/* //////////////////////////////////////////////////////////////////////////// -- Testing cgehrd */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; magmaFloatComplex *h_A, *h_R, *h_Q, *h_work, *tau, *twork; magmaFloatComplex_ptr dT; #if defined(PRECISION_z) || defined(PRECISION_c) float *rwork; #endif float eps, result[2]; magma_int_t N, n2, lda, nb, lwork, ltwork, info; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; eps = lapackf77_slamch( "E" ); magma_opts opts; parse_opts( argc, argv, &opts ); float tol = opts.tolerance * lapackf77_slamch("E"); printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) |A-QHQ'|/N|A| |I-QQ'|/N\n"); printf("=========================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; nb = magma_get_cgehrd_nb(N); /* We suppose the magma nb is bigger than lapack nb */ lwork = N*nb; gflops = FLOPS_CGEHRD( N ) / 1e9; TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( tau, magmaFloatComplex, N ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork ); TESTING_MALLOC_DEV( dT, magmaFloatComplex, nb*N ); /* Initialize the matrices */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_cgehrd( N, ione, N, h_R, lda, tau, h_work, lwork, dT, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_cgehrd returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the factorization =================================================================== */ if ( opts.check ) { ltwork = 2*(N*N); TESTING_MALLOC_PIN( h_Q, magmaFloatComplex, lda*N ); TESTING_MALLOC_CPU( twork, magmaFloatComplex, ltwork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_MALLOC_CPU( rwork, float, N ); #endif lapackf77_clacpy(MagmaUpperLowerStr, &N, &N, h_R, &lda, h_Q, &lda); for( int j = 0; j < N-1; ++j ) for( int i = j+2; i < N; ++i ) h_R[i+j*lda] = MAGMA_C_ZERO; magma_cunghr(N, ione, N, h_Q, lda, tau, dT, nb, &info); if (info != 0) { printf("magma_cunghr returned error %d: %s.\n", (int) info, magma_strerror( info )); exit(1); } #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_chst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, rwork, result); #else lapackf77_chst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, result); #endif TESTING_FREE_PIN( h_Q ); TESTING_FREE_CPU( twork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_FREE_CPU( rwork ); #endif } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_cgehrd(&N, &ione, &N, h_R, &lda, tau, h_work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_cgehrd returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Print performance and error. =================================================================== */ if ( opts.lapack ) { printf("%5d %7.2f (%7.2f) %7.2f (%7.2f)", (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time ); } else { printf("%5d --- ( --- ) %7.2f (%7.2f)", (int) N, gpu_perf, gpu_time ); } if ( opts.check ) { printf(" %8.2e %8.2e %s\n", result[0]*eps, result[1]*eps, ( ( (result[0]*eps < tol) && (result[1]*eps < tol) ) ? "ok" : "failed") ); status += ! (result[0]*eps < tol); status += ! (result[1]*eps < tol); } else { printf(" --- ---\n"); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( tau ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_work ); TESTING_FREE_DEV( dT ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_cgehrd(magma_int_t n, magma_int_t ilo, magma_int_t ihi, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex *dT, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= CGEHRD reduces a COMPLEX general matrix A to upper Hessenberg form H by an orthogonal similarity transformation: Q' * A * Q = H . This version stores the triangular matrices used in the factorization so that they can be applied directly (i.e., without being recomputed) later. As a result, the application of Q is much faster. Arguments ========= N (input) INTEGER The order of the matrix A. N >= 0. ILO (input) INTEGER IHI (input) INTEGER It is assumed that A is already upper triangular in rows and columns 1:ILO-1 and IHI+1:N. ILO and IHI are normally set by a previous call to CGEBAL; otherwise they should be set to 1 and N respectively. See Further Details. 1 <= ILO <= IHI <= N, if N > 0; ILO=1 and IHI=0, if N=0. A (input/output) COMPLEX array, dimension (LDA,N) On entry, the N-by-N general matrix to be reduced. On exit, the upper triangle and the first subdiagonal of A are overwritten with the upper Hessenberg matrix H, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). TAU (output) COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). Elements 1:ILO-1 and IHI:N-1 of TAU are set to zero. WORK (workspace/output) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= max(1,N). For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. dT (output) COMPLEX array on the GPU, dimension NB*N, where NB is the optimal blocksize. It stores the NB*NB blocks of the triangular T matrices used in the reduction. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value. Further Details =============== The matrix Q is represented as a product of (ihi-ilo) elementary reflectors Q = H(ilo) H(ilo+1) . . . H(ihi-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0, v(i+1) = 1 and v(ihi+1:n) = 0; v(i+2:ihi) is stored on exit in A(i+2:ihi,i), and tau in TAU(i). The contents of A are illustrated by the following example, with n = 7, ilo = 2 and ihi = 6: on entry, on exit, ( a a a a a a a ) ( a a h h h h a ) ( a a a a a a ) ( a h h h h a ) ( a a a a a a ) ( h h h h h h ) ( a a a a a a ) ( v2 h h h h h ) ( a a a a a a ) ( v2 v3 h h h h ) ( a a a a a a ) ( v2 v3 v4 h h h ) ( a ) ( a ) where a denotes an element of the original matrix A, h denotes a modified element of the upper Hessenberg matrix H, and vi denotes an element of the vector defining H(i). This implementation follows the hybrid algorithm and notations described in S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg form through hybrid GPU-based computing," University of Tennessee Computer Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219), May 24, 2009. This version stores the T matrices in dT, for later use in magma_cunghr. ===================================================================== */ #define A( i, j ) ( A + (i) + (j)*lda) #define dA( i, j ) (dA + (i) + (j-ilo)*ldda) magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; magma_int_t nb = magma_get_cgehrd_nb(n); magma_int_t ldda = n; // assumed in clahru magma_int_t nh, iws; magma_int_t iinfo; magma_int_t ldwork; magma_int_t lquery; *info = 0; iws = n*nb; MAGMA_C_SET2REAL( work[0], (float) iws ); lquery = lwork == -1; if (n < 0) { *info = -1; } else if (ilo < 1 || ilo > max(1,n)) { *info = -2; } else if (ihi < min(ilo,n) || ihi > n) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if (lwork < max(1,n) && ! lquery) { *info = -8; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; // Adjust from 1-based indexing ilo -= 1; // Quick return if possible nh = ihi - ilo; if (nh <= 1) { work[0] = c_one; return *info; } // GPU workspace is: // nb*ldda for dwork for clahru // nb*ldda for dV // n*ldda for dA magmaFloatComplex *dwork; if (MAGMA_SUCCESS != magma_cmalloc( &dwork, 2*nb*ldda + n*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magmaFloatComplex *dV = dwork + nb*ldda; magmaFloatComplex *dA = dwork + nb*ldda*2; ldwork = n; magma_int_t i; magmaFloatComplex *T, *dTi; magma_cmalloc_cpu( &T, nb*nb ); if ( T == NULL ) { magma_free( dwork ); *info = MAGMA_ERR_HOST_ALLOC; return *info; } // zero first block of V, which is lower triangular czero_nbxnb_block(nb, dV, ldda); // Set elements 0:ILO-1 and IHI-1:N-2 of TAU to zero for(i = 0; i < ilo; ++i) tau[i] = c_zero; for(i = max(0,ihi-1); i < n-1; ++i) tau[i] = c_zero; for(i=0; i < nb*nb; i += 4) T[i] = T[i+1] = T[i+2] = T[i+3] = c_zero; magmablas_claset( 'F', nb, n, dT, nb ); // If not enough workspace, use unblocked code if ( lwork < iws ) { nb = 1; } if (nb == 1 || nb > nh) { // Use unblocked code below i = ilo; } else { // Use blocked code // Copy the matrix to the GPU magma_csetmatrix( n, n-ilo, A(0,ilo), lda, dA, ldda ); for (i = ilo; i < ihi-1 - nb; i += nb) { // Reduce columns i:i+nb-1 to Hessenberg form, returning the // matrices V and T of the block reflector H = I - V*T*V' // which performs the reduction, and also the matrix Y = A*V*T // Get the current panel (no need for the 1st iteration) magma_cgetmatrix( ihi-i, nb, dA(i,i), ldda, A (i,i), lda ); // add 1 to i for 1-based index magma_clahr2( ihi, i+1, nb, dA(0,i), dV, A (0,i), lda, &tau[i], T, nb, work, ldwork); // Copy T from the CPU to dT on the GPU dTi = dT + (i - ilo)*nb; magma_csetmatrix( nb, nb, T, nb, dTi, nb ); magma_clahru( n, ihi, i, nb, A (0,i), lda, dA(0,i), // dA dA(i,i), // dY, stored over current panel dV, dTi, dwork ); } // Copy remainder to host magma_cgetmatrix( n, n-i, dA(0,i), ldda, A (0,i), lda ); } // Use unblocked code to reduce the rest of the matrix // add 1 to i for 1-based index i += 1; lapackf77_cgehd2(&n, &i, &ihi, A, &lda, tau, work, &iinfo); MAGMA_C_SET2REAL( work[0], (float) iws ); magma_free( dwork ); magma_free_cpu( T ); return *info; } /* magma_cgehrd */
/** Purpose ------- CGEHRD reduces a COMPLEX general matrix A to upper Hessenberg form H by an orthogonal similarity transformation: Q' * A * Q = H . This version stores the triangular matrices used in the factorization so that they can be applied directly (i.e., without being recomputed) later. As a result, the application of Q is much faster. Arguments --------- @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in] ilo INTEGER @param[in] ihi INTEGER It is assumed that A is already upper triangular in rows and columns 1:ILO-1 and IHI+1:N. ILO and IHI are normally set by a previous call to CGEBAL; otherwise they should be set to 1 and N respectively. See Further Details. 1 <= ILO <= IHI <= N, if N > 0; ILO=1 and IHI=0, if N=0. @param[in,out] A COMPLEX array, dimension (LDA,N) On entry, the N-by-N general matrix to be reduced. On exit, the upper triangle and the first subdiagonal of A are overwritten with the upper Hessenberg matrix H, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). Elements 1:ILO-1 and IHI:N-1 of TAU are set to zero. @param[out] work (workspace) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. LWORK >= max(1,N). For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] dT COMPLEX array on the GPU, dimension NB*N, where NB is the optimal blocksize. It stores the NB*NB blocks of the triangular T matrices used in the reduction. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value. Further Details --------------- The matrix Q is represented as a product of (ihi-ilo) elementary reflectors Q = H(ilo) H(ilo+1) . . . H(ihi-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0, v(i+1) = 1 and v(ihi+1:n) = 0; v(i+2:ihi) is stored on exit in A(i+2:ihi,i), and tau in TAU(i). The contents of A are illustrated by the following example, with n = 7, ilo = 2 and ihi = 6: @verbatim on entry, on exit, ( a a a a a a a ) ( a a h h h h a ) ( a a a a a a ) ( a h h h h a ) ( a a a a a a ) ( h h h h h h ) ( a a a a a a ) ( v2 h h h h h ) ( a a a a a a ) ( v2 v3 h h h h ) ( a a a a a a ) ( v2 v3 v4 h h h ) ( a ) ( a ) @endverbatim where a denotes an element of the original matrix A, h denotes a modified element of the upper Hessenberg matrix H, and vi denotes an element of the vector defining H(i). This implementation follows the hybrid algorithm and notations described in S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg form through hybrid GPU-based computing," University of Tennessee Computer Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219), May 24, 2009. This version stores the T matrices in dT, for later use in magma_cunghr. @ingroup magma_cgeev_comp ********************************************************************/ extern "C" magma_int_t magma_cgehrd( magma_int_t n, magma_int_t ilo, magma_int_t ihi, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex_ptr dT, magma_int_t *info) { #define A(i_,j_) ( A + (i_) + (j_)*lda) #define dA(i_,j_) (dA + (i_) + (j_)*ldda) magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; magma_int_t nb = magma_get_cgehrd_nb(n); magma_int_t ldda = ((n+31)/32)*32; magma_int_t i, nh, iws; magma_int_t iinfo; magma_int_t lquery; *info = 0; iws = n*nb; work[0] = MAGMA_C_MAKE( iws, 0 ); lquery = (lwork == -1); if (n < 0) { *info = -1; } else if (ilo < 1 || ilo > max(1,n)) { *info = -2; } else if (ihi < min(ilo,n) || ihi > n) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if (lwork < max(1,n) && ! lquery) { *info = -8; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; // Adjust from 1-based indexing ilo -= 1; // Quick return if possible nh = ihi - ilo; if (nh <= 1) { work[0] = c_one; return *info; } // If not enough workspace, use unblocked code if ( lwork < iws ) { nb = 1; } if (nb == 1 || nb > nh) { // Use unblocked code below i = ilo; } else { // Use blocked code // GPU workspace is: // nb*ldda for dwork for clahru // nb*ldda for dV // n*ldda for dA magmaFloatComplex *dwork; if (MAGMA_SUCCESS != magma_cmalloc( &dwork, 2*nb*ldda + n*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magmaFloatComplex *dV = dwork + nb*ldda; magmaFloatComplex *dA = dwork + nb*ldda*2; magmaFloatComplex *dTi; magmaFloatComplex *T; magma_cmalloc_cpu( &T, nb*nb ); if ( T == NULL ) { magma_free( dwork ); *info = MAGMA_ERR_HOST_ALLOC; return *info; } // zero first block of V, which is lower triangular magmablas_claset( MagmaFull, nb, nb, c_zero, c_zero, dV, ldda ); // Set elements 0:ILO-1 and IHI-1:N-2 of TAU to zero for (i = 0; i < ilo; ++i) tau[i] = c_zero; for (i = max(0,ihi-1); i < n-1; ++i) tau[i] = c_zero; assert( nb % 4 == 0 ); for (i=0; i < nb*nb; i += 4) T[i] = T[i+1] = T[i+2] = T[i+3] = c_zero; magmablas_claset( MagmaFull, nb, n, c_zero, c_zero, dT, nb ); // Copy the matrix to the GPU magma_csetmatrix( n, n-ilo, A(0,ilo), lda, dA, ldda ); for (i = ilo; i < ihi-1 - nb; i += nb) { // Reduce columns i:i+nb-1 to Hessenberg form, returning the // matrices V and T of the block reflector H = I - V*T*V' // which performs the reduction, and also the matrix Y = A*V*T // Get the current panel (no need for the 1st iteration) magma_cgetmatrix( ihi-i, nb, dA(i,i-ilo), ldda, A(i,i), lda ); // add 1 to i for 1-based index magma_clahr2( ihi, i+1, nb, dA(0,i-ilo), ldda, dV, ldda, A(0,i), lda, &tau[i], T, nb, work, n); // Copy T from the CPU to dT on the GPU dTi = dT + (i - ilo)*nb; magma_csetmatrix( nb, nb, T, nb, dTi, nb ); magma_clahru( n, ihi, i, nb, A(0,i), lda, dA(0,i-ilo), ldda, // dA dA(i,i-ilo), ldda, // dY, stored over current panel dV, ldda, dTi, dwork ); } // Copy remainder to host magma_cgetmatrix( n, n-i, dA(0,i-ilo), ldda, A(0,i), lda ); magma_free( dwork ); magma_free_cpu( T ); } // Use unblocked code to reduce the rest of the matrix // add 1 to i for 1-based index i += 1; lapackf77_cgehd2(&n, &i, &ihi, A, &lda, tau, work, &iinfo); work[0] = MAGMA_C_MAKE( iws, 0 ); return *info; } /* magma_cgehrd */
/***************************************************************************//** Purpose ------- CGEEV computes for an N-by-N complex 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)**H * A = lambda(j) * u(j)**H where u(j)**H denotes the conjugate 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 COMPLEX 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] w COMPLEX array, dimension (N) W contains the computed eigenvalues. @param[out] VL COMPLEX 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 COMPLEX 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) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= (1 + nb + nb*ngpu)*N. For optimal performance, LWORK >= (1 + 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 rwork (workspace) REAL array, dimension (2*N) @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_geev *******************************************************************************/ extern "C" magma_int_t magma_cgeev_m( magma_vec_t jobvl, magma_vec_t jobvr, magma_int_t n, magmaFloatComplex *A, magma_int_t lda, #ifdef COMPLEX magmaFloatComplex *w, #else float *wr, float *wi, #endif magmaFloatComplex *VL, magma_int_t ldvl, magmaFloatComplex *VR, magma_int_t ldvr, magmaFloatComplex *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; magmaFloatComplex tmp; float 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, irwork, lquery, wantvl, wantvr, select[1]; magma_side_t side = MagmaRight; magma_int_t ngpu = magma_num_gpus(); irwork = 0; *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 = -8; } else if ( (ldvr < 1) || (wantvr && (ldvr < n))) { *info = -10; } /* Compute workspace */ nb = magma_get_cgehrd_nb( n ); if (*info == 0) { minwrk = (1 + nb + nb*ngpu)*n; optwrk = (1 + 2*nb + nb*ngpu)*n; work[0] = magma_cmake_lwork( optwrk ); if (lwork < minwrk && ! lquery) { *info = -12; } } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } #if defined(Version3) magmaFloatComplex *dT; if (MAGMA_SUCCESS != magma_cmalloc( &dT, nb*n )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } #endif #if defined(Version5) magmaFloatComplex *T; if (MAGMA_SUCCESS != magma_cmalloc_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_clange( "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_clascl( "G", &izero, &izero, &anrm, &cscale, &n, &n, A, &lda, &ierr ); } /* Balance the matrix * (CWorkspace: none) * (RWorkspace: need N) * - this space is reserved until after gebak */ ibal = 0; lapackf77_cgebal( "B", &n, A, &lda, &ilo, &ihi, &rwork[ibal], &ierr ); /* Reduce to upper Hessenberg form * (CWorkspace: need 2*N, prefer N + N*NB + NB*NGPU) * (RWorkspace: N) * - added NB*NGPU needed for multi-GPU magma_cgehrd_m * - including N reserved for gebal/gebak, unused by cgehrd */ itau = 0; iwrk = itau + n; liwrk = lwork - iwrk; #if defined(Version1) // Version 1 - LAPACK lapackf77_cgehrd( &n, &ilo, &ihi, A, &lda, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version2) // Version 2 - LAPACK consistent HRD magma_cgehrd2( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored, magma_cgehrd( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, dT, &ierr ); #elif defined(Version5) // Version 4 - Multi-GPU, T on host magma_cgehrd_m( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, T, &ierr ); #endif if (wantvl) { /* Want left eigenvectors * Copy Householder vectors to VL */ side = MagmaLeft; lapackf77_clacpy( MagmaLowerStr, &n, &n, A, &lda, VL, &ldvl ); /* Generate unitary matrix in VL * (CWorkspace: need 2*N-1, prefer N + (N-1)*NB) * (RWorkspace: N) * - including N reserved for gebal/gebak, unused by cunghr */ #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_cunghr( &n, &ilo, &ihi, VL, &ldvl, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored magma_cunghr( n, ilo, ihi, VL, ldvl, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_cunghr_m( n, ilo, ihi, VL, ldvl, &work[itau], T, nb, &ierr ); #endif /* Perform QR iteration, accumulating Schur vectors in VL * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: N) * - including N reserved for gebal/gebak, unused by chseqr */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_chseqr( "S", "V", &n, &ilo, &ihi, A, &lda, w, VL, &ldvl, &work[iwrk], &liwrk, info ); if (wantvr) { /* Want left and right eigenvectors * Copy Schur vectors to VR */ side = MagmaBothSides; lapackf77_clacpy( "F", &n, &n, VL, &ldvl, VR, &ldvr ); } } else if (wantvr) { /* Want right eigenvectors * Copy Householder vectors to VR */ side = MagmaRight; lapackf77_clacpy( "L", &n, &n, A, &lda, VR, &ldvr ); /* Generate unitary matrix in VR * (CWorkspace: need 2*N-1, prefer N + (N-1)*NB) * (RWorkspace: N) * - including N reserved for gebal/gebak, unused by cunghr */ #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_cunghr( &n, &ilo, &ihi, VR, &ldvr, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + T matrices stored magma_cunghr( n, ilo, ihi, VR, ldvr, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_cunghr_m( n, ilo, ihi, VR, ldvr, &work[itau], T, nb, &ierr ); #endif /* Perform QR iteration, accumulating Schur vectors in VR * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: N) * - including N reserved for gebal/gebak, unused by chseqr */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_chseqr( "S", "V", &n, &ilo, &ihi, A, &lda, w, VR, &ldvr, &work[iwrk], &liwrk, info ); } else { /* Compute eigenvalues only * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: N) * - including N reserved for gebal/gebak, unused by chseqr */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_chseqr( "E", "N", &n, &ilo, &ihi, A, &lda, w, VR, &ldvr, &work[iwrk], &liwrk, info ); } /* If INFO > 0 from CHSEQR, then quit */ if (*info > 0) { goto CLEANUP; } if (wantvl || wantvr) { /* Compute left and/or right eigenvectors * (CWorkspace: need 2*N) * (RWorkspace: need 2*N) * - including N reserved for gebal/gebak, unused by ctrevc */ irwork = ibal + n; #if TREVC_VERSION == 1 lapackf77_ctrevc( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl, VR, &ldvr, &n, &nout, &work[iwrk], &rwork[irwork], &ierr ); #elif TREVC_VERSION == 2 liwrk = lwork - iwrk; lapackf77_ctrevc3( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl, VR, &ldvr, &n, &nout, &work[iwrk], &liwrk, &rwork[irwork], &ierr ); #elif TREVC_VERSION == 3 magma_ctrevc3( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr ); #elif TREVC_VERSION == 4 magma_ctrevc3_mt( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr ); #elif TREVC_VERSION == 5 magma_ctrevc3_mt_gpu( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl, VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr ); #else #error Unknown TREVC_VERSION #endif } if (wantvl) { /* Undo balancing of left eigenvectors * (CWorkspace: none) * (RWorkspace: need N) */ lapackf77_cgebak( "B", "L", &n, &ilo, &ihi, &rwork[ibal], &n, VL, &ldvl, &ierr ); /* Normalize left eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { scl = 1. / magma_cblas_scnrm2( n, VL(0,i), 1 ); blasf77_csscal( &n, &scl, VL(0,i), &ione ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = MAGMA_C_REAL( *VL(k,i) ); d__2 = MAGMA_C_IMAG( *VL(k,i) ); rwork[irwork + k] = d__1*d__1 + d__2*d__2; } k = blasf77_isamax( &n, &rwork[irwork], &ione ) - 1; // subtract 1; k is 0-based tmp = MAGMA_C_CONJ( *VL(k,i) ) / magma_ssqrt( rwork[irwork + k] ); blasf77_cscal( &n, &tmp, VL(0,i), &ione ); *VL(k,i) = MAGMA_C_MAKE( MAGMA_C_REAL( *VL(k,i) ), 0 ); } } if (wantvr) { /* Undo balancing of right eigenvectors * (CWorkspace: none) * (RWorkspace: need N) */ lapackf77_cgebak( "B", "R", &n, &ilo, &ihi, &rwork[ibal], &n, VR, &ldvr, &ierr ); /* Normalize right eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { scl = 1. / magma_cblas_scnrm2( n, VR(0,i), 1 ); blasf77_csscal( &n, &scl, VR(0,i), &ione ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = MAGMA_C_REAL( *VR(k,i) ); d__2 = MAGMA_C_IMAG( *VR(k,i) ); rwork[irwork + k] = d__1*d__1 + d__2*d__2; } k = blasf77_isamax( &n, &rwork[irwork], &ione ) - 1; // subtract 1; k is 0-based tmp = MAGMA_C_CONJ( *VR(k,i) ) / magma_ssqrt( rwork[irwork + k] ); blasf77_cscal( &n, &tmp, VR(0,i), &ione ); *VR(k,i) = MAGMA_C_MAKE( MAGMA_C_REAL( *VR(k,i) ), 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_clascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, w + (*info), &ld, &ierr ); if (*info > 0) { // first ilo columns were already upper triangular, // so the corresponding eigenvalues are also valid. nval = ilo - 1; lapackf77_clascl( "G", &izero, &izero, &cscale, &anrm, &nval, &ione, w, &n, &ierr ); } } #if defined(Version3) magma_free( dT ); #endif #if defined(Version5) magma_free_cpu( T ); #endif work[0] = magma_cmake_lwork( minwrk ); // TODO use optwrk as in dgeev return *info; } /* magma_cgeev */
extern "C" int calc_bounding_box(magmaFloatComplex *M, magma_int_t M_lead_dim, float *wReEig, float *wImEig) { magma_int_t rslt = 0; //magmaFloatComplex *AT = nullptr; magmaFloatComplex *dA = nullptr, *dAT = nullptr, *dreA = nullptr, *dimA = nullptr; float *dreEig = nullptr; float *dimEig = nullptr; //magma_int_t *ipiv = NULL; magma_int_t lda = M_lead_dim; //magma_int_t ldx = lda; magma_int_t info = 0; magma_int_t nb = 0; //magma_vec_t jobvl; //magma_vec_t jobvr; magmaFloatComplex *work = nullptr; magma_int_t lwork = 0; float *rwork = nullptr; magma_int_t lrwork = 0; magma_int_t *iwork = nullptr; magma_int_t liwork = 0; nb = magma_get_cgehrd_nb( M_lead_dim ); lwork = 2 * (M_lead_dim + M_lead_dim*nb); // MagmaNoVec //lwork = 2 * max(M_lead_dim + M_lead_dim*nb, 2*M_lead_dim + M_lead_dim*M_lead_dim); // MagmaVec lrwork = M_lead_dim; // MagmaNoVec //lrwork = 1 + 5 * M_lead_dim + 2*M_lead_dim*M_lead_dim; // MagmaVec liwork = 1; // MagmaNoVec //liwork = 3 + 5*M_lead_dim; // MagmaVec magma_imalloc_cpu(&iwork, liwork); magma_smalloc_cpu(&rwork, lrwork); //magma_cmalloc_cpu(&A, lda*M_lead_dim); //magma_cmalloc_cpu(&AT, lda*M_lead_dim); //magma_smalloc_cpu(&reEig, M_lead_dim); //magma_smalloc_cpu(&imEig, M_lead_dim); magma_cmalloc_pinned(&dA, lda*M_lead_dim); magma_cmalloc_pinned(&dAT, lda*M_lead_dim); magma_cmalloc_pinned(&dreA, lda*M_lead_dim); magma_cmalloc_pinned(&dimA, lda*M_lead_dim); //magma_cmalloc_pinned(&VL, lda*M_lead_dim); //magma_cmalloc_pinned(&VR, lda*M_lead_dim); magma_cmalloc_pinned(&work, lwork); magma_smalloc_pinned(&dreEig, M_lead_dim); magma_smalloc_pinned(&dimEig, M_lead_dim); //matrix_fillzero(AT, M_lead_dim); //vector_fillzero(reEig, M_lead_dim); //vector_fillzero(imEig, M_lead_dim); //prepare_matrix_2(M); magma_csetmatrix(M_lead_dim, M_lead_dim, M, lda, dA, M_lead_dim, queue); //magma_csetmatrix(M_lead_dim, M_lead_dim, AT, lda, dAT, M_lead_dim, queue); //magma_ssetvector(M_lead_dim, wReEig, 1, dreEig, 1, queue); //magma_ssetvector(M_lead_dim, wImEig, 1, dimEig, 1, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dA, lda); // reA = ( (A + A')/2.0 ) // A' magmablas_ctranspose(M_lead_dim, M_lead_dim, dA, M_lead_dim, dAT, M_lead_dim, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAT, lda); // AT = A + A' magmablas_cgeadd(M_lead_dim, M_lead_dim, MAGMA_C_MAKE(1.0f, 0.0f), dA, M_lead_dim, dAT, M_lead_dim, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAT, lda); // AT=AT*0.5 magma_cscal(lda*M_lead_dim, MAGMA_C_MAKE(0.5f, 0.0f), dAT, 1, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAT, lda); // reA = AT magma_ccopy(lda*M_lead_dim, dAT, 1, dreA, 1, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dreA, lda); magma_sync_wtime(queue); // imA = ( -1im*(A - A')/2.0 ) // A' magmablas_ctranspose(M_lead_dim, M_lead_dim, dA, M_lead_dim, dAT, M_lead_dim, queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAT, lda); // AT = A + A' magmablas_cgeadd(M_lead_dim, M_lead_dim, MAGMA_C_MAKE(-1.0f, 0.0f), dAT, M_lead_dim, dA, M_lead_dim, queue); // A=A*-1j*0.5 magma_cscal(lda*M_lead_dim, MAGMA_C_MAKE(0.0f, -0.5f), dA, 1, queue); // imA = A magma_ccopy(lda*M_lead_dim, dA, 1, dimA, 1, queue); magma_sync_wtime(queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dreA, lda); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dimA, lda); // reEig::Vector=eigvals(reA) rslt = magma_cheevd(MagmaNoVec, MagmaLower, M_lead_dim, dreA, lda, dreEig, work, lwork, rwork, lrwork, iwork, liwork, &info); // imEig::Vector=eigvals(imA) rslt = magma_cheevd(MagmaNoVec, MagmaLower, M_lead_dim, dimA, lda, dimEig, work, lwork, rwork, lrwork, iwork, liwork, &info); //magma_sprint_gpu(M_lead_dim, 1, dreEig, M_lead_dim); //magma_sprint_gpu(M_lead_dim, 1, dimEig, M_lead_dim); magma_sgetvector(M_lead_dim, dreEig, 1, wReEig, 1, queue); //magma_sync_wtime(queue); magma_sgetvector(M_lead_dim, dimEig, 1, wImEig, 1, queue); //magma_sync_wtime(queue); /* maxReIdx = magma_isamax(M_lead_dim, dreEig, 1, queue) - 1; minReIdx = magma_isamin(M_lead_dim, dreEig, 1, queue) - 1; maxImIdx = magma_isamax(M_lead_dim, dimEig, 1, queue) - 1; minImIdx = magma_isamin(M_lead_dim, dimEig, 1, queue) - 1; printf("max re idx = %d\nmin re idx = %d\n", maxReIdx, minReIdx); printf("%f %f\n", wReEig[maxReIdx], wReEig[minReIdx]); printf("max im idx = %d\nmin im idx = %d\n", maxImIdx, minImIdx); printf("%f %f\n", wImEig[maxImIdx], wImEig[minImIdx]); */ //printf("test wReEig: %f %f\n", wReEig[0], wReEig[1]); //printf("test wImEig: %f %f\n", wImEig[0], wImEig[1]); magma_free_cpu(iwork); magma_free_cpu(rwork); //magma_free_cpu(AT); magma_free_pinned(dA); magma_free_pinned(dAT); magma_free_pinned(dreA); magma_free_pinned(dimA); magma_free_pinned(work); magma_free_pinned(dreEig); magma_free_pinned(dimEig); return rslt; }
extern "C" int calc_numerical_range(magmaFloatComplex *M, magma_int_t M_lead_dim, float _from, float _step, magma_int_t _steps, magmaFloatComplex *pts) { magma_int_t idx = 0, rslt = 0; magmaFloatComplex p, scalar; std::complex<float> vtmp; float j; magmaFloatComplex *dA = nullptr; magmaFloatComplex *dAth = NULL, *dAthT = NULL, *dX = NULL, *dY = NULL; float *dE = NULL; //float *hE = NULL; //magma_int_t *ipiv = NULL; magma_int_t lda = M_lead_dim; //magma_int_t ldx = lda; magma_int_t info = 0; magma_int_t nb = 0; //magma_vec_t jobvl; //magma_vec_t jobvr; magmaFloatComplex *work = nullptr; magma_int_t lwork = 0; float *rwork = nullptr; magma_int_t lrwork = 0; magma_int_t *iwork = nullptr; magma_int_t liwork = 0; nb = magma_get_cgehrd_nb( M_lead_dim ); lwork = 2 * max(M_lead_dim + M_lead_dim*nb, 2 * M_lead_dim + M_lead_dim*M_lead_dim); // MagmaVec lrwork = 1 + 5 * M_lead_dim + 2 * M_lead_dim*M_lead_dim; // MagmaVec liwork = (3 + 5 * M_lead_dim); // MagmaVec magma_imalloc_cpu(&iwork, liwork); magma_smalloc_cpu(&rwork, lrwork); magma_cmalloc_pinned(&work, lwork); magma_cmalloc_pinned(&dA, lda*M_lead_dim); magma_cmalloc_pinned(&dAth, lda*M_lead_dim); magma_cmalloc_pinned(&dAthT, lda*M_lead_dim); magma_smalloc_pinned(&dE, M_lead_dim); //magma_smalloc_cpu(&hE, M_lead_dim); magma_cmalloc_pinned(&dX, M_lead_dim); magma_cmalloc_pinned(&dY, M_lead_dim); magma_csetmatrix(M_lead_dim, M_lead_dim, M, lda, dA, M_lead_dim, queue); // th=[0:resolution:2*pi] j = _from; for (idx = 0; idx < _steps; idx++) { //scalar = exp( 1im * -j); vtmp.real( 0.0f ); vtmp.imag( -j ); //vtmp = _FCbuild(0.0f, -j); //printf("vtmp = %f + i%f\n", vtmp._Val[0], vtmp._Val[1]); vtmp = exp(vtmp); scalar.x = vtmp.real(); scalar.y = vtmp.imag(); //printf("scalar = %f + i%f\n", scalar.x, scalar.y); magma_ccopy(lda * M_lead_dim, dA, 1, dAth, 1, queue); // Ath = exp(1im * -j) * As magma_cscal(lda * M_lead_dim, scalar, dAth, 1, queue); //magma_cprint_gpu(N, N, dA, lda); //magma_cprint_gpu(N, N, dAth, lda); // AthT = (Ath + Ath') magmablas_ctranspose_conj(M_lead_dim, M_lead_dim, dAth, M_lead_dim, dAthT, M_lead_dim, queue); magmablas_cgeadd(M_lead_dim, M_lead_dim, MAGMA_C_MAKE(1.0f, 0.0f), dAth, M_lead_dim, dAthT, M_lead_dim, queue); // AthT = AthT / 2 magma_cscal(lda*M_lead_dim, MAGMA_C_MAKE(0.5f, 0.0f), dAthT, 1, queue); magma_sync_wtime(queue); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAthT, lda); // e, r = eig(AthT) rslt = magma_cheevd(MagmaVec, MagmaLower, M_lead_dim, dAthT, lda, dE, work, lwork, rwork, lrwork, iwork, liwork, &info); magma_sync_wtime(queue); //printf("magma_cheevd info=%d\n", info); //magma_cprint_gpu(M_lead_dim, M_lead_dim, dAthT, lda); //magma_sprint_gpu(M_lead_dim, 1, dE, M_lead_dim); //magma_sgetvector(M_lead_dim, dE, 1, hE, 1, queue); //printf("%f %f\n", hE[0], hE[1]); // p = r[:,s]' * A * r[:,s] // r = r[:,s] magma_ccopy( M_lead_dim, dAthT + (M_lead_dim*(M_lead_dim-1)), 1, // dAthT + (N), where (N) is a column offset dX, 1, queue); magma_sync_wtime(queue); //magma_cprint_gpu(M_lead_dim, 1, dX, M_lead_dim); // pp = A * r[:,s] magma_cgemv(MagmaNoTrans, M_lead_dim, M_lead_dim, MAGMA_C_MAKE(1.0f, 0.0f), dA, lda, dX, 1, MAGMA_C_MAKE(0.0f, 0.0f), dY, 1, queue); magma_sync_wtime(queue); //magma_cprint_gpu(M_lead_dim, 1, dY, M_lead_dim); // p = r' * pp p = magma_cdotc(M_lead_dim, dX, 1, dY, 1, queue); magma_sync_wtime(queue); pts[idx] = p; //printf("p = %f %fi\n", p.x, p.y); j += _step; } // end of for (idx = 0; idx < _steps; idx++) magma_free_pinned(dY); magma_free_pinned(dX); //magma_free_cpu(hE); magma_free_pinned(dE); magma_free_pinned(dAthT); magma_free_pinned(dAth); magma_free_pinned(dA); magma_free_pinned(work); magma_free_cpu(rwork); magma_free_cpu(iwork); //magma_free_cpu(w); //magma_free_cpu(A); return rslt; }
extern "C" magma_int_t magma_cgehrd2(magma_int_t n, magma_int_t ilo, magma_int_t ihi, magmaFloatComplex *a, magma_int_t lda, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= CGEHRD2 reduces a COMPLEX general matrix A to upper Hessenberg form H by an orthogonal similarity transformation: Q' * A * Q = H . Arguments ========= N (input) INTEGER The order of the matrix A. N >= 0. ILO (input) INTEGER IHI (input) INTEGER It is assumed that A is already upper triangular in rows and columns 1:ILO-1 and IHI+1:N. ILO and IHI are normally set by a previous call to CGEBAL; otherwise they should be set to 1 and N respectively. See Further Details. 1 <= ILO <= IHI <= N, if N > 0; ILO=1 and IHI=0, if N=0. A (input/output) COMPLEX array, dimension (LDA,N) On entry, the N-by-N general matrix to be reduced. On exit, the upper triangle and the first subdiagonal of A are overwritten with the upper Hessenberg matrix H, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). TAU (output) COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). Elements 1:ILO-1 and IHI:N-1 of TAU are set to zero. WORK (workspace/output) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= max(1,N). For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value. Further Details =============== The matrix Q is represented as a product of (ihi-ilo) elementary reflectors Q = H(ilo) H(ilo+1) . . . H(ihi-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0, v(i+1) = 1 and v(ihi+1:n) = 0; v(i+2:ihi) is stored on exit in A(i+2:ihi,i), and tau in TAU(i). The contents of A are illustrated by the following example, with n = 7, ilo = 2 and ihi = 6: on entry, on exit, ( a a a a a a a ) ( a a h h h h a ) ( a a a a a a ) ( a h h h h a ) ( a a a a a a ) ( h h h h h h ) ( a a a a a a ) ( v2 h h h h h ) ( a a a a a a ) ( v2 v3 h h h h ) ( a a a a a a ) ( v2 v3 v4 h h h ) ( a ) ( a ) where a denotes an element of the original matrix A, h denotes a modified element of the upper Hessenberg matrix H, and vi denotes an element of the vector defining H(i). This implementation follows the hybrid algorithm and notations described in S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg form through hybrid GPU-based computing," University of Tennessee Computer Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219), May 24, 2009. ===================================================================== */ magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; magma_int_t nb = magma_get_cgehrd_nb(n); magma_int_t N = n, ldda = n; magma_int_t ib; magma_int_t nh, iws; magma_int_t nbmin, iinfo; magma_int_t ldwork; magma_int_t lquery; --tau; *info = 0; MAGMA_C_SET2REAL( work[0], (float) n * nb ); lquery = lwork == -1; if (n < 0) { *info = -1; } else if (ilo < 1 || ilo > max(1,n)) { *info = -2; } else if (ihi < min(ilo,n) || ihi > n) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if (lwork < max(1,n) && ! lquery) { *info = -8; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; /* Quick return if possible */ nh = ihi - ilo + 1; if (nh <= 1) { work[0] = c_one; return *info; } magmaFloatComplex *da; if (MAGMA_SUCCESS != magma_cmalloc( &da, N*ldda + 2*N*nb + nb*nb )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magmaFloatComplex *d_A = da; magmaFloatComplex *d_work = da + (N+nb)*ldda; magma_int_t i__; magmaFloatComplex *t, *d_t; magma_cmalloc_cpu( &t, nb*nb ); if ( t == NULL ) { magma_free( da ); *info = MAGMA_ERR_HOST_ALLOC; return *info; } d_t = d_work + nb * ldda; czero_nbxnb_block(nb, d_A+N*ldda, ldda); /* Set elements 1:ILO-1 and IHI:N-1 of TAU to zero */ for (i__ = 1; i__ < ilo; ++i__) tau[i__] = c_zero; for (i__ = max(1,ihi); i__ < n; ++i__) tau[i__] = c_zero; for(i__=0; i__< nb*nb; i__+=4) t[i__] = t[i__+1] = t[i__+2] = t[i__+3] = c_zero; nbmin = 2; iws = 1; if (nb > 1 && nb < nh) { /* Determine when to cross over from blocked to unblocked code (last block is always handled by unblocked code) */ if (nb < nh) { /* Determine if workspace is large enough for blocked code */ iws = n * nb; if (lwork < iws) { /* Not enough workspace to use optimal NB: determine the minimum value of NB, and reduce NB or force use of unblocked code */ nbmin = nb; if (lwork >= n * nbmin) nb = lwork / n; else nb = 1; } } } ldwork = n; if (nb < nbmin || nb >= nh) { /* Use unblocked code below */ i__ = ilo; } else { /* Use blocked code */ /* Copy the matrix to the GPU */ magma_csetmatrix( N, N-ilo+1, a+(ilo-1)*(lda), lda, d_A, ldda ); for (i__ = ilo; i__ < ihi - nb; i__ += nb) { /* Computing MIN */ ib = min(nb, ihi - i__); /* Reduce columns i:i+ib-1 to Hessenberg form, returning the matrices V and T of the block reflector H = I - V*T*V' which performs the reduction, and also the matrix Y = A*V*T */ /* Get the current panel (no need for the 1st iteration) */ magma_cgetmatrix( ihi-i__+1, ib, d_A + (i__ - ilo)*ldda + i__ - 1, ldda, a + (i__ - 1 )*lda + i__ - 1, lda ); magma_clahr2(ihi, i__, ib, d_A + (i__ - ilo)*ldda, d_A + N*ldda + 1, a + (i__ - 1 )*(lda) , lda, &tau[i__], t, nb, work, ldwork); /* Copy T from the CPU to D_T on the GPU */ magma_csetmatrix( nb, nb, t, nb, d_t, nb ); magma_clahru(n, ihi, i__ - 1, ib, a + (i__ - 1 )*(lda), lda, d_A + (i__ - ilo)*ldda, d_A + (i__ - ilo)*ldda + i__ - 1, d_A + N*ldda, d_t, d_work); } } /* Use unblocked code to reduce the rest of the matrix */ if (!(nb < nbmin || nb >= nh)) { magma_cgetmatrix( n, n-i__+1, d_A+ (i__-ilo)*ldda, ldda, a + (i__-1)*(lda), lda ); } lapackf77_cgehd2(&n, &i__, &ihi, a, &lda, &tau[1], work, &iinfo); MAGMA_C_SET2REAL( work[0], (float) iws ); magma_free( da ); magma_free_cpu(t); return *info; } /* magma_cgehrd2 */
/** Purpose ------- CGEHRD reduces a COMPLEX general matrix A to upper Hessenberg form H by an orthogonal similarity transformation: Q' * A * Q = H . This version stores the triangular matrices used in the factorization so that they can be applied directly (i.e., without being recomputed) later. As a result, the application of Q is much faster. Arguments --------- @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in] ilo INTEGER @param[in] ihi INTEGER It is assumed that A is already upper triangular in rows and columns 1:ILO-1 and IHI+1:N. ILO and IHI are normally set by a previous call to CGEBAL; otherwise they should be set to 1 and N respectively. See Further Details. 1 <= ILO <= IHI <= N, if N > 0; ILO=1 and IHI=0, if N=0. @param[in,out] A COMPLEX array, dimension (LDA,N) On entry, the N-by-N general matrix to be reduced. On exit, the upper triangle and the first subdiagonal of A are overwritten with the upper Hessenberg matrix H, and the elements below the first subdiagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). Elements 1:ILO-1 and IHI:N-1 of TAU are set to zero. @param[out] work (workspace) COMPLEX array, dimension (LWORK) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. LWORK >= N*NB. where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] T COMPLEX array, dimension NB*N, where NB is the optimal blocksize. It stores the NB*NB blocks of the triangular T matrices used in the reduction. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value. Further Details --------------- The matrix Q is represented as a product of (ihi-ilo) elementary reflectors Q = H(ilo) H(ilo+1) . . . H(ihi-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0, v(i+1) = 1 and v(ihi+1:n) = 0; v(i+2:ihi) is stored on exit in A(i+2:ihi,i), and tau in TAU(i). The contents of A are illustrated by the following example, with n = 7, ilo = 2 and ihi = 6: @verbatim on entry, on exit, ( a a a a a a a ) ( a a h h h h a ) ( a a a a a a ) ( a h h h h a ) ( a a a a a a ) ( h h h h h h ) ( a a a a a a ) ( v2 h h h h h ) ( a a a a a a ) ( v2 v3 h h h h ) ( a a a a a a ) ( v2 v3 v4 h h h ) ( a ) ( a ) @endverbatim where a denotes an element of the original matrix A, h denotes a modified element of the upper Hessenberg matrix H, and vi denotes an element of the vector defining H(i). This implementation follows the hybrid algorithm and notations described in S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg form through hybrid GPU-based computing," University of Tennessee Computer Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219), May 24, 2009. This version stores the T matrices, for later use in magma_cunghr. @ingroup magma_cgeev_comp ********************************************************************/ extern "C" magma_int_t magma_cgehrd_m( magma_int_t n, magma_int_t ilo, magma_int_t ihi, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex *T, magma_int_t *info) { #define A( i, j ) (A + (i) + (j)*lda) #define dA( d, i, j ) (data.A[d] + (i) + (j)*ldda) magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; magma_int_t nb = magma_get_cgehrd_nb(n); magma_int_t nh, iws, ldda, min_lblocks, max_lblocks, last_dev, d; magma_int_t dpanel, di, nlocal, i, i2, ib, ldwork; magma_int_t iinfo; magma_int_t lquery; struct cgehrd_data data; magma_int_t ngpu = magma_num_gpus(); *info = 0; iws = n*(nb + nb*ngpu); work[0] = magma_cmake_lwork( iws ); lquery = (lwork == -1); if (n < 0) { *info = -1; } else if (ilo < 1 || ilo > max(1,n)) { *info = -2; } else if (ihi < min(ilo,n) || ihi > n) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if (lwork < iws && ! lquery) { *info = -8; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) return *info; // Adjust from 1-based indexing ilo -= 1; // Quick return if possible nh = ihi - ilo; if (nh <= 1) { work[0] = c_one; return *info; } magma_device_t orig_dev; magma_getdevice( &orig_dev ); // Set elements 0:ILO-1 and IHI-1:N-2 of TAU to zero for (i = 0; i < ilo; ++i) tau[i] = c_zero; for (i = max(0,ihi-1); i < n-1; ++i) tau[i] = c_zero; // set T to zero lapackf77_claset( "Full", &nb, &n, &c_zero, &c_zero, T, &nb ); // set to null, to simplify cleanup code for( d = 0; d < ngpu; ++d ) { data.A[d] = NULL; data.queues[d] = NULL; } // Now requires lwork >= iws; else dT won't be computed in unblocked code. // If not enough workspace, use unblocked code //if ( lwork < iws ) { // nb = 1; //} if (nb == 1 || nb >= nh) { // Use unblocked code below i = ilo; } else { // Use blocked code // allocate memory on GPUs for A and workspaces ldda = magma_roundup( n, 32 ); min_lblocks = (n / nb) / ngpu; max_lblocks = ((n-1) / nb) / ngpu + 1; last_dev = (n / nb) % ngpu; // V and Vd need to be padded for copying in mclahr2 data.ngpu = ngpu; data.ldda = ldda; data.ldv = nb*max_lblocks*ngpu; data.ldvd = nb*max_lblocks; for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); nlocal = min_lblocks*nb; if ( d < last_dev ) { nlocal += nb; } else if ( d == last_dev ) { nlocal += (n % nb); } ldwork = nlocal*ldda // A + nb*data.ldv // V + nb*data.ldvd // Vd + nb*ldda // Y + nb*ldda // W + nb*nb; // Ti if ( MAGMA_SUCCESS != magma_cmalloc( &data.A[d], ldwork )) { *info = MAGMA_ERR_DEVICE_ALLOC; goto CLEANUP; } data.V [d] = data.A [d] + nlocal*ldda; data.Vd[d] = data.V [d] + nb*data.ldv; data.Y [d] = data.Vd[d] + nb*data.ldvd; data.W [d] = data.Y [d] + nb*ldda; data.Ti[d] = data.W [d] + nb*ldda; magma_queue_create( d, &data.queues[d] ); } // Copy the matrix to GPUs magma_csetmatrix_1D_col_bcyclic( n, n, A, lda, data.A, ldda, ngpu, nb, data.queues ); // round ilo down to block boundary ilo = (ilo/nb)*nb; for (i = ilo; i < ihi - 1 - nb; i += nb) { // Reduce columns i:i+nb-1 to Hessenberg form, returning the // matrices V and T of the block reflector H = I - V*T*V' // which performs the reduction, and also the matrix Y = A*V*T // Get the current panel (no need for the 1st iteration) dpanel = (i / nb) % ngpu; di = ((i / nb) / ngpu) * nb; if ( i > ilo ) { magma_setdevice( dpanel ); magma_cgetmatrix( ihi-i, nb, dA(dpanel, i, di), ldda, A(i,i), lda, data.queues[dpanel] ); } // add 1 to i for 1-based index magma_clahr2_m( ihi, i+1, nb, A(0,i), lda, &tau[i], &T[i*nb], nb, work, n, &data ); magma_clahru_m( n, ihi, i, nb, A, lda, &data ); // copy first i rows above panel to host magma_setdevice( dpanel ); magma_cgetmatrix_async( i, nb, dA(dpanel, 0, di), ldda, A(0,i), lda, data.queues[dpanel] ); } // Copy remainder to host, block-by-block for( i2 = i; i2 < n; i2 += nb ) { ib = min( nb, n-i2 ); d = (i2 / nb) % ngpu; di = (i2 / nb) / ngpu * nb; magma_setdevice( d ); magma_cgetmatrix( n, ib, dA(d, 0, di), ldda, A(0,i2), lda, data.queues[d] ); } } // Use unblocked code to reduce the rest of the matrix // add 1 to i for 1-based index i += 1; lapackf77_cgehd2(&n, &i, &ihi, A, &lda, tau, work, &iinfo); work[0] = magma_cmake_lwork( iws ); CLEANUP: for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); magma_free( data.A[d] ); magma_queue_destroy( data.queues[d] ); } magma_setdevice( orig_dev ); return *info; } /* magma_cgehrd */