_sync_wtime(0) - gpu_time; gpu_perf = gbytes / gpu_time; if (info != 0) printf("magmablas_dlat2s returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_sgetmatrix( n, n, dSA, ldda, SR, lda ); if ( opts.verbose ) { printf( "A= " ); magma_dprint( n, n, A, lda ); printf( "SA= " ); magma_sprint( n, n, SA, lda ); printf( "dA= " ); magma_dprint_gpu( n, n, dA, ldda ); printf( "dSA=" ); magma_sprint_gpu( n, n, dSA, ldda ); } /* ===================================================================== compute error |SA_magma - SA_lapack| should be zero if both are IEEE compliant =================================================================== */ blasf77_saxpy( &size, &s_neg_one, SA, &ione, SR, &ione ); serror = lapackf77_slange( "Fro", &n, &n, SR, &lda, swork ); printf( "dlat2s %5s %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", lapack_uplo_const(uplo[iuplo]), (int) n, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., serror, (serror == 0 ? "ok" : "failed") ); status += ! (serror == 0); /* ===================================================================== Reset matrices =================================================================== */ lapackf77_dlarnv( &ione, ISEED, &size, A ); lapackf77_slarnv( &ione, ISEED, &size, SA ); magma_dsetmatrix( n, n, A, lda, dA, ldda ); magma_ssetmatrix( n, n, SA, lda, dSA, ldda ); /* ===================================================================== Performs operation using LAPACK slat2d LAPACK doesn't implement slat2d; use our own simple implementation. =================================================================== */ cpu_time = magma_wtime(); if ( uplo[iuplo] == MagmaLower ) { for( int j=0; j < n; ++j ) { for( int i=j; i < n; ++i ) { *A(i,j) = MAGMA_D_MAKE( real(*SA(i,j)), imag(*SA(i,j)) ); } } } else { // upper for( int j=0; j < n; ++j ) { for( int i=0; i <= j; ++i ) { *A(i,j) = MAGMA_D_MAKE( real(*SA(i,j)), imag(*SA(i,j)) ); } } } cpu_time = magma_wtime() - cpu_time; cpu_perf = gbytes / cpu_time; if (info != 0) printf("lapackf77_slat2d returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ==================================================================== Performs operation using MAGMA slat2d =================================================================== */ magma_ssetmatrix( n, n, SA, lda, dSA, ldda ); gpu_time = magma_sync_wtime(0); magmablas_slat2d( uplo[iuplo], n, dSA, ldda, dA, ldda, &info ); gpu_time = magma_sync_wtime(0) - gpu_time; gpu_perf = gbytes / gpu_time; if (info != 0) printf("magmablas_slat2d returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_dgetmatrix( n, n, dA, ldda, R, lda ); if ( opts.verbose ) { printf( "A= " ); magma_dprint( n, n, A, lda ); printf( "SA= " ); magma_sprint( n, n, SA, lda ); printf( "dA= " ); magma_dprint_gpu( n, n, dA, ldda ); printf( "dSA=" ); magma_sprint_gpu( n, n, dSA, ldda ); } /* ===================================================================== compute error |A_magma - A_lapack| should be zero if both are IEEE compliant =================================================================== */ blasf77_daxpy( &size, &c_neg_one, A, &ione, R, &ione ); error = lapackf77_dlange( "Fro", &n, &n, R, &lda, work ); printf( "slat2d %5s %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", lapack_uplo_const(uplo[iuplo]), (int) n, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., error, (error == 0 ? "ok" : "failed") ); status += ! (error == 0); TESTING_FREE_CPU( SA ); TESTING_FREE_CPU( A ); TESTING_FREE_CPU( SR ); TESTING_FREE_CPU( R ); TESTING_FREE_DEV( dSA ); TESTING_FREE_DEV( dA ); printf( "\n" ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } printf( "\n" ); } TESTING_FINALIZE(); return status; }