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; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dtrsm */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0; double magma_error, cublas_error, work[1]; magma_int_t M, N, info; magma_int_t Ak; magma_int_t sizeA, sizeB; magma_int_t lda, ldb, ldda, lddb; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t *piv; magma_err_t err; double *h_A, *h_B, *h_Bcublas, *h_Bmagma, *h_B1, *h_X1, *h_X2, *LU, *LUT; double *d_A, *d_B; double c_neg_one = MAGMA_D_NEG_ONE; double c_one = MAGMA_D_ONE; double alpha = MAGMA_D_MAKE( 0.29, -0.86 ); magma_opts opts; parse_opts( argc, argv, &opts ); printf("If running lapack (option --lapack), MAGMA and CUBLAS error are both computed\n" "relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.\n\n" "side = %c, uplo = %c, transA = %c, diag = %c \n", opts.side, opts.uplo, opts.transA, opts.diag ); printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n"); printf("==================================================================================================\n"); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[i]; N = opts.nsize[i]; gflops = FLOPS_DTRSM(opts.side, M, N) / 1e9; if ( opts.side == MagmaLeft ) { lda = M; Ak = M; } else { lda = N; Ak = N; } ldb = M; ldda = ((lda+31)/32)*32; lddb = ((ldb+31)/32)*32; sizeA = lda*Ak; sizeB = ldb*N; TESTING_MALLOC( h_A, double, lda*Ak ); TESTING_MALLOC( LU, double, lda*Ak ); TESTING_MALLOC( LUT, double, lda*Ak ); TESTING_MALLOC( h_B, double, ldb*N ); TESTING_MALLOC( h_B1, double, ldb*N ); TESTING_MALLOC( h_X1, double, ldb*N ); TESTING_MALLOC( h_X2, double, ldb*N ); TESTING_MALLOC( h_Bcublas, double, ldb*N ); TESTING_MALLOC( h_Bmagma, double, ldb*N ); TESTING_DEVALLOC( d_A, double, ldda*Ak ); TESTING_DEVALLOC( d_B, double, lddb*N ); /* Initialize the matrices */ lapackf77_dlarnv( &ione, ISEED, &sizeA, LU ); err = magma_malloc_cpu( (void**) &piv, Ak*sizeof(magma_int_t) ); assert( err == 0 ); lapackf77_dgetrf( &Ak, &Ak, LU, &lda, piv, &info ); int i, j; for(i=0;i<Ak;i++){ for(j=0;j<Ak;j++){ LUT[j+i*lda] = LU[i+j*lda]; } } lapackf77_dlacpy(MagmaUpperStr, &Ak, &Ak, LUT, &lda, LU, &lda); if(opts.uplo == MagmaLower){ lapackf77_dlacpy(MagmaLowerStr, &Ak, &Ak, LU, &lda, h_A, &lda); }else{ lapackf77_dlacpy(MagmaUpperStr, &Ak, &Ak, LU, &lda, h_A, &lda); } lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B ); memcpy(h_B1, h_B, sizeB*sizeof(double)); /* ===================================================================== Performs operation using MAGMA-BLAS =================================================================== */ magma_dsetmatrix( Ak, Ak, h_A, lda, d_A, ldda ); magma_dsetmatrix( M, N, h_B, ldb, d_B, lddb ); magma_time = magma_sync_wtime( NULL ); magmablas_dtrsm( opts.side, opts.uplo, opts.transA, opts.diag, M, N, alpha, d_A, ldda, d_B, lddb ); magma_time = magma_sync_wtime( NULL ) - magma_time; magma_perf = gflops / magma_time; magma_dgetmatrix( M, N, d_B, lddb, h_Bmagma, ldb ); /* ===================================================================== Performs operation using CUDA-BLAS =================================================================== */ magma_dsetmatrix( M, N, h_B, ldb, d_B, lddb ); cublas_time = magma_sync_wtime( NULL ); cublasDtrsm( opts.side, opts.uplo, opts.transA, opts.diag, M, N, alpha, d_A, ldda, d_B, lddb ); cublas_time = magma_sync_wtime( NULL ) - cublas_time; cublas_perf = gflops / cublas_time; magma_dgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb ); /* ===================================================================== Performs operation using CPU BLAS =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); blasf77_dtrsm( &opts.side, &opts.uplo, &opts.transA, &opts.diag, &M, &N, &alpha, h_A, &lda, h_B, &ldb ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; } /* ===================================================================== Check the result =================================================================== */ // ||b - Ax|| / (||A||*||x||) memcpy(h_X1, h_Bmagma, sizeB*sizeof(double)); double alpha2 = MAGMA_D_DIV( c_one, alpha ); blasf77_dtrmm( &opts.side, &opts.uplo, &opts.transA, &opts.diag, &M, &N, &alpha2, h_A, &lda, h_X1, &ldb ); blasf77_daxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X1, &ione ); double norm1 = lapackf77_dlange( "M", &M, &N, h_X1, &ldb, work ); double normx = lapackf77_dlange( "M", &M, &N, h_Bmagma, &ldb, work ); double normA = lapackf77_dlange( "M", &Ak, &Ak, h_A, &lda, work ); magma_error = norm1/(normx*normA); memcpy(h_X2, h_Bcublas, sizeB*sizeof(double)); blasf77_dtrmm( &opts.side, &opts.uplo, &opts.transA, &opts.diag, &M, &N, &alpha2, h_A, &lda, h_X2, &ldb ); blasf77_daxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X2, &ione ); norm1 = lapackf77_dlange( "M", &M, &N, h_X2, &ldb, work ); normx = lapackf77_dlange( "M", &M, &N, h_Bcublas, &ldb, work ); normA = lapackf77_dlange( "M", &Ak, &Ak, h_A, &lda, work ); cublas_error = norm1/(normx*normA); if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n", (int) M, (int) N, magma_perf, 1000.*magma_time, cublas_perf, 1000.*cublas_time, cpu_perf, 1000.*cpu_time, magma_error, cublas_error ); } else { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e %8.2e\n", (int) M, (int) N, magma_perf, 1000.*magma_time, cublas_perf, 1000.*cublas_time, magma_error, cublas_error ); } TESTING_FREE( h_A ); TESTING_FREE( LU ); TESTING_FREE( LUT ); TESTING_FREE( h_B ); TESTING_FREE( h_Bcublas ); TESTING_FREE( h_Bmagma ); TESTING_FREE( h_B1 ); TESTING_FREE( h_X1 ); TESTING_FREE( h_X2 ); TESTING_DEVFREE( d_A ); TESTING_DEVFREE( d_B ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return 0; }