extern "C" magma_int_t magma_sgeev(magma_vec_t jobvl, magma_vec_t jobvr, magma_int_t n, float *a, magma_int_t lda, float *WR, float *WI, float *vl, magma_int_t ldvl, float *vr, magma_int_t ldvr, float *work, magma_int_t lwork, magma_int_t *info, magma_queue_t queue) { /* -- clMAGMA (version 1.0.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver September 2012 Purpose ======= SGEEV computes for an N-by-N real nonsymmetric matrix A, the eigenvalues and, optionally, the left and/or right eigenvectors. The right eigenvector v(j) of A satisfies A * v(j) = lambda(j) * v(j) where lambda(j) is its eigenvalue. The left eigenvector u(j) of A satisfies u(j)**T * A = lambda(j) * u(j)**T where u(j)**T denotes the transpose of u(j). The computed eigenvectors are normalized to have Euclidean norm equal to 1 and largest component real. Arguments ========= JOBVL (input) CHARACTER*1 = 'N': left eigenvectors of A are not computed; = 'V': left eigenvectors of are computed. JOBVR (input) CHARACTER*1 = 'N': right eigenvectors of A are not computed; = 'V': right eigenvectors of A are computed. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) DOUBLE PRECISION array, dimension (LDA,N) On entry, the N-by-N matrix A. On exit, A has been overwritten. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). WR (output) DOUBLE PRECISION array, dimension (N) WI (output) DOUBLE PRECISION array, dimension (N) WR and WI contain the real and imaginary parts, respectively, of the computed eigenvalues. Complex conjugate pairs of eigenvalues appear consecutively with the eigenvalue having the positive imaginary part first. VL (output) DOUBLE PRECISION array, dimension (LDVL,N) If JOBVL = 'V', the left eigenvectors u(j) are stored one after another in the columns of VL, in the same order as their eigenvalues. If JOBVL = 'N', VL is not referenced. u(j) = VL(:,j), the j-th column of VL. LDVL (input) INTEGER The leading dimension of the array VL. LDVL >= 1; if JOBVL = 'V', LDVL >= N. VR (output) DOUBLE PRECISION array, dimension (LDVR,N) If JOBVR = 'V', the right eigenvectors v(j) are stored one after another in the columns of VR, in the same order as their eigenvalues. If JOBVR = 'N', VR is not referenced. v(j) = VR(:,j), the j-th column of VR. LDVR (input) INTEGER The leading dimension of the array VR. LDVR >= 1; if JOBVR = 'V', LDVR >= N. WORK (workspace/output) DOUBLE PRECISION 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+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. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value. > 0: if INFO = i, the QR algorithm failed to compute all the eigenvalues, and no eigenvectors have been computed; elements and i+1:N of W contain eigenvalues which have converged. ===================================================================== */ magma_int_t c__1 = 1; magma_int_t c__0 = 0; magma_int_t c_n1 = -1; magma_int_t a_dim1, a_offset, vl_dim1, vl_offset, vr_dim1, vr_offset, i__1, i__2, i__3; float d__1, d__2; magma_int_t i__, k, ihi, ilo; float r__, cs, sn, scl; float dum[1], eps; magma_int_t ibal; float anrm; magma_int_t ierr, itau, iwrk, nout; magma_int_t scalea; float cscale; float bignum; magma_int_t minwrk; magma_int_t wantvl; float smlnum; magma_int_t lquery, wantvr, select[1]; magma_int_t nb = 0; magmaFloat_ptr dT; //magma_timestr_t start, end; char side[2] = {0, 0}; magma_vec_t jobvl_ = jobvl; magma_vec_t jobvr_ = jobvr; *info = 0; lquery = lwork == -1; wantvl = lapackf77_lsame(lapack_const(jobvl_), "V"); wantvr = lapackf77_lsame(lapack_const(jobvr_), "V"); if (! wantvl && ! lapackf77_lsame(lapack_const(jobvl_), "N")) { *info = -1; } else if (! wantvr && ! lapackf77_lsame(lapack_const(jobvr_), "N")) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if ( (ldvl < 1) || (wantvl && (ldvl < n))) { *info = -9; } else if ( (ldvr < 1) || (wantvr && (ldvr < n))) { *info = -11; } /* Compute workspace */ if (*info == 0) { nb = magma_get_sgehrd_nb(n); minwrk = (2+nb)*n; work[0] = (float) minwrk; if (lwork < minwrk && ! lquery) { *info = -13; } } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } // if eigenvectors are needed #if defined(VERSION3) if (MAGMA_SUCCESS != magma_malloc( &dT, nb*n*sizeof(float) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } #endif // subtract row and col for 1-based indexing a_dim1 = lda; a_offset = 1 + a_dim1; a -= a_offset; vl_dim1 = ldvl; vl_offset = 1 + vl_dim1; vl -= vl_offset; vr_dim1 = ldvr; vr_offset = 1 + vr_dim1; vr -= vr_offset; --work; /* Get machine constants */ eps = lapackf77_slamch("P"); smlnum = lapackf77_slamch("S"); bignum = 1. / smlnum; lapackf77_slabad(&smlnum, &bignum); smlnum = magma_ssqrt(smlnum) / eps; bignum = 1. / smlnum; /* Scale A if max element outside range [SMLNUM,BIGNUM] */ anrm = lapackf77_slange("M", &n, &n, &a[a_offset], &lda, dum); scalea = 0; if (anrm > 0. && anrm < smlnum) { scalea = 1; cscale = smlnum; } else if (anrm > bignum) { scalea = 1; cscale = bignum; } if (scalea) { lapackf77_slascl("G", &c__0, &c__0, &anrm, &cscale, &n, &n, &a[a_offset], &lda, &ierr); } /* Balance the matrix (Workspace: need N) */ ibal = 1; lapackf77_sgebal("B", &n, &a[a_offset], &lda, &ilo, &ihi, &work[ibal], &ierr); /* Reduce to upper Hessenberg form (Workspace: need 3*N, prefer 2*N+N*NB) */ itau = ibal + n; iwrk = itau + n; i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) /* * Version 1 - LAPACK */ lapackf77_sgehrd(&n, &ilo, &ihi, &a[a_offset], &lda, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION2) /* * Version 2 - LAPACK consistent HRD */ magma_sgehrd2(n, ilo, ihi, &a[a_offset], lda, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored, */ magma_sgehrd(n, ilo, ihi, &a[a_offset], lda, &work[itau], &work[iwrk], i__1, dT, 0, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for sgehrd = %5.2f sec\n", GetTimerValue(start,end)/1000.); if (wantvl) { /* Want left eigenvectors Copy Householder vectors to VL */ side[0] = 'L'; lapackf77_slacpy(MagmaLowerStr, &n, &n, &a[a_offset], &lda, &vl[vl_offset], &ldvl); /* * Generate orthogonal matrix in VL * (Workspace: need 3*N-1, prefer 2*N+(N-1)*NB) */ i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) || defined(VERSION2) /* * Version 1 & 2 - LAPACK */ lapackf77_sorghr(&n, &ilo, &ihi, &vl[vl_offset], &ldvl, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored */ magma_sorghr(n, ilo, ihi, &vl[vl_offset], ldvl, &work[itau], dT, 0, nb, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for sorghr = %5.2f sec\n", GetTimerValue(start,end)/1000.); /* * Perform QR iteration, accumulating Schur vectors in VL * (Workspace: need N+1, prefer N+HSWORK (see comments) ) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_shseqr("S", "V", &n, &ilo, &ihi, &a[a_offset], &lda, WR, WI, &vl[vl_offset], &ldvl, &work[iwrk], &i__1, info); if (wantvr) { /* Want left and right eigenvectors Copy Schur vectors to VR */ side[0] = 'B'; lapackf77_slacpy("F", &n, &n, &vl[vl_offset], &ldvl, &vr[vr_offset], &ldvr); } } else if (wantvr) { /* Want right eigenvectors Copy Householder vectors to VR */ side[0] = 'R'; lapackf77_slacpy("L", &n, &n, &a[a_offset], &lda, &vr[vr_offset], &ldvr); /* * Generate orthogonal matrix in VR * (Workspace: need 3*N-1, prefer 2*N+(N-1)*NB) */ i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) || defined(VERSION2) /* * Version 1 & 2 - LAPACK */ lapackf77_sorghr(&n, &ilo, &ihi, &vr[vr_offset], &ldvr, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored */ magma_sorghr(n, ilo, ihi, &vr[vr_offset], ldvr, &work[itau], dT, 0, nb, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for sorghr = %5.2f sec\n", GetTimerValue(start,end)/1000.); /* * Perform QR iteration, accumulating Schur vectors in VR * (Workspace: need N+1, prefer N+HSWORK (see comments) ) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_shseqr("S", "V", &n, &ilo, &ihi, &a[a_offset], &lda, WR, WI, &vr[vr_offset], &ldvr, &work[iwrk], &i__1, info); } else { /* * Compute eigenvalues only * (Workspace: need N+1, prefer N+HSWORK (see comments) ) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_shseqr("E", "N", &n, &ilo, &ihi, &a[a_offset], &lda, WR, WI, &vr[vr_offset], &ldvr, &work[iwrk], &i__1, info); } /* If INFO > 0 from SHSEQR, then quit */ if (*info > 0) { fprintf(stderr, "SHSEQR returned with info = %d\n", (int) *info); goto L50; } if (wantvl || wantvr) { /* * Compute left and/or right eigenvectors * (Workspace: need 4*N) */ lapackf77_strevc(side, "B", select, &n, &a[a_offset], &lda, &vl[vl_offset], &ldvl, &vr[vr_offset], &ldvr, &n, &nout, &work[iwrk], &ierr); } if (wantvl) { /* * Undo balancing of left eigenvectors * (Workspace: need N) */ lapackf77_sgebak("B", "L", &n, &ilo, &ihi, &work[ibal], &n, &vl[vl_offset], &ldvl, &ierr); /* Normalize left eigenvectors and make largest component real */ for (i__ = 1; i__ <= n; ++i__) { if ( WI[i__-1] == 0.) { scl = cblas_snrm2(n, &vl[i__ * vl_dim1 + 1], 1); scl = 1. / scl; cblas_sscal(n, (scl), &vl[i__ * vl_dim1 + 1], 1); } else if (WI[i__-1] > 0.) { d__1 = cblas_snrm2(n, &vl[ i__ * vl_dim1 + 1], 1); d__2 = cblas_snrm2(n, &vl[(i__ + 1) * vl_dim1 + 1], 1); scl = lapackf77_slapy2(&d__1, &d__2); scl = 1. / scl; cblas_sscal(n, (scl), &vl[ i__ * vl_dim1 + 1], 1); cblas_sscal(n, (scl), &vl[(i__ + 1) * vl_dim1 + 1], 1); i__2 = n; for (k = 1; k <= i__2; ++k) { /* Computing 2nd power */ d__1 = vl[k + i__ * vl_dim1]; /* Computing 2nd power */ d__2 = vl[k + (i__ + 1) * vl_dim1]; work[iwrk + k - 1] = d__1 * d__1 + d__2 * d__2; } /* Comment: Fortran BLAS does not have to add 1 C BLAS must add one to cblas_isamax */ k = cblas_isamax(n, &work[iwrk], 1)+1; lapackf77_slartg(&vl[k + i__ * vl_dim1], &vl[k + (i__ + 1) * vl_dim1], &cs, &sn, &r__); cblas_srot(n, &vl[ i__ * vl_dim1 + 1], 1, &vl[(i__ + 1) * vl_dim1 + 1], 1, cs, (sn)); vl[k + (i__ + 1) * vl_dim1] = 0.; } } } if (wantvr) { /* * Undo balancing of right eigenvectors * (Workspace: need N) */ lapackf77_sgebak("B", "R", &n, &ilo, &ihi, &work[ibal], &n, &vr[vr_offset], &ldvr, &ierr); /* Normalize right eigenvectors and make largest component real */ for (i__ = 1; i__ <= n; ++i__) { if (WI[i__-1] == 0.) { scl = 1. / cblas_snrm2(n, &vr[i__ * vr_dim1 + 1], 1); cblas_sscal(n, (scl), &vr[i__ * vr_dim1 + 1], 1); } else if (WI[i__-1] > 0.) { d__1 = cblas_snrm2(n, &vr[ i__ * vr_dim1 + 1], 1); d__2 = cblas_snrm2(n, &vr[(i__ + 1) * vr_dim1 + 1], 1); scl = lapackf77_slapy2(&d__1, &d__2); scl = 1. / scl; cblas_sscal(n, (scl), &vr[ i__ * vr_dim1 + 1], 1); cblas_sscal(n, (scl), &vr[(i__ + 1) * vr_dim1 + 1], 1); i__2 = n; for (k = 1; k <= i__2; ++k) { /* Computing 2nd power */ d__1 = vr[k + i__ * vr_dim1]; /* Computing 2nd power */ d__2 = vr[k + (i__ + 1) * vr_dim1]; work[iwrk + k - 1] = d__1 * d__1 + d__2 * d__2; } /* Comment: Fortran BLAS does not have to add 1 C BLAS must add one to cblas_isamax */ k = cblas_isamax(n, &work[iwrk], 1)+1; lapackf77_slartg(&vr[k + i__ * vr_dim1], &vr[k + (i__ + 1) * vr_dim1], &cs, &sn, &r__); cblas_srot(n, &vr[ i__ * vr_dim1 + 1], 1, &vr[(i__ + 1) * vr_dim1 + 1], 1, cs, (sn)); vr[k + (i__ + 1) * vr_dim1] = 0.; } } } /* Undo scaling if necessary */ L50: if (scalea) { i__1 = n - *info; /* Computing MAX */ i__3 = n - *info; i__2 = max(i__3,1); lapackf77_slascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, WR + (*info), &i__2, &ierr); i__1 = n - *info; /* Computing MAX */ i__3 = n - *info; i__2 = max(i__3,1); lapackf77_slascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, WI + (*info), &i__2, &ierr); if (*info > 0) { i__1 = ilo - 1; lapackf77_slascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, WR, &n, &ierr); i__1 = ilo - 1; lapackf77_slascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, WI, &n, &ierr); } } #if defined(VERSION3) magma_free( dT ); #endif return *info; } /* magma_sgeev */
/** Purpose ------- SSYEVD computes all eigenvalues and, optionally, eigenvectors of a real symmetric 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 REAL array, dimension (LDA, N) On entry, the symmetric 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) REAL 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 >= 2*N + N*NB. If JOBZ = MagmaVec and N > 1, LWORK >= max( 2*N + N*NB, 1 + 6*N + 2*N**2 ). NB can be obtained through magma_get_ssytrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK and IWORK arrays, returns these values as the first entries of the WORK and IWORK arrays, and no error message related to LWORK 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 and IWORK arrays, returns these values as the first entries of the WORK and IWORK arrays, and no error message related to LWORK 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_ssyev_driver ********************************************************************/ extern "C" magma_int_t magma_ssyevd(magma_vec_t jobz, magma_uplo_t uplo, magma_int_t n, float *A, magma_int_t lda, float *w, float *work, magma_int_t lwork, 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; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indwrk, liwmin; magma_int_t llwork; float smlnum; magma_int_t lquery; float* dwork; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); lquery = (lwork == -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_ssytrd_nb( n ); if ( n <= 1 ) { lwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( 2*n + n*nb, 1 + 6*n + 2*n*n ); liwmin = 3 + 5*n; } else { lwmin = 2*n + n*nb; liwmin = 1; } // multiply by 1+eps to ensure length gets rounded up, // if it cannot be exactly represented in floating point. float one_eps = 1. + lapackf77_slamch("Epsilon"); work[0] = lwmin * one_eps; iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -8; } else if ((liwork < liwmin) && ! lquery) { *info = -10; } 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] = A[0]; if (wantz) { A[0] = 1.; } 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_ssyevd(jobz_, uplo_, &n, A, &lda, w, work, &lwork, 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_slansy("M", uplo_, &n, A, &lda, work ); 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_slascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, A, &lda, info); } /* Call SSYTRD to reduce symmetric matrix to tridiagonal form. */ // ssytrd work: e (n) + tau (n) + llwork (n*nb) ==> 2n + n*nb // sstedx work: e (n) + tau (n) + z (n*n) + llwrk2 (1 + 4*n + n^2) ==> 1 + 6n + 2n^2 inde = 0; indtau = inde + n; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; magma_ssytrd(uplo, n, A, lda, w, &work[inde], &work[indtau], &work[indwrk], llwork, &iinfo); /* For eigenvalues only, call SSTERF. For eigenvectors, first call SSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call SORMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &work[inde], info); } else { if (MAGMA_SUCCESS != magma_smalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } // TTT Possible bug for n < 128 magma_sstedx(311, n, 0., 0., 0, 0, w, &work[inde], &work[indwrk], n, &work[indwk2], llwrk2, iwork, liwork, dwork, info); magma_free( dwork ); magma_sormtr(MagmaLeft, uplo, MagmaNoTrans, n, n, A, lda, &work[indtau], &work[indwrk], n, &work[indwk2], llwrk2, &iinfo); lapackf77_slacpy("A", &n, &n, &work[indwrk], &n, A, &lda); } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { d__1 = 1. / sigma; blasf77_sscal(&n, &d__1, w, &ione); } work[0] = lwmin * one_eps; // round up iwork[0] = liwmin; return *info; } /* magma_ssyevd */
extern "C" magma_int_t magma_ssyevdx_2stage(char jobz, char range, char uplo, magma_int_t n, float *a, magma_int_t lda, float vl, float vu, magma_int_t il, magma_int_t iu, magma_int_t *m, float *w, float *work, magma_int_t lwork, 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 ======= ZHEEVD_2STAGE computes all eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. It uses a two-stage algorithm for the tridiagonalization. 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_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, 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) 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'. 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 required m 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 >= LQ2 + N * (NB + 2). If JOBZ = 'V' and N > 1, LWORK >= LQ2 + 1 + 6*N + 2*N**2. where LQ2 is the size needed to store the Q2 matrix and is returned by MAGMA_BULGE_GET_LQ2. 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. 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: 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}; float d_one = 1.; magma_int_t ione = 1; magma_int_t izero = 0; float d__1; float eps; float anrm; magma_int_t imax; float rmin, rmax; float sigma; magma_int_t lwmin, liwmin; magma_int_t lower; magma_int_t wantz; magma_int_t iscale; float safmin; float bignum; float smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; float* dwork; /* determine the number of threads */ magma_int_t threads = magma_get_numthreads(); magma_setlapack_numthreads(threads); 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 || 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_sbulge_nb(n, threads); magma_int_t Vblksiz = magma_sbulge_get_Vblksiz(n, nb, threads); magma_int_t ldt = Vblksiz; magma_int_t ldv = nb + Vblksiz; magma_int_t blkcnt = magma_bulge_get_blkcnt(n, nb, Vblksiz); magma_int_t lq2 = magma_sbulge_get_lq2(n, threads); if (wantz) { lwmin = lq2 + 1 + 6 * n + 2 * n * n; liwmin = 5 * n + 3; } else { lwmin = lq2 + n * (nb + 1); liwmin = 1; } work[0] = lwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -14; } else if ((liwork < liwmin) && ! lquery) { *info = -16; } 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] = a[0]; if (wantz) { a[0] = MAGMA_S_ONE; } return *info; } #ifdef ENABLE_TIMER printf("using %d threads\n", threads); #endif /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ magma_int_t ntiles = n/nb; if( ( ntiles < 2 ) || ( 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_ssyevd(jobz_, uplo_, &n, a, &lda, w, work, &lwork, iwork, &liwork, info); *m = n; 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_slansy("M", uplo_, &n, a, &lda, work); 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_slascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a, &lda, info); } magma_int_t inde = 0; magma_int_t indT2 = inde + n; magma_int_t indTAU2 = indT2 + blkcnt*ldt*Vblksiz; magma_int_t indV2 = indTAU2+ blkcnt*Vblksiz; magma_int_t indtau1 = indV2 + blkcnt*ldv*Vblksiz; magma_int_t indwrk = indtau1+ n; magma_int_t indwk2 = indwrk + n * n; magma_int_t llwork = lwork - indwrk; magma_int_t llwrk2 = lwork - indwk2; #ifdef ENABLE_TIMER magma_timestr_t start, st1, st2, end; start = get_current_time(); #endif float *dT1; if (MAGMA_SUCCESS != magma_smalloc( &dT1, n*nb)) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_ssytrd_sy2sb(uplo, n, nb, a, lda, &work[indtau1], &work[indwrk], llwork, dT1, threads, info); #ifdef ENABLE_TIMER st1 = get_current_time(); printf(" time ssytrd_sy2sb = %6.2f\n" , GetTimerValue(start,st1)/1000.); #endif /* copy the input matrix into WORK(INDWRK) with band storage */ /* PAY ATTENTION THAT work[indwrk] should be able to be of size lda2*n which it should be checked in any future modification of lwork.*/ magma_int_t lda2 = 2*nb; //nb+1+(nb-1); float* A2 = &work[indwrk]; memset(A2 , 0, n*lda2*sizeof(float)); for (magma_int_t j = 0; j < n-nb; j++) { cblas_scopy(nb+1, &a[j*(lda+1)], 1, &A2[j*lda2], 1); memset(&a[j*(lda+1)], 0, (nb+1)*sizeof(float)); a[nb + j*(lda+1)] = d_one; } for (magma_int_t j = 0; j < nb; j++) { cblas_scopy(nb-j, &a[(j+n-nb)*(lda+1)], 1, &A2[(j+n-nb)*lda2], 1); memset(&a[(j+n-nb)*(lda+1)], 0, (nb-j)*sizeof(float)); } #ifdef ENABLE_TIMER st2 = get_current_time(); printf(" time ssytrd_convert = %6.2f\n" , GetTimerValue(st1,st2)/1000.); #endif magma_ssytrd_sb2st(threads, uplo, n, nb, Vblksiz, A2, lda2, w, &work[inde], &work[indV2], ldv, &work[indTAU2], wantz, &work[indT2], ldt); #ifdef ENABLE_TIMER end = get_current_time(); printf(" time ssytrd_sy2st = %6.2f\n" , GetTimerValue(st2,end)/1000.); printf(" time ssytrd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif /* For eigenvalues only, call SSTERF. 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) { #ifdef ENABLE_TIMER start = get_current_time(); #endif lapackf77_ssterf(&n, w, &work[inde], info); magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); #ifdef ENABLE_TIMER end = get_current_time(); printf(" time sstedc = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } 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_sstedx(range, n, vl, vu, il, iu, w, &work[inde], &work[indwrk], n, &work[indwk2], llwrk2, iwork, liwork, dwork, info); magma_free( dwork ); #ifdef ENABLE_TIMER end = get_current_time(); printf(" time sstedx = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif float *dZ; magma_int_t lddz = n; float *da; magma_int_t ldda = n; magma_smove_eig(range, n, w, &il, &iu, vl, vu, m); if (MAGMA_SUCCESS != magma_smalloc( &dZ, *m*lddz)) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } if (MAGMA_SUCCESS != magma_smalloc( &da, n*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_sbulge_back(threads, uplo, n, nb, *m, Vblksiz, &work[indwrk + n * (il-1)], n, dZ, lddz, &work[indV2], ldv, &work[indTAU2], &work[indT2], ldt, info); #ifdef ENABLE_TIMER st1 = get_current_time(); printf(" time sbulge_back = %6.2f\n" , GetTimerValue(start,st1)/1000.); #endif magma_ssetmatrix( n, n, a, lda, da, ldda ); magma_sormqr_gpu_2stages(MagmaLeft, MagmaNoTrans, n-nb, *m, n-nb, da+nb, ldda, dZ+nb, n, dT1, nb, info); magma_sgetmatrix( n, *m, dZ, lddz, a, lda ); magma_free(dT1); magma_free(dZ); magma_free(da); #ifdef ENABLE_TIMER end = get_current_time(); printf(" time sormqr + copy = %6.2f\n", GetTimerValue(st1,end)/1000.); printf(" time eigenvectors backtransf. = %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] = lwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; return *info; } /* magma_zheevdx_2stage */
/** Purpose ------- SSTEDX computes some eigenvalues and, optionally, eigenvectors of a symmetric tridiagonal matrix using the divide and conquer method. This code 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. See SLAEX3 for details. Arguments --------- @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] n INTEGER The dimension of the symmetric tridiagonal matrix. N >= 0. @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,out] d REAL array, dimension (N) On entry, the diagonal elements of the tridiagonal matrix. On exit, if INFO = 0, the eigenvalues in ascending order. @param[in,out] e REAL array, dimension (N-1) On entry, the subdiagonal elements of the tridiagonal matrix. On exit, E has been destroyed. @param[in,out] Z REAL array, dimension (LDZ,N) On exit, if INFO = 0, Z contains the orthonormal eigenvectors of the symmetric tridiagonal matrix. @param[in] ldz INTEGER The leading dimension of the array Z. LDZ >= max(1,N). @param[out] work (workspace) REAL array, dimension (LWORK) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. If N > 1 then LWORK >= ( 1 + 4*N + N**2 ). Note that if N is less than or equal to the minimum divide size, usually 25, then LWORK need only be max(1,2*(N-1)). \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] 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. LIWORK >= ( 3 + 5*N ). Note that if N is less than or equal to the minimum divide size, usually 25, then LIWORK need only be 1. \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 dwork (workspace) REAL array, dimension (3*N*N/2+3*N) @param[out] info INTEGER - = 0: successful exit. - < 0: if INFO = -i, the i-th argument had an illegal value. - > 0: 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 by Francoise Tisseur, University of Tennessee. @ingroup magma_ssyev_comp ********************************************************************/ extern "C" magma_int_t magma_sstedx( magma_range_t range, magma_int_t n, float vl, float vu, magma_int_t il, magma_int_t iu, float *d, float *e, float *Z, magma_int_t ldz, float *work, magma_int_t lwork, magma_int_t *iwork, magma_int_t liwork, magmaFloat_ptr dwork, magma_int_t *info) { #define Z(i_,j_) (Z + (i_) + (j_)*ldz) float d_zero = 0.; float d_one = 1.; magma_int_t izero = 0; magma_int_t ione = 1; magma_int_t alleig, indeig, valeig, lquery; magma_int_t i, j, k, m; magma_int_t liwmin, lwmin; magma_int_t start, end, smlsiz; float eps, orgnrm, p, tiny; // Test the input parameters. alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1 || liwork == -1); *info = 0; if (! (alleig || valeig || indeig)) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldz < max(1,n)) { *info = -10; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -4; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -5; } else if (iu < min(n,il) || iu > n) { *info = -6; } } } if (*info == 0) { // Compute the workspace requirements smlsiz = magma_get_smlsize_divideconquer(); if ( n <= 1 ) { lwmin = 1; liwmin = 1; } else { lwmin = 1 + 4*n + n*n; liwmin = 3 + 5*n; } work[0] = magma_smake_lwork( lwmin ); iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -12; } else if (liwork < liwmin && ! lquery) { *info = -14; } } 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) { *Z = 1.; return *info; } /* determine the number of threads *///not needed here to be checked Azzam //magma_int_t threads = magma_get_parallel_numthreads(); //magma_int_t mklth = magma_get_lapack_numthreads(); //magma_set_lapack_numthreads(mklth); #ifdef ENABLE_DEBUG //printf(" D&C is using %d threads\n", threads); #endif // If N is smaller than the minimum divide size (SMLSIZ+1), then // solve the problem with another solver. if (n < smlsiz) { lapackf77_ssteqr("I", &n, d, e, Z, &ldz, work, info); } else { lapackf77_slaset("F", &n, &n, &d_zero, &d_one, Z, &ldz); //Scale. orgnrm = lapackf77_slanst("M", &n, d, e); if (orgnrm == 0) { work[0] = magma_smake_lwork( lwmin ); iwork[0] = liwmin; return *info; } eps = lapackf77_slamch( "Epsilon" ); if (alleig) { start = 0; while ( start < n ) { // Let FINISH be the position of the next subdiagonal entry // such that E( END ) <= TINY or FINISH = N if no such // subdiagonal exists. The matrix identified by the elements // between START and END constitutes an independent // sub-problem. for (end = start+1; end < n; ++end) { tiny = eps * sqrt( MAGMA_S_ABS(d[end-1]*d[end])); if (MAGMA_S_ABS(e[end-1]) <= tiny) break; } // (Sub) Problem determined. Compute its size and solve it. m = end - start; if (m == 1) { start = end; continue; } if (m > smlsiz) { // Scale orgnrm = lapackf77_slanst("M", &m, &d[start], &e[start]); lapackf77_slascl("G", &izero, &izero, &orgnrm, &d_one, &m, &ione, &d[start], &m, info); magma_int_t mm = m-1; lapackf77_slascl("G", &izero, &izero, &orgnrm, &d_one, &mm, &ione, &e[start], &mm, info); magma_slaex0( m, &d[start], &e[start], Z(start, start), ldz, work, iwork, dwork, MagmaRangeAll, vl, vu, il, iu, info); if ( *info != 0) { return *info; } // Scale Back lapackf77_slascl("G", &izero, &izero, &d_one, &orgnrm, &m, &ione, &d[start], &m, info); } else { lapackf77_ssteqr( "I", &m, &d[start], &e[start], Z(start, start), &ldz, work, info); if (*info != 0) { *info = (start+1) *(n+1) + end; } } start = end; } // If the problem split any number of times, then the eigenvalues // will not be properly ordered. Here we permute the eigenvalues // (and the associated eigenvectors) into ascending order. if (m < n) { // Use Selection Sort to minimize swaps of eigenvectors for (i = 1; i < n; ++i) { k = i-1; p = d[i-1]; for (j = i; j < n; ++j) { if (d[j] < p) { k = j; p = d[j]; } } if (k != i-1) { d[k] = d[i-1]; d[i-1] = p; blasf77_sswap(&n, Z(0,i-1), &ione, Z(0,k), &ione); } } } } else { // Scale lapackf77_slascl("G", &izero, &izero, &orgnrm, &d_one, &n, &ione, d, &n, info); magma_int_t nm = n-1; lapackf77_slascl("G", &izero, &izero, &orgnrm, &d_one, &nm, &ione, e, &nm, info); magma_slaex0( n, d, e, Z, ldz, work, iwork, dwork, range, vl, vu, il, iu, info); if ( *info != 0) { return *info; } // Scale Back lapackf77_slascl("G", &izero, &izero, &d_one, &orgnrm, &n, &ione, d, &n, info); } } work[0] = magma_smake_lwork( lwmin ); iwork[0] = liwmin; return *info; } /* magma_sstedx */
/** Purpose ------- SGEEV computes for an N-by-N real nonsymmetric matrix A, the eigenvalues and, optionally, the left and/or right eigenvectors. The right eigenvector v(j) of A satisfies A * v(j) = lambda(j) * v(j) where lambda(j) is its eigenvalue. The left eigenvector u(j) of A satisfies u(j)**T * A = lambda(j) * u(j)**T where u(j)**T denotes the transpose of u(j). The computed eigenvectors are normalized to have Euclidean norm equal to 1 and largest component real. Arguments --------- @param[in] jobvl magma_vec_t - = MagmaNoVec: left eigenvectors of A are not computed; - = MagmaVec: left eigenvectors of are computed. @param[in] jobvr magma_vec_t - = MagmaNoVec: right eigenvectors of A are not computed; - = MagmaVec: right eigenvectors of A are computed. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A REAL array, dimension (LDA,N) On entry, the N-by-N matrix A. On exit, A has been overwritten. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] wr REAL array, dimension (N) @param[out] wi REAL array, dimension (N) WR and WI contain the real and imaginary parts, respectively, of the computed eigenvalues. Complex conjugate pairs of eigenvalues appear consecutively with the eigenvalue having the positive imaginary part first. @param[out] VL REAL array, dimension (LDVL,N) If JOBVL = MagmaVec, the left eigenvectors u(j) are stored one after another in the columns of VL, in the same order as their eigenvalues. If JOBVL = MagmaNoVec, VL is not referenced. u(j) = VL(:,j), the j-th column of VL. @param[in] ldvl INTEGER The leading dimension of the array VL. LDVL >= 1; if JOBVL = MagmaVec, LDVL >= N. @param[out] VR REAL array, dimension (LDVR,N) If JOBVR = MagmaVec, the right eigenvectors v(j) are stored one after another in the columns of VR, in the same order as their eigenvalues. If JOBVR = MagmaNoVec, VR is not referenced. v(j) = VR(:,j), the j-th column of VR. @param[in] ldvr INTEGER The leading dimension of the array VR. LDVR >= 1; if JOBVR = MagmaVec, LDVR >= N. @param[out] work (workspace) REAL 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 >= (2 + nb + nb*ngpu)*N. For optimal performance, LWORK >= (2 + 2*nb + nb*ngpu)*N. \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. - > 0: if INFO = i, the QR algorithm failed to compute all the eigenvalues, and no eigenvectors have been computed; elements and i+1:N of W contain eigenvalues which have converged. @ingroup magma_sgeev_driver ********************************************************************/ extern "C" magma_int_t magma_sgeev_m( magma_vec_t jobvl, magma_vec_t jobvr, magma_int_t n, float *A, magma_int_t lda, #ifdef COMPLEX float *w, #else float *wr, float *wi, #endif float *VL, magma_int_t ldvl, float *VR, magma_int_t ldvr, float *work, magma_int_t lwork, #ifdef COMPLEX float *rwork, #endif magma_int_t *info ) { #define VL(i,j) (VL + (i) + (j)*ldvl) #define VR(i,j) (VR + (i) + (j)*ldvr) const magma_int_t ione = 1; const magma_int_t izero = 0; float d__1, d__2; float r, cs, sn, scl; float dum[1], eps; float anrm, cscale, bignum, smlnum; magma_int_t i, k, ilo, ihi; magma_int_t ibal, ierr, itau, iwrk, nout, liwrk, nb; magma_int_t scalea, minwrk, optwrk, lquery, wantvl, wantvr, select[1]; magma_side_t side = MagmaRight; magma_int_t ngpu = magma_num_gpus(); magma_timer_t time_total=0, time_gehrd=0, time_unghr=0, time_hseqr=0, time_trevc=0, time_sum=0; magma_flops_t flop_total=0, flop_gehrd=0, flop_unghr=0, flop_hseqr=0, flop_trevc=0, flop_sum=0; timer_start( time_total ); flops_start( flop_total ); *info = 0; lquery = (lwork == -1); wantvl = (jobvl == MagmaVec); wantvr = (jobvr == MagmaVec); if (! wantvl && jobvl != MagmaNoVec) { *info = -1; } else if (! wantvr && jobvr != MagmaNoVec) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if ( (ldvl < 1) || (wantvl && (ldvl < n))) { *info = -9; } else if ( (ldvr < 1) || (wantvr && (ldvr < n))) { *info = -11; } /* Compute workspace */ nb = magma_get_sgehrd_nb( n ); if (*info == 0) { minwrk = (2 + nb + nb*ngpu)*n; optwrk = (2 + 2*nb + nb*ngpu)*n; work[0] = magma_smake_lwork( optwrk ); if (lwork < minwrk && ! lquery) { *info = -13; } } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } #if defined(Version3) float *dT; if (MAGMA_SUCCESS != magma_smalloc( &dT, nb*n )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } #endif #if defined(Version5) float *T; if (MAGMA_SUCCESS != magma_smalloc_cpu( &T, nb*n )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } #endif /* Get machine constants */ eps = lapackf77_slamch( "P" ); smlnum = lapackf77_slamch( "S" ); bignum = 1. / smlnum; lapackf77_slabad( &smlnum, &bignum ); smlnum = magma_ssqrt( smlnum ) / eps; bignum = 1. / smlnum; /* Scale A if max element outside range [SMLNUM,BIGNUM] */ anrm = lapackf77_slange( "M", &n, &n, A, &lda, dum ); scalea = 0; if (anrm > 0. && anrm < smlnum) { scalea = 1; cscale = smlnum; } else if (anrm > bignum) { scalea = 1; cscale = bignum; } if (scalea) { lapackf77_slascl( "G", &izero, &izero, &anrm, &cscale, &n, &n, A, &lda, &ierr ); } /* Balance the matrix * (Workspace: need N) * - this space is reserved until after gebak */ ibal = 0; lapackf77_sgebal( "B", &n, A, &lda, &ilo, &ihi, &work[ibal], &ierr ); /* Reduce to upper Hessenberg form * (Workspace: need 3*N, prefer 2*N + N*NB + NB*NGPU) * - added NB*NGPU needed for multi-GPU magma_sgehrd_m * - including N reserved for gebal/gebak, unused by sgehrd */ itau = ibal + n; iwrk = itau + n; liwrk = lwork - iwrk; timer_start( time_gehrd ); flops_start( flop_gehrd ); #if defined(Version1) // Version 1 - LAPACK lapackf77_sgehrd( &n, &ilo, &ihi, A, &lda, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version2) // Version 2 - LAPACK consistent HRD magma_sgehrd2( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored, magma_sgehrd( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, dT, &ierr ); #elif defined(Version5) // Version 4 - Multi-GPU, T on host magma_sgehrd_m( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, T, &ierr ); #endif time_sum += timer_stop( time_gehrd ); flop_sum += flops_stop( flop_gehrd ); if (wantvl) { /* Want left eigenvectors * Copy Householder vectors to VL */ side = MagmaLeft; lapackf77_slacpy( MagmaLowerStr, &n, &n, A, &lda, VL, &ldvl ); /* Generate orthogonal matrix in VL * (Workspace: need 3*N-1, prefer 2*N + (N-1)*NB) * - including N reserved for gebal/gebak, unused by sorghr */ timer_start( time_unghr ); flops_start( flop_unghr ); #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_sorghr( &n, &ilo, &ihi, VL, &ldvl, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored magma_sorghr( n, ilo, ihi, VL, ldvl, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_sorghr_m( n, ilo, ihi, VL, ldvl, &work[itau], T, nb, &ierr ); #endif time_sum += timer_stop( time_unghr ); flop_sum += flops_stop( flop_unghr ); timer_start( time_hseqr ); flops_start( flop_hseqr ); /* Perform QR iteration, accumulating Schur vectors in VL * (Workspace: need N+1, prefer N+HSWORK (see comments) ) * - including N reserved for gebal/gebak, unused by shseqr */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_shseqr( "S", "V", &n, &ilo, &ihi, A, &lda, wr, wi, VL, &ldvl, &work[iwrk], &liwrk, info ); time_sum += timer_stop( time_hseqr ); flop_sum += flops_stop( flop_hseqr ); if (wantvr) { /* Want left and right eigenvectors * Copy Schur vectors to VR */ side = MagmaBothSides; lapackf77_slacpy( "F", &n, &n, VL, &ldvl, VR, &ldvr ); } } else if (wantvr) { /* Want right eigenvectors * Copy Householder vectors to VR */ side = MagmaRight; lapackf77_slacpy( "L", &n, &n, A, &lda, VR, &ldvr ); /* Generate orthogonal matrix in VR * (Workspace: need 3*N-1, prefer 2*N + (N-1)*NB) * - including N reserved for gebal/gebak, unused by sorghr */ timer_start( time_unghr ); flops_start( flop_unghr ); #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_sorghr( &n, &ilo, &ihi, VR, &ldvr, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored magma_sorghr( n, ilo, ihi, VR, ldvr, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_sorghr_m( n, ilo, ihi, VR, ldvr, &work[itau], T, nb, &ierr ); #endif time_sum += timer_stop( time_unghr ); flop_sum += flops_stop( flop_unghr ); /* Perform QR iteration, accumulating Schur vectors in VR * (Workspace: need N+1, prefer N+HSWORK (see comments) ) * - including N reserved for gebal/gebak, unused by shseqr */ timer_start( time_hseqr ); flops_start( flop_hseqr ); iwrk = itau; liwrk = lwork - iwrk; lapackf77_shseqr( "S", "V", &n, &ilo, &ihi, A, &lda, wr, wi, VR, &ldvr, &work[iwrk], &liwrk, info ); time_sum += timer_stop( time_hseqr ); flop_sum += flops_stop( flop_hseqr ); } else { /* Compute eigenvalues only * (Workspace: need N+1, prefer N+HSWORK (see comments) ) * - including N reserved for gebal/gebak, unused by shseqr */ timer_start( time_hseqr ); flops_start( flop_hseqr ); iwrk = itau; liwrk = lwork - iwrk; lapackf77_shseqr( "E", "N", &n, &ilo, &ihi, A, &lda, wr, wi, VR, &ldvr, &work[iwrk], &liwrk, info ); time_sum += timer_stop( time_hseqr ); flop_sum += flops_stop( flop_hseqr ); } /* If INFO > 0 from SHSEQR, then quit */ if (*info > 0) { goto CLEANUP; } timer_start( time_trevc ); flops_start( flop_trevc ); if (wantvl || wantvr) { /* Compute left and/or right eigenvectors * (Workspace: need 4*N, prefer (2 + 2*nb)*N) * - including N reserved for gebal/gebak, unused by strevc */ liwrk = lwork - iwrk; #if TREVC_VERSION == 1 lapackf77_strevc( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl, VR, &ldvr, &n, &nout, &work[iwrk], &ierr ); #elif TREVC_VERSION == 2 lapackf77_strevc3( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl, VR, &ldvr, &n, &nout, &work[iwrk], &liwrk, &ierr ); #elif TREVC_VERSION == 3 magma_strevc3( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &ierr ); #elif TREVC_VERSION == 4 magma_strevc3_mt( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &ierr ); #elif TREVC_VERSION == 5 magma_strevc3_mt_gpu( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &ierr ); #else #error Unknown TREVC_VERSION #endif } time_sum += timer_stop( time_trevc ); flop_sum += flops_stop( flop_trevc ); if (wantvl) { /* Undo balancing of left eigenvectors * (Workspace: need N) */ lapackf77_sgebak( "B", "L", &n, &ilo, &ihi, &work[ibal], &n, VL, &ldvl, &ierr ); /* Normalize left eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { if ( wi[i] == 0. ) { scl = 1. / magma_cblas_snrm2( n, VL(0,i), 1 ); blasf77_sscal( &n, &scl, VL(0,i), &ione ); } else if ( wi[i] > 0. ) { d__1 = magma_cblas_snrm2( n, VL(0,i), 1 ); d__2 = magma_cblas_snrm2( n, VL(0,i+1), 1 ); scl = 1. / lapackf77_slapy2( &d__1, &d__2 ); blasf77_sscal( &n, &scl, VL(0,i), &ione ); blasf77_sscal( &n, &scl, VL(0,i+1), &ione ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = *VL(k,i); d__2 = *VL(k,i+1); work[iwrk + k] = d__1*d__1 + d__2*d__2; } k = blasf77_isamax( &n, &work[iwrk], &ione ) - 1; // subtract 1; k is 0-based lapackf77_slartg( VL(k,i), VL(k,i+1), &cs, &sn, &r ); blasf77_srot( &n, VL(0,i), &ione, VL(0,i+1), &ione, &cs, &sn ); *VL(k,i+1) = 0.; } } } if (wantvr) { /* Undo balancing of right eigenvectors * (Workspace: need N) */ lapackf77_sgebak( "B", "R", &n, &ilo, &ihi, &work[ibal], &n, VR, &ldvr, &ierr ); /* Normalize right eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { if ( wi[i] == 0. ) { scl = 1. / magma_cblas_snrm2( n, VR(0,i), 1 ); blasf77_sscal( &n, &scl, VR(0,i), &ione ); } else if ( wi[i] > 0. ) { d__1 = magma_cblas_snrm2( n, VR(0,i), 1 ); d__2 = magma_cblas_snrm2( n, VR(0,i+1), 1 ); scl = 1. / lapackf77_slapy2( &d__1, &d__2 ); blasf77_sscal( &n, &scl, VR(0,i), &ione ); blasf77_sscal( &n, &scl, VR(0,i+1), &ione ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = *VR(k,i); d__2 = *VR(k,i+1); work[iwrk + k] = d__1*d__1 + d__2*d__2; } k = blasf77_isamax( &n, &work[iwrk], &ione ) - 1; // subtract 1; k is 0-based lapackf77_slartg( VR(k,i), VR(k,i+1), &cs, &sn, &r ); blasf77_srot( &n, VR(0,i), &ione, VR(0,i+1), &ione, &cs, &sn ); *VR(k,i+1) = 0.; } } } CLEANUP: /* Undo scaling if necessary */ if (scalea) { // converged eigenvalues, stored in wr[i+1:n] and wi[i+1:n] for i = INFO magma_int_t nval = n - (*info); magma_int_t ld = max( nval, 1 ); lapackf77_slascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, wr + (*info), &ld, &ierr ); lapackf77_slascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, wi + (*info), &ld, &ierr ); if (*info > 0) { // first ilo columns were already upper triangular, // so the corresponding eigenvalues are also valid. nval = ilo - 1; lapackf77_slascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, wr, &n, &ierr ); lapackf77_slascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, wi, &n, &ierr ); } } #if defined(Version3) magma_free( dT ); #endif #if defined(Version5) magma_free_cpu( T ); #endif timer_stop( time_total ); flops_stop( flop_total ); timer_printf( "sgeev times n %5d, gehrd %7.3f, unghr %7.3f, hseqr %7.3f, trevc %7.3f, total %7.3f, sum %7.3f\n", (int) n, time_gehrd, time_unghr, time_hseqr, time_trevc, time_total, time_sum ); timer_printf( "sgeev flops n %5d, gehrd %7lld, unghr %7lld, hseqr %7lld, trevc %7lld, total %7lld, sum %7lld\n", (int) n, flop_gehrd, flop_unghr, flop_hseqr, flop_trevc, flop_total, flop_sum ); work[0] = magma_smake_lwork( optwrk ); return *info; } /* magma_sgeev */
extern "C" magma_int_t magma_sstedx(magma_vec_t range, magma_int_t n, float vl, float vu, magma_int_t il, magma_int_t iu, float* d, float* e, float* z, magma_int_t ldz, float* work, magma_int_t lwork, magma_int_t* iwork, magma_int_t liwork, magmaFloat_ptr dwork, magma_int_t* info, magma_queue_t queue) { /* -- MAGMA (version 1.1.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver @date January 2014 .. Scalar Arguments .. CHARACTER RANGE INTEGER IL, IU, INFO, LDZ, LIWORK, LWORK, N REAL VL, VU .. .. Array Arguments .. INTEGER IWORK( * ) REAL D( * ), E( * ), WORK( * ), Z( LDZ, * ), $ DWORK ( * ) .. Purpose ======= SSTEDX computes some eigenvalues and, optionally, eigenvectors of a symmetric tridiagonal matrix using the divide and conquer method. This code 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. See SLAEX3 for details. Arguments ========= 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. N (input) INTEGER The dimension of the symmetric tridiagonal matrix. N >= 0. 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'. D (input/output) REAL array, dimension (N) On entry, the diagonal elements of the tridiagonal matrix. On exit, if INFO = 0, the eigenvalues in ascending order. E (input/output) REAL array, dimension (N-1) On entry, the subdiagonal elements of the tridiagonal matrix. On exit, E has been destroyed. Z (input/output) REAL array, dimension (LDZ,N) On exit, if INFO = 0, Z contains the orthonormal eigenvectors of the symmetric tridiagonal matrix. LDZ (input) INTEGER The leading dimension of the array Z. LDZ >= max(1,N). WORK (workspace/output) REAL array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. If N > 1 then LWORK must be at least ( 1 + 4*N + N**2 ). Note that if N is less than or equal to the minimum divide size, usually 25, then LWORK need only be max(1,2*(N-1)). 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. 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. LIWORK must be at least ( 3 + 5*N ). Note that if N is less than or equal to the minimum divide size, usually 25, then LIWORK need only be 1. 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. DWORK (device workspace) REAL array, dimension (3*N*N/2+3*N) INFO (output) INTEGER = 0: successful exit. < 0: if INFO = -i, the i-th argument had an illegal value. > 0: 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 by Francoise Tisseur, University of Tennessee. ===================================================================== */ magma_vec_t range_ = range; float d_zero = 0.; float d_one = 1.; magma_int_t izero = 0; magma_int_t ione = 1; magma_int_t alleig, indeig, valeig, lquery; magma_int_t i, j, k, m; magma_int_t liwmin, lwmin; magma_int_t start, end, smlsiz; float eps, orgnrm, p, tiny; // Test the input parameters. alleig = lapackf77_lsame(lapack_const(range_), "A"); valeig = lapackf77_lsame(lapack_const(range_), "V"); indeig = lapackf77_lsame(lapack_const(range_), "I"); lquery = lwork == -1 || liwork == -1; *info = 0; if (! (alleig || valeig || indeig)) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldz < max(1,n)) { *info = -10; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -4; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -5; } else if (iu < min(n,il) || iu > n) { *info = -6; } } } if (*info == 0) { // Compute the workspace requirements smlsiz = get_sstedx_smlsize(); if( n <= 1 ){ lwmin = 1; liwmin = 1; } else { lwmin = 1 + 4*n + n*n; liwmin = 3 + 5*n; } work[0] = lwmin; iwork[0] = liwmin; if (lwork < lwmin && ! lquery) { *info = -12; } else if (liwork < liwmin && ! lquery) { *info = -14; } } if (*info != 0) { magma_xerbla( __func__, -(*info)); return MAGMA_ERR_ILLEGAL_VALUE; } else if (lquery) { return MAGMA_SUCCESS; } // Quick return if possible if(n==0) return MAGMA_SUCCESS; if(n==1){ *z = 1.; return MAGMA_SUCCESS; } // If N is smaller than the minimum divide size (SMLSIZ+1), then // solve the problem with another solver. if (n < smlsiz){ char char_I[]= {'I', 0}; lapackf77_ssteqr(char_I, &n, d, e, z, &ldz, work, info); } else { char char_F[]= {'F', 0}; lapackf77_slaset(char_F, &n, &n, &d_zero, &d_one, z, &ldz); //Scale. char char_M[]= {'M', 0}; orgnrm = lapackf77_slanst(char_M, &n, d, e); if (orgnrm == 0){ work[0] = lwmin; iwork[0] = liwmin; return MAGMA_SUCCESS; } eps = lapackf77_slamch( "Epsilon" ); if (alleig){ start = 0; while ( start < n ){ // Let FINISH be the position of the next subdiagonal entry // such that E( END ) <= TINY or FINISH = N if no such // subdiagonal exists. The matrix identified by the elements // between START and END constitutes an independent // sub-problem. for(end = start+1; end < n; ++end){ tiny = eps * sqrt( MAGMA_S_ABS(d[end-1]*d[end])); if (MAGMA_S_ABS(e[end-1]) <= tiny) break; } // (Sub) Problem determined. Compute its size and solve it. m = end - start; if (m==1){ start = end; continue; } if (m > smlsiz){ // Scale char char_G[] = {'G', 0}; orgnrm = lapackf77_slanst(char_M, &m, &d[start], &e[start]); lapackf77_slascl(char_G, &izero, &izero, &orgnrm, &d_one, &m, &ione, &d[start], &m, info); magma_int_t mm = m-1; lapackf77_slascl(char_G, &izero, &izero, &orgnrm, &d_one, &mm, &ione, &e[start], &mm, info); magma_slaex0( m, &d[start], &e[start], Z(start, start), ldz, work, iwork, dwork, MagmaAllVec, vl, vu, il, iu, info, queue); if( *info != 0) { return MAGMA_SUCCESS; } // Scale Back lapackf77_slascl(char_G, &izero, &izero, &d_one, &orgnrm, &m, &ione, &d[start], &m, info); } else { char char_I[]= {'I', 0}; lapackf77_ssteqr( char_I, &m, &d[start], &e[start], Z(start, start), &ldz, work, info); if (*info != 0){ *info = (start+1) *(n+1) + end; } } start = end; } // If the problem split any number of times, then the eigenvalues // will not be properly ordered. Here we permute the eigenvalues // (and the associated eigenvectors) into ascending order. if (m < n){ // Use Selection Sort to minimize swaps of eigenvectors for (i = 1; i < n; ++i){ k = i-1; p = d[i-1]; for (j = i; j < n; ++j){ if (d[j] < p){ k = j; p = d[j]; } } if(k != i-1) { d[k] = d[i-1]; d[i-1] = p; blasf77_sswap(&n, Z(0,i-1), &ione, Z(0,k), &ione); } } } } else { // Scale char char_G[] = {'G', 0}; lapackf77_slascl(char_G, &izero, &izero, &orgnrm, &d_one, &n, &ione, d, &n, info); magma_int_t nm = n-1; lapackf77_slascl(char_G, &izero, &izero, &orgnrm, &d_one, &nm, &ione, e, &nm, info); magma_slaex0( n, d, e, z, ldz, work, iwork, dwork, range, vl, vu, il, iu, info, queue); if( *info != 0) { return MAGMA_SUCCESS; } // Scale Back lapackf77_slascl(char_G, &izero, &izero, &d_one, &orgnrm, &n, &ione, d, &n, info); } } work[0] = lwmin; iwork[0] = liwmin; return MAGMA_SUCCESS; } /* sstedx */
extern "C" magma_int_t magma_ssyevd_m(magma_int_t nrgpu, char jobz, char uplo, magma_int_t n, float *a, magma_int_t lda, float *w, float *work, magma_int_t lwork, 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 ======= SSYEVD computes all eigenvalues and, optionally, eigenvectors of a real symmetric 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) REAL array, dimension (LDA, N) On entry, the symmetric 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) DOUBLE PRECISION array, dimension (N) If INFO = 0, the eigenvalues in ascending order. WORK (workspace/output) REAL 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 * (NB + 1). If JOBZ = 'V' and N > 1, LWORK >= 2*N + 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 (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: 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}; 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; float rmin, rmax; float sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; float safmin; float bignum; magma_int_t indtau; magma_int_t indwrk, liwmin; magma_int_t llwork; float smlnum; magma_int_t lquery; wantz = lapackf77_lsame(jobz_, MagmaVecStr); lower = lapackf77_lsame(uplo_, MagmaLowerStr); lquery = lwork == -1 || liwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -1; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } magma_int_t nb = magma_get_ssytrd_nb( n ); if ( n <= 1 ) { lwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = 1 + 6*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = 2*n + n*nb; liwmin = 1; } // multiply by 1+eps to ensure length gets rounded up, // if it cannot be exactly represented in floating point. work[0] = lwmin * (1. + lapackf77_slamch("Epsilon")); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -8; } else if ((liwork < liwmin) && ! lquery) { *info = -10; } 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] = a[0]; if (wantz) { a[0] = 1.; } 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_ssyevd(jobz_, uplo_, &n, a, &lda, w, work, &lwork, 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_slansy("M", uplo_, &n, a, &lda, work); 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_slascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a, &lda, info); } /* Call SSYTRD to reduce symmetric matrix to tridiagonal form. */ // ssytrd work: e (n) + tau (n) + llwork (n*nb) ==> 2n + n*nb // sstedx work: e (n) + tau (n) + z (n*n) + llwrk2 (1 + 4*n + n^2) ==> 1 + 6n + 2n^2 inde = 0; indtau = inde + n; 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_ssytrd_mgpu(nrgpu, 1, uplo, n, a, lda, w, &work[inde], &work[indtau], &work[indwrk], llwork, &iinfo); #ifdef ENABLE_TIMER end = get_current_time(); printf("time ssytrd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif /* For eigenvalues only, call SSTERF. For eigenvectors, first call SSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call SORMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_ssterf(&n, w, &work[inde], info); } else { #ifdef ENABLE_TIMER start = get_current_time(); #endif #ifdef USE_SINGLE_GPU if (MAGMA_SUCCESS != magma_smalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_sstedx('A', n, 0., 0., 0, 0, w, &work[inde], &work[indwrk], n, &work[indwk2], llwrk2, iwork, liwork, dwork, info); magma_free( dwork ); #else magma_sstedx_m(nrgpu, 'A', n, 0., 0., 0, 0, w, &work[inde], &work[indwrk], n, &work[indwk2], llwrk2, iwork, liwork, info); #endif #ifdef ENABLE_TIMER end = get_current_time(); printf("time sstedc = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif magma_sormtr_m(nrgpu, MagmaLeft, uplo, MagmaNoTrans, n, n, a, lda, &work[indtau], &work[indwrk], n, &work[indwk2], llwrk2, &iinfo); lapackf77_slacpy("A", &n, &n, &work[indwrk], &n, a, &lda); #ifdef ENABLE_TIMER end = get_current_time(); printf("time sormtr + copy = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { d__1 = 1. / sigma; blasf77_sscal(&n, &d__1, w, &ione); } work[0] = lwmin * (1. + lapackf77_slamch("Epsilon")); // round up iwork[0] = liwmin; return *info; } /* magma_ssyevd_m */