extern "C" magma_int_t magma_zheevx_gpu(char jobz, char range, char uplo, magma_int_t n, magmaDoubleComplex *da, magma_int_t ldda, double vl, double vu, magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m, double *w, magmaDoubleComplex *dz, magma_int_t lddz, magmaDoubleComplex *wa, magma_int_t ldwa, magmaDoubleComplex *wz, magma_int_t ldwz, magmaDoubleComplex *work, magma_int_t lwork, double *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 ======= ZHEEVX 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. DA (device input/output) COMPLEX_16 array, dimension (LDDA, 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. LDDA (input) INTEGER The leading dimension of the array DA. LDDA >= 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'. ABSTOL (input) DOUBLE PRECISION 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*DLAMCH('S'), not zero. If this routine returns with INFO>0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*DLAMCH('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) DOUBLE PRECISION array, dimension (N) On normal exit, the first M elements contain the selected eigenvalues in ascending order. DZ (device output) COMPLEX_16 array, dimension (LDDZ, 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. ********* (workspace) If FAST_HEMV is defined DZ should be (LDDZ, max(1,N)) in both cases. LDDZ (input) INTEGER The leading dimension of the array DZ. LDDZ >= 1, and if JOBZ = 'V', LDDZ >= max(1,N). WA (workspace) COMPLEX_16 array, dimension (LDWA, N) LDWA (input) INTEGER The leading dimension of the array WA. LDWA >= max(1,N). WZ (workspace) COMPLEX_16 array, dimension (LDWZ, max(1,M)) LDWZ (input) INTEGER The leading dimension of the array DZ. LDWZ >= 1, and if JOBZ = 'V', LDWZ >= max(1,N). WORK (workspace/output) COMPLEX_16 array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= (NB+1)*N, where NB is the max of the blocksize for ZHETRD. 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) DOUBLE PRECISION 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 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; double safmin; double bignum; double smlnum; double eps, tmp1; double anrm; double sigma, d__1; double rmin, rmax; double *dwork; /* 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 (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_zhetrd_nb(n); lopt = n * (nb + 1); work[0] = MAGMA_Z_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 magmaDoubleComplex *a = (magmaDoubleComplex *) malloc( n * n * sizeof(magmaDoubleComplex) ); magma_zgetmatrix(n, n, da, ldda, a, n); lapackf77_zheevx(jobz_, range_, uplo_, &n, a, &n, &vl, &vu, &il, &iu, &abstol, m, w, wz, &ldwz, work, &lwork, rwork, iwork, ifail, info); magma_zsetmatrix( n, n, a, n, da, ldda); magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz); free(a); return *info; } if (MAGMA_SUCCESS != magma_dmalloc( &dwork, n )) { fprintf (stderr, "!!!! device memory allocation error (magma_zheevx_gpu)\n"); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } --w; --work; --rwork; --iwork; --ifail; /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_zlanhe('M', 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_zlascl(uplo, 0, 0, 1., sigma, n, n, da, ldda, info); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call ZHETRD 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_zhetrd2_gpu(uplo, n, da, ldda, &rwork[indd], &rwork[inde], &work[indtau], wa, ldwa, &work[indwrk], llwork, dz, lddz*n, &iinfo); #else magma_zhetrd_gpu (uplo, n, da, ldda, &rwork[indd], &rwork[inde], &work[indtau], wa, ldwa, &work[indwrk], llwork, &iinfo); #endif lopt = n + (magma_int_t)MAGMA_Z_REAL(work[indwrk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call DSTERF or ZUNGTR and ZSTEQR. If this fails for some eigenvalue, then try DSTEBZ. */ if ((alleig || (indeig && il == 1 && iu == n)) && abstol <= 0.) { blasf77_dcopy(&n, &rwork[indd], &ione, &w[1], &ione); indee = indrwk + 2*n; if (! wantz) { i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_dsterf(&n, &w[1], &rwork[indee], info); } else { lapackf77_zlacpy("A", &n, &n, wa, &ldwa, wz, &ldwz); lapackf77_zungtr(uplo_, &n, wz, &ldwz, &work[indtau], &work[indwrk], &llwork, &iinfo); i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_zsteqr(jobz_, &n, &w[1], &rwork[indee], wz, &ldwz, &rwork[indrwk], info); if (*info == 0) { for (i = 1; i <= n; ++i) { ifail[i] = 0; } magma_zsetmatrix( n, n, wz, ldwz, dz, lddz ); } } if (*info == 0) { *m = n; } } /* Otherwise, call DSTEBZ and, if eigenvectors are desired, ZSTEIN. */ 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_dstebz(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_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp], wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info); magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz ); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by ZSTEIN. */ magma_zunmtr_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_dscal(&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_zswap(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(1) to optimal complex workspace size. */ work[1] = MAGMA_Z_MAKE( lopt, 0 ); return *info; } /* magma_zheevx_gpu */
/* //////////////////////////////////////////////////////////////////////////// -- Testing zlanhe */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time; magmaDoubleComplex *h_A; double *h_work; magmaDoubleComplex_ptr d_A; magmaDouble_ptr d_work; magma_int_t i, j, N, n2, lda, ldda; magma_int_t idist = 3; // normal distribution (otherwise max norm is always ~ 1) magma_int_t ISEED[4] = {0,0,0,1}; double error, norm_magma, norm_lapack; magma_int_t status = 0; magma_int_t lapack_nan_fail = 0; magma_int_t lapack_inf_fail = 0; bool mkl_warning = false; magma_opts opts; opts.parse_opts( argc, argv ); double tol = opts.tolerance * lapackf77_dlamch("E"); double tol2; magma_uplo_t uplo[] = { MagmaLower, MagmaUpper }; magma_norm_t norm[] = { MagmaInfNorm, MagmaOneNorm, MagmaMaxNorm, MagmaFrobeniusNorm }; // Double-Complex inf-norm not supported on Tesla (CUDA arch 1.x) #if defined(PRECISION_z) magma_int_t arch = magma_getdevice_arch(); if ( arch < 200 ) { printf("!!!! NOTE: Double-Complex %s and %s norm are not supported\n" "!!!! on CUDA architecture %d; requires arch >= 200.\n" "!!!! It should report \"parameter number 1 had an illegal value\" below.\n\n", MagmaInfNormStr, MagmaOneNormStr, (int) arch ); for( int inorm = 0; inorm < 2; ++inorm ) { for( int iuplo = 0; iuplo < 2; ++iuplo ) { printf( "Testing that magmablas_zlanhe( %s, %s, ... ) returns -1 error...\n", lapack_norm_const( norm[inorm] ), lapack_uplo_const( uplo[iuplo] )); norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], 1, NULL, 1, NULL, 1 ); if ( norm_magma != -1 ) { printf( "expected magmablas_zlanhe to return -1 error, but got %f\n", norm_magma ); status = 1; } }} printf( "...return values %s\n\n", (status == 0 ? "ok" : "failed") ); } #endif #ifdef MAGMA_WITH_MKL // MKL 11.1 has bug in multi-threaded zlanhe; use single thread to work around. // MKL 11.2 corrects it for inf, one, max norm. // MKL 11.2 still segfaults for Frobenius norm, which is not tested here // because MAGMA doesn't implement Frobenius norm yet. MKLVersion mkl_version; mkl_get_version( &mkl_version ); magma_int_t la_threads = magma_get_lapack_numthreads(); bool mkl_single_thread = (mkl_version.MajorVersion <= 11 && mkl_version.MinorVersion < 2); if ( mkl_single_thread ) { printf( "\nNote: using single thread to work around MKL zlanhe bug.\n\n" ); } #endif printf("%% N norm uplo CPU GByte/s (ms) GPU GByte/s (ms) error nan inf\n"); printf("%%=================================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int inorm = 0; inorm < 3; ++inorm ) { /* < 4 for Frobenius */ for( int iuplo = 0; iuplo < 2; ++iuplo ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; ldda = magma_roundup( N, opts.align ); // read upper or lower triangle gbytes = 0.5*(N+1)*N*sizeof(magmaDoubleComplex) / 1e9; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_work, double, N ); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( d_work, double, N ); /* Initialize the matrix */ lapackf77_zlarnv( &idist, ISEED, &n2, h_A ); magma_zsetmatrix( N, N, h_A, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gbytes / gpu_time; if (norm_magma == -1) { printf( "%5d %4c skipped because %s norm isn't supported\n", (int) N, lapacke_norm_const( norm[inorm] ), lapack_norm_const( norm[inorm] )); goto cleanup; } else if (norm_magma < 0) { printf("magmablas_zlanhe returned error %f: %s.\n", norm_magma, magma_strerror( (int) norm_magma )); } /* ===================================================================== Performs operation using LAPACK =================================================================== */ #ifdef MAGMA_WITH_MKL if ( mkl_single_thread ) { // work around MKL bug in multi-threaded zlanhe magma_set_lapack_numthreads( 1 ); } #endif cpu_time = magma_wtime(); norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ), lapack_uplo_const( uplo[iuplo] ), &N, h_A, &lda, h_work ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gbytes / cpu_time; if (norm_lapack < 0) { printf("lapackf77_zlanhe returned error %f: %s.\n", norm_lapack, magma_strerror( (int) norm_lapack )); } /* ===================================================================== Check the result compared to LAPACK =================================================================== */ error = fabs( norm_magma - norm_lapack ) / norm_lapack; tol2 = tol; if ( norm[inorm] == MagmaMaxNorm ) { // max-norm depends on only one element, so for Real precisions, // MAGMA and LAPACK should exactly agree (tol2 = 0), // while Complex precisions incur roundoff in cuCabs. #ifdef REAL tol2 = 0; #endif } bool okay; okay = (error <= tol2); status += ! okay; mkl_warning |= ! okay; /* ==================================================================== Check for NAN and INF propagation =================================================================== */ #define h_A(i_, j_) (h_A + (i_) + (j_)*lda) #define d_A(i_, j_) (d_A + (i_) + (j_)*ldda) i = rand() % N; j = rand() % N; magma_int_t tmp; if ( uplo[iuplo] == MagmaLower && i < j ) { tmp = i; i = j; j = tmp; } else if ( uplo[iuplo] == MagmaUpper && i > j ) { tmp = i; i = j; j = tmp; } *h_A(i,j) = MAGMA_Z_NAN; magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 ); norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N ); norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ), lapack_uplo_const( uplo[iuplo] ), &N, h_A, &lda, h_work ); bool nan_okay; nan_okay = isnan(norm_magma); bool la_nan_okay; la_nan_okay = isnan(norm_lapack); lapack_nan_fail += ! la_nan_okay; status += ! nan_okay; *h_A(i,j) = MAGMA_Z_INF; magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 ); norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N ); norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ), lapack_uplo_const( uplo[iuplo] ), &N, h_A, &lda, h_work ); bool inf_okay; inf_okay = isinf(norm_magma); bool la_inf_okay; la_inf_okay = isinf(norm_lapack); lapack_inf_fail += ! la_inf_okay; status += ! inf_okay; #ifdef MAGMA_WITH_MKL if ( mkl_single_thread ) { // end single thread to work around MKL bug magma_set_lapack_numthreads( la_threads ); } #endif printf("%5d %4c %4c %7.2f (%7.2f) %7.2f (%7.2f) %#9.3g %-6s %6s%1s %6s%1s\n", (int) N, lapacke_norm_const( norm[inorm] ), lapacke_uplo_const( uplo[iuplo] ), cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., error, (okay ? "ok" : "failed"), (nan_okay ? "ok" : "failed"), (la_nan_okay ? " " : "*"), (inf_okay ? "ok" : "failed"), (la_inf_okay ? " " : "*")); cleanup: TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_work ); fflush( stdout ); } // end iter if ( opts.niter > 1 ) { printf( "\n" ); } }} // end iuplo, inorm printf( "\n" ); } // don't print "failed" here because then run_tests.py thinks MAGMA failed if ( lapack_nan_fail ) { printf( "* Warning: LAPACK did not pass NAN propagation test; upgrade to LAPACK version >= 3.4.2 (Sep. 2012)\n" ); } if ( lapack_inf_fail ) { printf( "* Warning: LAPACK did not pass INF propagation test\n" ); } if ( mkl_warning ) { printf("* MKL (e.g., 11.1) has a bug in zlanhe with multiple threads;\n" " corrected in 11.2 for one, inf, max norms, but still in Frobenius norm.\n" " Try again with MKL_NUM_THREADS=1.\n" ); } opts.cleanup(); TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_zcposv_gpu(char uplo, magma_int_t n, magma_int_t nrhs, magmaDoubleComplex *dA, magma_int_t ldda, magmaDoubleComplex *dB, magma_int_t lddb, magmaDoubleComplex *dX, magma_int_t lddx, magmaDoubleComplex *dworkd, magmaFloatComplex *dworks, magma_int_t *iter, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= ZCPOSV computes the solution to a complex system of linear equations A * X = B, where A is an N-by-N Hermitian positive definite matrix and X and B are N-by-NRHS matrices. ZCPOSV first attempts to factorize the matrix in complex SINGLE PRECISION and use this factorization within an iterative refinement procedure to produce a solution with complex DOUBLE PRECISION norm-wise backward error quality (see below). If the approach fails the method switches to a complex DOUBLE PRECISION factorization and solve. The iterative refinement is not going to be a winning strategy if the ratio complex SINGLE PRECISION performance over complex DOUBLE PRECISION performance is too small. A reasonable strategy should take the number of right-hand sides and the size of the matrix into account. This might be done with a call to ILAENV in the future. Up to now, we always try iterative refinement. The iterative refinement process is stopped if ITER > ITERMAX or for all the RHS we have: RNRM < SQRT(N)*XNRM*ANRM*EPS*BWDMAX where o ITER is the number of the current iteration in the iterative refinement process o RNRM is the infinity-norm of the residual o XNRM is the infinity-norm of the solution o ANRM is the infinity-operator-norm of the matrix A o EPS is the machine epsilon returned by DLAMCH('Epsilon') The value ITERMAX and BWDMAX are fixed to 30 and 1.0D+00 respectively. Arguments ========= UPLO (input) CHARACTER = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The number of linear equations, i.e., the order of the matrix A. N >= 0. NRHS (input) INTEGER The number of right hand sides, i.e., the number of columns of the matrix B. NRHS >= 0. dA (input or input/output) COMPLEX_16 array on the GPU, dimension (LDDA,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 iterative refinement has been successfully used (INFO.EQ.0 and ITER.GE.0, see description below), then A is unchanged, if double factorization has been used (INFO.EQ.0 and ITER.LT.0, see description below), then the array dA contains the factor U or L from the Cholesky factorization A = U**T*U or A = L*L**T. LDDA (input) INTEGER The leading dimension of the array dA. LDDA >= max(1,N). dB (input) COMPLEX_16 array on the GPU, dimension (LDDB,NRHS) The N-by-NRHS right hand side matrix B. LDDB (input) INTEGER The leading dimension of the array dB. LDDB >= max(1,N). dX (output) COMPLEX_16 array on the GPU, dimension (LDDX,NRHS) If INFO = 0, the N-by-NRHS solution matrix X. LDDX (input) INTEGER The leading dimension of the array dX. LDDX >= max(1,N). dworkd (workspace) COMPLEX_16 array on the GPU, dimension (N*NRHS) This array is used to hold the residual vectors. dworks (workspace) COMPLEX array on the GPU, dimension (N*(N+NRHS)) This array is used to store the complex single precision matrix and the right-hand sides or solutions in single precision. ITER (output) INTEGER < 0: iterative refinement has failed, double precision factorization has been performed -1 : the routine fell back to full precision for implementation- or machine-specific reasons -2 : narrowing the precision induced an overflow, the routine fell back to full precision -3 : failure of SPOTRF -31: stop the iterative refinement after the 30th iteration > 0: iterative refinement has been successfully used. Returns the number of iterations INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i, the leading minor of order i of (DOUBLE PRECISION) A is not positive definite, so the factorization could not be completed, and the solution has not been computed. ===================================================================== */ #define dB(i,j) (dB + (i) + (j)*lddb) #define dX(i,j) (dX + (i) + (j)*lddx) #define dR(i,j) (dR + (i) + (j)*lddr) #define dSX(i,j) (dSX + (i) + (j)*lddsx) magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex c_one = MAGMA_Z_ONE; magma_int_t ione = 1; magmaDoubleComplex *dR; magmaFloatComplex *dSA, *dSX; magmaDoubleComplex Xnrmv, Rnrmv; double Anrm, Xnrm, Rnrm, cte, eps; magma_int_t i, j, iiter, lddsa, lddsx, lddr; /* Check arguments */ *iter = 0; *info = 0; if ( n < 0 ) *info = -1; else if ( nrhs < 0 ) *info = -2; else if ( ldda < max(1,n)) *info = -4; else if ( lddb < max(1,n)) *info = -7; else if ( lddx < max(1,n)) *info = -9; if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } if ( n == 0 || nrhs == 0 ) return *info; lddsa = n; lddsx = n; lddr = n; dSA = dworks; dSX = dSA + lddsa*n; dR = dworkd; eps = lapackf77_dlamch("Epsilon"); Anrm = magmablas_zlanhe('I', uplo, n, dA, ldda, (double*)dworkd ); cte = Anrm * eps * pow((double)n, 0.5) * BWDMAX; /* * Convert to single precision */ magmablas_zlag2c( n, nrhs, dB, lddb, dSX, lddsx, info ); if (*info != 0) { *iter = -2; goto FALLBACK; } magmablas_zlat2c( uplo, n, dA, ldda, dSA, lddsa, info ); if (*info != 0) { *iter = -2; goto FALLBACK; } // factor dSA in single precision magma_cpotrf_gpu( uplo, n, dSA, lddsa, info ); if (*info != 0) { *iter = -3; goto FALLBACK; } // solve dSA*dSX = dB in single precision magma_cpotrs_gpu( uplo, n, nrhs, dSA, lddsa, dSX, lddsx, info ); // residual dR = dB - dA*dX in double precision magmablas_clag2z( n, nrhs, dSX, lddsx, dX, lddx, info ); magmablas_zlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dR, lddr ); if ( nrhs == 1 ) { magma_zhemv( uplo, n, c_neg_one, dA, ldda, dX, 1, c_one, dR, 1 ); } else { magma_zhemm( MagmaLeft, uplo, n, nrhs, c_neg_one, dA, ldda, dX, lddx, c_one, dR, lddr ); } // TODO: use MAGMA_Z_ABS( dX(i,j) ) instead of zlange? for( j=0; j < nrhs; j++ ) { i = magma_izamax( n, dX(0,j), 1) - 1; magma_zgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 ); Xnrm = lapackf77_zlange( "F", &ione, &ione, &Xnrmv, &ione, NULL ); i = magma_izamax ( n, dR(0,j), 1 ) - 1; magma_zgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 ); Rnrm = lapackf77_zlange( "F", &ione, &ione, &Rnrmv, &ione, NULL ); if ( Rnrm > Xnrm*cte ) { goto REFINEMENT; } } *iter = 0; return *info; REFINEMENT: for( iiter=1; iiter < ITERMAX; ) { *info = 0; // convert residual dR to single precision dSX magmablas_zlag2c( n, nrhs, dR, lddr, dSX, lddsx, info ); if (*info != 0) { *iter = -2; goto FALLBACK; } // solve dSA*dSX = R in single precision magma_cpotrs_gpu( uplo, n, nrhs, dSA, lddsa, dSX, lddsx, info ); // Add correction and setup residual // dX += dSX [including conversion] --and-- // dR = dB for( j=0; j < nrhs; j++ ) { magmablas_zcaxpycp( n, dSX(0,j), dX(0,j), dB(0,j), dR(0,j) ); } // residual dR = dB - dA*dX in double precision if ( nrhs == 1 ) { magma_zhemv( uplo, n, c_neg_one, dA, ldda, dX, 1, c_one, dR, 1 ); } else { magma_zhemm( MagmaLeft, uplo, n, nrhs, c_neg_one, dA, ldda, dX, lddx, c_one, dR, lddr ); } /* Check whether the nrhs normwise backward errors satisfy the * stopping criterion. If yes, set ITER=IITER>0 and return. */ for( j=0; j < nrhs; j++ ) { i = magma_izamax( n, dX(0,j), 1) - 1; magma_zgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 ); Xnrm = lapackf77_zlange( "F", &ione, &ione, &Xnrmv, &ione, NULL ); i = magma_izamax ( n, dR(0,j), 1 ) - 1; magma_zgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 ); Rnrm = lapackf77_zlange( "F", &ione, &ione, &Rnrmv, &ione, NULL ); if ( Rnrm > Xnrm*cte ) { goto L20; } } /* If we are here, the nrhs normwise backward errors satisfy * the stopping criterion, we are good to exit. */ *iter = iiter; return *info; L20: iiter++; } /* If we are at this place of the code, this is because we have * performed ITER=ITERMAX iterations and never satisified the * stopping criterion. Set up the ITER flag accordingly and follow * up on double precision routine. */ *iter = -ITERMAX - 1; FALLBACK: /* Single-precision iterative refinement failed to converge to a * satisfactory solution, so we resort to double precision. */ magma_zpotrf_gpu( uplo, n, dA, ldda, info ); if (*info == 0) { magmablas_zlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dX, lddx ); magma_zpotrs_gpu( uplo, n, nrhs, dA, ldda, dX, lddx, info ); } return *info; }
/** Purpose ------- ZHEEVDX_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_16 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 mout 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 DOUBLE PRECISION @param[in] vu DOUBLE PRECISION 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] mout INTEGER The total number of eigenvalues found. 0 <= MOUT <= N. If RANGE = MagmaRangeAll, MOUT = N, and if RANGE = MagmaRangeI, MOUT = IU-IL+1. @param[out] w DOUBLE PRECISION array, dimension (N) If INFO = 0, the required mout eigenvalues in ascending order. @param wA (workspace) COMPLEX_16 array, dimension (LDWA, N) @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param[out] work (workspace) COMPLEX_16 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_zhetrd_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) DOUBLE PRECISION 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_zheev_driver ********************************************************************/ extern "C" magma_int_t magma_zheevdx_gpu( magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaDoubleComplex_ptr dA, magma_int_t ldda, double vl, double vu, magma_int_t il, magma_int_t iu, magma_int_t *mout, double *w, magmaDoubleComplex *wA, magma_int_t ldwa, magmaDoubleComplex *work, magma_int_t lwork, #ifdef COMPLEX double *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; double d__1; double eps; magma_int_t inde; double anrm; magma_int_t imax; double rmin, rmax; double sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; //magma_int_t indwk2; magma_int_t iscale; double safmin; double bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; double smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; magmaDouble_ptr dwork; magmaDoubleComplex_ptr 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_zhetrd_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_dlamch("Epsilon"); work[0] = MAGMA_Z_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; } /* If matrix is very small, then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { magma_int_t lda = n; magmaDoubleComplex *A; magma_zmalloc_cpu( &A, lda*n ); magma_zgetmatrix( n, n, dA, ldda, A, lda ); lapackf77_zheevd( jobz_, uplo_, &n, A, &lda, w, work, &lwork, rwork, &lrwork, iwork, &liwork, info ); magma_zsetmatrix( n, n, A, lda, dA, ldda ); magma_free_cpu( A ); *mout = 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 double). // (this is easier in dsyevd_gpu where everything is double.) // zhetrd2_gpu requires ldda*ceildiv(n,64) + 2*ldda*nb, in double-complex. // zunmtr_gpu requires lddc*n, in double-complex. // zlanhe requires n, in double. magma_int_t ldwork = max( ldda*ceildiv(n,64) + 2*ldda*nb, lddc*n ); magma_int_t ldwork_real = max( ldwork*2, n ); if ( wantz ) { // zstedx requrise 3n^2/2, in double ldwork_real = max( ldwork_real, 3*n*(n/2 + 1) ); } if (MAGMA_SUCCESS != magma_dmalloc( &dwork, ldwork_real )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } dC = (magmaDoubleComplex*) dwork; /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt( smlnum ); rmax = magma_dsqrt( bignum ); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_zlanhe( 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_zlascl( uplo, 0, 0, 1., sigma, n, n, dA, ldda, info ); } /* Call ZHETRD to reduce Hermitian matrix to tridiagonal form. */ // zhetrd rwork: e (n) // zstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // zhetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // zstedx work: tau (n) + z (n^2) // zunmtr 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_zhetrd2_gpu( uplo, n, dA, ldda, w, &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, dC, ldwork, &iinfo ); #else magma_zhetrd_gpu ( uplo, n, dA, ldda, w, &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, &iinfo ); #endif timer_stop( time ); timer_printf( "time zhetrd_gpu = %6.2f\n", time ); /* For eigenvalues only, call DSTERF. For eigenvectors, first call ZSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call ZUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_dsterf( &n, w, &rwork[inde], info ); magma_dmove_eig( range, n, w, &il, &iu, vl, vu, mout ); } else { timer_start( time ); magma_zstedx( 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 zstedx = %6.2f\n", time ); timer_start( time ); magma_dmove_eig( range, n, w, &il, &iu, vl, vu, mout ); magma_zsetmatrix( n, *mout, &work[indwrk + n * (il-1) ], n, dC, lddc ); magma_zunmtr_gpu( MagmaLeft, uplo, MagmaNoTrans, n, *mout, dA, ldda, &work[indtau], dC, lddc, wA, ldwa, &iinfo ); magma_zcopymatrix( n, *mout, dC, lddc, dA, ldda ); timer_stop( time ); timer_printf( "time zunmtr_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_dscal( &imax, &d__1, w, &ione ); } work[0] = MAGMA_Z_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_zheevdx_gpu */
/** Purpose ------- ZHEEVR computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix T. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. Whenever possible, ZHEEVR calls ZSTEGR to compute the eigenspectrum using Relatively Robust Representations. ZSTEGR 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, 1. 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, 2. Compute the eigenvalues, lambda_j, of L_i D_i L_i^T to high relative accuracy by the dqds algorithm, 3. If there is a cluster of close eigenvalues, "choose" sigma_i close to the cluster, and go to step (a), 4. 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 : ZHEEVR calls ZSTEGR when the full spectrum is requested on machines which conform to the ieee-754 floating point standard. ZHEEVR calls DSTEBZ and ZSTEIN on non-ieee machines and when partial spectrum requests are made. Normal execution of ZSTEGR 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 --------- @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_16 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, DA is destroyed. @param[in] ldda INTEGER The leading dimension of the array A. LDDA >= max(1,N). @param[in] vl DOUBLE PRECISION @param[in] vu DOUBLE PRECISION 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 DOUBLE PRECISION 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 See "Computing Small Singular Values of Bidiagonal Matrices with Guaranteed High Relative Accuracy," by Demmel and Kahan, LAPACK Working Note #3. \n If high relative accuracy is important, set ABSTOL to DLAMCH( '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. @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 DOUBLE PRECISION array, dimension (N) The first M elements contain the selected eigenvalues in ascending order. @param[out] dZ COMPLEX_16 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 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 Z. LDDZ >= 1, and if JOBZ = MagmaVec, LDDZ >= max(1,N). @param[out] isuppz 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 = MagmaRangeAll or MagmaRangeI and IU - IL = N - 1 @param wA (workspace) COMPLEX_16 array, dimension (LDWA, N) @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param wZ (workspace) COMPLEX_16 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_16 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 ZHETRD \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] rwork (workspace) DOUBLE PRECISION array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal (and minimal) LRWORK. @param[in] lrwork INTEGER The length of the array RWORK. LRWORK >= max(1,24*N). \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. @param[out] iwork (workspace) INTEGER array, dimension (LIWORK) On exit, if INFO = 0, IWORK[0] returns the optimal (and minimal) LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. LIWORK >= max(1,10*N). \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. @param[out] info 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 @ingroup magma_zheev_driver ********************************************************************/ extern "C" magma_int_t magma_zheevr_gpu( magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaDoubleComplex_ptr dA, magma_int_t ldda, double vl, double vu, magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m, double *w, magmaDoubleComplex_ptr dZ, magma_int_t lddz, magma_int_t *isuppz, magmaDoubleComplex *wA, magma_int_t ldwa, magmaDoubleComplex *wZ, magma_int_t ldwz, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { /* Constants */ magma_int_t ione = 1; float szero = 0.; float sone = 1.; /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); const char* range_ = lapack_range_const( range ); magma_int_t indrd, indre; magma_int_t imax; magma_int_t lopt, itmp1, indree, indrdd; magma_int_t tryrac; magma_int_t i, j, jj, i__1; magma_int_t iscale, indibl, indifl; magma_int_t indiwo, indisp, indtau; magma_int_t indrwk, indwk; magma_int_t llwork, llrwork, nsplit; magma_int_t ieeeok; magma_int_t iinfo; magma_int_t lwmin, lrwmin, liwmin; double safmin; double bignum; double smlnum; double eps, tmp1; double anrm; double sigma, d__1; double rmin, rmax; magmaDouble_ptr dwork; bool lower = (uplo == MagmaLower); bool wantz = (jobz == MagmaVec); bool alleig = (range == MagmaRangeAll); bool valeig = (range == MagmaRangeV); bool indeig = (range == MagmaRangeI); bool 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 (lddz < 1 || (wantz && lddz < n)) { *info = -15; } else if (ldwa < max(1,n)) { *info = -18; } else if (ldwz < 1 || (wantz && ldwz < n)) { *info = -20; } 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_zhetrd_nb(n); lwmin = n * (nb + 1); lrwmin = 24 * n; liwmin = 10 * n; work[0] = magma_zmake_lwork( lwmin ); rwork[0] = magma_dmake_lwork( lrwmin ); iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -22; } else if ((lrwork < lrwmin) && ! lquery) { *info = -24; } else if ((liwork < liwmin) && ! lquery) { *info = -26; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } *m = 0; magma_queue_t queue; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queue ); /* 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 magmaDoubleComplex *A; magma_zmalloc_cpu( &A, n*n ); magma_zgetmatrix( n, n, dA, ldda, A, n, queue ); lapackf77_zheevr(jobz_, range_, uplo_, &n, A, &n, &vl, &vu, &il, &iu, &abstol, m, w, wZ, &ldwz, isuppz, work, &lwork, rwork, &lrwork, iwork, &liwork, info); magma_zsetmatrix( n, n, A, n, dA, ldda, queue ); magma_zsetmatrix( n, *m, wZ, ldwz, dZ, lddz, queue ); magma_free_cpu( A ); magma_queue_destroy( queue ); return *info; } if (MAGMA_SUCCESS != magma_dmalloc( &dwork, n )) { fprintf (stderr, "!!!! device memory allocation error (magma_zheevr_gpu)\n"); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } --w; --work; --rwork; --iwork; --isuppz; /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_zlanhe( MagmaMaxNorm, uplo, n, dA, ldda, dwork, n, queue ); 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_zlascl( uplo, 0, 0, 1., sigma, n, n, dA, ldda, queue, info ); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call ZHETRD to reduce Hermitian matrix to tridiagonal form. */ indtau = 1; indwk = indtau + n; indre = 1; indrd = indre + n; indree = indrd + n; indrdd = indree + n; indrwk = indrdd + n; llwork = lwork - indwk + 1; llrwork = lrwork - indrwk + 1; indifl = 1; indibl = indifl + n; indisp = indibl + n; indiwo = indisp + n; #ifdef FAST_HEMV magma_zhetrd2_gpu(uplo, n, dA, ldda, &rwork[indrd], &rwork[indre], &work[indtau], wA, ldwa, &work[indwk], llwork, dZ, lddz*n, &iinfo); #else magma_zhetrd_gpu (uplo, n, dA, ldda, &rwork[indrd], &rwork[indre], &work[indtau], wA, ldwa, &work[indwk], llwork, &iinfo); #endif lopt = n + (magma_int_t)MAGMA_Z_REAL(work[indwk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call DSTERF or ZUNGTR and ZSTEQR. If this fails for some eigenvalue, then try DSTEBZ. */ ieeeok = lapackf77_ieeeck( &ione, &szero, &sone); /* If only the eigenvalues are required call DSTERF for all or DSTEBZ for a part */ if (! wantz) { blasf77_dcopy(&n, &rwork[indrd], &ione, &w[1], &ione); i__1 = n - 1; if (alleig || (indeig && il == 1 && iu == n)) { lapackf77_dsterf(&n, &w[1], &rwork[indre], info); *m = n; } else { lapackf77_dstebz(range_, "E", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m, &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwo], info); } /* Otherwise call ZSTEMR if infinite and NaN arithmetic is supported */ } else if (ieeeok == 1) { //printf("MRRR\n"); i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[indre], &ione, &rwork[indree], &ione); blasf77_dcopy(&n, &rwork[indrd], &ione, &rwork[indrdd], &ione); if (abstol < 2*n*eps) tryrac=1; else tryrac=0; lapackf77_zstemr(jobz_, range_, &n, &rwork[indrdd], &rwork[indree], &vl, &vu, &il, &iu, m, &w[1], wZ, &ldwz, &n, &isuppz[1], &tryrac, &rwork[indrwk], &llrwork, &iwork[1], &liwork, info); if (*info == 0 && wantz) { magma_zsetmatrix( n, *m, wZ, ldwz, dZ, lddz, queue ); magma_zunmtr_gpu(MagmaLeft, uplo, MagmaNoTrans, n, *m, dA, ldda, &work[indtau], dZ, lddz, wA, ldwa, &iinfo); } } /* Call DSTEBZ and ZSTEIN if infinite and NaN arithmetic is not supported or ZSTEMR didn't converge. */ if (wantz && (ieeeok == 0 || *info != 0)) { //printf("B/I\n"); *info = 0; lapackf77_dstebz(range_, "B", &n, &vl, &vu, &il, &iu, &abstol, &rwork[indrd], &rwork[indre], m, &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwo], info); lapackf77_zstein(&n, &rwork[indrd], &rwork[indre], m, &w[1], &iwork[indibl], &iwork[indisp], wZ, &ldwz, &rwork[indrwk], &iwork[indiwo], &iwork[indifl], info); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by ZSTEIN. */ magma_zsetmatrix( n, *m, wZ, ldwz, dZ, lddz, queue ); magma_zunmtr_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_dscal(&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_zswap( n, dZ + (i-1)*lddz, ione, dZ + (j-1)*lddz, ione, queue ); } } } /* Set WORK[0] to optimal complex workspace size. */ work[1] = magma_zmake_lwork( lopt ); rwork[1] = magma_dmake_lwork( lrwmin ); iwork[1] = liwmin; magma_queue_destroy( queue ); return *info; } /* magma_zheevr_gpu */
/** Purpose ------- ZHEEVX 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_16 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 DOUBLE PRECISION @param[in] vu DOUBLE PRECISION 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 DOUBLE PRECISION 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*DLAMCH('S'), not zero. If this routine returns with INFO > 0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*DLAMCH('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 DOUBLE PRECISION array, dimension (N) On normal exit, the first M elements contain the selected eigenvalues in ascending order. @param[out] dZ COMPLEX_16 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_16 array, dimension (LDWA, N) @param[in] ldwa INTEGER The leading dimension of the array wA. LDWA >= max(1,N). @param wZ (workspace) COMPLEX_16 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_16 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 ZHETRD. \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) DOUBLE PRECISION 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_zheev_driver ********************************************************************/ extern "C" magma_int_t magma_zheevx_gpu(magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaDoubleComplex *dA, magma_int_t ldda, double vl, double vu, magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m, double *w, magmaDoubleComplex *dZ, magma_int_t lddz, magmaDoubleComplex *wA, magma_int_t ldwa, magmaDoubleComplex *wZ, magma_int_t ldwz, magmaDoubleComplex *work, magma_int_t lwork, double *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; double safmin; double bignum; double smlnum; double eps, tmp1; double anrm; double sigma, d__1; double rmin, rmax; double *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_zhetrd_nb(n); lopt = n * (nb + 1); work[0] = MAGMA_Z_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 magmaDoubleComplex *a; magma_zmalloc_cpu( &a, n*n ); magma_zgetmatrix(n, n, dA, ldda, a, n); lapackf77_zheevx(jobz_, range_, uplo_, &n, a, &n, &vl, &vu, &il, &iu, &abstol, m, w, wZ, &ldwz, work, &lwork, rwork, iwork, ifail, info); magma_zsetmatrix( n, n, a, n, dA, ldda); magma_zsetmatrix( n, *m, wZ, ldwz, dZ, lddz); magma_free_cpu(a); return *info; } if (MAGMA_SUCCESS != magma_dmalloc( &dwork, n )) { fprintf (stderr, "!!!! device memory allocation error (magma_zheevx_gpu)\n"); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } --w; --work; --rwork; --iwork; --ifail; /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_zlanhe(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_zlascl(uplo, 0, 0, 1., sigma, n, n, dA, ldda, info); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call ZHETRD 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_zhetrd2_gpu(uplo, n, dA, ldda, &rwork[indd], &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, dZ, lddz*n, &iinfo); #else magma_zhetrd_gpu (uplo, n, dA, ldda, &rwork[indd], &rwork[inde], &work[indtau], wA, ldwa, &work[indwrk], llwork, &iinfo); #endif lopt = n + (magma_int_t)MAGMA_Z_REAL(work[indwrk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call DSTERF or ZUNGTR and ZSTEQR. If this fails for some eigenvalue, then try DSTEBZ. */ if ((alleig || (indeig && il == 1 && iu == n)) && abstol <= 0.) { blasf77_dcopy(&n, &rwork[indd], &ione, &w[1], &ione); indee = indrwk + 2*n; if (! wantz) { i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_dsterf(&n, &w[1], &rwork[indee], info); } else { lapackf77_zlacpy("A", &n, &n, wA, &ldwa, wZ, &ldwz); lapackf77_zungtr(uplo_, &n, wZ, &ldwz, &work[indtau], &work[indwrk], &llwork, &iinfo); i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_zsteqr(jobz_, &n, &w[1], &rwork[indee], wZ, &ldwz, &rwork[indrwk], info); if (*info == 0) { for (i = 1; i <= n; ++i) { ifail[i] = 0; } magma_zsetmatrix( n, n, wZ, ldwz, dZ, lddz ); } } if (*info == 0) { *m = n; } } /* Otherwise, call DSTEBZ and, if eigenvectors are desired, ZSTEIN. */ if (*m == 0) { *info = 0; if (wantz) { order_ = "B"; } else { order_ = "E"; } indibl = 1; indisp = indibl + n; indiwk = indisp + n; lapackf77_dstebz(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_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp], wZ, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info); magma_zsetmatrix( n, *m, wZ, ldwz, dZ, lddz ); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by ZSTEIN. */ magma_zunmtr_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_dscal(&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_zswap(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_Z_MAKE( lopt, 0 ); return *info; } /* magma_zheevx_gpu */