/** Purpose ------- CHEEVX computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. Arguments --------- @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] dA COMPLEX array, dimension (LDDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] ldda INTEGER The leading dimension of the array DA. LDDA >= max(1,N). @param[in] vl REAL @param[in] vu REAL If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[in] abstol REAL The absolute error tolerance for the eigenvalues. An approximate eigenvalue is accepted as converged when it is determined to lie in an interval [a,b] of width less than or equal to ABSTOL + EPS * max( |a|,|b| ), \n where EPS is the machine precision. If ABSTOL is less than or equal to zero, then EPS*|T| will be used in its place, where |T| is the 1-norm of the tridiagonal matrix obtained by reducing A to tridiagonal form. \n Eigenvalues will be computed most accurately when ABSTOL is set to twice the underflow threshold 2*SLAMCH('S'), not zero. If this routine returns with INFO > 0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*SLAMCH('S'). \n See "Computing Small Singular Values of Bidiagonal Matrices with Guaranteed High Relative Accuracy," by Demmel and Kahan, LAPACK Working Note #3. @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w REAL array, dimension (N) On normal exit, the first M elements contain the selected eigenvalues in ascending order. @param[out] dZ COMPLEX array, dimension (LDDZ, max(1,M)) If JOBZ = MagmaVec, then if INFO = 0, the first M columns of Z contain the orthonormal eigenvectors of the matrix A corresponding to the selected eigenvalues, with the i-th column of Z holding the eigenvector associated with W(i). If an eigenvector fails to converge, then that column of Z contains the latest approximation to the eigenvector, and the index of the eigenvector is returned in IFAIL. If JOBZ = MagmaNoVec, then Z is not referenced. Note: the user must ensure that at least max(1,M) columns are supplied in the array Z; if RANGE = MagmaRangeV, the exact value of M is not known in advance and an upper bound must be used. ********* (workspace) If FAST_HEMV is defined DZ should be (LDDZ, max(1,N)) in both cases. @param[in] lddz INTEGER The leading dimension of the array DZ. LDDZ >= 1, and if JOBZ = MagmaVec, LDDZ >= max(1,N). @param wA (workspace) COMPLEX array, dimension (LDWA, N) @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param wZ (workspace) COMPLEX array, dimension (LDWZ, max(1,M)) @param[in] ldwz INTEGER The leading dimension of the array wZ. LDWZ >= 1, and if JOBZ = MagmaVec, LDWZ >= max(1,N). @param[out] work (workspace) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. LWORK >= (NB+1)*N, where NB is the max of the blocksize for CHETRD. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param rwork (workspace) REAL array, dimension (7*N) @param iwork (workspace) INTEGER array, dimension (5*N) @param[out] ifail INTEGER array, dimension (N) If JOBZ = MagmaVec, then if INFO = 0, the first M elements of IFAIL are zero. If INFO > 0, then IFAIL contains the indices of the eigenvectors that failed to converge. If JOBZ = MagmaNoVec, then IFAIL is not referenced. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i, then i eigenvectors failed to converge. Their indices are stored in array IFAIL. @ingroup magma_cheev_driver ********************************************************************/ extern "C" magma_int_t magma_cheevx_gpu( magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex_ptr dA, magma_int_t ldda, float vl, float vu, magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m, float *w, magmaFloatComplex_ptr dZ, magma_int_t lddz, magmaFloatComplex *wA, magma_int_t ldwa, magmaFloatComplex *wZ, magma_int_t ldwz, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); const char* range_ = lapack_range_const( range ); magma_int_t ione = 1; const char* order_; magma_int_t indd, inde; magma_int_t imax; magma_int_t lopt, itmp1, indee; magma_int_t lower, wantz; magma_int_t i, j, jj, i__1; magma_int_t alleig, valeig, indeig; magma_int_t iscale, indibl; magma_int_t indiwk, indisp, indtau; magma_int_t indrwk, indwrk; magma_int_t llwork, nsplit; magma_int_t lquery; magma_int_t iinfo; float safmin; float bignum; float smlnum; float eps, tmp1; float anrm; float sigma, d__1; float rmin, rmax; magmaFloat_ptr dwork; /* Function Body */ lower = (uplo == MagmaLower); wantz = (jobz == MagmaVec); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1); *info = 0; if (! (wantz || (jobz == MagmaNoVec))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || (uplo == MagmaUpper))) { *info = -3; } else if (n < 0) { *info = -4; } else if (ldda < max(1,n)) { *info = -6; } else if (lddz < 1 || (wantz && lddz < n)) { *info = -15; } else if (ldwa < max(1,n)) { *info = -17; } else if (ldwz < 1 || (wantz && ldwz < n)) { *info = -19; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_chetrd_nb(n); lopt = n * (nb + 1); work[0] = MAGMA_C_MAKE( lopt, 0 ); if (lwork < lopt && ! lquery) { *info = -21; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } *m = 0; /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif magmaFloatComplex *a; magma_cmalloc_cpu( &a, n*n ); magma_cgetmatrix(n, n, dA, ldda, a, n); lapackf77_cheevx(jobz_, range_, uplo_, &n, a, &n, &vl, &vu, &il, &iu, &abstol, m, w, wZ, &ldwz, work, &lwork, rwork, iwork, ifail, info); magma_csetmatrix( n, n, a, n, dA, ldda); magma_csetmatrix( n, *m, wZ, ldwz, dZ, lddz); magma_free_cpu(a); return *info; } if (MAGMA_SUCCESS != magma_smalloc( &dwork, n )) { fprintf (stderr, "!!!! device memory allocation error (magma_cheevx_gpu)\n"); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } --w; --work; --rwork; --iwork; --ifail; /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_clanhe(MagmaMaxNorm, uplo, n, dA, ldda, dwork); iscale = 0; sigma = 1; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { d__1 = 1.; magmablas_clascl(uplo, 0, 0, 1., sigma, n, n, dA, ldda, info); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ indd = 1; inde = indd + n; indrwk = inde + n; indtau = 1; indwrk = indtau + n; llwork = lwork - indwrk + 1; #ifdef FAST_HEMV magma_chetrd2_gpu(uplo, n, dA, ldda, &rwork[indd], &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, dZ, lddz*n, &iinfo); #else magma_chetrd_gpu (uplo, n, dA, ldda, &rwork[indd], &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, &iinfo); #endif lopt = n + (magma_int_t)MAGMA_C_REAL(work[indwrk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call SSTERF or CUNGTR and CSTEQR. If this fails for some eigenvalue, then try SSTEBZ. */ if ((alleig || (indeig && il == 1 && iu == n)) && abstol <= 0.) { blasf77_scopy(&n, &rwork[indd], &ione, &w[1], &ione); indee = indrwk + 2*n; if (! wantz) { i__1 = n - 1; blasf77_scopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_ssterf(&n, &w[1], &rwork[indee], info); } else { lapackf77_clacpy("A", &n, &n, wA, &ldwa, wZ, &ldwz); lapackf77_cungtr(uplo_, &n, wZ, &ldwz, &work[indtau], &work[indwrk], &llwork, &iinfo); i__1 = n - 1; blasf77_scopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_csteqr(jobz_, &n, &w[1], &rwork[indee], wZ, &ldwz, &rwork[indrwk], info); if (*info == 0) { for (i = 1; i <= n; ++i) { ifail[i] = 0; } magma_csetmatrix( n, n, wZ, ldwz, dZ, lddz ); } } if (*info == 0) { *m = n; } } /* Otherwise, call SSTEBZ and, if eigenvectors are desired, CSTEIN. */ if (*m == 0) { *info = 0; if (wantz) { order_ = "B"; } else { order_ = "E"; } indibl = 1; indisp = indibl + n; indiwk = indisp + n; lapackf77_sstebz(range_, order_, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m, &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info); if (wantz) { lapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp], wZ, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info); magma_csetmatrix( n, *m, wZ, ldwz, dZ, lddz ); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by CSTEIN. */ magma_cunmtr_gpu(MagmaLeft, uplo, MagmaNoTrans, n, *m, dA, ldda, &work[indtau], dZ, lddz, wA, ldwa, &iinfo); } } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = *m; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, &w[1], &ione); } /* If eigenvalues are not in order, then sort them, along with eigenvectors. */ if (wantz) { for (j = 1; j <= *m-1; ++j) { i = 0; tmp1 = w[j]; for (jj = j + 1; jj <= *m; ++jj) { if (w[jj] < tmp1) { i = jj; tmp1 = w[jj]; } } if (i != 0) { itmp1 = iwork[indibl + i - 1]; w[i] = w[j]; iwork[indibl + i - 1] = iwork[indibl + j - 1]; w[j] = tmp1; iwork[indibl + j - 1] = itmp1; magma_cswap(n, dZ + (i-1)*lddz, ione, dZ + (j-1)*lddz, ione); if (*info != 0) { itmp1 = ifail[i]; ifail[i] = ifail[j]; ifail[j] = itmp1; } } } } /* Set WORK[0] to optimal complex workspace size. */ work[1] = MAGMA_C_MAKE( lopt, 0 ); return *info; } /* magma_cheevx_gpu */
/* //////////////////////////////////////////////////////////////////////////// -- Testing chetrd_he2hb */ int main( int argc, char** argv) { TESTING_INIT_MGPU(); real_Double_t gpu_time, gpu_perf, gflops; magmaFloatComplex *h_A, *h_R, *h_work, *dT1; magmaFloatComplex *tau; float *D, *E; /* Matrix size */ magma_int_t N, n2, lda, lwork, ldt, lwork0; magma_int_t info; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; #if defined(CHECKEIG) #if defined(PRECISION_z) || defined(PRECISION_d) magma_int_t WANTZ=0; magma_int_t THREADS=1; #endif #endif magma_int_t NE = 0; magma_int_t NB = 0; magma_int_t ngpu = 1; magma_opts opts; parse_opts( argc, argv, &opts ); NB = opts.nb; if (NB < 1) NB = 64; //64; //magma_get_chetrd_he2hb_nb(N); // what is NE ? if (NE < 1) NE = 64; //N; //magma_get_chetrd_he2hb_nb(N); // N not yet initialized printf(" N GPU GFlop/s \n"); printf("=====================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; ldt = N; n2 = N*lda; gflops = FLOPS_CHETRD( N ) / 1e9; /* We suppose the magma NB is bigger than lapack NB */ lwork0 = N*NB; /* Allocate host memory for the matrix */ TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N ); TESTING_MALLOC_CPU( tau, magmaFloatComplex, N-1 ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, lda*N ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork0 ); TESTING_MALLOC_PIN( D, float, N ); TESTING_MALLOC_PIN( E, float, N ); //TESTING_MALLOC_DEV( dT1, magmaFloatComplex, (2*min(N,N)+(N+31)/32*32)*NB ); TESTING_MALLOC_DEV( dT1, magmaFloatComplex, (N*NB) ); // if (WANTZ) gflops = 2.0*gflops; /* ==================================================================== Initialize the matrix =================================================================== */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); magma_cmake_hermitian( N, h_A, lda ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_device_t cdev; magma_getdevice( &cdev ); gpu_time = magma_wtime(); /* magma_chetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork0, dT1, THREADS, &info); tband = magma_wtime - gpu_time(); printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband); magma_chetrd_bhe2trc_v5(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt); */ /* magma_chetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork, dT1, THREADS, &info); tband = magma_wtime - gpu_time(); printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband); magma_chetrd_bhe2trc(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt); */ magma_range_t range = MagmaRangeAll; magma_int_t fraction_ev = 100; magma_int_t il, iu, m1; float vl=0., vu=0.; if (fraction_ev == 0) { il = N / 10; iu = N / 5+il; } else { il = 1; iu = (int)(fraction_ev*N); if (iu < 1) iu = 1; } magmaFloatComplex *hh_work; magma_int_t *iwork; magma_int_t nb, /*lwork,*/ liwork; magma_int_t threads = magma_get_parallel_numthreads(); #if defined(PRECISION_z) || defined(PRECISION_c) float *rwork; magma_int_t lrwork; lwork = magma_cbulge_get_lq2(N, threads) + 2*N + N*N; lrwork = 1 + 5*N +2*N*N; TESTING_MALLOC_PIN( rwork, float, lrwork ); #else lwork = magma_cbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N; #endif liwork = 3 + 5*N; nb = magma_get_chetrd_nb(N); TESTING_MALLOC_PIN( hh_work, magmaFloatComplex, lwork ); TESTING_MALLOC_CPU( iwork, magma_int_t, liwork ); if (ngpu == 1) { printf("calling cheevdx_2stage 1 GPU\n"); magma_cheevdx_2stage( opts.jobz, range, opts.uplo, N, h_R, lda, vl, vu, il, iu, &m1, D, hh_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); } else { printf("calling cheevdx_2stage_m %d GPU\n", (int) ngpu); magma_cheevdx_2stage_m(ngpu, opts.jobz, range, opts.uplo, N, h_R, lda, vl, vu, il, iu, &m1, D, hh_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); } magma_setdevice( cdev ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; /* ===================================================================== Check the factorization =================================================================== */ /* if ( opts.check ) { FILE *fp ; printf("Writing input matrix in matlab_i_mat.txt ...\n"); fp = fopen ("matlab_i_mat.txt", "w") ; if ( fp == NULL ) { printf("Couldn't open output file\n"); exit(1); } for (j=0; j < N; j++) { for (k=0; k < N; k++) { #if defined(PRECISION_z) || defined(PRECISION_c) fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1, h_A[k+j*lda].x, h_A[k+j*lda].y); #else fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_A[k+j*lda]); #endif } } fclose( fp ) ; printf("Writing output matrix in matlab_o_mat.txt ...\n"); fp = fopen ("matlab_o_mat.txt", "w") ; if ( fp == NULL ) { printf("Couldn't open output file\n"); exit(1); } for (j=0; j < N; j++) { for (k=0; k < N; k++) { #if defined(PRECISION_z) || defined(PRECISION_c) fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1, h_R[k+j*lda].x, h_R[k+j*lda].y); #else fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_R[k+j*lda]); #endif } } fclose( fp ) ; } */ /* ===================================================================== Print performance and error. =================================================================== */ #if defined(CHECKEIG) #if defined(PRECISION_z) || defined(PRECISION_d) if ( opts.check ) { printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time ); char JOBZ; if (WANTZ == 0) JOBZ = 'N'; else JOBZ = 'V'; float nrmI=0.0, nrm1=0.0, nrm2=0.0; int lwork2 = 256*N; magmaFloatComplex *work2, *AINIT; float *rwork2, *D2; // TODO free this memory ! magma_cmalloc_cpu( &work2, lwork2 ); magma_smalloc_cpu( &rwork2, N ); magma_smalloc_cpu( &D2, N ); magma_cmalloc_cpu( &AINIT, N*lda ); memcpy(AINIT, h_A, N*lda*sizeof(magmaFloatComplex)); /* compute the eigenvalues using lapack routine to be able to compare to it and used as ref */ cpu_time = magma_wtime(); i= min(12, THREADS); #if defined(USEMKL) mkl_set_num_threads( i ); #endif #if defined(USEACML) omp_set_num_threads(i); #endif lapackf77_cheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, #if defined(PRECISION_z) || defined (PRECISION_c) rwork2, #endif &info ); ///* call eigensolver for our resulting tridiag [D E] and for Q */ //dstedc_withZ('V', N, D, E, h_R, lda); ////ssterf_( &N, D, E, &info); //// cpu_time = magma_wtime() - cpu_time; printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, i); /* for (i=0; i < 10; i++) printf(" voici lpk D[%d] %8.2e\n", i, D2[i]); */ //magmaFloatComplex mydz=0.0, mydo=1.0; //magmaFloatComplex *Z; // magma_cmalloc_cpu( &Z, N*lda ); // dgemm_("N", "N", &N, &N, &N, &mydo, h_R, &lda, h_A, &lda, &mydz, Z, &lda); /* compare result */ cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2); magmaFloatComplex *WORKAJETER; float *RWORKAJETER, *RESU; // TODO free this memory ! magma_cmalloc_cpu( &WORKAJETER, (2* N * N + N) ); magma_smalloc_cpu( &RWORKAJETER, N ); magma_smalloc_cpu( &RESU, 10 ); int MATYPE; memset(RESU, 0, 10*sizeof(float)); MATYPE=3; float NOTHING=0.0; cpu_time = magma_wtime(); // check results ccheck_eig_(&JOBZ, &MATYPE, &N, &NB, AINIT, &lda, &NOTHING, &NOTHING, D2, D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU ); cpu_time = magma_wtime() - cpu_time; printf(" Finish CHECK - results timing= %f\n", cpu_time); #if defined(USEMKL) mkl_set_num_threads( 1 ); #endif #if defined(USEACML) omp_set_num_threads(1); #endif printf("\n"); printf(" ================================================================================================================\n"); printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) THREADS, (int) N, (int) NB, (int) WANTZ); printf(" ================================================================================================================\n"); printf(" DSBTRD : %15s \n", "STATblgv9withQ "); printf(" ================================================================================================================\n"); if (WANTZ > 0) printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]); if (WANTZ > 0) printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]); printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]); printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]); printf(" ================================================================================================================\n\n\n"); printf(" ****************************************************************************************************************\n"); printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2); printf(" ****************************************************************************************************************\n\n"); } #endif #endif printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time ); printf("============================================================================\n\n\n"); /* Memory clean up */ TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( tau ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_work ); TESTING_FREE_PIN( D ); TESTING_FREE_PIN( E ); TESTING_FREE_DEV( dT1 ); /* TODO - not all memory has been freed inside loop */ fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE_MGPU(); return EXIT_SUCCESS; }
/** Purpose ------- CHEGVD computes all the eigenvalues, and optionally, the eigenvectors of a complex generalized Hermitian-definite eigenproblem, of the form A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and B are assumed to be Hermitian and B is also positive definite. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] nrgpu INTEGER Number of GPUs to use. @param[in] itype INTEGER Specifies the problem type to be solved: = 1: A*x = (lambda)*B*x = 2: A*B*x = (lambda)*x = 3: B*A*x = (lambda)*x @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangles of A and B are stored; - = MagmaLower: Lower triangles of A and B are stored. @param[in] n INTEGER The order of the matrices A and B. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. \n On exit, if JOBZ = MagmaVec, then if INFO = 0, A contains the matrix Z of eigenvectors. The eigenvectors are normalized as follows: if ITYPE = 1 or 2, Z**H*B*Z = I; if ITYPE = 3, Z**H*inv(B)*Z = I. If JOBZ = MagmaNoVec, then on exit the upper triangle (if UPLO=MagmaUpper) or the lower triangle (if UPLO=MagmaLower) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[in,out] B COMPLEX array, dimension (LDB, N) On entry, the Hermitian matrix B. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of B contains the upper triangular part of the matrix B. If UPLO = MagmaLower, the leading N-by-N lower triangular part of B contains the lower triangular part of the matrix B. \n On exit, if INFO <= N, the part of B containing the matrix is overwritten by the triangular factor U or L from the Cholesky factorization B = U**H*U or B = L*L**H. @param[in] ldb INTEGER The leading dimension of the array B. LDB >= max(1,N). @param[in] vl REAL @param[in] vu REAL If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LWORK >= N + 1. If JOBZ = MagmaVec and N > 1, LWORK >= 2*N*nb + N**2. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) REAL array, dimension (MAX(1,LRWORK)) On exit, if INFO = 0, RWORK(1) returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: CPOTRF or CHEEVD returned an error code: <= N: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1); > N: if INFO = N + i, for 1 <= i <= N, then the leading minor of order i of B is not positive definite. The factorization of B could not be completed and no eigenvalues or eigenvectors were computed. Further Details --------------- Based on contributions by Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA Modified so that no backsubstitution is performed if CHEEVD fails to converge (NEIG in old code could be greater than N causing out of bounds reference to A - reported by Ralf Meyer). Also corrected the description of INFO and the test on ITYPE. Sven, 16 Feb 05. @ingroup magma_chegv_driver ********************************************************************/ extern "C" magma_int_t magma_chegvdx_m(magma_int_t nrgpu, magma_int_t itype, magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *B, magma_int_t ldb, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magmaFloatComplex c_one = MAGMA_C_ONE; magma_int_t lower; magma_trans_t trans; magma_int_t wantz; magma_int_t lquery; magma_int_t alleig, valeig, indeig; magma_int_t lwmin; magma_int_t liwmin; magma_int_t lrwmin; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (itype < 1 || itype > 3) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (wantz || (jobz == MagmaNoVec))) { *info = -3; } else if (! (lower || (uplo == MagmaUpper))) { *info = -4; } else if (n < 0) { *info = -5; } else if (lda < max(1,n)) { *info = -7; } else if (ldb < max(1,n)) { *info = -9; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -11; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -12; } else if (iu < min(n,il) || iu > n) { *info = -13; } } } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = 2*n + n*n; lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps (in Double!) to ensure length gets rounded up, // if it cannot be exactly represented in floating point. real_Double_t one_eps = 1. + lapackf77_slamch("Epsilon"); work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0.); // round up rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -17; } else if (lrwork < lrwmin && ! lquery) { *info = -19; } else if (liwork < liwmin && ! lquery) { *info = -21; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_chegvd(&itype, jobz_, uplo_, &n, A, &lda, B, &ldb, w, work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); *m = n; return *info; } magma_timer_t time=0; timer_start( time ); magma_cpotrf_m(nrgpu, uplo, n, B, ldb, info); if (*info != 0) { *info = n + *info; return *info; } timer_stop( time ); timer_printf("time cpotrf = %6.2f\n", time ); timer_start( time ); /* Transform problem to standard eigenvalue problem and solve. */ magma_chegst_m(nrgpu, itype, uplo, n, A, lda, B, ldb, info); timer_stop( time ); timer_printf( "time chegst = %6.2f\n", time ); timer_start( time ); magma_cheevdx_m(nrgpu, jobz, range, uplo, n, A, lda, vl, vu, il, iu, m, w, work, lwork, rwork, lrwork, iwork, liwork, info); timer_stop( time ); timer_printf( "time cheevd = %6.2f\n", time ); if (wantz && *info == 0) { timer_start( time ); /* Backtransform eigenvectors to the original problem. */ if (itype == 1 || itype == 2) { /* For A*x=(lambda)*B*x and A*B*x=(lambda)*x; backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */ if (lower) { trans = MagmaConjTrans; } else { trans = MagmaNoTrans; } magma_ctrsm_m(nrgpu, MagmaLeft, uplo, trans, MagmaNonUnit, n, *m, c_one, B, ldb, A, lda); } else if (itype == 3) { /* For B*A*x=(lambda)*x; backtransform eigenvectors: x = L*y or U'*y */ if (lower) { trans = MagmaNoTrans; } else { trans = MagmaConjTrans; } //magma_ctrmm(MagmaLeft, uplo, trans, MagmaNonUnit, // n, n, c_one, db, lddb, da, ldda); } timer_stop( time ); timer_printf( "time setmatrices trsm/mm + getmatrices = %6.2f\n", time ); } work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0.); // round up rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; return *info; } /* magma_chegvd_m */
extern "C" magma_int_t magma_chetrd2_gpu(char uplo, magma_int_t n, magmaFloatComplex *da, magma_int_t ldda, float *d, float *e, magmaFloatComplex *tau, magmaFloatComplex *wa, magma_int_t ldwa, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex *dwork, magma_int_t ldwork, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= CHETRD2_GPU reduces a complex Hermitian matrix A to real symmetric tridiagonal form T by an orthogonal similarity transformation: Q**H * A * Q = T. This version passes a workspace that is used in an optimized GPU matrix-vector product. Arguments ========= UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. DA (device input/output) COMPLEX array, dimension (LDA,N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit, if UPLO = 'U', the diagonal and first superdiagonal of A are overwritten by the corresponding elements of the tridiagonal matrix T, and the elements above the first superdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = 'L', the diagonal and first subdiagonal of A are over- written by the corresponding elements of the tridiagonal matrix T, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. LDDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). D (output) COMPLEX array, dimension (N) The diagonal elements of the tridiagonal matrix T: D(i) = A(i,i). E (output) COMPLEX array, dimension (N-1) The off-diagonal elements of the tridiagonal matrix T: E(i) = A(i,i+1) if UPLO = 'U', E(i) = A(i+1,i) if UPLO = 'L'. TAU (output) COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). WA (workspace/output) COMPLEX array, dimension (LDA,N) On exit the diagonal, the upper part (UPLO='U') or the lower part (UPLO='L') are copies of DA LDWA (input) INTEGER The leading dimension of the array WA. LDWA >= max(1,N). WORK (workspace/output) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. LWORK >= 1. For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. DWORK (workspace/output) COMPLEX array on the GPU, dim (MAX(1,LDWORK)) LDWORK (input) INTEGER The dimension of the array DWORK. LDWORK >= (n*n+64-1)/64 + 2*n*nb, where nb = magma_get_chetrd_nb(n) INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value Further Details =============== If UPLO = 'U', the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = 'L', the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = 'U': if UPLO = 'L': ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). ===================================================================== */ char uplo_[2] = {uplo, 0}; magma_int_t nb = magma_get_chetrd_nb(n); magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex c_one = MAGMA_C_ONE; float d_one = MAGMA_D_ONE; magma_int_t kk, nx; magma_int_t i, j, i_n; magma_int_t iinfo; magma_int_t ldw, lddw, lwkopt; magma_int_t lquery; *info = 0; int upper = lapackf77_lsame(uplo_, "U"); lquery = lwork == -1; if (! upper && ! lapackf77_lsame(uplo_, "L")) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } else if (ldwa < max(1,n)) { *info = -9; } else if (lwork < 1 && ! lquery) { *info = -11; } /* Determine the block size. */ ldw = lddw = n; lwkopt = n * nb; if (*info == 0) { MAGMA_C_SET2REAL( work[0], lwkopt ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; /* Quick return if possible */ if (n == 0) { work[0] = c_one; return *info; } if (n < 1024) nx = n; else nx = 300; if (ldwork<(ldw*n+64-1)/64 + 2*ldw*nb) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } if (upper) { /* Reduce the upper triangle of A. Columns 1:kk are handled by the unblocked method. */ kk = n - (n - nx + nb - 1) / nb * nb; for (i = n - nb; i >= kk; i -= nb) { /* Reduce columns i:i+nb-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel */ magma_cgetmatrix( i+nb, nb, dA(0, i), ldda, A(0, i), ldwa ); magma_clatrd2(uplo, i+nb, nb, A(0, 0), ldwa, e, tau, work, ldw, dA(0, 0), ldda, dwork, lddw, dwork + 2*ldw*nb, ldwork - 2*ldw*nb); /* Update the unreduced submatrix A(0:i-2,0:i-2), using an update of the form: A := A - V*W' - W*V' */ magma_csetmatrix( i + nb, nb, work, ldw, dwork, lddw ); magma_cher2k(uplo, MagmaNoTrans, i, nb, c_neg_one, dA(0, i), ldda, dwork, lddw, d_one, dA(0, 0), ldda); /* Copy superdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+nb; ++j) { MAGMA_C_SET2REAL( *A(j-1, j), e[j - 1] ); d[j] = MAGMA_C_REAL( *A(j, j) ); } } magma_cgetmatrix( kk, kk, dA(0, 0), ldda, A(0, 0), ldwa ); /* Use CPU code to reduce the last or only block */ lapackf77_chetrd(uplo_, &kk, A(0, 0), &ldwa, d, e, tau, work, &lwork, &iinfo); magma_csetmatrix( kk, kk, A(0, 0), ldwa, dA(0, 0), ldda ); } else { /* Reduce the lower triangle of A */ for (i = 0; i < n-nx; i += nb) { /* Reduce columns i:i+nb-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel */ magma_cgetmatrix( n-i, nb, dA(i, i), ldda, A(i, i), ldwa ); magma_clatrd2(uplo, n-i, nb, A(i, i), ldwa, &e[i], &tau[i], work, ldw, dA(i, i), ldda, dwork, lddw, dwork + 2*ldw*nb, ldwork - 2*ldw*nb); /* Update the unreduced submatrix A(i+ib:n,i+ib:n), using an update of the form: A := A - V*W' - W*V' */ magma_csetmatrix( n-i, nb, work, ldw, dwork, lddw ); magma_cher2k(MagmaLower, MagmaNoTrans, n-i-nb, nb, c_neg_one, dA(i+nb, i), ldda, &dwork[nb], lddw, d_one, dA(i+nb, i+nb), ldda); /* Copy subdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+nb; ++j) { MAGMA_C_SET2REAL( *A(j+1, j), e[j] ); d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* Use unblocked code to reduce the last or only block */ magma_cgetmatrix( n-i, n-i, dA(i, i), ldda, A(i, i), ldwa ); i_n = n-i; lapackf77_chetrd(uplo_, &i_n, A(i, i), &ldwa, &d[i], &e[i], &tau[i], work, &lwork, &iinfo); magma_csetmatrix( n-i, n-i, A(i, i), ldwa, dA(i, i), ldda ); } MAGMA_C_SET2REAL( work[0], lwkopt ); return *info; } /* chetrd2_gpu */
extern "C" magma_int_t magma_cheevdx(char jobz, char range, char uplo, magma_int_t n, magmaFloatComplex *a, magma_int_t lda, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= CHEEVDX computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments ========= JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. RANGE (input) CHARACTER*1 = 'A': all eigenvalues will be found. = 'V': all eigenvalues in the half-open interval (VL,VU] will be found. = 'I': the IL-th through IU-th eigenvalues will be found. UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = 'V', then if INFO = 0, the first m columns of A contains the required orthonormal eigenvectors of the matrix A. If JOBZ = 'N', then on exit the lower triangle (if UPLO='L') or the upper triangle (if UPLO='U') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). VL (input) DOUBLE PRECISION VU (input) DOUBLE PRECISION If RANGE='V', the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = 'A' or 'I'. IL (input) INTEGER IU (input) INTEGER If RANGE='I', the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = 'A' or 'V'. M (output) INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1. W (output) DOUBLE PRECISION array, dimension (N) If INFO = 0, the required m eigenvalues in ascending order. WORK (workspace/output) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = 'N' and N > 1, LWORK >= N + N*NB. If JOBZ = 'V' and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_chetrd_nb(N). If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. RWORK (workspace/output) DOUBLE PRECISION array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. LRWORK (input) INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = 'N' and N > 1, LRWORK >= N. If JOBZ = 'V' and N > 1, LRWORK >= 1 + 5*N + 2*N**2. If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. LIWORK (input) INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = 'N' and N > 1, LIWORK >= 1. If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N. If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i and JOBZ = 'N', then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = 'V', then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details =============== Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; char range_[2] = {range, 0}; magma_int_t ione = 1; magma_int_t izero = 0; float d_one = 1.; float d__1; float eps; magma_int_t inde; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; float smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; float* dwork; wantz = lapackf77_lsame(jobz_, MagmaVecStr); lower = lapackf77_lsame(uplo_, MagmaLowerStr); alleig = lapackf77_lsame( range_, "A" ); valeig = lapackf77_lsame( range_, "V" ); indeig = lapackf77_lsame( range_, "I" ); lquery = lwork == -1 || lrwork == -1 || liwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -3; } else if (n < 0) { *info = -4; } else if (lda < max(1,n)) { *info = -6; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps to ensure length gets rounded up, // if it cannot be exactly represented in floating point. work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -14; } else if ((lrwork < lrwmin) && ! lquery) { *info = -16; } else if ((liwork < liwmin) && ! lquery) { *info = -18; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_C_REAL(a[0]); if (wantz) { a[0] = MAGMA_C_ONE; } return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128){ #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_cheevd(jobz_, uplo_, &n, a, &lda, w, work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); return *info; } /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_clanhe("M", uplo_, &n, a, &lda, rwork); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_clascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a, &lda, info); } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ // chetrd rwork: e (n) // cstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // chetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // cstedx work: tau (n) + z (n^2) // cunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; // #ifdef ENABLE_TIMER magma_timestr_t start, end; start = get_current_time(); #endif magma_chetrd(uplo_[0], n, a, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo); #ifdef ENABLE_TIMER end = get_current_time(); printf("time chetrd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif /* For eigenvalues only, call SSTERF. For eigenvectors, first call CSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call CUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &rwork[inde], info); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); } else { #ifdef ENABLE_TIMER start = get_current_time(); #endif if (MAGMA_SUCCESS != magma_smalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_cstedx(range, n, vl, vu, il, iu, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info); magma_free( dwork ); #ifdef ENABLE_TIMER end = get_current_time(); printf("time cstedx = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); magma_cunmtr(MagmaLeft, uplo, MagmaNoTrans, n, *m, a, lda, &work[indtau], &work[indwrk + n * (il-1) ], n, &work[indwk2], llwrk2, &iinfo); lapackf77_clacpy("A", &n, m, &work[indwrk + n * (il-1)] , &n, a, &lda); #ifdef ENABLE_TIMER end = get_current_time(); printf("time cunmtr + copy = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, w, &ione); } work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); // round up rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; return *info; } /* magma_cheevdx */
extern "C" magma_int_t magma_cheevx(char jobz, char range, char uplo, magma_int_t n, magmaFloatComplex *a, magma_int_t lda, float vl, float vu, magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m, float *w, magmaFloatComplex *z, magma_int_t ldz, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= CHEEVX computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. Arguments ========= JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. RANGE (input) CHARACTER*1 = 'A': all eigenvalues will be found. = 'V': all eigenvalues in the half-open interval (VL,VU] will be found. = 'I': the IL-th through IU-th eigenvalues will be found. UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, the lower triangle (if UPLO='L') or the upper triangle (if UPLO='U') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). VL (input) REAL VU (input) REAL If RANGE='V', the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = 'A' or 'I'. IL (input) INTEGER IU (input) INTEGER If RANGE='I', the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = 'A' or 'V'. ABSTOL (input) REAL The absolute error tolerance for the eigenvalues. An approximate eigenvalue is accepted as converged when it is determined to lie in an interval [a,b] of width less than or equal to ABSTOL + EPS * max( |a|,|b| ) , where EPS is the machine precision. If ABSTOL is less than or equal to zero, then EPS*|T| will be used in its place, where |T| is the 1-norm of the tridiagonal matrix obtained by reducing A to tridiagonal form. Eigenvalues will be computed most accurately when ABSTOL is set to twice the underflow threshold 2*SLAMCH('S'), not zero. If this routine returns with INFO>0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*SLAMCH('S'). See "Computing Small Singular Values of Bidiagonal Matrices with Guaranteed High Relative Accuracy," by Demmel and Kahan, LAPACK Working Note #3. M (output) INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1. W (output) REAL array, dimension (N) On normal exit, the first M elements contain the selected eigenvalues in ascending order. Z (output) COMPLEX array, dimension (LDZ, max(1,M)) If JOBZ = 'V', then if INFO = 0, the first M columns of Z contain the orthonormal eigenvectors of the matrix A corresponding to the selected eigenvalues, with the i-th column of Z holding the eigenvector associated with W(i). If an eigenvector fails to converge, then that column of Z contains the latest approximation to the eigenvector, and the index of the eigenvector is returned in IFAIL. If JOBZ = 'N', then Z is not referenced. Note: the user must ensure that at least max(1,M) columns are supplied in the array Z; if RANGE = 'V', the exact value of M is not known in advance and an upper bound must be used. LDZ (input) INTEGER The leading dimension of the array Z. LDZ >= 1, and if JOBZ = 'V', LDZ >= max(1,N). WORK (workspace/output) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= max(1,2*N-1). For optimal efficiency, LWORK >= (NB+1)*N, where NB is the max of the blocksize for CHETRD and for CUNMTR as returned by ILAENV. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. RWORK (workspace) REAL array, dimension (7*N) IWORK (workspace) INTEGER array, dimension (5*N) IFAIL (output) INTEGER array, dimension (N) If JOBZ = 'V', then if INFO = 0, the first M elements of IFAIL are zero. If INFO > 0, then IFAIL contains the indices of the eigenvectors that failed to converge. If JOBZ = 'N', then IFAIL is not referenced. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i, then i eigenvectors failed to converge. Their indices are stored in array IFAIL. ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; char range_[2] = {range, 0}; magma_int_t izero = 0; magma_int_t ione = 1; char order[1]; magma_int_t indd, inde; magma_int_t imax; magma_int_t lopt, itmp1, indee; magma_int_t lower, wantz; magma_int_t i, j, jj, i__1; magma_int_t alleig, valeig, indeig; magma_int_t iscale, indibl; magma_int_t indiwk, indisp, indtau; magma_int_t indrwk, indwrk; magma_int_t llwork, nsplit; magma_int_t lquery; magma_int_t iinfo; float safmin; float bignum; float smlnum; float eps, tmp1; float anrm; float sigma, d__1; float rmin, rmax; /* Function Body */ lower = lapackf77_lsame(uplo_, MagmaLowerStr); wantz = lapackf77_lsame(jobz_, MagmaVecStr); alleig = lapackf77_lsame(range_, "A"); valeig = lapackf77_lsame(range_, "V"); indeig = lapackf77_lsame(range_, "I"); lquery = lwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -3; } else if (n < 0) { *info = -4; } else if (lda < max(1,n)) { *info = -6; } else if (ldz < 1 || (wantz && ldz < n)) { *info = -15; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_chetrd_nb(n); lopt = n * (nb + 1); work[0] = MAGMA_C_MAKE( lopt, 0 ); if (lwork < lopt && ! lquery) { *info = -17; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } *m = 0; /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_cheevx(jobz_, range_, uplo_, &n, a, &lda, &vl, &vu, &il, &iu, &abstol, m, w, z, &ldz, work, &lwork, rwork, iwork, ifail, info); return *info; } --w; --work; --rwork; --iwork; --ifail; /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_clanhe("M", uplo_, &n, a, &lda, &rwork[1]); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { d__1 = 1.; lapackf77_clascl(uplo_, &izero, &izero, &d__1, &sigma, &n, &n, a, &lda, info); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ indd = 1; inde = indd + n; indrwk = inde + n; indtau = 1; indwrk = indtau + n; llwork = lwork - indwrk + 1; magma_chetrd(uplo, n, a, lda, &rwork[indd], &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo); lopt = n + (magma_int_t)MAGMA_C_REAL(work[indwrk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call SSTERF or CUNGTR and CSTEQR. If this fails for some eigenvalue, then try SSTEBZ. */ if ((alleig || (indeig && il == 1 && iu == n)) && abstol <= 0.) { blasf77_scopy(&n, &rwork[indd], &ione, &w[1], &ione); indee = indrwk + 2*n; if (! wantz) { i__1 = n - 1; blasf77_scopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_ssterf(&n, &w[1], &rwork[indee], info); } else { lapackf77_clacpy("A", &n, &n, a, &lda, z, &ldz); lapackf77_cungtr(uplo_, &n, z, &ldz, &work[indtau], &work[indwrk], &llwork, &iinfo); i__1 = n - 1; blasf77_scopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_csteqr(jobz_, &n, &w[1], &rwork[indee], z, &ldz, &rwork[indrwk], info); if (*info == 0) { for (i = 1; i <= n; ++i) { ifail[i] = 0; } } } if (*info == 0) { *m = n; } } /* Otherwise, call SSTEBZ and, if eigenvectors are desired, CSTEIN. */ if (*m == 0) { *info = 0; if (wantz) { *(unsigned char *)order = 'B'; } else { *(unsigned char *)order = 'E'; } indibl = 1; indisp = indibl + n; indiwk = indisp + n; lapackf77_sstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m, &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info); if (wantz) { lapackf77_cstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp], z, &ldz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by CSTEIN. */ magma_cunmtr(MagmaLeft, uplo, MagmaNoTrans, n, *m, a, lda, &work[indtau], z, ldz, &work[indwrk], llwork, &iinfo); } } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = *m; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, &w[1], &ione); } /* If eigenvalues are not in order, then sort them, along with eigenvectors. */ if (wantz) { for (j = 1; j <= *m-1; ++j) { i = 0; tmp1 = w[j]; for (jj = j + 1; jj <= *m; ++jj) { if (w[jj] < tmp1) { i = jj; tmp1 = w[jj]; } } if (i != 0) { itmp1 = iwork[indibl + i - 1]; w[i] = w[j]; iwork[indibl + i - 1] = iwork[indibl + j - 1]; w[j] = tmp1; iwork[indibl + j - 1] = itmp1; blasf77_cswap(&n, z + (i-1)*ldz, &ione, z + (j-1)*ldz, &ione); if (*info != 0) { itmp1 = ifail[i]; ifail[i] = ifail[j]; ifail[j] = itmp1; } } } } /* Set WORK(1) to optimal complex workspace size. */ work[1] = MAGMA_C_MAKE( lopt, 0 ); return *info; } /* magma_cheevx */
/** Purpose ------- CHETRD2_GPU reduces a complex Hermitian matrix A to real symmetric tridiagonal form T by an orthogonal similarity transformation: Q**H * A * Q = T. This version passes a workspace that is used in an optimized GPU matrix-vector product. Arguments --------- @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] dA COMPLEX array, dimension (LDA,N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal of A are overwritten by the corresponding elements of the tridiagonal matrix T, and the elements above the first superdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = MagmaLower, the diagonal and first subdiagonal of A are over- written by the corresponding elements of the tridiagonal matrix T, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. @param[in] ldda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] d COMPLEX array, dimension (N) The diagonal elements of the tridiagonal matrix T: D(i) = A(i,i). @param[out] e COMPLEX array, dimension (N-1) The off-diagonal elements of the tridiagonal matrix T: E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower. @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). @param[out] wA (workspace) COMPLEX array, dimension (LDA,N) On exit the diagonal, the upper part (UPLO=MagmaUpper) or the lower part (UPLO=MagmaLower) are copies of DA @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= 1. For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] dwork (workspace) COMPLEX array on the GPU, dim (MAX(1,LDWORK)) @param[in] ldwork INTEGER The dimension of the array DWORK. LDWORK >= (n*n+64-1)/64 + 2*n*nb, where nb = magma_get_chetrd_nb(n) @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). @ingroup magma_cheev_comp ********************************************************************/ extern "C" magma_int_t magma_chetrd2_gpu(magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *dA, magma_int_t ldda, float *d, float *e, magmaFloatComplex *tau, magmaFloatComplex *wA, magma_int_t ldwa, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex *dwork, magma_int_t ldwork, magma_int_t *info) { #define A(i, j) (wA + (j)*ldwa + (i)) #define dA(i, j) (dA + (j)*ldda + (i)) const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t nb = magma_get_chetrd_nb(n); magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex c_one = MAGMA_C_ONE; float d_one = MAGMA_D_ONE; magma_int_t kk, nx; magma_int_t i, j, i_n; magma_int_t iinfo; magma_int_t ldw, lddw, lwkopt; magma_int_t lquery; *info = 0; int upper = (uplo == MagmaUpper); lquery = (lwork == -1); if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } else if (ldwa < max(1,n)) { *info = -9; } else if (lwork < 1 && ! lquery) { *info = -11; } /* Determine the block size. */ ldw = lddw = n; lwkopt = n * nb; if (*info == 0) { work[0] = MAGMA_C_MAKE( lwkopt, 0 ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; /* Quick return if possible */ if (n == 0) { work[0] = c_one; return *info; } if (n < 1024) nx = n; else nx = 300; if (ldwork < (ldw*n+64-1)/64 + 2*ldw*nb) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } if (upper) { /* Reduce the upper triangle of A. Columns 1:kk are handled by the unblocked method. */ kk = n - (n - nx + nb - 1) / nb * nb; for (i = n - nb; i >= kk; i -= nb) { /* Reduce columns i:i+nb-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel */ magma_cgetmatrix( i+nb, nb, dA(0, i), ldda, A(0, i), ldwa ); magma_clatrd2(uplo, i+nb, nb, A(0, 0), ldwa, e, tau, work, ldw, dA(0, 0), ldda, dwork, lddw, dwork + 2*ldw*nb, ldwork - 2*ldw*nb); /* Update the unreduced submatrix A(0:i-2,0:i-2), using an update of the form: A := A - V*W' - W*V' */ magma_csetmatrix( i + nb, nb, work, ldw, dwork, lddw ); magma_cher2k(uplo, MagmaNoTrans, i, nb, c_neg_one, dA(0, i), ldda, dwork, lddw, d_one, dA(0, 0), ldda); /* Copy superdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+nb; ++j) { *A(j-1,j) = MAGMA_C_MAKE( e[j - 1], 0 ); d[j] = MAGMA_C_REAL( *A(j, j) ); } } magma_cgetmatrix( kk, kk, dA(0, 0), ldda, A(0, 0), ldwa ); /* Use CPU code to reduce the last or only block */ lapackf77_chetrd(uplo_, &kk, A(0, 0), &ldwa, d, e, tau, work, &lwork, &iinfo); magma_csetmatrix( kk, kk, A(0, 0), ldwa, dA(0, 0), ldda ); } else { /* Reduce the lower triangle of A */ for (i = 0; i < n-nx; i += nb) { /* Reduce columns i:i+nb-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel */ magma_cgetmatrix( n-i, nb, dA(i, i), ldda, A(i, i), ldwa ); magma_clatrd2(uplo, n-i, nb, A(i, i), ldwa, &e[i], &tau[i], work, ldw, dA(i, i), ldda, dwork, lddw, dwork + 2*ldw*nb, ldwork - 2*ldw*nb); /* Update the unreduced submatrix A(i+ib:n,i+ib:n), using an update of the form: A := A - V*W' - W*V' */ magma_csetmatrix( n-i, nb, work, ldw, dwork, lddw ); magma_cher2k(MagmaLower, MagmaNoTrans, n-i-nb, nb, c_neg_one, dA(i+nb, i), ldda, &dwork[nb], lddw, d_one, dA(i+nb, i+nb), ldda); /* Copy subdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+nb; ++j) { *A(j+1,j) = MAGMA_C_MAKE( e[j], 0 ); d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* Use unblocked code to reduce the last or only block */ magma_cgetmatrix( n-i, n-i, dA(i, i), ldda, A(i, i), ldwa ); i_n = n-i; lapackf77_chetrd(uplo_, &i_n, A(i, i), &ldwa, &d[i], &e[i], &tau[i], work, &lwork, &iinfo); magma_csetmatrix( n-i, n-i, A(i, i), ldwa, dA(i, i), ldda ); } work[0] = MAGMA_C_MAKE( lwkopt, 0 ); return *info; } /* magma_chetrd2_gpu */
extern "C" magma_int_t magma_cheevd(magma_vec_t jobz, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *a, magma_int_t lda, float *w, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info, magma_queue_t queue) { /* -- clMAGMA (version 1.1.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver @date January 2014 Purpose ======= CHEEVD computes all eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments ========= JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = 'V', then if INFO = 0, A contains the orthonormal eigenvectors of the matrix A. If JOBZ = 'N', then on exit the lower triangle (if UPLO='L') or the upper triangle (if UPLO='U') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). W (output) REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. WORK (workspace/output) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. If N <= 1, LWORK must be at least 1. If JOBZ = 'N' and N > 1, LWORK must be at least N + N*NB. If JOBZ = 'V' and N > 1, LWORK must be at least 2*N + N**2. NB can be obtained through magma_get_chetrd_nb(N). If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. RWORK (workspace/output) REAL array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. LRWORK (input) INTEGER The dimension of the array RWORK. If N <= 1, LRWORK must be at least 1. If JOBZ = 'N' and N > 1, LRWORK must be at least N. If JOBZ = 'V' and N > 1, LRWORK must be at least 1 + 5*N + 2*N**2. If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. LIWORK (input) INTEGER The dimension of the array IWORK. If N <= 1, LIWORK must be at least 1. If JOBZ = 'N' and N > 1, LIWORK must be at least 1. If JOBZ = 'V' and N > 1, LIWORK must be at least 3 + 5*N. If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i and JOBZ = 'N', then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = 'V', then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details =============== Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. ===================================================================== */ magma_uplo_t uplo_ = uplo; magma_vec_t jobz_ = jobz; magma_int_t ione = 1; magma_int_t izero = 0; float d_one = 1.; float d__1; float eps; magma_int_t inde; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; float smlnum; magma_int_t lquery; magmaFloat_ptr dwork; wantz = lapackf77_lsame(lapack_const(jobz_), MagmaVecStr); lower = lapackf77_lsame(lapack_const(uplo_), MagmaLowerStr); lquery = lwork == -1 || lrwork == -1 || liwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(lapack_const(jobz_), MagmaNoVecStr))) { *info = -1; } else if (! (lower || lapackf77_lsame(lapack_const(uplo_), MagmaUpperStr))) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = 2*n + n*n; lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps to ensure length gets rounded up, // if it cannot be exactly represented in floating point. work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -8; } else if ((lrwork < lrwmin) && ! lquery) { *info = -10; } else if ((liwork < liwmin) && ! lquery) { *info = -12; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_C_REAL(a[0]); if (wantz) { a[0] = MAGMA_C_ONE; } return *info; } /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_clanhe("M", lapack_const(uplo_), &n, a, &lda, rwork); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_clascl(lapack_const(uplo_), &izero, &izero, &d_one, &sigma, &n, &n, a, &lda, info); } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ // chetrd rwork: e (n) // cstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // chetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // cstedx work: tau (n) + z (n^2) // cunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; //#define ENABLE_TIMER #ifdef ENABLE_TIMER magma_timestr_t start, end; start = get_current_time(); #endif magma_chetrd(lapack_const(uplo)[0], n, a, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo, queue); #ifdef ENABLE_TIMER end = get_current_time(); printf("time chetrd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif /* For eigenvalues only, call SSTERF. For eigenvectors, first call CSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call CUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &rwork[inde], info); } else { #ifdef ENABLE_TIMER start = get_current_time(); #endif if (MAGMA_SUCCESS != magma_smalloc( &dwork, (3*n*(n/2 + 1) ) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_cstedx(MagmaAllVec, n, 0., 0., 0, 0, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info, queue); magma_free( dwork ); #ifdef ENABLE_TIMER end = get_current_time(); printf("time cstedx = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif magma_cunmtr(MagmaLeft, uplo, MagmaNoTrans, n, n, a, lda, &work[indtau], &work[indwrk], n, &work[indwk2], llwrk2, &iinfo, queue); lapackf77_clacpy("A", &n, &n, &work[indwrk], &n, a, &lda); #ifdef ENABLE_TIMER end = get_current_time(); printf("time cunmtr + copy = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, w, &ione); } work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); // round up rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; return *info; } /* magma_cheevd */
/** Purpose ------- CHETRD reduces a complex Hermitian matrix A to real symmetric tridiagonal form T by an orthogonal similarity transformation: Q**H * A * Q = T. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @param[in] nqueue INTEGER The number of GPU queues used for update. 10 >= nqueue > 0. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA,N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal of A are overwritten by the corresponding elements of the tridiagonal matrix T, and the elements above the first superdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = MagmaLower, the diagonal and first subdiagonal of A are over- written by the corresponding elements of the tridiagonal matrix T, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] d COMPLEX array, dimension (N) The diagonal elements of the tridiagonal matrix T: D(i) = A(i,i). @param[out] e COMPLEX array, dimension (N-1) The off-diagonal elements of the tridiagonal matrix T: E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower. @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= N*NB, where NB is the optimal blocksize given by magma_get_chetrd_nb(). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). @ingroup magma_cheev_comp ********************************************************************/ extern "C" magma_int_t magma_chetrd_mgpu( magma_int_t ngpu, magma_int_t nqueue, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, float *d, float *e, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define dA(id, i, j) (dA[(id)] + (j)*ldda + (i)) #define dW(id, i, j) (dW[(id)] + (j)*ldda + (i)) /* Constants */ const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; const magmaFloatComplex c_one = MAGMA_C_ONE; const float d_one = MAGMA_D_ONE; /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t nlocal, ldda; magma_int_t nb = magma_get_chetrd_nb(n), ib, ib2; #ifdef PROFILE_SY2RK float mv_time = 0.0; float up_time = 0.0; #endif magma_int_t kk, nx; magma_int_t i, ii, iii, j, dev, i_n; magma_int_t iinfo; magma_int_t ldwork, lddw, lwkopt, ldwork2, lhwork; // set pointers to NULL so it is safe to goto CLEANUP if any malloc fails. magma_queue_t queues[MagmaMaxGPUs][10] = { { NULL, NULL } }; magma_queue_t queues0[MagmaMaxGPUs] = { NULL }; magmaFloatComplex *hwork = NULL; magmaFloatComplex_ptr dwork2[MagmaMaxGPUs] = { NULL }; magmaFloatComplex_ptr dA[MagmaMaxGPUs] = { NULL }; magmaFloatComplex_ptr dW[MagmaMaxGPUs] = { NULL }; *info = 0; bool upper = (uplo == MagmaUpper); bool lquery = (lwork == -1); if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,n)) { *info = -4; } else if (lwork < nb*n && ! lquery) { *info = -9; } else if ( nqueue > 2 ) { *info = 2; // TODO fix } /* Determine the block size. */ ldwork = n; lwkopt = n * nb; if (*info == 0) { work[0] = magma_cmake_lwork( lwkopt ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { work[0] = c_one; return *info; } magma_device_t orig_dev; magma_getdevice( &orig_dev ); //#define PROFILE_SY2RK #ifdef PROFILE_SY2RK float times[11] = { 0 }; magma_event_t start, stop; float etime; magma_setdevice( 0 ); magma_event_create( &start ); magma_event_create( &stop ); #endif ldda = magma_roundup( lda, 32 ); lddw = ldda; nlocal = nb*(1 + n/(nb*ngpu)); ldwork2 = ldda*( magma_ceildiv( n, nb ) + 1); // i.e., ldda*(blocks + 1) for( dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); // TODO fix memory leak if ( MAGMA_SUCCESS != magma_cmalloc( &dA[dev], nlocal*ldda + 3*lddw*nb ) || MAGMA_SUCCESS != magma_cmalloc( &dwork2[dev], ldwork2 ) ) { *info = MAGMA_ERR_DEVICE_ALLOC; goto CLEANUP; } dW[dev] = dA[dev] + nlocal*ldda; for( kk=0; kk < nqueue; kk++ ) { magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queues[dev][kk] ); } queues0[dev] = queues[dev][0]; } lhwork = nqueue*ngpu*n; if ( MAGMA_SUCCESS != magma_cmalloc_pinned( &hwork, lhwork ) ) { *info = MAGMA_ERR_HOST_ALLOC; goto CLEANUP; } // nx <= n is required // use LAPACK for n < 3000, otherwise switch at 512 if (n < 3000) nx = n; else nx = 512; if (upper) { /* Copy the matrix to the GPU */ if (1 <= n-nx) { magma_chtodhe( ngpu, uplo, n, nb, A, lda, dA, ldda, queues, &iinfo ); } /* Reduce the upper triangle of A. Columns 1:kk are handled by the unblocked method. */ for (i = nb*((n-1)/nb); i >= nx; i -= nb) { ib = min(nb, n-i); ii = nb*(i/(nb*ngpu)); dev = (i/nb)%ngpu; /* wait for the next panel */ if (i != nb*((n-1)/nb)) { magma_setdevice( dev ); magma_queue_sync( queues[dev][0] ); } magma_clatrd_mgpu( ngpu, uplo, i+ib, ib, nb, A(0, 0), lda, e, tau, work, ldwork, dA, ldda, 0, dW, i+ib, hwork, lhwork, dwork2, ldwork2, queues0 ); magma_cher2k_mgpu( ngpu, MagmaUpper, MagmaNoTrans, nb, i, ib, c_neg_one, dW, i+ib, 0, d_one, dA, ldda, 0, nqueue, queues ); /* get the next panel */ if (i-nb >= nx ) { ib2 = min(nb, n-(i-nb)); ii = nb*((i-nb)/(nb*ngpu)); dev = ((i-nb)/nb)%ngpu; magma_setdevice( dev ); magma_cgetmatrix_async( (i-nb)+ib2, ib2, dA(dev, 0, ii), ldda, A(0, i-nb), lda, queues[dev][0] ); } /* Copy superdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+ib; ++j) { if ( j > 0 ) { *A(j-1,j) = MAGMA_C_MAKE( e[j - 1], 0 ); } d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* end of for i=... */ if ( nx > 0 ) { if (1 <= n-nx) { /* else A is already on CPU */ for (i=0; i < nx; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*ngpu)); dev = (i/nb)%ngpu; magma_setdevice( dev ); magma_cgetmatrix_async( nx, ib, dA(dev, 0, ii), ldda, A(0, i), lda, queues[dev][0] ); } } for( dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); magma_queue_sync( queues[dev][0] ); } /* Use CPU code to reduce the last or only block */ lapackf77_chetrd( uplo_, &nx, A(0, 0), &lda, d, e, tau, work, &lwork, &iinfo ); } } else { trace_init( 1, ngpu, nqueue, queues ); /* Copy the matrix to the GPU */ if (1 <= n-nx) { magma_chtodhe( ngpu, uplo, n, nb, A, lda, dA, ldda, queues, &iinfo ); } /* Reduce the lower triangle of A */ for (i = 0; i < n-nx; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*ngpu)); dev = (i/nb)%ngpu; /* Reduce columns i:i+ib-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel (no need for the 1st iteration) */ if (i != 0) { magma_setdevice( dev ); trace_gpu_start( dev, 0, "comm", "get" ); magma_cgetmatrix_async( n-i, ib, dA(dev, i, ii), ldda, A(i,i), lda, queues[dev][0] ); trace_gpu_end( dev, 0 ); magma_queue_sync( queues[dev][0] ); magma_setdevice( 0 ); } magma_clatrd_mgpu( ngpu, uplo, n-i, ib, nb, A(i, i), lda, &e[i], &tau[i], work, ldwork, dA, ldda, i, dW, n-i, hwork, lhwork, dwork2, ldwork2, queues0 ); #ifdef PROFILE_SY2RK magma_setdevice( 0 ); if ( i > 0 ) { cudaEventElapsedTime( &etime, start, stop ); up_time += (etime/1000.0); } magma_event_record( start, 0 ); #endif magma_cher2k_mgpu( ngpu, MagmaLower, MagmaNoTrans, nb, n-i-ib, ib, c_neg_one, dW, n-i, ib, d_one, dA, ldda, i+ib, nqueue, queues ); #ifdef PROFILE_SY2RK magma_setdevice( 0 ); magma_event_record( stop, 0 ); #endif /* Copy subdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+ib; ++j) { if ( j+1 < n ) { *A(j+1,j) = MAGMA_C_MAKE( e[j], 0 ); } d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* for i=... */ /* Use CPU code to reduce the last or only block */ if ( i < n ) { iii = i; i_n = n-i; if ( i > 0 ) { for (; i < n; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*ngpu)); dev = (i/nb)%ngpu; magma_setdevice( dev ); magma_cgetmatrix_async( i_n, ib, dA(dev, iii, ii), ldda, A(iii, i), lda, queues[dev][0] ); } for( dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); magma_queue_sync( queues[dev][0] ); } } lapackf77_chetrd( uplo_, &i_n, A(iii, iii), &lda, &d[iii], &e[iii], &tau[iii], work, &lwork, &iinfo ); } } for( dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); for( kk=0; kk < nqueue; kk++ ) { magma_queue_sync( queues[dev][kk] ); } } #ifdef PROFILE_SY2RK magma_setdevice( 0 ); if ( n > nx ) { cudaEventElapsedTime( &etime, start, stop ); up_time += (etime/1000.0); } magma_event_destroy( start ); magma_event_destroy( stop ); #endif trace_finalize( "chetrd.svg", "trace.css" ); #ifdef PROFILE_SY2RK printf( " n=%d nb=%d\n", n, nb ); printf( " Time in CLARFG: %.2e seconds\n", times[0] ); //printf( " Time in CHEMV : %.2e seconds\n", mv_time ); printf( " Time in CHER2K: %.2e seconds\n", up_time ); #endif CLEANUP: for( dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); for( kk=0; kk < nqueue; kk++ ) { magma_queue_destroy( queues[dev][kk] ); } magma_free( dA[dev] ); magma_free( dwork2[dev] ); } magma_free_pinned( hwork ); magma_setdevice( orig_dev ); work[0] = magma_cmake_lwork( lwkopt ); return *info; } /* magma_chetrd */
/* //////////////////////////////////////////////////////////////////////////// -- Testing chegvd */ int main( int argc, char** argv) { TESTING_INIT_MGPU(); magmaFloatComplex *h_A, *h_Ainit, *h_B, *h_Binit, *h_work; #if defined(PRECISION_z) || defined(PRECISION_c) float *rwork; #endif float *w1, *w2, result; magma_int_t *iwork; real_Double_t mgpu_time, gpu_time, cpu_time; /* Matrix size */ magma_int_t N, n2, nb; magma_int_t info; magma_int_t ione = 1; magmaFloatComplex c_zero = MAGMA_C_ZERO; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); float tol = opts.tolerance * lapackf77_slamch("E"); float tolulp = opts.tolerance * lapackf77_slamch("P"); if ( opts.check && opts.jobz == MagmaNoVec ) { fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" ); opts.jobz = MagmaVec; } printf("using: ngpu = %d, itype = %d, jobz = %s, uplo = %s, check = %d\n", (int) opts.ngpu, (int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), (int) opts.check); printf(" N CPU Time (sec) GPU Time (sec) MGPU Time (sec)\n"); printf("=========================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { // TODO define lda N = opts.nsize[itest]; n2 = N*N; nb = magma_get_chetrd_nb(N); #if defined(PRECISION_z) || defined(PRECISION_c) magma_int_t lwork = max( N + N*nb, 2*N + N*N ); magma_int_t lrwork = 1 + 5*N +2*N*N; #else magma_int_t lwork = max( 2*N + N*nb, 1 + 6*N + 2*N*N ); #endif magma_int_t liwork = 3 + 5*N; TESTING_MALLOC_PIN( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_PIN( h_B, magmaFloatComplex, n2 ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_MALLOC_PIN( rwork, float, lrwork ); #endif TESTING_MALLOC_CPU( w1, float, N ); TESTING_MALLOC_CPU( w2, float, N ); TESTING_MALLOC_CPU( iwork, magma_int_t, liwork ); /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clarnv( &ione, ISEED, &n2, h_B ); magma_cmake_hpd( N, h_B, N ); magma_cmake_hermitian( N, h_A, N ); if ( opts.warmup || opts.check ) { TESTING_MALLOC_CPU( h_Ainit, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_Binit, magmaFloatComplex, n2 ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N ); } if (opts.warmup) { // ================================================================== // Warmup using MAGMA. // ================================================================== magma_chegvd_m( opts.ngpu, opts.itype, opts.jobz, opts.uplo, N, h_A, N, h_B, N, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N ); } // =================================================================== // Performs operation using MAGMA // =================================================================== mgpu_time = magma_wtime(); magma_chegvd_m( opts.ngpu, opts.itype, opts.jobz, opts.uplo, N, h_A, N, h_B, N, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); mgpu_time = magma_wtime() - mgpu_time; if (info != 0) printf("magma_chegvd_m returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.check ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvd routine. A x = lambda B x is solved and the following 3 tests computed: (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) | A B Z - Z D | / ( |A||Z| N ) (itype = 2) | B A Z - Z D | / ( |A||Z| N ) (itype = 3) =================================================================== */ #if defined(PRECISION_d) || defined(PRECISION_s) float *rwork = h_work + N*N; #endif result = 1.; result /= lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, rwork); result /= lapackf77_clange("1", &N, &N, h_A, &N, rwork); if (opts.itype == 1) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N); result *= lapackf77_clange("1", &N, &N, h_work, &N, rwork)/N; } else if (opts.itype == 2) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N); result *= lapackf77_clange("1", &N, &N, h_A, &N, rwork)/N; } else if (opts.itype == 3) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N); result *= lapackf77_clange("1", &N, &N, h_A, &N, rwork)/N; } lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_chegvd(opts.itype, opts.jobz, opts.uplo, N, h_A, N, h_B, N, w2, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); gpu_time = magma_wtime() - gpu_time; if (info != 0) printf("magma_chegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_chegvd(&opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, h_Binit, &N, w2, h_work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, &info); cpu_time = magma_wtime() - cpu_time; if (info != 0) printf("lapackf77_chegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); float temp1 = 0; float temp2 = 0; for(int j=0; j<N; j++) { temp1 = max(temp1, fabs(w1[j])); temp1 = max(temp1, fabs(w2[j])); temp2 = max(temp2, fabs(w1[j]-w2[j])); } float result2 = temp2 / (((float)N)*temp1); /* ===================================================================== Print execution time =================================================================== */ printf("%5d %7.2f %7.2f %7.2f\n", (int) N, cpu_time, gpu_time, mgpu_time); printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if (opts.itype==1) { printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") ); } else if (opts.itype==2) { printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") ); } else if (opts.itype==3) { printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") ); } printf( "(3) | D(MGPU)-D(LAPACK) |/ |D| = %8.2e %s\n\n", result2, (result2 < tolulp ? "ok" : "failed") ); status += ! (result < tol && result2 < tolulp); } else { printf("%5d --- --- %7.2f\n", (int) N, mgpu_time); } /* Memory clean up */ TESTING_FREE_PIN( h_A ); TESTING_FREE_PIN( h_B ); TESTING_FREE_PIN( h_work ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_FREE_PIN( rwork ); #endif TESTING_FREE_CPU( w1 ); TESTING_FREE_CPU( w2 ); TESTING_FREE_CPU( iwork ); if ( opts.warmup || opts.check ) { TESTING_FREE_CPU( h_Ainit ); TESTING_FREE_CPU( h_Binit ); } fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } /* Shutdown */ TESTING_FINALIZE_MGPU(); return status; }
/** Purpose ------- CHEEVD computes all eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = MagmaVec, then if INFO = 0, A contains the orthonormal eigenvectors of the matrix A. If JOBZ = MagmaNoVec, then on exit the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] w REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_chetrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) REAL array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details --------------- Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. @ingroup magma_cheev_driver ********************************************************************/ extern "C" magma_int_t magma_cheevd( magma_vec_t jobz, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, float *w, magmaFloatComplex *work, magma_int_t lwork, #ifdef COMPLEX float *rwork, magma_int_t lrwork, #endif magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magma_int_t ione = 1; magma_int_t izero = 0; float d_one = 1.; float d__1; float eps; magma_int_t inde; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; float smlnum; magma_int_t lquery; float* dwork; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (! (wantz || (jobz == MagmaNoVec))) { *info = -1; } else if (! (lower || (uplo == MagmaUpper))) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps (in Double!) to ensure length gets rounded up, // if it cannot be exactly represented in floating point. real_Double_t one_eps = 1. + lapackf77_slamch("Epsilon"); work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0 ); rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -8; } else if ((lrwork < lrwmin) && ! lquery) { *info = -10; } else if ((liwork < liwmin) && ! lquery) { *info = -12; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_C_REAL( A[0] ); if (wantz) { A[0] = MAGMA_C_ONE; } return *info; } /* If matrix is very small, then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { lapackf77_cheevd( jobz_, uplo_, &n, A, &lda, w, work, &lwork, #ifdef COMPLEX rwork, &lrwork, #endif iwork, &liwork, info ); return *info; } /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt( smlnum ); rmax = magma_ssqrt( bignum ); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_clanhe( "M", uplo_, &n, A, &lda, rwork ); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_clascl( uplo_, &izero, &izero, &d_one, &sigma, &n, &n, A, &lda, info ); } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ // chetrd rwork: e (n) // cstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // chetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // cstedx work: tau (n) + z (n^2) // cunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; magma_timer_t time=0; timer_start( time ); magma_chetrd( uplo, n, A, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo ); timer_stop( time ); timer_printf( "time chetrd = %6.2f\n", time ); /* For eigenvalues only, call SSTERF. For eigenvectors, first call * CSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the * tridiagonal matrix, then call CUNMTR to multiply it to the Householder * transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf( &n, w, &rwork[inde], info ); } else { timer_start( time ); if (MAGMA_SUCCESS != magma_smalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_cstedx( MagmaRangeAll, n, 0., 0., 0, 0, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info ); magma_free( dwork ); timer_stop( time ); timer_printf( "time cstedx = %6.2f\n", time ); timer_start( time ); magma_cunmtr( MagmaLeft, uplo, MagmaNoTrans, n, n, A, lda, &work[indtau], &work[indwrk], n, &work[indwk2], llwrk2, &iinfo ); lapackf77_clacpy( "A", &n, &n, &work[indwrk], &n, A, &lda ); timer_stop( time ); timer_printf( "time cunmtr + copy = %6.2f\n", time ); } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal( &imax, &d__1, w, &ione ); } work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0 ); // round up rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; return *info; } /* magma_cheevd */
/* //////////////////////////////////////////////////////////////////////////// -- Testing chegvdx */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gpu_time /*cpu_time*/; magmaFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work; float *w1, *w2, vl=0, vu=0; float result[2] = {0}; magma_int_t *iwork; magma_int_t N, n2, info, il, iu, m1, nb, lwork, liwork; magmaFloatComplex c_zero = MAGMA_C_ZERO; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; #ifdef COMPLEX float *rwork; magma_int_t lrwork; #endif //float d_one = 1.; //float d_ten = 10.; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); float tol = opts.tolerance * lapackf77_slamch("E"); //float tolulp = opts.tolerance * lapackf77_slamch("P"); // TODO: how to check NoVec? This doesn't have an equivalent call to LAPACK. if ( opts.check && opts.jobz == MagmaNoVec ) { fprintf( stderr, "WARNING: cannot currently check results when not computing vectors (option -JN).\n" ); } printf("using: itype = %d, jobz = %s, uplo = %s, check = %d, fraction = %6.4f\n", (int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), (int) opts.check, opts.fraction); printf(" N M GPU Time (sec)\n"); printf("============================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; n2 = N*N; nb = magma_get_chetrd_nb(N); #ifdef COMPLEX lwork = max( N + N*nb, 2*N + N*N ); lrwork = 1 + 5*N +2*N*N; #else lwork = max( 2*N + N*nb, 1 + 6*N + 2*N*N ); #endif liwork = 3 + 5*N; if ( opts.fraction == 0 ) { il = N / 10; iu = N / 5+il; } else { il = 1; iu = (int) (opts.fraction*N); if (iu < 1) iu = 1; } TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_B, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( w1, float, N ); TESTING_MALLOC_CPU( w2, float, N ); TESTING_MALLOC_CPU( iwork, magma_int_t, liwork ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_PIN( h_S, magmaFloatComplex, n2 ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork ); #ifdef COMPLEX TESTING_MALLOC_PIN( rwork, float, lrwork); #endif /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clarnv( &ione, ISEED, &n2, h_B ); magma_cmake_hpd( N, h_B, N ); magma_cmake_hermitian( N, h_A, N ); // ================================================================== // Warmup using MAGMA // ================================================================== if ( opts.warmup ) { lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N ); magma_chegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo, N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #ifdef COMPLEX rwork, lrwork, #endif iwork, liwork, &info ); if (info != 0) printf("magma_chegvdx returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ==================================================================== Performs operation using MAGMA =================================================================== */ lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N ); gpu_time = magma_wtime(); magma_chegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo, N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #ifdef COMPLEX rwork, lrwork, #endif iwork, liwork, &info ); gpu_time = magma_wtime() - gpu_time; if (info != 0) printf("magma_chegvdx returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.check && opts.jobz != MagmaNoVec ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvdx routine. A x = lambda B x is solved and the following 3 tests computed: (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) | A B Z - Z D | / ( |A||Z| N ) (itype = 2) | B A Z - Z D | / ( |A||Z| N ) (itype = 3) (2) | D(with V) - D(w/o V) | / | D | =================================================================== */ #ifdef REAL float *rwork = h_work + N*N; #endif result[0] = 1.; result[0] /= lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork); result[0] /= lapackf77_clange("1", &N, &m1, h_R, &N, rwork); if (opts.itype == 1) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i < m1; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N); result[0] *= lapackf77_clange("1", &N, &m1, h_work, &N, rwork)/N; } else if (opts.itype == 2) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i < m1; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_clange("1", &N, &m1, h_R, &N, rwork)/N; } else if (opts.itype == 3) { blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i < m1; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_clange("1", &N, &m1, h_R, &N, rwork)/N; } // Disable eigenvalue check which calls routine again -- // it obscures whether error occurs in first call above or in this call. // TODO: add comparison to LAPACK, as in testing_chegvd.cpp? // //lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &N, h_R, &N ); //lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &N, h_S, &N ); // //magma_int_t m2; //magma_chegvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo, // N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2, // h_work, lwork, // #ifdef COMPLEX // rwork, lrwork, // #endif // iwork, liwork, // &info ); //if (info != 0) // printf("magma_chegvdx returned error %d: %s.\n", // (int) info, magma_strerror( info )); // //float maxw=0, diff=0; //for(int j=0; j < m2; j++) { // maxw = max(maxw, fabs(w1[j])); // maxw = max(maxw, fabs(w2[j])); // diff = max(diff, fabs(w1[j]-w2[j])); //} //result[1] = diff / (m2*maxw); } /* ===================================================================== Print execution time =================================================================== */ printf("%5d %5d %7.2f\n", (int) N, (int) m1, gpu_time); if ( opts.check && opts.jobz != MagmaNoVec ) { printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if (opts.itype == 1) { printf(" | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } else if (opts.itype == 2) { printf(" | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } else if (opts.itype == 3) { printf(" | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } //printf( " | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed")); status += ! (result[0] < tol); // && result[1] < tolulp); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( w1 ); TESTING_FREE_CPU( w2 ); TESTING_FREE_CPU( iwork ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_S ); TESTING_FREE_PIN( h_work ); #ifdef COMPLEX TESTING_FREE_PIN( rwork ); #endif fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_chegvr(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n, magmaFloatComplex *a, magma_int_t lda, magmaFloatComplex *b, magma_int_t ldb, float vl, float vu, magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m, float *w, magmaFloatComplex *z, magma_int_t ldz, magma_int_t *isuppz, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= CHEGVR computes all the eigenvalues, and optionally, the eigenvectors of a complex generalized Hermitian-definite eigenproblem, of the form A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and B are assumed to be Hermitian and B is also positive definite. Whenever possible, CHEEVR calls CSTEGR to compute the eigenspectrum using Relatively Robust Representations. CSTEGR computes eigenvalues by the dqds algorithm, while orthogonal eigenvectors are computed from various "good" L D L^T representations (also known as Relatively Robust Representations). Gram-Schmidt orthogonalization is avoided as far as possible. More specifically, the various steps of the algorithm are as follows. For the i-th unreduced block of T, (a) Compute T - sigma_i = L_i D_i L_i^T, such that L_i D_i L_i^T is a relatively robust representation, (b) Compute the eigenvalues, lambda_j, of L_i D_i L_i^T to high relative accuracy by the dqds algorithm, (c) If there is a cluster of close eigenvalues, "choose" sigma_i close to the cluster, and go to step (a), (d) Given the approximate eigenvalue lambda_j of L_i D_i L_i^T, compute the corresponding eigenvector by forming a rank-revealing twisted factorization. The desired accuracy of the output can be specified by the input parameter ABSTOL. For more details, see "A new O(n^2) algorithm for the symmetric tridiagonal eigenvalue/eigenvector problem", by Inderjit Dhillon, Computer Science Division Technical Report No. UCB//CSD-97-971, UC Berkeley, May 1997. Note 1 : CHEEVR calls CSTEGR when the full spectrum is requested on machines which conform to the ieee-754 floating point standard. CHEEVR calls SSTEBZ and CSTEIN on non-ieee machines and when partial spectrum requests are made. Normal execution of CSTEGR may create NaNs and infinities and hence may abort due to a floating point exception in environments which do not handle NaNs and infinities in the ieee standard default manner. Arguments ========= ITYPE (input) INTEGER Specifies the problem type to be solved: = 1: A*x = (lambda)*B*x = 2: A*B*x = (lambda)*x = 3: B*A*x = (lambda)*x RANGE (input) CHARACTER*1 = 'A': all eigenvalues will be found. = 'V': all eigenvalues in the half-open interval (VL,VU] will be found. = 'I': the IL-th through IU-th eigenvalues will be found. JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. UPLO (input) CHARACTER*1 = 'U': Upper triangles of A and B are stored; = 'L': Lower triangles of A and B are stored. N (input) INTEGER The order of the matrices A and B. N >= 0. A (input/output) COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = 'V', then if INFO = 0, A contains the matrix Z of eigenvectors. The eigenvectors are normalized as follows: if ITYPE = 1 or 2, Z**H*B*Z = I; if ITYPE = 3, Z**H*inv(B)*Z = I. If JOBZ = 'N', then on exit the upper triangle (if UPLO='U') or the lower triangle (if UPLO='L') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). B (input/output) COMPLEX array, dimension (LDB, N) On entry, the Hermitian matrix B. If UPLO = 'U', the leading N-by-N upper triangular part of B contains the upper triangular part of the matrix B. If UPLO = 'L', the leading N-by-N lower triangular part of B contains the lower triangular part of the matrix B. On exit, if INFO <= N, the part of B containing the matrix is overwritten by the triangular factor U or L from the Cholesky factorization B = U**H*U or B = L*L**H. LDB (input) INTEGER The leading dimension of the array B. LDB >= max(1,N). VL (input) REAL VU (input) REAL If RANGE='V', the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = 'A' or 'I'. IL (input) INTEGER IU (input) INTEGER If RANGE='I', the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = 'A' or 'V'. ABSTOL (input) REAL The absolute error tolerance for the eigenvalues. An approximate eigenvalue is accepted as converged when it is determined to lie in an interval [a,b] of width less than or equal to ABSTOL + EPS * max( |a|,|b| ) , where EPS is the machine precision. If ABSTOL is less than or equal to zero, then EPS*|T| will be used in its place, where |T| is the 1-norm of the tridiagonal matrix obtained by reducing A to tridiagonal form. See "Computing Small Singular Values of Bidiagonal Matrices with Guaranteed High Relative Accuracy," by Demmel and Kahan, LAPACK Working Note #3. If high relative accuracy is important, set ABSTOL to SLAMCH( 'Safe minimum' ). Doing so will guarantee that eigenvalues are computed to high relative accuracy when possible in future releases. The current code does not make any guarantees about high relative accuracy, but furutre releases will. See J. Barlow and J. Demmel, "Computing Accurate Eigensystems of Scaled Diagonally Dominant Matrices", LAPACK Working Note #7, for a discussion of which matrices define their eigenvalues to high relative accuracy. M (output) INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1. W (output) REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. Z (output) COMPLEX array, dimension (LDZ, max(1,M)) If JOBZ = 'V', then if INFO = 0, the first M columns of Z contain the orthonormal eigenvectors of the matrix A corresponding to the selected eigenvalues, with the i-th column of Z holding the eigenvector associated with W(i). If JOBZ = 'N', then Z is not referenced. Note: the user must ensure that at least max(1,M) columns are supplied in the array Z; if RANGE = 'V', the exact value of M is not known in advance and an upper bound must be used. LDZ (input) INTEGER The leading dimension of the array Z. LDZ >= 1, and if JOBZ = 'V', LDZ >= max(1,N). ISUPPZ (output) INTEGER ARRAY, dimension ( 2*max(1,M) ) The support of the eigenvectors in Z, i.e., the indices indicating the nonzero elements in Z. The i-th eigenvector is nonzero only in elements ISUPPZ( 2*i-1 ) through ISUPPZ( 2*i ). ********* Implemented only for RANGE = 'A' or 'I' and IU - IL = N - 1 WORK (workspace/output) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= max(1,2*N). For optimal efficiency, LWORK >= (NB+1)*N, where NB is the max of the blocksize for CHETRD and for CUNMTR as returned by ILAENV. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. RWORK (workspace/output) REAL array, dimension (LRWORK) On exit, if INFO = 0, RWORK(1) returns the optimal (and minimal) LRWORK. LRWORK (input) INTEGER The length of the array RWORK. LRWORK >= max(1,24*N). If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the RWORK array, returns this value as the first entry of the RWORK array, and no error message related to LRWORK is issued by XERBLA. IWORK (workspace/output) INTEGER array, dimension (LIWORK) On exit, if INFO = 0, IWORK(1) returns the optimal (and minimal) LIWORK. LIWORK (input) INTEGER The dimension of the array IWORK. LIWORK >= max(1,10*N). If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the IWORK array, returns this value as the first entry of the IWORK array, and no error message related to LIWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: Internal error Further Details =============== Based on contributions by Inderjit Dhillon, IBM Almaden, USA Osni Marques, LBNL/NERSC, USA Ken Stanley, Computer Science Division, University of California at Berkeley, USA ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; char range_[2] = {range, 0}; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex *da; magmaFloatComplex *db; magmaFloatComplex *dz; magma_int_t ldda = n; magma_int_t lddb = n; magma_int_t lddz = n; magma_int_t lower; char trans[1]; magma_int_t wantz; magma_int_t lquery; magma_int_t alleig, valeig, indeig; magma_int_t lwmin, lrwmin, liwmin; magma_queue_t stream; magma_queue_create( &stream ); wantz = lapackf77_lsame(jobz_, MagmaVecStr); lower = lapackf77_lsame(uplo_, MagmaLowerStr); alleig = lapackf77_lsame(range_, "A"); valeig = lapackf77_lsame(range_, "V"); indeig = lapackf77_lsame(range_, "I"); lquery = lwork == -1; *info = 0; if (itype < 1 || itype > 3) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -3; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -4; } else if (n < 0) { *info = -5; } else if (lda < max(1,n)) { *info = -7; } else if (ldb < max(1,n)) { *info = -9; } else if (ldz < 1 || (wantz && ldz < n)) { *info = -18; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -11; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -12; } else if (iu < min(n,il) || iu > n) { *info = -13; } } } magma_int_t nb = magma_get_chetrd_nb(n); lwmin = n * (nb + 1); lrwmin = 24 * n; liwmin = 10 * n; work[0] = MAGMA_C_MAKE( lwmin, 0 ); rwork[0] = lrwmin; iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -21; } else if ((lrwork < lrwmin) && ! lquery) { *info = -23; } else if ((liwork < liwmin) && ! lquery) { *info = -25; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (MAGMA_SUCCESS != magma_cmalloc( &da, n*ldda ) || MAGMA_SUCCESS != magma_cmalloc( &db, n*lddb ) || MAGMA_SUCCESS != magma_cmalloc( &dz, n*lddz )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } /* Form a Cholesky factorization of B. */ magma_csetmatrix( n, n, b, ldb, db, lddb ); magma_csetmatrix_async( n, n, a, lda, da, ldda, stream ); magma_cpotrf_gpu(uplo_[0], n, db, lddb, info); if (*info != 0) { *info = n + *info; return *info; } magma_queue_sync( stream ); magma_cgetmatrix_async( n, n, db, lddb, b, ldb, stream ); /* Transform problem to standard eigenvalue problem and solve. */ magma_chegst_gpu(itype, uplo, n, da, ldda, db, lddb, info); magma_cheevr_gpu(jobz, range, uplo, n, da, ldda, vl, vu, il, iu, abstol, m, w, dz, lddz, isuppz, a, lda, z, ldz, work, lwork, rwork, lrwork, iwork, liwork, info); if (wantz && *info == 0) { /* Backtransform eigenvectors to the original problem. */ if (itype == 1 || itype == 2) { /* For A*x=(lambda)*B*x and A*B*x=(lambda)*x; backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */ if (lower) { *(unsigned char *)trans = MagmaConjTrans; } else { *(unsigned char *)trans = MagmaNoTrans; } magma_ctrsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz); } else if (itype == 3) { /* For B*A*x=(lambda)*x; backtransform eigenvectors: x = L*y or U'*y */ if (lower) { *(unsigned char *)trans = MagmaNoTrans; } else { *(unsigned char *)trans = MagmaConjTrans; } magma_ctrmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, c_one, db, lddb, dz, lddz); } magma_cgetmatrix( n, *m, dz, lddz, z, ldz ); } magma_queue_sync( stream ); magma_queue_destroy( stream ); magma_free( da ); magma_free( db ); magma_free( dz ); return *info; } /* chegvr */
/** Purpose ------- CHEGVD computes all the eigenvalues, and optionally, the eigenvectors of a complex generalized Hermitian-definite eigenproblem, of the form A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and B are assumed to be Hermitian and B is also positive definite. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @param[in] itype INTEGER Specifies the problem type to be solved: = 1: A*x = (lambda)*B*x = 2: A*B*x = (lambda)*x = 3: B*A*x = (lambda)*x @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangles of A and B are stored; - = MagmaLower: Lower triangles of A and B are stored. @param[in] n INTEGER The order of the matrices A and B. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. \n On exit, if JOBZ = MagmaVec, then if INFO = 0, A contains the matrix Z of eigenvectors. The eigenvectors are normalized as follows: if ITYPE = 1 or 2, Z**H*B*Z = I; if ITYPE = 3, Z**H*inv(B)*Z = I. If JOBZ = MagmaNoVec, then on exit the upper triangle (if UPLO=MagmaUpper) or the lower triangle (if UPLO=MagmaLower) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[in,out] B COMPLEX array, dimension (LDB, N) On entry, the Hermitian matrix B. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of B contains the upper triangular part of the matrix B. If UPLO = MagmaLower, the leading N-by-N lower triangular part of B contains the lower triangular part of the matrix B. \n On exit, if INFO <= N, the part of B containing the matrix is overwritten by the triangular factor U or L from the Cholesky factorization B = U**H*U or B = L*L**H. @param[in] ldb INTEGER The leading dimension of the array B. LDB >= max(1,N). @param[out] w REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_chetrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) REAL array, dimension (MAX(1,LRWORK)) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: CPOTRF or CHEEVD returned an error code: <= N: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1); > N: if INFO = N + i, for 1 <= i <= N, then the leading minor of order i of B is not positive definite. The factorization of B could not be completed and no eigenvalues or eigenvectors were computed. Further Details --------------- Based on contributions by Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA Modified so that no backsubstitution is performed if CHEEVD fails to converge (NEIG in old code could be greater than N causing out of bounds reference to A - reported by Ralf Meyer). Also corrected the description of INFO and the test on ITYPE. Sven, 16 Feb 05. @ingroup magma_chegv_driver ********************************************************************/ extern "C" magma_int_t magma_chegvd_m( magma_int_t ngpu, magma_int_t itype, magma_vec_t jobz, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *B, magma_int_t ldb, float *w, magmaFloatComplex *work, magma_int_t lwork, #ifdef COMPLEX float *rwork, magma_int_t lrwork, #endif magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magmaFloatComplex c_one = MAGMA_C_ONE; magma_int_t lower; magma_trans_t trans; magma_int_t wantz; magma_int_t lquery; magma_int_t lwmin; magma_int_t liwmin; magma_int_t lrwmin; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (itype < 1 || itype > 3) { *info = -1; } else if (! (wantz || (jobz == MagmaNoVec))) { *info = -2; } else if (! (lower || (uplo == MagmaUpper))) { *info = -3; } else if (n < 0) { *info = -4; } else if (lda < max(1,n)) { *info = -6; } else if (ldb < max(1,n)) { *info = -8; } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } work[0] = magma_cmake_lwork( lwmin ); rwork[0] = magma_smake_lwork( lrwmin ); iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -11; } else if (lrwork < lrwmin && ! lquery) { *info = -13; } else if (liwork < liwmin && ! lquery) { *info = -15; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } /* If matrix is very small, then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { lapackf77_chegvd( &itype, jobz_, uplo_, &n, A, &lda, B, &ldb, w, work, &lwork, #ifdef COMPLEX rwork, &lrwork, #endif iwork, &liwork, info); return *info; } magma_timer_t time=0; timer_start( time ); magma_cpotrf_m( ngpu, uplo, n, B, ldb, info ); if (*info != 0) { *info = n + *info; return *info; } timer_stop( time ); timer_printf( "time cpotrf = %6.2f\n", time ); timer_start( time ); /* Transform problem to standard eigenvalue problem and solve. */ magma_chegst_m( ngpu, itype, uplo, n, A, lda, B, ldb, info ); timer_stop( time ); timer_printf( "time chegst = %6.2f\n", time ); timer_start( time ); magma_cheevd_m( ngpu, jobz, uplo, n, A, lda, w, work, lwork, rwork, lrwork, iwork, liwork, info ); timer_stop( time ); timer_printf( "time cheevd = %6.2f\n", time ); if (wantz && *info == 0) { timer_start( time ); /* Backtransform eigenvectors to the original problem. */ if (itype == 1 || itype == 2) { /* For A*x=(lambda)*B*x and A*B*x=(lambda)*x; backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */ if (lower) { trans = MagmaConjTrans; } else { trans = MagmaNoTrans; } magma_ctrsm_m( ngpu, MagmaLeft, uplo, trans, MagmaNonUnit, n, n, c_one, B, ldb, A, lda ); } else if (itype == 3) { /* For B*A*x=(lambda)*x; backtransform eigenvectors: x = L*y or U'*y */ if (lower) { trans = MagmaNoTrans; } else { trans = MagmaConjTrans; } #ifdef ENABLE_DEBUG printf("--- the multi GPU version is falling back to 1 GPU to perform the last TRMM since there is no TRMM_mgpu --- \n"); #endif magmaFloatComplex *dA=NULL, *dB=NULL; magma_int_t ldda = magma_roundup( n, 32 ); magma_int_t lddb = ldda; if (MAGMA_SUCCESS != magma_cmalloc( &dA, n*ldda ) || MAGMA_SUCCESS != magma_cmalloc( &dB, n*lddb )) { magma_free( dA ); magma_free( dB ); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_queue_t queue; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queue ); magma_csetmatrix( n, n, B, ldb, dB, lddb, queue ); magma_csetmatrix( n, n, A, lda, dA, ldda, queue ); magma_ctrmm( MagmaLeft, uplo, trans, MagmaNonUnit, n, n, c_one, dB, lddb, dA, ldda, queue ); magma_cgetmatrix( n, n, dA, ldda, A, lda, queue ); magma_queue_destroy( queue ); magma_free( dA ); magma_free( dB ); } timer_stop( time ); timer_printf( "time setmatrices trsm/mm + getmatrices = %6.2f\n", time ); } work[0] = magma_cmake_lwork( lwmin ); rwork[0] = magma_smake_lwork( lrwmin ); iwork[0] = liwmin; return *info; } /* magma_chegvd_m */
/** Purpose ------- CHEEVDX_GPU computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] dA COMPLEX array on the GPU, dimension (LDDA, N). On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = MagmaVec, then if INFO = 0, the first m columns of A contains the required orthonormal eigenvectors of the matrix A. If JOBZ = MagmaNoVec, then on exit the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] ldda INTEGER The leading dimension of the array DA. LDDA >= max(1,N). @param[in] vl REAL @param[in] vu REAL If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w REAL array, dimension (N) If INFO = 0, the required m eigenvalues in ascending order. @param wA (workspace) COMPLEX array, dimension (LDWA, N) @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_chetrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) REAL array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details --------------- Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. @ingroup magma_cheev_driver ********************************************************************/ extern "C" magma_int_t magma_cheevdx_gpu(magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *dA, magma_int_t ldda, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, magmaFloatComplex *wA, magma_int_t ldwa, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magma_int_t ione = 1; float d__1; float eps; magma_int_t inde; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; float smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; float *dwork; magmaFloatComplex *dC; magma_int_t lddc = ldda; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (! (wantz || (jobz == MagmaNoVec))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || (uplo == MagmaUpper))) { *info = -3; } else if (n < 0) { *info = -4; } else if (ldda < max(1,n)) { *info = -6; } else if (ldwa < max(1,n)) { *info = -14; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps (in Double!) to ensure length gets rounded up, // if it cannot be exactly represented in floating point. real_Double_t one_eps = 1. + lapackf77_slamch("Epsilon"); work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0.); rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -16; } else if ((lrwork < lrwmin) && ! lquery) { *info = -18; } else if ((liwork < liwmin) && ! lquery) { *info = -20; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif magmaFloatComplex *A; magma_cmalloc_cpu( &A, n*n ); magma_cgetmatrix(n, n, dA, ldda, A, n); lapackf77_cheevd(jobz_, uplo_, &n, A, &n, w, work, &lwork, rwork, &lrwork, iwork, &liwork, info); magma_csetmatrix( n, n, A, n, dA, ldda); magma_free_cpu(A); *m=n; return *info; } magma_queue_t stream; magma_queue_create( &stream ); // dC and dwork are never used together, so use one buffer for both; // unfortunately they're different types (complex and float). // (this works better in dsyevd_gpu where they're both float). // n*lddc for chetrd2_gpu, *2 for complex // n for clanhe magma_int_t ldwork = n*lddc*2; if ( wantz ) { // need 3n^2/2 for cstedx ldwork = max( ldwork, 3*n*(n/2 + 1) ); } if (MAGMA_SUCCESS != magma_smalloc( &dwork, ldwork )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } dC = (magmaFloatComplex*) dwork; /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_clanhe(MagmaMaxNorm, uplo, n, dA, ldda, dwork); iscale = 0; sigma = 1; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { magmablas_clascl(uplo, 0, 0, 1., sigma, n, n, dA, ldda, info); } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ // chetrd rwork: e (n) // cstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // chetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // cstedx work: tau (n) + z (n^2) // cunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; magma_timer_t time=0; timer_start( time ); #ifdef FAST_HEMV magma_chetrd2_gpu(uplo, n, dA, ldda, w, &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, dC, n*lddc, &iinfo); #else magma_chetrd_gpu (uplo, n, dA, ldda, w, &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, &iinfo); #endif timer_stop( time ); timer_printf( "time chetrd_gpu = %6.2f\n", time ); /* For eigenvalues only, call SSTERF. For eigenvectors, first call CSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call CUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &rwork[inde], info); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); } else { timer_start( time ); magma_cstedx(range, n, vl, vu, il, iu, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info); timer_stop( time ); timer_printf( "time cstedx = %6.2f\n", time ); timer_start( time ); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); magma_csetmatrix( n, *m, &work[indwrk + n * (il-1) ], n, dC, lddc ); magma_cunmtr_gpu(MagmaLeft, uplo, MagmaNoTrans, n, *m, dA, ldda, &work[indtau], dC, lddc, wA, ldwa, &iinfo); magma_ccopymatrix( n, *m, dC, lddc, dA, ldda ); timer_stop( time ); timer_printf( "time cunmtr_gpu + copy = %6.2f\n", time ); } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, w, &ione); } work[0] = MAGMA_C_MAKE( lwmin * one_eps, 0.); // round up rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; magma_queue_destroy( stream ); magma_free( dwork ); return *info; } /* magma_cheevdx_gpu */
/***************************************************************************//** Purpose ------- CHEEVD computes all eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = MagmaVec, then if INFO = 0, A contains the orthonormal eigenvectors of the matrix A. If JOBZ = MagmaNoVec, then on exit the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[in] vl REAL @param[in] vu REAL If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w REAL array, dimension (N) If INFO = 0, the eigenvalues in ascending order. @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. - If N <= 1, LWORK >= 1. - If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. - If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_chetrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) REAL array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. - If N <= 1, LRWORK >= 1. - If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. - If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. - If N <= 1, LIWORK >= 1. - If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. - If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details --------------- Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. @ingroup magma_heevdx *******************************************************************************/ extern "C" magma_int_t magma_cheevdx_m( magma_int_t ngpu, magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, magmaFloatComplex *work, magma_int_t lwork, #ifdef COMPLEX float *rwork, magma_int_t lrwork, #endif magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magma_int_t ione = 1; magma_int_t izero = 0; float d_one = 1.; float d__1; float eps; magma_int_t inde; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; float smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (! (wantz || (jobz == MagmaNoVec))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || (uplo == MagmaUpper))) { *info = -3; } else if (n < 0) { *info = -4; } else if (lda < max(1,n)) { *info = -6; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } work[0] = magma_cmake_lwork( lwmin ); rwork[0] = magma_smake_lwork( lrwmin ); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -14; } else if ((lrwork < lrwmin) && ! lquery) { *info = -16; } else if ((liwork < liwmin) && ! lquery) { *info = -18; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_C_REAL(A[0]); if (wantz) { A[0] = MAGMA_C_ONE; } return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%lld NB=%lld, calling lapack on CPU\n", (long long) n, (long long) nb ); printf("--------------------------------------------------------------\n"); #endif lapackf77_cheevd(jobz_, uplo_, &n, A, &lda, w, work, &lwork, #ifdef COMPLEX rwork, &lrwork, #endif iwork, &liwork, info); return *info; } /* Get machine constants. */ safmin = lapackf77_slamch("Safe minimum"); eps = lapackf77_slamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_ssqrt(smlnum); rmax = magma_ssqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_clanhe("M", uplo_, &n, A, &lda, rwork); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_clascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, A, &lda, info); } /* Call CHETRD to reduce Hermitian matrix to tridiagonal form. */ inde = 0; indtau = 0; indwrk = indtau + n; indrwk = inde + n; indwk2 = indwrk + n * n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; llrwk = lrwork - indrwk; magma_timer_t time=0; timer_start( time ); magma_chetrd_mgpu(ngpu, 1, uplo, n, A, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo); timer_stop( time ); timer_printf( "time chetrd = %6.2f\n", time ); /* For eigenvalues only, call SSTERF. For eigenvectors, first call CSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call CUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &rwork[inde], info); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); } else { timer_start( time ); magma_cstedx_m(ngpu, range, n, vl, vu, il, iu, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, info); timer_stop( time ); timer_printf( "time cstedc = %6.2f\n", time ); timer_start( time ); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); magma_cunmtr_m(ngpu, MagmaLeft, uplo, MagmaNoTrans, n, *m, A, lda, &work[indtau], &work[indwrk + n * (il-1)], n, &work[indwk2], llwrk2, &iinfo); lapackf77_clacpy("A", &n, m, &work[indwrk + n * (il-1)], &n, A, &lda); timer_stop( time ); timer_printf( "time cunmtr + copy = %6.2f\n", time ); } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_sscal(&imax, &d__1, w, &ione); } work[0] = magma_cmake_lwork( lwmin ); rwork[0] = magma_smake_lwork( lrwmin ); iwork[0] = liwmin; return *info; } /* magma_cheevd_m */
/** Purpose ------- CHETRD reduces a complex Hermitian matrix A to real symmetric tridiagonal form T by an orthogonal similarity transformation: Q\*\*H * A * Q = T. Arguments --------- @param[in] num_gpus INTEGER The number of GPUs. num_gpus > 0. @param[in] num_streams INTEGER The number of GPU streams used for update. 10 >= num_streams > 0. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA,N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal of A are overwritten by the corresponding elements of the tridiagonal matrix T, and the elements above the first superdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = MagmaLower, the diagonal and first subdiagonal of A are over- written by the corresponding elements of the tridiagonal matrix T, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] d COMPLEX array, dimension (N) The diagonal elements of the tridiagonal matrix T: D(i) = A(i,i). @param[out] e COMPLEX array, dimension (N-1) The off-diagonal elements of the tridiagonal matrix T: E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower. @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= 1. For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). @ingroup magma_cheev_comp ********************************************************************/ extern "C" magma_int_t magma_chetrd_mgpu( magma_int_t num_gpus, magma_int_t num_streams, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, float *d, float *e, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define dA(id, i, j) (dA[(id)] + (j)*ldda + (i)) #define dW(id, i, j) (dwork[(id)] + (j)*ldda + (i)) const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t ln, ldda; magma_int_t nb = magma_get_chetrd_nb(n), ib; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex c_one = MAGMA_C_ONE; float d_one = MAGMA_D_ONE; float mv_time = 0.0; #ifdef PROFILE_SY2RK float up_time = 0.0; #endif magma_int_t kk, nx; magma_int_t i = 0, ii, iii, j, did, i_n; magma_int_t iinfo; magma_int_t ldwork, lddwork, lwkopt, ldwork2; magma_int_t lquery; magma_queue_t stream[MagmaMaxGPUs][10]; magmaFloatComplex *dx[MagmaMaxGPUs], *dy[MagmaMaxGPUs], *hwork; magmaFloatComplex *dwork2[MagmaMaxGPUs]; *info = 0; int upper = (uplo == MagmaUpper); lquery = (lwork == -1); if (! upper && uplo != MagmaLower) { printf( " uplo = %c\n",uplo ); *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,n)) { *info = -4; } else if (lwork < nb*n && ! lquery) { *info = -9; } else if ( num_streams > 2 ) { *info = 2; } /* Determine the block size. */ ldwork = lddwork = n; lwkopt = n * nb; if (*info == 0) { work[0] = MAGMA_C_MAKE( lwkopt, 0 ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return MAGMA_ERR_ILLEGAL_VALUE; } else if (lquery) return 0; /* Quick return if possible */ if (n == 0) { work[0] = c_one; return 0; } magmaFloatComplex *dA[MagmaMaxGPUs]; magmaFloatComplex *dwork[MagmaMaxGPUs]; float times[11]; for( did=0; did < 11; did++ ) times[did] = 0; //#define PROFILE_SY2RK #ifdef PROFILE_SY2RK magma_event_t start, stop; float etime; magma_setdevice(0); magma_event_create( &start ); magma_event_create( &stop ); #endif ldda = lda; ln = ((nb*(1+n/(nb*num_gpus))+31)/32)*32; ldwork2 = (1+ n / nb + (n % nb != 0)) * ldda; for( did=0; did < num_gpus; did++ ) { magma_setdevice(did); // TODO fix memory leak if ( MAGMA_SUCCESS != magma_cmalloc(&dA[did], ln*ldda+3*lddwork*nb) || MAGMA_SUCCESS != magma_cmalloc(&dx[did], num_streams*n) || MAGMA_SUCCESS != magma_cmalloc(&dy[did], num_streams*n) || MAGMA_SUCCESS != magma_cmalloc(&dwork2[did], ldwork2 ) ) { for( i=0; i < did; i++ ) { magma_setdevice(i); magma_free(dA[i]); magma_free(dx[i]); magma_free(dy[i]); } *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } dwork[did] = dA[did] + ln*ldda; for( kk=0; kk < num_streams; kk++ ) magma_queue_create(&stream[did][kk]); } magma_setdevice(0); // TODO fix memory leak dwork2 if ( MAGMA_SUCCESS != magma_cmalloc_pinned( &hwork, num_streams*num_gpus*n ) ) { for( i=0; i < num_gpus; i++ ) { magma_setdevice(i); magma_free(dA[i]); magma_free(dx[i]); magma_free(dy[i]); } *info = MAGMA_ERR_HOST_ALLOC; return *info; } if (n < 2048) nx = n; else nx = 512; if (upper) { /* Copy the matrix to the GPU */ if (1 <= n-nx) { magma_chtodhe(num_gpus, uplo, n, nb, A, lda, dA, ldda, stream, &iinfo ); } /* Reduce the upper triangle of A. Columns 1:kk are handled by the unblocked method. */ for (i = nb*((n-1)/nb); i >= nx; i -= nb) { ib = min(nb, n-i); ii = nb*(i/(nb*num_gpus)); did = (i/nb)%num_gpus; /* wait for the next panel */ if (i != nb*((n-1)/nb)) { magma_setdevice(did); magma_queue_sync(stream[did][0]); } magma_clatrd_mgpu(num_gpus, uplo, n, i+ib, ib, nb, A(0, 0), lda, e, tau, work, ldwork, dA, ldda, 0, dwork, i+ib, dwork2, ldwork2, 1, dx, dy, hwork, stream, times); magma_cher2k_mgpu(num_gpus, MagmaUpper, MagmaNoTrans, nb, i, ib, c_neg_one, dwork, i+ib, 0, d_one, dA, ldda, 0, num_streams, stream); /* get the next panel */ if (i-nb >= nx ) { ib = min(nb, n-(i-nb)); ii = nb*((i-nb)/(nb*num_gpus)); did = ((i-nb)/nb)%num_gpus; magma_setdevice(did); magma_cgetmatrix_async( (i-nb)+ib, ib, dA(did, 0, ii), ldda, A(0, i-nb), lda, stream[did][0] ); } /* Copy superdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+ib; ++j) { if ( j > 0 ) { *A(j-1,j) = MAGMA_C_MAKE( e[j - 1], 0 ); } d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* end of for i=... */ if ( nx > 0 ) { if (1 <= n-nx) { /* else A is already on CPU */ for (i=0; i < nx; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*num_gpus)); did = (i/nb)%num_gpus; magma_setdevice(did); magma_cgetmatrix_async( nx, ib, dA(did, 0, ii), ldda, A(0, i), lda, stream[did][0] ); } } for( did=0; did < num_gpus; did++ ) { magma_setdevice(did); magma_queue_sync(stream[did][0]); } /* Use unblocked code to reduce the last or only block */ lapackf77_chetd2(uplo_, &nx, A(0, 0), &lda, d, e, tau, &iinfo); } } else { trace_init( 1, num_gpus, num_streams, (CUstream_st**)stream ); /* Copy the matrix to the GPU */ if (1 <= n-nx) { magma_chtodhe(num_gpus, uplo, n, nb, A, lda, dA, ldda, stream, &iinfo ); } /* Reduce the lower triangle of A */ for (i = 0; i < n-nx; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*num_gpus)); did = (i/nb)%num_gpus; /* Reduce columns i:i+ib-1 to tridiagonal form and form the matrix W which is needed to update the unreduced part of the matrix */ /* Get the current panel (no need for the 1st iteration) */ if (i != 0) { magma_setdevice(did); trace_gpu_start( did, 0, "comm", "get" ); magma_cgetmatrix_async( n-i, ib, dA(did, i, ii), ldda, A(i,i), lda, stream[did][0] ); trace_gpu_end( did, 0 ); magma_queue_sync(stream[did][0]); magma_setdevice(0); } mv_time += magma_clatrd_mgpu(num_gpus, uplo, n, n-i, ib, nb, A(i, i), lda, &e[i], &tau[i], work, ldwork, dA, ldda, i, dwork, (n-i), dwork2, ldwork2, 1, dx, dy, hwork, stream, times ); #ifdef PROFILE_SY2RK magma_setdevice(0); if ( i > 0 ) { cudaEventElapsedTime(&etime, start, stop); up_time += (etime/1000.0); } magma_event_record(start, 0); #endif magma_cher2k_mgpu(num_gpus, MagmaLower, MagmaNoTrans, nb, n-i-ib, ib, c_neg_one, dwork, n-i, ib, d_one, dA, ldda, i+ib, num_streams, stream); #ifdef PROFILE_SY2RK magma_setdevice(0); magma_event_record(stop, 0); #endif /* Copy subdiagonal elements back into A, and diagonal elements into D */ for (j = i; j < i+ib; ++j) { if ( j+1 < n ) { *A(j+1,j) = MAGMA_C_MAKE( e[j], 0 ); } d[j] = MAGMA_C_REAL( *A(j, j) ); } } /* for i=... */ /* Use unblocked code to reduce the last or only block */ if ( i < n ) { iii = i; i_n = n-i; if ( i > 0 ) { for (; i < n; i += nb) { ib = min(nb, n-i); ii = nb*(i/(nb*num_gpus)); did = (i/nb)%num_gpus; magma_setdevice(did); magma_cgetmatrix_async( i_n, ib, dA(did, iii, ii), ldda, A(iii, i), lda, stream[did][0] ); } for( did=0; did < num_gpus; did++ ) { magma_setdevice(did); magma_queue_sync(stream[did][0]); } } lapackf77_chetrd(uplo_, &i_n, A(iii, iii), &lda, &d[iii], &e[iii], &tau[iii], work, &lwork, &iinfo); } } #ifdef PROFILE_SY2RK magma_setdevice(0); if ( n > nx ) { cudaEventElapsedTime(&etime, start, stop); up_time += (etime/1000.0); } magma_event_destroy( start ); magma_event_destroy( stop ); #endif trace_finalize( "chetrd.svg", "trace.css" ); for( did=0; did < num_gpus; did++ ) { magma_setdevice(did); for( kk=0; kk < num_streams; kk++ ) magma_queue_sync(stream[did][kk]); for( kk=0; kk < num_streams; kk++ ) magma_queue_destroy(stream[did][kk]); magma_free(dA[did]); magma_free(dx[did]); magma_free(dy[did]); magma_free(dwork2[did]); } magma_setdevice(0); magma_free_pinned(hwork); work[0] = MAGMA_C_MAKE( lwkopt, 0 ); #ifdef PROFILE_SY2RK printf( " n=%d nb=%d\n",n,nb ); printf( " Time in CLARFG: %.2e seconds\n",times[0] ); printf( " Time in CHEMV : %.2e seconds\n",mv_time ); printf( " Time in CHER2K: %.2e seconds\n",up_time ); #endif return MAGMA_SUCCESS; } /* magma_chetrd */
/** Purpose ------- CHEGVX computes selected eigenvalues, and optionally, eigenvectors of a complex generalized Hermitian-definite eigenproblem, of the form A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and B are assumed to be Hermitian and B is also positive definite. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. Arguments --------- @param[in] itype INTEGER Specifies the problem type to be solved: = 1: A*x = (lambda)*B*x = 2: A*B*x = (lambda)*x = 3: B*A*x = (lambda)*x @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangles of A and B are stored; - = MagmaLower: Lower triangles of A and B are stored. @param[in] n INTEGER The order of the matrices A and B. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. \n On exit, the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[in,out] B COMPLEX array, dimension (LDB, N) On entry, the Hermitian matrix B. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of B contains the upper triangular part of the matrix B. If UPLO = MagmaLower, the leading N-by-N lower triangular part of B contains the lower triangular part of the matrix B. \n On exit, if INFO <= N, the part of B containing the matrix is overwritten by the triangular factor U or L from the Cholesky factorization B = U**H*U or B = L*L**H. @param[in] ldb INTEGER The leading dimension of the array B. LDB >= max(1,N). @param[in] vl REAL @param[in] vu REAL If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[in] abstol REAL The absolute error tolerance for the eigenvalues. An approximate eigenvalue is accepted as converged when it is determined to lie in an interval [a,b] of width less than or equal to \n ABSTOL + EPS * max( |a|,|b| ), \n where EPS is the machine precision. If ABSTOL is less than or equal to zero, then EPS*|T| will be used in its place, where |T| is the 1-norm of the tridiagonal matrix obtained by reducing A to tridiagonal form. \n Eigenvalues will be computed most accurately when ABSTOL is set to twice the underflow threshold 2*SLAMCH('S'), not zero. If this routine returns with INFO > 0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*SLAMCH('S'). @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w REAL array, dimension (N) The first M elements contain the selected eigenvalues in ascending order. @param[out] Z COMPLEX array, dimension (LDZ, max(1,M)) If JOBZ = MagmaNoVec, then Z is not referenced. If JOBZ = MagmaVec, then if INFO = 0, the first M columns of Z contain the orthonormal eigenvectors of the matrix A corresponding to the selected eigenvalues, with the i-th column of Z holding the eigenvector associated with W(i). The eigenvectors are normalized as follows: if ITYPE = 1 or 2, Z**T*B*Z = I; if ITYPE = 3, Z**T*inv(B)*Z = I. \n If an eigenvector fails to converge, then that column of Z contains the latest approximation to the eigenvector, and the index of the eigenvector is returned in IFAIL. Note: the user must ensure that at least max(1,M) columns are supplied in the array Z; if RANGE = MagmaRangeV, the exact value of M is not known in advance and an upper bound must be used. @param[in] ldz INTEGER The leading dimension of the array Z. LDZ >= 1, and if JOBZ = MagmaVec, LDZ >= max(1,N). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. LWORK >= max(1,2*N). For optimal efficiency, LWORK >= (NB+1)*N, where NB is the blocksize for CHETRD returned by ILAENV. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param rwork (workspace) REAL array, dimension (7*N) @param iwork (workspace) INTEGER array, dimension (5*N) @param[out] ifail INTEGER array, dimension (N) If JOBZ = MagmaVec, then if INFO = 0, the first M elements of IFAIL are zero. If INFO > 0, then IFAIL contains the indices of the eigenvectors that failed to converge. If JOBZ = MagmaNoVec, then IFAIL is not referenced. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: CPOTRF or CHEEVX returned an error code: <= N: if INFO = i, CHEEVX failed to converge; i eigenvectors failed to converge. Their indices are stored in array IFAIL. > N: if INFO = N + i, for 1 <= i <= N, then the leading minor of order i of B is not positive definite. The factorization of B could not be completed and no eigenvalues or eigenvectors were computed. Further Details --------------- Based on contributions by Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA @ingroup magma_chegv_driver ********************************************************************/ extern "C" magma_int_t magma_chegvx( magma_int_t itype, magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *B, magma_int_t ldb, float vl, float vu, magma_int_t il, magma_int_t iu, float abstol, magma_int_t *m, float *w, magmaFloatComplex *Z, magma_int_t ldz, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info) { magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex *dA; magmaFloatComplex *dB; magmaFloatComplex *dZ; magma_int_t ldda = n; magma_int_t lddb = n; magma_int_t lddz = n; magma_int_t lower; magma_trans_t trans; magma_int_t wantz; magma_int_t lquery; magma_int_t alleig, valeig, indeig; magma_int_t lwmin; magma_queue_t stream; magma_queue_create( &stream ); wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1); *info = 0; if (itype < 1 || itype > 3) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (wantz || (jobz == MagmaNoVec))) { *info = -3; } else if (! (lower || (uplo == MagmaUpper))) { *info = -4; } else if (n < 0) { *info = -5; } else if (lda < max(1,n)) { *info = -7; } else if (ldb < max(1,n)) { *info = -9; } else if (ldz < 1 || (wantz && ldz < n)) { *info = -18; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -11; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -12; } else if (iu < min(n,il) || iu > n) { *info = -13; } } } magma_int_t nb = magma_get_chetrd_nb(n); lwmin = n * (nb + 1); work[0] = MAGMA_C_MAKE( lwmin, 0 ); if (lwork < lwmin && ! lquery) { *info = -20; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (MAGMA_SUCCESS != magma_cmalloc( &dA, n*ldda ) || MAGMA_SUCCESS != magma_cmalloc( &dB, n*lddb ) || MAGMA_SUCCESS != magma_cmalloc( &dZ, n*lddz )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } /* Form a Cholesky factorization of B. */ magma_csetmatrix( n, n, B, ldb, dB, lddb ); magma_csetmatrix_async( n, n, A, lda, dA, ldda, stream ); magma_cpotrf_gpu(uplo, n, dB, lddb, info); if (*info != 0) { *info = n + *info; return *info; } magma_queue_sync( stream ); magma_cgetmatrix_async( n, n, dB, lddb, B, ldb, stream ); /* Transform problem to standard eigenvalue problem and solve. */ magma_chegst_gpu(itype, uplo, n, dA, ldda, dB, lddb, info); magma_cheevx_gpu(jobz, range, uplo, n, dA, ldda, vl, vu, il, iu, abstol, m, w, dZ, lddz, A, lda, Z, ldz, work, lwork, rwork, iwork, ifail, info); if (wantz && *info == 0) { /* Backtransform eigenvectors to the original problem. */ if (itype == 1 || itype == 2) { /* For A*x=(lambda)*B*x and A*B*x=(lambda)*x; backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */ if (lower) { trans = MagmaConjTrans; } else { trans = MagmaNoTrans; } magma_ctrsm(MagmaLeft, uplo, trans, MagmaNonUnit, n, *m, c_one, dB, lddb, dZ, lddz); } else if (itype == 3) { /* For B*A*x=(lambda)*x; backtransform eigenvectors: x = L*y or U'*y */ if (lower) { trans = MagmaNoTrans; } else { trans = MagmaConjTrans; } magma_ctrmm(MagmaLeft, uplo, trans, MagmaNonUnit, n, *m, c_one, dB, lddb, dZ, lddz); } magma_cgetmatrix( n, *m, dZ, lddz, Z, ldz ); } magma_queue_sync( stream ); magma_queue_destroy( stream ); magma_free( dA ); magma_free( dB ); magma_free( dZ ); return *info; } /* magma_chegvx */
extern "C" magma_int_t magma_chegvdx_m(magma_int_t nrgpu, magma_int_t itype, char jobz, char range, char uplo, magma_int_t n, magmaFloatComplex *a, magma_int_t lda, magmaFloatComplex *b, magma_int_t ldb, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, magmaFloatComplex *work, magma_int_t lwork, float *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= CHEGVD computes all the eigenvalues, and optionally, the eigenvectors of a complex generalized Hermitian-definite eigenproblem, of the form A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and B are assumed to be Hermitian and B is also positive definite. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments ========= ITYPE (input) INTEGER Specifies the problem type to be solved: = 1: A*x = (lambda)*B*x = 2: A*B*x = (lambda)*x = 3: B*A*x = (lambda)*x RANGE (input) CHARACTER*1 = 'A': all eigenvalues will be found. = 'V': all eigenvalues in the half-open interval (VL,VU] will be found. = 'I': the IL-th through IU-th eigenvalues will be found. JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. UPLO (input) CHARACTER*1 = 'U': Upper triangles of A and B are stored; = 'L': Lower triangles of A and B are stored. N (input) INTEGER The order of the matrices A and B. N >= 0. A (input/output) COMPLEX*16 array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = 'V', then if INFO = 0, A contains the matrix Z of eigenvectors. The eigenvectors are normalized as follows: if ITYPE = 1 or 2, Z**H*B*Z = I; if ITYPE = 3, Z**H*inv(B)*Z = I. If JOBZ = 'N', then on exit the upper triangle (if UPLO='U') or the lower triangle (if UPLO='L') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). B (input/output) COMPLEX*16 array, dimension (LDB, N) On entry, the Hermitian matrix B. If UPLO = 'U', the leading N-by-N upper triangular part of B contains the upper triangular part of the matrix B. If UPLO = 'L', the leading N-by-N lower triangular part of B contains the lower triangular part of the matrix B. On exit, if INFO <= N, the part of B containing the matrix is overwritten by the triangular factor U or L from the Cholesky factorization B = U**H*U or B = L*L**H. LDB (input) INTEGER The leading dimension of the array B. LDB >= max(1,N). VL (input) DOUBLE PRECISION VU (input) DOUBLE PRECISION If RANGE='V', the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = 'A' or 'I'. IL (input) INTEGER IU (input) INTEGER If RANGE='I', the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = 'A' or 'V'. M (output) INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1. W (output) DOUBLE PRECISION array, dimension (N) If INFO = 0, the eigenvalues in ascending order. WORK (workspace/output) COMPLEX*16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = 'N' and N > 1, LWORK >= N + 1. If JOBZ = 'V' and N > 1, LWORK >= 2*N*nb + N**2. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. RWORK (workspace/output) DOUBLE PRECISION array, dimension (MAX(1,LRWORK)) On exit, if INFO = 0, RWORK(1) returns the optimal LRWORK. LRWORK (input) INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = 'N' and N > 1, LRWORK >= N. If JOBZ = 'V' and N > 1, LRWORK >= 1 + 5*N + 2*N**2. If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK. LIWORK (input) INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = 'N' and N > 1, LIWORK >= 1. If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N. If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: CPOTRF or CHEEVD returned an error code: <= N: if INFO = i and JOBZ = 'N', then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = 'V', then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1); > N: if INFO = N + i, for 1 <= i <= N, then the leading minor of order i of B is not positive definite. The factorization of B could not be completed and no eigenvalues or eigenvectors were computed. Further Details =============== Based on contributions by Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA Modified so that no backsubstitution is performed if CHEEVD fails to converge (NEIG in old code could be greater than N causing out of bounds reference to A - reported by Ralf Meyer). Also corrected the description of INFO and the test on ITYPE. Sven, 16 Feb 05. ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; char range_[2] = {range, 0}; magmaFloatComplex c_one = MAGMA_C_ONE; magma_int_t lower; char trans[1]; magma_int_t wantz; magma_int_t lquery; magma_int_t alleig, valeig, indeig; magma_int_t lwmin; magma_int_t liwmin; magma_int_t lrwmin; wantz = lapackf77_lsame(jobz_, MagmaVecStr); lower = lapackf77_lsame(uplo_, MagmaLowerStr); alleig = lapackf77_lsame(range_, "A"); valeig = lapackf77_lsame(range_, "V"); indeig = lapackf77_lsame(range_, "I"); lquery = lwork == -1 || lrwork == -1 || liwork == -1; *info = 0; if (itype < 1 || itype > 3) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -3; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -4; } else if (n < 0) { *info = -5; } else if (lda < max(1,n)) { *info = -7; } else if (ldb < max(1,n)) { *info = -9; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -11; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -12; } else if (iu < min(n,il) || iu > n) { *info = -13; } } } magma_int_t nb = magma_get_chetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = 2*n + n*n; lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); // round up rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -17; } else if (lrwork < lrwmin && ! lquery) { *info = -19; } else if (liwork < liwmin && ! lquery) { *info = -21; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128){ #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_chegvd(&itype, jobz_, uplo_, &n, a, &lda, b, &ldb, w, work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); *m = n; return *info; } // #ifdef ENABLE_TIMER magma_timestr_t start, end; start = get_current_time(); #endif magma_cpotrf_m(nrgpu, uplo_[0], n, b, ldb, info); if (*info != 0) { *info = n + *info; return *info; } #ifdef ENABLE_TIMER end = get_current_time(); printf("time cpotrf = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif /* Transform problem to standard eigenvalue problem and solve. */ magma_chegst_m(nrgpu, itype, uplo_[0], n, a, lda, b, ldb, info); #ifdef ENABLE_TIMER end = get_current_time(); printf("time chegst = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif magma_cheevdx_m(nrgpu, jobz, range, uplo, n, a, lda, vl, vu, il, iu, m, w, work, lwork, rwork, lrwork, iwork, liwork, info); #ifdef ENABLE_TIMER end = get_current_time(); printf("time cheevd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif if (wantz && *info == 0) { #ifdef ENABLE_TIMER start = get_current_time(); #endif /* Backtransform eigenvectors to the original problem. */ if (itype == 1 || itype == 2) { /* For A*x=(lambda)*B*x and A*B*x=(lambda)*x; backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */ if (lower) { *(unsigned char *)trans = MagmaConjTrans; } else { *(unsigned char *)trans = MagmaNoTrans; } magma_ctrsm_m(nrgpu, MagmaLeft, uplo_[0], *trans, MagmaNonUnit, n, *m, c_one, b, ldb, a, lda); } else if (itype == 3) { /* For B*A*x=(lambda)*x; backtransform eigenvectors: x = L*y or U'*y */ if (lower) { *(unsigned char *)trans = MagmaNoTrans; } else { *(unsigned char *)trans = MagmaConjTrans; } //magma_ctrmm(MagmaLeft, uplo_[0], *trans, MagmaNonUnit, // n, n, c_one, db, lddb, da, ldda); } #ifdef ENABLE_TIMER end = get_current_time(); printf("time setmatrices trsm/mm + getmatrices = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } work[0] = MAGMA_C_MAKE( lwmin * (1. + lapackf77_slamch("Epsilon")), 0.); // round up rwork[0] = lrwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; return *info; } /* magma_chegvd_m */
/* //////////////////////////////////////////////////////////////////////////// -- Testing chegvd */ int main( int argc, char** argv) { TESTING_CUDA_INIT(); cuFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work; float *rwork, *w1, *w2; magma_int_t *iwork; float gpu_time, cpu_time; magma_timestr_t start, end; /* Matrix size */ magma_int_t N=0, n2; magma_int_t size[4] = {1024,2048,4100,6001}; magma_int_t i, itype, info; magma_int_t ione = 1, izero = 0; magma_int_t five = 5; cuFloatComplex c_zero = MAGMA_C_ZERO; cuFloatComplex c_one = MAGMA_C_ONE; cuFloatComplex c_neg_one = MAGMA_C_NEG_ONE; float d_one = 1.; float d_neg_one = -1.; float d_ten = 10.; magma_int_t ISEED[4] = {0,0,0,1}; const char *uplo = MagmaLowerStr; const char *jobz = MagmaVectorsStr; itype = 1; magma_int_t checkres; float result[4]; int flagN = 0; if (argc != 1){ for(i = 1; i<argc; i++){ if (strcmp("-N", argv[i])==0){ N = atoi(argv[++i]); if (N>0){ printf(" testing_chegvd -N %d\n\n", (int) N); flagN=1; } else { printf("\nUsage: \n"); printf(" testing_chegvd -N %d\n\n", (int) N); exit(1); } } if (strcmp("-itype", argv[i])==0){ itype = atoi(argv[++i]); if (itype>0 && itype <= 3){ printf(" testing_chegvd -itype %d\n\n", (int) itype); } else { printf("\nUsage: \n"); printf(" testing_chegvd -itype %d\n\n", (int) itype); exit(1); } } if (strcmp("-L", argv[i])==0){ uplo = MagmaLowerStr; printf(" testing_chegvd -L"); } if (strcmp("-U", argv[i])==0){ uplo = MagmaUpperStr; printf(" testing_chegvd -U"); } } } else { printf("\nUsage: \n"); printf(" testing_chegvd -L/U -N %d -itype %d\n\n", 1024, 1); } if(!flagN) N = size[3]; checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL; n2 = N * N; /* Allocate host memory for the matrix */ TESTING_MALLOC( h_A, cuFloatComplex, n2); TESTING_MALLOC( h_B, cuFloatComplex, n2); TESTING_MALLOC( w1, float , N); TESTING_MALLOC( w2, float , N); TESTING_HOSTALLOC(h_R, cuFloatComplex, n2); TESTING_HOSTALLOC(h_S, cuFloatComplex, n2); magma_int_t nb = magma_get_chetrd_nb(N); magma_int_t lwork = 2*N*nb + N*N; magma_int_t lrwork = 1 + 5*N +2*N*N; magma_int_t liwork = 3 + 5*N; TESTING_HOSTALLOC(h_work, cuFloatComplex, lwork); TESTING_MALLOC( rwork, float, lrwork); TESTING_MALLOC( iwork, magma_int_t, liwork); printf(" N CPU Time(s) GPU Time(s) \n"); printf("===================================\n"); for(i=0; i<4; i++){ if (!flagN){ N = size[i]; n2 = N*N; } /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); //lapackf77_clatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten, // &d_one, &N, &N, uplo, h_B, &N, h_work, &info); //lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_B, &N); lapackf77_clarnv( &ione, ISEED, &n2, h_B ); /* increase the diagonal */ { magma_int_t i, j; for(i=0; i<N; i++) { MAGMA_C_SET2REAL( h_B[i*N+i], MAGMA_C_REAL(h_B[i*N+i]) + 1.*N ); MAGMA_C_SET2REAL( h_A[i*N+i], MAGMA_C_REAL(h_A[i*N+i]) ); } } lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); magma_chegvd(itype, jobz[0], uplo[0], N, h_R, N, h_S, N, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ start = get_current_time(); magma_chegvd(itype, jobz[0], uplo[0], N, h_R, N, h_S, N, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); end = get_current_time(); gpu_time = GetTimerValue(start,end)/1000.; if ( checkres ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvd routine. A x = lambda B x is solved and the following 3 tests computed: (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) | A B Z - Z D | / ( |A||Z| N ) (itype = 2) | B A Z - Z D | / ( |A||Z| N ) (itype = 3) (2) | I - V V' B | / ( N ) (itype = 1,2) | B - V V' | / ( |B| N ) (itype = 3) (3) | S(with V) - S(w/o V) | / | S | =================================================================== */ float temp1, temp2; cuFloatComplex *tau; if (itype == 1 || itype == 2){ lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_S, &N); blasf77_cgemm("N", "C", &N, &N, &N, &c_one, h_R, &N, h_R, &N, &c_zero, h_work, &N); blasf77_chemm("R", uplo, &N, &N, &c_neg_one, h_B, &N, h_work, &N, &c_one, h_S, &N); result[1]= lapackf77_clange("1", &N, &N, h_S, &N, rwork) / N; } else if (itype == 3){ lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N); blasf77_cherk(uplo, "N", &N, &N, &d_neg_one, h_R, &N, &d_one, h_S, &N); result[1]= lapackf77_clanhe("1",uplo, &N, h_S, &N, rwork) / N / lapackf77_clanhe("1",uplo, &N, h_B, &N, rwork); } result[0] = 1.; result[0] /= lapackf77_clanhe("1",uplo, &N, h_A, &N, rwork); result[0] /= lapackf77_clange("1",&N , &N, h_R, &N, rwork); if (itype == 1){ blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", uplo, &N, &N, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N); result[0] *= lapackf77_clange("1", &N, &N, h_work, &N, rwork)/N; } else if (itype == 2){ blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N; } else if (itype == 3){ blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N; } /* lapackf77_chet21(&ione, uplo, &N, &izero, h_A, &N, w1, w1, h_R, &N, h_R, &N, tau, h_work, rwork, &result[0]); */ lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); magma_chegvd(itype, 'N', uplo[0], N, h_R, N, h_S, N, w2, h_work, lwork, rwork, lrwork, iwork, liwork, &info); temp1 = temp2 = 0; for(int j=0; j<N; j++){ temp1 = max(temp1, absv(w1[j])); temp1 = max(temp1, absv(w2[j])); temp2 = max(temp2, absv(w1[j]-w2[j])); } result[2] = temp2 / temp1; } /* ===================================================================== Performs operation using LAPACK =================================================================== */ start = get_current_time(); lapackf77_chegvd(&itype, jobz, uplo, &N, h_A, &N, h_B, &N, w2, h_work, &lwork, rwork, &lrwork, iwork, &liwork, &info); end = get_current_time(); if (info < 0) printf("Argument %d of chegvd had an illegal value.\n", (int) -info); cpu_time = GetTimerValue(start,end)/1000.; /* ===================================================================== Print execution time =================================================================== */ printf("%5d %6.2f %6.2f\n", (int) N, cpu_time, gpu_time); if ( checkres ){ printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if(itype==1) printf("(1) | A Z - B Z D | / (|A| |Z| N) = %e\n", result[0]); else if(itype==2) printf("(1) | A B Z - Z D | / (|A| |Z| N) = %e\n", result[0]); else if(itype==3) printf("(1) | B A Z - Z D | / (|A| |Z| N) = %e\n", result[0]); if(itype==1 || itype ==2) printf("(2) | I - Z Z' B | / N = %e\n", result[1]); else printf("(2) | B - Z Z' | / (|B| N) = %e\n", result[1]); printf("(3) | D(w/ Z)-D(w/o Z)|/ |D| = %e\n\n", result[2]); } if (flagN) break; } /* Memory clean up */ TESTING_FREE( h_A); TESTING_FREE( h_B); TESTING_FREE( w1); TESTING_FREE( w2); TESTING_FREE( rwork); TESTING_FREE( iwork); TESTING_HOSTFREE(h_work); TESTING_HOSTFREE( h_R); TESTING_HOSTFREE( h_S); /* Shutdown */ TESTING_CUDA_FINALIZE(); }