/** Purpose ------- SLAHR2 reduces the first NB columns of a real general n-BY-(n-k+1) matrix A so that elements below the k-th subdiagonal are zero. The reduction is performed by an orthogonal similarity transformation Q' * A * Q. The routine returns the matrices V and T which determine Q as a block reflector I - V*T*V', and also the matrix Y = A * V. (Note this is different than LAPACK, which computes Y = A * V * T.) This is an auxiliary routine called by SGEHRD. Arguments --------- @param[in] n INTEGER The order of the matrix A. @param[in] k INTEGER The offset for the reduction. Elements below the k-th subdiagonal in the first NB columns are reduced to zero. K < N. @param[in] nb INTEGER The number of columns to be reduced. @param[in,out] dA REAL array on the GPU, dimension (LDDA,N-K+1) On entry, the n-by-(n-k+1) general matrix A. On exit, the elements in rows K:N of the first NB columns are overwritten with the matrix Y. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,N). @param[out] dV REAL array on the GPU, dimension (LDDV, NB) On exit this n-by-nb array contains the Householder vectors of the transformation. @param[in] lddv INTEGER The leading dimension of the array dV. LDDV >= max(1,N). @param[in,out] A REAL array, dimension (LDA,N-K+1) On entry, the n-by-(n-k+1) general matrix A. On exit, the elements on and above the k-th subdiagonal in the first NB columns are overwritten with the corresponding elements of the reduced matrix; the elements below the k-th subdiagonal, with the array TAU, represent the matrix Q as a product of elementary reflectors. The other columns of A are unchanged. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] tau REAL array, dimension (NB) The scalar factors of the elementary reflectors. See Further Details. @param[out] T REAL array, dimension (LDT,NB) The upper triangular matrix T. @param[in] ldt INTEGER The leading dimension of the array T. LDT >= NB. @param[out] Y REAL array, dimension (LDY,NB) The n-by-nb matrix Y. @param[in] ldy INTEGER The leading dimension of the array Y. LDY >= N. Further Details --------------- The matrix Q is represented as a product of nb elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i+k-1) = 0, v(i+k) = 1; v(i+k+1:n) is stored on exit in A(i+k+1:n,i), and tau in TAU(i). The elements of the vectors v together form the (n-k+1)-by-nb matrix V which is needed, with T and Y, to apply the transformation to the unreduced part of the matrix, using an update of the form: A := (I - V*T*V') * (A - Y*T*V'). The contents of A on exit are illustrated by the following example with n = 7, k = 3 and nb = 2: @verbatim ( a a a a a ) ( a a a a a ) ( a a a a a ) ( h h a a a ) ( v1 h a a a ) ( v1 v2 a a a ) ( v1 v2 a 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. @ingroup magma_sgeev_aux ********************************************************************/ extern "C" magma_int_t magma_slahr2( magma_int_t n, magma_int_t k, magma_int_t nb, magmaFloat_ptr dA, magma_int_t ldda, magmaFloat_ptr dV, magma_int_t lddv, float *A, magma_int_t lda, float *tau, float *T, magma_int_t ldt, float *Y, magma_int_t ldy ) { #define A(i_,j_) ( A + (i_) + (j_)*lda) #define Y(i_,j_) ( Y + (i_) + (j_)*ldy) #define T(i_,j_) ( T + (i_) + (j_)*ldt) #define dA(i_,j_) (dA + (i_) + (j_)*ldda) #define dV(i_,j_) (dV + (i_) + (j_)*lddv) float c_zero = MAGMA_S_ZERO; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; magma_int_t ione = 1; magma_int_t n_k_i_1, n_k; float scale; magma_int_t i; float ei = MAGMA_S_ZERO; magma_int_t info = 0; if (n < 0) { info = -1; } else if (k < 0 || k > n) { info = -2; } else if (nb < 1 || nb > n) { info = -3; } else if (ldda < max(1,n)) { info = -5; } else if (lddv < max(1,n)) { info = -7; } else if (lda < max(1,n)) { info = -9; } else if (ldt < max(1,nb)) { info = -12; } else if (ldy < max(1,n)) { info = -13; } if (info != 0) { magma_xerbla( __func__, -(info) ); return info; } // adjust from 1-based indexing k -= 1; if (n <= 1) return info; for (i = 0; i < nb; ++i) { n_k_i_1 = n - k - i - 1; n_k = n - k; if (i > 0) { // Update A(k:n-1,i); Update i-th column of A - Y * T * V' // This updates one more row than LAPACK does (row k), // making the block above the panel an even multiple of nb. // Use last column of T as workspace, w. // w(0:i-1, nb-1) = VA(k+i, 0:i-1)' blasf77_scopy( &i, A(k+i,0), &lda, T(0,nb-1), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) // If real, conjugate row of V. lapackf77_slacgv(&i, T(0,nb-1), &ione); #endif // w = T(0:i-1, 0:i-1) * w blasf77_strmv( "Upper", "No trans", "No trans", &i, T(0,0), &ldt, T(0,nb-1), &ione ); // A(k:n-1, i) -= Y(k:n-1, 0:i-1) * w blasf77_sgemv( "No trans", &n_k, &i, &c_neg_one, Y(k,0), &ldy, T(0,nb-1), &ione, &c_one, A(k,i), &ione ); // Apply I - V * T' * V' to this column (call it b) from the // left, using the last column of T as workspace, w. // // Let V = ( V1 ) and b = ( b1 ) (first i-1 rows) // ( V2 ) ( b2 ) // where V1 is unit lower triangular // w := b1 = A(k+1:k+i, i) blasf77_scopy( &i, A(k+1,i), &ione, T(0,nb-1), &ione ); // w := V1' * b1 = VA(k+1:k+i, 0:i-1)' * w blasf77_strmv( "Lower", "Conj", "Unit", &i, A(k+1,0), &lda, T(0,nb-1), &ione ); // w := w + V2'*b2 = w + VA(k+i+1:n-1, 0:i-1)' * A(k+i+1:n-1, i) blasf77_sgemv( "Conj", &n_k_i_1, &i, &c_one, A(k+i+1,0), &lda, A(k+i+1,i), &ione, &c_one, T(0,nb-1), &ione ); // w := T'*w = T(0:i-1, 0:i-1)' * w blasf77_strmv( "Upper", "Conj", "Non-unit", &i, T(0,0), &ldt, T(0,nb-1), &ione ); // b2 := b2 - V2*w = A(k+i+1:n-1, i) - VA(k+i+1:n-1, 0:i-1) * w blasf77_sgemv( "No trans", &n_k_i_1, &i, &c_neg_one, A(k+i+1,0), &lda, T(0,nb-1), &ione, &c_one, A(k+i+1,i), &ione ); // w := V1*w = VA(k+1:k+i, 0:i-1) * w blasf77_strmv( "Lower", "No trans", "Unit", &i, A(k+1,0), &lda, T(0,nb-1), &ione ); // b1 := b1 - w = A(k+1:k+i-1, i) - w blasf77_saxpy( &i, &c_neg_one, T(0,nb-1), &ione, A(k+1,i), &ione ); // Restore diagonal element, saved below during previous iteration *A(k+i,i-1) = ei; } // Generate the elementary reflector H(i) to annihilate A(k+i+1:n-1,i) lapackf77_slarfg( &n_k_i_1, A(k+i+1,i), A(k+i+2,i), &ione, &tau[i] ); // Save diagonal element and set to one, to simplify multiplying by V ei = *A(k+i+1,i); *A(k+i+1,i) = c_one; // dV(i+1:n-k-1, i) = VA(k+i+1:n-1, i) magma_ssetvector( n_k_i_1, A(k+i+1,i), 1, dV(i+1,i), 1 ); // Compute Y(k+1:n,i) = A vi // dA(k:n-1, i) = dA(k:n-1, i+1:n-k-1) * dV(i+1:n-k-1, i) magma_sgemv( MagmaNoTrans, n_k, n_k_i_1, c_one, dA(k,i+1), ldda, dV(i+1,i), ione, c_zero, dA(k,i), ione ); // Compute T(0:i,i) = [ -tau T V' vi ] // [ tau ] // T(0:i-1, i) = -tau VA(k+i+1:n-1, 0:i-1)' VA(k+i+1:n-1, i) scale = MAGMA_S_NEGATE( tau[i]); blasf77_sgemv( "Conj", &n_k_i_1, &i, &scale, A(k+i+1,0), &lda, A(k+i+1,i), &ione, &c_zero, T(0,i), &ione ); // T(0:i-1, i) = T(0:i-1, 0:i-1) * T(0:i-1, i) blasf77_strmv( "Upper", "No trans", "Non-unit", &i, T(0,0), &ldt, T(0,i), &ione ); *T(i,i) = tau[i]; // Y(k:n-1, i) = dA(k:n-1, i) magma_sgetvector( n-k, dA(k,i), 1, Y(k,i), 1 ); } // Restore diagonal element *A(k+nb,nb-1) = ei; return info; } /* magma_slahr2 */
extern "C" magma_int_t magma_slahr2( magma_int_t n, magma_int_t k, magma_int_t nb, magmaFloat_ptr da, size_t da_offset, magma_int_t ldda, magmaFloat_ptr dv, size_t dv_offset, magma_int_t lddv, float *a, magma_int_t lda, float *tau, float *t, magma_int_t ldt, float *y, magma_int_t ldy, magma_queue_t queue) { /* -- clMAGMA auxiliary routine (version 0.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver @date November 2014 Purpose ======= SLAHR2 reduces the first NB columns of a real general n-BY-(n-k+1) matrix A so that elements below the k-th subdiagonal are zero. The reduction is performed by an orthogonal similarity transformation Q' * A * Q. The routine returns the matrices V and T which determine Q as a block reflector I - V*T*V', and also the matrix Y = A * V. This is an auxiliary routine called by SGEHRD. Arguments ========= N (input) INTEGER The order of the matrix A. K (input) INTEGER The offset for the reduction. Elements below the k-th subdiagonal in the first NB columns are reduced to zero. K < N. NB (input) INTEGER The number of columns to be reduced. DA (input/output) REAL array on the GPU, dimension (LDA,N-K+1) On entry, the n-by-(n-k+1) general matrix A. On exit, the elements on and above the k-th subdiagonal in the first NB columns are overwritten with the corresponding elements of the reduced matrix; the elements below the k-th subdiagonal, with the array TAU, represent the matrix Q as a product of elementary reflectors. The other columns of A are unchanged. See Further Details. DV (output) REAL array on the GPU, dimension (N, NB) On exit this contains the Householder vectors of the transformation. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). TAU (output) REAL array, dimension (NB) The scalar factors of the elementary reflectors. See Further Details. T (output) REAL array, dimension (LDT,NB) The upper triangular matrix T. LDT (input) INTEGER The leading dimension of the array T. LDT >= NB. Y (output) REAL array, dimension (LDY,NB) The n-by-nb matrix Y. LDY (input) INTEGER The leading dimension of the array Y. LDY >= N. Further Details =============== The matrix Q is represented as a product of nb elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i+k-1) = 0, v(i+k) = 1; v(i+k+1:n) is stored on exit in A(i+k+1:n,i), and tau in TAU(i). The elements of the vectors v together form the (n-k+1)-by-nb matrix V which is needed, with T and Y, to apply the transformation to the unreduced part of the matrix, using an update of the form: A := (I - V*T*V') * (A - Y*T*V'). The contents of A on exit are illustrated by the following example with n = 7, k = 3 and nb = 2: ( a a a a a ) ( a a a a a ) ( a a a a a ) ( h h a a a ) ( v1 h a a a ) ( v1 v2 a a a ) ( v1 v2 a 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. ===================================================================== */ float c_zero = MAGMA_S_ZERO; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; magma_int_t c__1 = 1; magma_int_t a_dim1, a_offset, t_dim1, t_offset, y_dim1, y_offset, i__2, i__3; float d__1; magma_int_t i__; float ei; --tau; a_dim1 = lda; a_offset = 1 + a_dim1; a -= a_offset; t_dim1 = ldt; t_offset = 1 + t_dim1; t -= t_offset; y_dim1 = ldy; y_offset = 1 + y_dim1; y -= y_offset; if (n <= 1) return MAGMA_SUCCESS; for (i__ = 1; i__ <= nb; ++i__) { if (i__ > 1) { /* Update A(K+1:N,I); Update I-th column of A - Y * V' */ i__2 = n - k + 1; i__3 = i__ - 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i__3, &a[k+i__-1+a_dim1], &lda); #endif blasf77_scopy(&i__3, &a[k+i__-1+a_dim1], &lda, &t[nb*t_dim1+1], &c__1); blasf77_strmv("u","n","n",&i__3,&t[t_offset], &ldt, &t[nb*t_dim1+1], &c__1); blasf77_sgemv("NO TRANSPOSE", &i__2, &i__3, &c_neg_one, &y[k + y_dim1], &ldy, &t[nb*t_dim1+1], &c__1, &c_one, &a[k+i__*a_dim1],&c__1); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i__3, &a[k+i__-1+a_dim1], &lda); #endif /* Apply I - V * T' * V' to this column (call it b) from the left, using the last column of T as workspace Let V = ( V1 ) and b = ( b1 ) (first I-1 rows) ( V2 ) ( b2 ) where V1 is unit lower triangular w := V1' * b1 */ i__2 = i__ - 1; blasf77_scopy(&i__2, &a[k+1+i__*a_dim1], &c__1, &t[nb*t_dim1+1], &c__1); blasf77_strmv("Lower", MagmaConjTransStr, "UNIT", &i__2, &a[k + 1 + a_dim1], &lda, &t[nb * t_dim1 + 1], &c__1); /* w := w + V2'*b2 */ i__2 = n - k - i__ + 1; i__3 = i__ - 1; blasf77_sgemv(MagmaConjTransStr, &i__2, &i__3, &c_one, &a[k + i__ + a_dim1], &lda, &a[k+i__+i__*a_dim1], &c__1, &c_one, &t[nb*t_dim1+1], &c__1); /* w := T'*w */ i__2 = i__ - 1; blasf77_strmv("U", MagmaConjTransStr, "N", &i__2, &t[t_offset], &ldt, &t[nb*t_dim1+1], &c__1); /* b2 := b2 - V2*w */ i__2 = n - k - i__ + 1; i__3 = i__ - 1; blasf77_sgemv("N", &i__2, &i__3, &c_neg_one, &a[k + i__ + a_dim1], &lda, &t[nb*t_dim1+1], &c__1, &c_one, &a[k+i__+i__*a_dim1], &c__1); /* b1 := b1 - V1*w */ i__2 = i__ - 1; blasf77_strmv("L","N","U",&i__2,&a[k+1+a_dim1],&lda,&t[nb*t_dim1+1],&c__1); blasf77_saxpy(&i__2, &c_neg_one, &t[nb * t_dim1 + 1], &c__1, &a[k + 1 + i__ * a_dim1], &c__1); a[k + i__ - 1 + (i__ - 1) * a_dim1] = ei; } /* Generate the elementary reflector H(I) to annihilate A(K+I+1:N,I) */ i__2 = n - k - i__ + 1; i__3 = k + i__ + 1; lapackf77_slarfg(&i__2, &a[k + i__ + i__ * a_dim1], &a[min(i__3,n) + i__ * a_dim1], &c__1, &tau[i__]); ei = a[k + i__ + i__ * a_dim1]; a[k + i__ + i__ * a_dim1] = c_one; /* Compute Y(K+1:N,I) */ i__2 = n - k; i__3 = n - k - i__ + 1; magma_ssetvector( i__3, &a[k + i__ + i__*a_dim1], 1, dv, dv_offset+(i__-1)*(lddv+1), 1, queue ); magma_sgemv(MagmaNoTrans, i__2+1, i__3, c_one, da, da_offset + (-1 + k + i__ * ldda), ldda, dv, dv_offset + (i__-1)*(lddv+1), c__1, c_zero, da, da_offset + (-1 + k + (i__-1)*ldda), c__1, queue); i__2 = n - k - i__ + 1; i__3 = i__ - 1; blasf77_sgemv(MagmaConjTransStr, &i__2, &i__3, &c_one, &a[k + i__ + a_dim1], &lda, &a[k+i__+i__*a_dim1], &c__1, &c_zero, &t[i__*t_dim1+1], &c__1); /* Compute T(1:I,I) */ i__2 = i__ - 1; d__1 = MAGMA_S_NEGATE( tau[i__] ); blasf77_sscal(&i__2, &d__1, &t[i__ * t_dim1 + 1], &c__1); blasf77_strmv("U","N","N", &i__2, &t[t_offset], &ldt, &t[i__*t_dim1+1], &c__1); t[i__ + i__ * t_dim1] = tau[i__]; magma_sgetvector( n - k + 1, da, da_offset+(-1+ k+(i__-1)*ldda), 1, y+ k + i__*y_dim1, 1, queue ); } a[k + nb + nb * a_dim1] = ei; return MAGMA_SUCCESS; } /* magma_slahr2 */
/***************************************************************************//** Purpose ------- SGELQF computes an LQ factorization of a REAL M-by-N matrix dA: dA = L * Q. Arguments --------- @param[in] m INTEGER The number of rows of the matrix A. M >= 0. @param[in] n INTEGER The number of columns of the matrix A. N >= 0. @param[in,out] dA REAL array on the GPU, dimension (LDDA,N) On entry, the M-by-N matrix dA. On exit, the elements on and below the diagonal of the array contain the m-by-min(m,n) lower trapezoidal matrix L (L is lower triangular if m <= n); the elements above the diagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors (see Further Details). @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,M). @param[out] tau REAL array, dimension (min(M,N)) The scalar factors of the elementary reflectors (see Further Details). @param[out] work (workspace) REAL array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. \n Higher performance is achieved if WORK is in pinned memory, e.g. allocated using magma_malloc_pinned. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= max(1,M). For optimum performance LWORK >= M*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. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value or another error occured, such as memory allocation failed. Further Details --------------- The matrix Q is represented as a product of elementary reflectors Q = H(k) . . . H(2) H(1), where k = min(m,n). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i-1) = 0 and v(i) = 1; v(i+1:n) is stored on exit in A(i,i+1:n), and tau in TAU(i). @ingroup magma_gelqf *******************************************************************************/ extern "C" magma_int_t magma_sgelqf_gpu( magma_int_t m, magma_int_t n, magmaFloat_ptr dA, magma_int_t ldda, float *tau, float *work, magma_int_t lwork, magma_int_t *info) { /* Constants */ const float c_one = MAGMA_S_ONE; const magma_int_t ione = 1; MAGMA_UNUSED( ione ); // used only for real /* Local variables */ magmaFloat_ptr dAT=NULL; magma_int_t min_mn, maxm, maxn, nb; magma_int_t iinfo; *info = 0; nb = magma_get_sgelqf_nb( m, n ); min_mn = min( m, n ); work[0] = magma_smake_lwork( m*nb ); bool lquery = (lwork == -1); if (m < 0) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,m)) { *info = -4; } else if (lwork < max(1,m) && ! lquery) { *info = -7; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (min_mn == 0) { work[0] = c_one; return *info; } maxm = magma_roundup( m, 32 ); maxn = magma_roundup( n, 32 ); magma_int_t lddat = maxn; magma_queue_t queue; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queue ); if ( m == n ) { dAT = dA; lddat = ldda; magmablas_stranspose_inplace( m, dAT, ldda, queue ); } else { if (MAGMA_SUCCESS != magma_smalloc( &dAT, maxm*maxn ) ) { *info = MAGMA_ERR_DEVICE_ALLOC; goto cleanup; } magmablas_stranspose( m, n, dA, ldda, dAT, lddat, queue ); } magma_sgeqrf2_gpu( n, m, dAT, lddat, tau, &iinfo ); assert( iinfo >= 0 ); if ( iinfo > 0 ) { *info = iinfo; } // conjugate tau #ifdef COMPLEX lapackf77_slacgv( &min_mn, tau, &ione ); #endif if ( m == n ) { magmablas_stranspose_inplace( m, dAT, lddat, queue ); } else { magmablas_stranspose( n, m, dAT, lddat, dA, ldda, queue ); magma_free( dAT ); } cleanup: magma_queue_destroy( queue ); return *info; } /* magma_sgelqf_gpu */
/** Purpose ------- SLATRD reduces NB rows and columns of a real symmetric matrix A to symmetric tridiagonal form by an orthogonal similarity transformation Q' * A * Q, and returns the matrices V and W which are needed to apply the transformation to the unreduced part of A. If UPLO = MagmaUpper, SLATRD reduces the last NB rows and columns of a matrix, of which the upper triangle is supplied; if UPLO = MagmaLower, SLATRD reduces the first NB rows and columns of a matrix, of which the lower triangle is supplied. This is an auxiliary routine called by SSYTRD. Arguments --------- @param[in] uplo magma_uplo_t Specifies whether the upper or lower triangular part of the symmetric matrix A is stored: - = MagmaUpper: Upper triangular - = MagmaLower: Lower triangular @param[in] n INTEGER The order of the matrix A. @param[in] nb INTEGER The number of rows and columns to be reduced. @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, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading n-by-n lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit: - if UPLO = MagmaUpper, the last NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements above the diagonal with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; - if UPLO = MagmaLower, the first NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements below the diagonal 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 >= (1,N). @param[out] e REAL array, dimension (N-1) If UPLO = MagmaUpper, E(n-nb:n-1) contains the superdiagonal elements of the last NB columns of the reduced matrix; if UPLO = MagmaLower, E(1:nb) contains the subdiagonal elements of the first NB columns of the reduced matrix. @param[out] tau REAL array, dimension (N-1) The scalar factors of the elementary reflectors, stored in TAU(n-nb:n-1) if UPLO = MagmaUpper, and in TAU(1:nb) if UPLO = MagmaLower. See Further Details. @param[out] W REAL array, dimension (LDW,NB) The n-by-nb matrix W required to update the unreduced part of A. @param[in] ldw INTEGER The leading dimension of the array W. LDW >= max(1,N). @param dA @param ldda @param dW @param lddw Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n) H(n-1) . . . H(n-nb+1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(i:n) = 0 and v(i-1) = 1; v(1:i-1) is stored on exit in A(1:i-1,i), and tau in TAU(i-1). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i) = 0 and v(i+1) = 1; v(i+1:n) is stored on exit in A(i+1:n,i), and tau in TAU(i). The elements of the vectors v together form the n-by-nb matrix V which is needed, with W, to apply the transformation to the unreduced part of the matrix, using a symmetric rank-2k update of the form: A := A - V*W' - W*V'. The contents of A on exit are illustrated by the following examples with n = 5 and nb = 2: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( a a a v4 v5 ) ( d ) ( a a v4 v5 ) ( 1 d ) ( a 1 v5 ) ( v1 1 a ) ( d 1 ) ( v1 v2 a a ) ( d ) ( v1 v2 a a a ) where d denotes a diagonal element of the reduced matrix, a denotes an element of the original matrix that is unchanged, and vi denotes an element of the vector defining H(i). @ingroup magma_ssyev_aux ********************************************************************/ extern "C" magma_int_t magma_slatrd( magma_uplo_t uplo, magma_int_t n, magma_int_t nb, float *A, magma_int_t lda, float *e, float *tau, float *W, magma_int_t ldw, magmaFloat_ptr dA, magma_int_t ldda, magmaFloat_ptr dW, magma_int_t lddw) { #define A(i_, j_) (A + (i_) + (j_)*lda) #define W(i_, j_) (W + (i_) + (j_)*ldw) #define dA(i_, j_) (dA + (i_) + (j_)*ldda) #define dW(i_, j_) (dW + (i_) + (j_)*lddw) const float c_neg_one = MAGMA_S_NEG_ONE; const float c_one = MAGMA_S_ONE; const float c_zero = MAGMA_S_ZERO; const magma_int_t ione = 1; float alpha, value; magma_int_t i, i_n, i_1, iw; /* Check arguments */ magma_int_t info = 0; if ( uplo != MagmaLower && uplo != MagmaUpper ) { info = -1; } else if ( n < 0 ) { info = -2; } else if ( nb < 1 ) { info = -3; } else if ( lda < max(1,n) ) { info = -5; } else if ( ldw < max(1,n) ) { info = -9; } else if ( ldda < max(1,n) ) { info = -11; } else if ( lddw < max(1,n) ) { info = -13; } if (info != 0) { magma_xerbla( __func__, -(info) ); return info; } /* Quick return if possible */ if (n == 0) { return info; } magma_queue_t stream; magma_queue_create( &stream ); float *f; magma_smalloc_cpu( &f, n ); if ( f == NULL ) { info = MAGMA_ERR_HOST_ALLOC; return info; } if (uplo == MagmaUpper) { /* Reduce last NB columns of upper triangle */ for (i = n-1; i >= n - nb; --i) { i_1 = i + 1; i_n = n - i - 1; iw = i - n + nb; if (i < n-1) { /* Update A(1:i,i) */ #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i_n, W(i, iw+1), &ldw ); #endif blasf77_sgemv( "No transpose", &i_1, &i_n, &c_neg_one, A(0, i+1), &lda, W(i, iw+1), &ldw, &c_one, A(0, i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i_n, W(i, iw+1), &ldw ); lapackf77_slacgv( &i_n, A(i, i+1), &lda ); #endif blasf77_sgemv( "No transpose", &i_1, &i_n, &c_neg_one, W(0, iw+1), &ldw, A(i, i+1), &lda, &c_one, A(0, i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i_n, A(i, i+1), &lda ); #endif } if (i > 0) { /* Generate elementary reflector H(i) to annihilate A(1:i-2,i) */ alpha = *A(i-1, i); lapackf77_slarfg( &i, &alpha, A(0, i), &ione, &tau[i - 1] ); e[i-1] = MAGMA_S_REAL( alpha ); *A(i-1,i) = MAGMA_S_ONE; /* Compute W(1:i-1,i) */ // 1. Send the block reflector A(0:n-i-1,i) to the GPU magma_ssetvector( i, A(0, i), 1, dA(0, i), 1 ); magma_ssymv( MagmaUpper, i, c_one, dA(0, 0), ldda, dA(0, i), ione, c_zero, dW(0, iw), ione ); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i, 1, dW(0, iw), lddw, W(0, iw), ldw, stream ); if (i < n-1) { blasf77_sgemv( MagmaConjTransStr, &i, &i_n, &c_one, W(0, iw+1), &ldw, A(0, i), &ione, &c_zero, W(i+1, iw), &ione ); } // 3. Here is where we need it // TODO find the right place magma_queue_sync( stream ); if (i < n-1) { blasf77_sgemv( "No transpose", &i, &i_n, &c_neg_one, A(0, i+1), &lda, W(i+1, iw), &ione, &c_one, W(0, iw), &ione ); blasf77_sgemv( MagmaConjTransStr, &i, &i_n, &c_one, A(0, i+1), &lda, A(0, i), &ione, &c_zero, W(i+1, iw), &ione ); blasf77_sgemv( "No transpose", &i, &i_n, &c_neg_one, W(0, iw+1), &ldw, W(i+1, iw), &ione, &c_one, W(0, iw), &ione ); } blasf77_sscal( &i, &tau[i - 1], W(0, iw), &ione ); value = magma_cblas_sdot( i, W(0,iw), ione, A(0,i), ione ); alpha = tau[i - 1] * -0.5f * value; blasf77_saxpy( &i, &alpha, A(0, i), &ione, W(0, iw), &ione ); } } } else { /* Reduce first NB columns of lower triangle */ for (i = 0; i < nb; ++i) { /* Update A(i:n,i) */ i_n = n - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i, W(i, 0), &ldw ); #endif blasf77_sgemv( "No transpose", &i_n, &i, &c_neg_one, A(i, 0), &lda, W(i, 0), &ldw, &c_one, A(i, i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i, W(i, 0), &ldw ); lapackf77_slacgv( &i, A(i, 0), &lda ); #endif blasf77_sgemv( "No transpose", &i_n, &i, &c_neg_one, W(i, 0), &ldw, A(i, 0), &lda, &c_one, A(i, i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i, A(i, 0), &lda ); #endif if (i < n-1) { /* Generate elementary reflector H(i) to annihilate A(i+2:n,i) */ i_n = n - i - 1; alpha = *A(i+1, i); lapackf77_slarfg( &i_n, &alpha, A(min(i+2,n-1), i), &ione, &tau[i] ); e[i] = MAGMA_S_REAL( alpha ); *A(i+1,i) = MAGMA_S_ONE; /* Compute W(i+1:n,i) */ // 1. Send the block reflector A(i+1:n,i) to the GPU magma_ssetvector( i_n, A(i+1, i), 1, dA(i+1, i), 1 ); magma_ssymv( MagmaLower, i_n, c_one, dA(i+1, i+1), ldda, dA(i+1, i), ione, c_zero, dW(i+1, i), ione ); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i_n, 1, dW(i+1, i), lddw, W(i+1, i), ldw, stream ); blasf77_sgemv( MagmaConjTransStr, &i_n, &i, &c_one, W(i+1, 0), &ldw, A(i+1, i), &ione, &c_zero, W(0, i), &ione ); blasf77_sgemv( "No transpose", &i_n, &i, &c_neg_one, A(i+1, 0), &lda, W(0, i), &ione, &c_zero, f, &ione ); blasf77_sgemv( MagmaConjTransStr, &i_n, &i, &c_one, A(i+1, 0), &lda, A(i+1, i), &ione, &c_zero, W(0, i), &ione ); // 3. Here is where we need it magma_queue_sync( stream ); if (i != 0) blasf77_saxpy( &i_n, &c_one, f, &ione, W(i+1, i), &ione ); blasf77_sgemv( "No transpose", &i_n, &i, &c_neg_one, W(i+1, 0), &ldw, W(0, i), &ione, &c_one, W(i+1, i), &ione ); blasf77_sscal( &i_n, &tau[i], W(i+1,i), &ione ); value = magma_cblas_sdot( i_n, W(i+1,i), ione, A(i+1,i), ione ); alpha = tau[i] * -0.5f * value; blasf77_saxpy( &i_n, &alpha, A(i+1, i), &ione, W(i+1,i), &ione ); } } } magma_free_cpu( f ); magma_queue_destroy( stream ); return info; } /* magma_slatrd */
extern "C" magma_err_t magma_slatrd(char uplo, magma_int_t n, magma_int_t nb, float *a, magma_int_t lda, float *e, float *tau, float *w, magma_int_t ldw, magmaFloat_ptr da, size_t da_offset, magma_int_t ldda, magmaFloat_ptr dw, size_t dw_offset, magma_int_t lddw, magma_queue_t queue) { /* -- clMAGMA (version 1.1.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver @date January 2014 Purpose ======= SLATRD reduces NB rows and columns of a real symmetric matrix A to symmetric tridiagonal form by an orthogonal similarity transformation Q' * A * Q, and returns the matrices V and W which are needed to apply the transformation to the unreduced part of A. If UPLO = 'U', SLATRD reduces the last NB rows and columns of a matrix, of which the upper triangle is supplied; if UPLO = 'L', SLATRD reduces the first NB rows and columns of a matrix, of which the lower triangle is supplied. This is an auxiliary routine called by SSYTRD. Arguments ========= UPLO (input) CHARACTER*1 Specifies whether the upper or lower triangular part of the symmetric matrix A is stored: = 'U': Upper triangular = 'L': Lower triangular N (input) INTEGER The order of the matrix A. NB (input) INTEGER The number of rows and columns to be reduced. 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, and the strictly lower triangular part of A is not referenced. If UPLO = 'L', the leading n-by-n lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit: if UPLO = 'U', the last NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements above the diagonal with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = 'L', the first NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements below the diagonal 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 >= (1,N). E (output) REAL array, dimension (N-1) If UPLO = 'U', E(n-nb:n-1) contains the superdiagonal elements of the last NB columns of the reduced matrix; if UPLO = 'L', E(1:nb) contains the subdiagonal elements of the first NB columns of the reduced matrix. TAU (output) REAL array, dimension (N-1) The scalar factors of the elementary reflectors, stored in TAU(n-nb:n-1) if UPLO = 'U', and in TAU(1:nb) if UPLO = 'L'. See Further Details. W (output) REAL array, dimension (LDW,NB) The n-by-nb matrix W required to update the unreduced part of A. LDW (input) INTEGER The leading dimension of the array W. LDW >= max(1,N). Further Details =============== If UPLO = 'U', the matrix Q is represented as a product of elementary reflectors Q = H(n) H(n-1) . . . H(n-nb+1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(i:n) = 0 and v(i-1) = 1; v(1:i-1) is stored on exit in A(1:i-1,i), and tau in TAU(i-1). If UPLO = 'L', the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i) = 0 and v(i+1) = 1; v(i+1:n) is stored on exit in A(i+1:n,i), and tau in TAU(i). The elements of the vectors v together form the n-by-nb matrix V which is needed, with W, to apply the transformation to the unreduced part of the matrix, using a symmetric rank-2k update of the form: A := A - V*W' - W*V'. The contents of A on exit are illustrated by the following examples with n = 5 and nb = 2: if UPLO = 'U': if UPLO = 'L': ( a a a v4 v5 ) ( d ) ( a a v4 v5 ) ( 1 d ) ( a 1 v5 ) ( v1 1 a ) ( d 1 ) ( v1 v2 a a ) ( d ) ( v1 v2 a a a ) where d denotes a diagonal element of the reduced matrix, a denotes an element of the original matrix that is unchanged, and vi denotes an element of the vector defining H(i). ===================================================================== */ char uplo_[2] = {uplo, 0}; magma_int_t i; float c_neg_one = MAGMA_S_NEG_ONE; float c_one = MAGMA_S_ONE; float c_zero = MAGMA_S_ZERO; float value = MAGMA_S_ZERO; magma_int_t ione = 1; magma_int_t i_n, i_1, iw; float alpha; float *f; magma_smalloc_cpu( &f, n ); if (n <= 0) { return 0; } magma_event_t event = NULL; if (lapackf77_lsame(uplo_, "U")) { /* Reduce last NB columns of upper triangle */ for (i = n-1; i >= n - nb ; --i) { i_1 = i + 1; i_n = n - i - 1; iw = i - n + nb; if (i < n-1) { /* Update A(1:i,i) */ #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, W(i, iw+1), &ldw); #endif blasf77_sgemv("No transpose", &i_1, &i_n, &c_neg_one, A(0, i+1), &lda, W(i, iw+1), &ldw, &c_one, A(0, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, W(i, iw+1), &ldw); lapackf77_slacgv(&i_n, A(i, i+1), &lda); #endif blasf77_sgemv("No transpose", &i_1, &i_n, &c_neg_one, W(0, iw+1), &ldw, A(i, i+1), &lda, &c_one, A(0, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, A(i, i+1), &lda); #endif } if (i > 0) { /* Generate elementary reflector H(i) to annihilate A(1:i-2,i) */ alpha = *A(i-1, i); lapackf77_slarfg(&i, &alpha, A(0, i), &ione, &tau[i - 1]); e[i-1] = MAGMA_S_REAL( alpha ); MAGMA_S_SET2REAL(*A(i-1, i), 1.); /* Compute W(1:i-1,i) */ // 1. Send the block reflector A(0:n-i-1,i) to the GPU magma_ssetvector( i, A(0, i), 0, 1, dA(0, i), 1, queue ); magma_ssymv(MagmaUpper, i, c_one, dA(0, 0), ldda, dA(0, i), ione, c_zero, dW(0, iw), ione, queue); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i, 1, dW(0, iw), lddw, W(0, iw)/*test*/, 0, ldw, queue, &event ); if (i < n-1) { blasf77_sgemv(MagmaTransStr, &i, &i_n, &c_one, W(0, iw+1), &ldw, A(0, i), &ione, &c_zero, W(i+1, iw), &ione); } // 3. Here is where we need it // TODO find the right place magma_event_sync(event); if (i < n-1) { blasf77_sgemv("No transpose", &i, &i_n, &c_neg_one, A(0, i+1), &lda, W(i+1, iw), &ione, &c_one, W(0, iw), &ione); blasf77_sgemv(MagmaTransStr, &i, &i_n, &c_one, A(0, i+1), &lda, A(0, i), &ione, &c_zero, W(i+1, iw), &ione); blasf77_sgemv("No transpose", &i, &i_n, &c_neg_one, W(0, iw+1), &ldw, W(i+1, iw), &ione, &c_one, W(0, iw), &ione); } blasf77_sscal(&i, &tau[i - 1], W(0, iw), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) cblas_sdot_sub( i, W(0,iw), ione, A(0,i), ione, &value ); #else value = cblas_sdot( i, W(0,iw), ione, A(0,i), ione ); #endif alpha = tau[i - 1] * -0.5f * value; blasf77_saxpy(&i, &alpha, A(0, i), &ione, W(0, iw), &ione); } } } else { /* Reduce first NB columns of lower triangle */ for (i = 0; i < nb; ++i) { /* Update A(i:n,i) */ i_n = n - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, W(i, 0), &ldw); #endif blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, A(i, 0), &lda, W(i, 0), &ldw, &c_one, A(i, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, W(i, 0), &ldw); lapackf77_slacgv(&i, A(i ,0), &lda); #endif blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, W(i, 0), &ldw, A(i, 0), &lda, &c_one, A(i, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, A(i, 0), &lda); #endif if (i < n-1) { /* Generate elementary reflector H(i) to annihilate A(i+2:n,i) */ i_n = n - i - 1; alpha = *A(i+1, i); lapackf77_slarfg(&i_n, &alpha, A(min(i+2,n-1), i), &ione, &tau[i]); e[i] = MAGMA_S_REAL( alpha ); MAGMA_S_SET2REAL(*A(i+1, i), 1.); /* Compute W(i+1:n,i) */ // 1. Send the block reflector A(i+1:n,i) to the GPU magma_ssetvector( i_n, A(i+1, i), 0, 1, dA(i+1, i), 1, queue ); magma_ssymv(MagmaLower, i_n, c_one, dA(i+1, i+1), ldda, dA(i+1, i), ione, c_zero, dW(i+1, i), ione, queue); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i_n, 1, dW(i+1, i), lddw, W(i+1, i), 0, ldw, queue, &event ); blasf77_sgemv(MagmaTransStr, &i_n, &i, &c_one, W(i+1, 0), &ldw, A(i+1, i), &ione, &c_zero, W(0, i), &ione); blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, A(i+1, 0), &lda, W(0, i), &ione, &c_zero, f, &ione); blasf77_sgemv(MagmaTransStr, &i_n, &i, &c_one, A(i+1, 0), &lda, A(i+1, i), &ione, &c_zero, W(0, i), &ione); // 3. Here is where we need it magma_event_sync(event); if (i!=0) blasf77_saxpy(&i_n, &c_one, f, &ione, W(i+1, i), &ione); blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, W(i+1, 0), &ldw, W(0, i), &ione, &c_one, W(i+1, i), &ione); blasf77_sscal(&i_n, &tau[i], W(i+1,i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) cblas_sdot_sub( i_n, W(i+1,i), ione, A(i+1,i), ione, &value ); #else value = cblas_sdot( i_n, W(i+1,i), ione, A(i+1,i), ione ); #endif alpha = tau[i] * -0.5f * value; blasf77_saxpy(&i_n, &alpha, A(i+1, i), &ione, W(i+1,i), &ione); } } } magma_free_cpu(f); return 0; } /* slatrd_ */
/** Purpose ------- SLABRD reduces the first NB rows and columns of a real general m by n matrix A to upper or lower bidiagonal form by an orthogonal transformation Q' * A * P, and returns the matrices X and Y which are needed to apply the transformation to the unreduced part of A. If m >= n, A is reduced to upper bidiagonal form; if m < n, to lower bidiagonal form. This is an auxiliary routine called by SGEBRD. Arguments --------- @param[in] m INTEGER The number of rows in the matrix A. @param[in] n INTEGER The number of columns in the matrix A. @param[in] nb INTEGER The number of leading rows and columns of A to be reduced. @param[in,out] A REAL array, dimension (LDA,N) On entry, the m by n general matrix to be reduced. On exit, the first NB rows and columns of the matrix are overwritten; the rest of the array is unchanged. If m >= n, elements on and below the diagonal in the first NB columns, with the array TAUQ, represent the orthogonal matrix Q as a product of elementary reflectors; and elements above the diagonal in the first NB rows, with the array TAUP, represent the orthogonal matrix P as a product of elementary reflectors. \n If m < n, elements below the diagonal in the first NB columns, with the array TAUQ, represent the orthogonal matrix Q as a product of elementary reflectors, and elements on and above the diagonal in the first NB rows, with the array TAUP, represent the orthogonal matrix P as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,M). @param[in,out] dA REAL array, dimension (LDDA,N) Copy of A on GPU. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,M). @param[out] d REAL array, dimension (NB) The diagonal elements of the first NB rows and columns of the reduced matrix. D(i) = A(i,i). @param[out] e REAL array, dimension (NB) The off-diagonal elements of the first NB rows and columns of the reduced matrix. @param[out] tauq REAL array dimension (NB) The scalar factors of the elementary reflectors which represent the orthogonal matrix Q. See Further Details. @param[out] taup REAL array, dimension (NB) The scalar factors of the elementary reflectors which represent the orthogonal matrix P. See Further Details. @param[out] X REAL array, dimension (LDX,NB) The m-by-nb matrix X required to update the unreduced part of A. @param[in] ldx INTEGER The leading dimension of the array X. LDX >= M. @param[out] dX REAL array, dimension (LDDX,NB) Copy of X on GPU. @param[in] lddx INTEGER The leading dimension of the array dX. LDDX >= M. @param[out] Y REAL array, dimension (LDY,NB) The n-by-nb matrix Y required to update the unreduced part of A. @param[in] ldy INTEGER The leading dimension of the array Y. LDY >= N. @param[out] dY REAL array, dimension (LDDY,NB) Copy of Y on GPU. @param[in] lddy INTEGER The leading dimension of the array dY. LDDY >= N. Further Details --------------- The matrices Q and P are represented as products of elementary reflectors: Q = H(1) H(2) . . . H(nb) and P = G(1) G(2) . . . G(nb) Each H(i) and G(i) has the form: H(i) = I - tauq * v * v' and G(i) = I - taup * u * u' where tauq and taup are real scalars, and v and u are real vectors. If m >= n, v(1:i-1) = 0, v(i) = 1, and v(i:m) is stored on exit in A(i:m,i); u(1:i) = 0, u(i+1) = 1, and u(i+1:n) is stored on exit in A(i,i+1:n); tauq is stored in TAUQ(i) and taup in TAUP(i). If m < n, v(1:i) = 0, v(i+1) = 1, and v(i+1:m) is stored on exit in A(i+2:m,i); u(1:i-1) = 0, u(i) = 1, and u(i:n) is stored on exit in A(i,i+1:n); tauq is stored in TAUQ(i) and taup in TAUP(i). The elements of the vectors v and u together form the m-by-nb matrix V and the nb-by-n matrix U' which are needed, with X and Y, to apply the transformation to the unreduced part of the matrix, using a block update of the form: A := A - V*Y' - X*U'. The contents of A on exit are illustrated by the following examples with nb = 2: @verbatim m = 6 and n = 5 (m > n): m = 5 and n = 6 (m < n): ( 1 1 u1 u1 u1 ) ( 1 u1 u1 u1 u1 u1 ) ( v1 1 1 u2 u2 ) ( 1 1 u2 u2 u2 u2 ) ( v1 v2 a a a ) ( v1 1 a a a a ) ( v1 v2 a a a ) ( v1 v2 a a a a ) ( v1 v2 a a a ) ( v1 v2 a a a a ) ( v1 v2 a a a ) @endverbatim where a denotes an element of the original matrix which is unchanged, vi denotes an element of the vector defining H(i), and ui an element of the vector defining G(i). @ingroup magma_sgesvd_aux ********************************************************************/ extern "C" magma_int_t magma_slabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, float *A, magma_int_t lda, float *dA, magma_int_t ldda, float *d, float *e, float *tauq, float *taup, float *X, magma_int_t ldx, float *dX, magma_int_t lddx, float *Y, magma_int_t ldy, float *dY, magma_int_t lddy) { #define A(i_,j_) (A + (i_) + (j_)*lda) #define X(i_,j_) (X + (i_) + (j_)*ldx) #define Y(i_,j_) (Y + (i_) + (j_)*ldy) #define dA(i_,j_) (dA + (i_) + (j_)*ldda) #define dY(i_,j_) (dY + (i_) + (j_)*lddy) #define dX(i_,j_) (dX + (i_) + (j_)*lddx) float c_neg_one = MAGMA_S_NEG_ONE; float c_one = MAGMA_S_ONE; float c_zero = MAGMA_S_ZERO; magma_int_t ione = 1; magma_int_t i__2, i__3; magma_int_t i; float alpha; A -= 1 + lda; X -= 1 + ldx; dX -= 1 + lddx; Y -= 1 + ldy; dY -= 1 + lddy; --d; --e; --tauq; --taup; /* Quick return if possible */ magma_int_t info = 0; if (m <= 0 || n <= 0) { return info; } float *f; magma_queue_t stream; magma_queue_create( &stream ); magma_smalloc_cpu( &f, max(n,m) ); if ( f == NULL ) { info = MAGMA_ERR_HOST_ALLOC; return info; } if (m >= n) { /* Reduce to upper bidiagonal form */ for (i = 1; i <= nb; ++i) { /* Update A(i:m,i) */ i__2 = m - i + 1; i__3 = i - 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__3, Y(i,1), &ldy ); #endif blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, A(i,1), &lda, Y(i,1), &ldy, &c_one, A(i,i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__3, Y(i,1), &ldy ); #endif blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, X(i,1), &ldx, A(1,i), &ione, &c_one, A(i,i), &ione ); /* Generate reflection Q(i) to annihilate A(i+1:m,i) */ alpha = *A(i,i); i__2 = m - i + 1; i__3 = i + 1; lapackf77_slarfg( &i__2, &alpha, A(min(i__3,m),i), &ione, &tauq[i] ); d[i] = MAGMA_S_REAL( alpha ); if (i < n) { *A(i,i) = c_one; /* Compute Y(i+1:n,i) */ i__2 = m - i + 1; i__3 = n - i; // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( i__2, A(i,i), 1, dA(i-1,i-1), 1 ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaConjTrans, i__2, i__3, c_one, dA(i-1,i), ldda, dA(i-1,i-1), ione, c_zero, dY(i+1,i), ione ); // 3. Put the result back ---------------------------------- magma_sgetmatrix_async( i__3, 1, dY(i+1,i), lddy, Y(i+1,i), ldy, stream ); i__2 = m - i + 1; i__3 = i - 1; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_one, A(i,1), &lda, A(i,i), &ione, &c_zero, Y(1,i), &ione ); i__2 = n - i; i__3 = i - 1; blasf77_sgemv( "N", &i__2, &i__3, &c_neg_one, Y(i+1,1), &ldy, Y(1,i), &ione, &c_zero, f, &ione ); i__2 = m - i + 1; i__3 = i - 1; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_one, X(i,1), &ldx, A(i,i), &ione, &c_zero, Y(1,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( stream ); if (i__3 != 0) { i__2 = n - i; blasf77_saxpy( &i__2, &c_one, f, &ione, Y(i+1,i), &ione ); } i__2 = i - 1; i__3 = n - i; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_neg_one, A(1,i+1), &lda, Y(1,i), &ione, &c_one, Y(i+1,i), &ione ); i__2 = n - i; blasf77_sscal( &i__2, &tauq[i], Y(i+1,i), &ione ); /* Update A(i,i+1:n) */ i__2 = n - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__2, A(i,i+1), &lda ); lapackf77_slacgv( &i, A(i,1), &lda ); #endif blasf77_sgemv( "No transpose", &i__2, &i, &c_neg_one, Y(i+1,1), &ldy, A(i,1), &lda, &c_one, A(i,i+1), &lda ); i__2 = i - 1; i__3 = n - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i, A(i,1), &lda ); lapackf77_slacgv( &i__2, X(i,1), &ldx ); #endif blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_neg_one, A(1,i+1), &lda, X(i,1), &ldx, &c_one, A(i,i+1), &lda ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__2, X(i,1), &ldx ); #endif /* Generate reflection P(i) to annihilate A(i,i+2:n) */ i__2 = n - i; i__3 = i + 2; alpha = *A(i,i+1); lapackf77_slarfg( &i__2, &alpha, A(i,min(i__3,n)), &lda, &taup[i] ); e[i] = MAGMA_S_REAL( alpha ); *A(i,i+1) = c_one; /* Compute X(i+1:m,i) */ i__2 = m - i; i__3 = n - i; // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( i__3, A(i,i+1), lda, dA(i-1,i), ldda ); // 2. Multiply --------------------------------------------- //magma_scopy( i__3, dA(i-1,i), ldda, dY(1,1), 1 ); magma_sgemv( MagmaNoTrans, i__2, i__3, c_one, dA(i,i), ldda, dA(i-1,i), ldda, //dY(1,1), 1, c_zero, dX(i+1,i), ione ); // 3. Put the result back ---------------------------------- magma_sgetmatrix_async( i__2, 1, dX(i+1,i), lddx, X(i+1,i), ldx, stream ); i__2 = n - i; blasf77_sgemv( MagmaConjTransStr, &i__2, &i, &c_one, Y(i+1,1), &ldy, A(i,i+1), &lda, &c_zero, X(1,i), &ione ); i__2 = m - i; blasf77_sgemv( "N", &i__2, &i, &c_neg_one, A(i+1,1), &lda, X(1,i), &ione, &c_zero, f, &ione ); i__2 = i - 1; i__3 = n - i; blasf77_sgemv( "N", &i__2, &i__3, &c_one, A(1,i+1), &lda, A(i,i+1), &lda, &c_zero, X(1,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( stream ); if (i != 0) { i__2 = m - i; blasf77_saxpy( &i__2, &c_one, f, &ione, X(i+1,i), &ione ); } i__2 = m - i; i__3 = i - 1; blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, X(i+1,1), &ldx, X(1,i), &ione, &c_one, X(i+1,i), &ione ); i__2 = m - i; blasf77_sscal( &i__2, &taup[i], X(i+1,i), &ione ); #if defined(PRECISION_z) || defined(PRECISION_c) i__2 = n - i; lapackf77_slacgv( &i__2, A(i,i+1), &lda ); // 4. Send the block reflector A(i+1:m,i) to the GPU after SLACGV() magma_ssetvector( i__2, A(i,i+1), lda, dA(i-1,i), ldda ); #endif } } } else { /* Reduce to lower bidiagonal form */ for (i = 1; i <= nb; ++i) { /* Update A(i,i:n) */ i__2 = n - i + 1; i__3 = i - 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__2, A(i,i), &lda ); lapackf77_slacgv( &i__3, A(i,1), &lda ); #endif blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, Y(i,1), &ldy, A(i,1), &lda, &c_one, A(i,i), &lda ); i__2 = i - 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__3, A(i,1), &lda ); lapackf77_slacgv( &i__3, X(i,1), &ldx ); #endif i__3 = n - i + 1; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_neg_one, A(1,i), &lda, X(i,1), &ldx, &c_one, A(i,i), &lda ); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__2, X(i,1), &ldx ); #endif /* Generate reflection P(i) to annihilate A(i,i+1:n) */ i__2 = n - i + 1; i__3 = i + 1; alpha = *A(i,i); lapackf77_slarfg( &i__2, &alpha, A(i,min(i__3,n)), &lda, &taup[i] ); d[i] = MAGMA_S_REAL( alpha ); if (i < m) { *A(i,i) = c_one; /* Compute X(i+1:m,i) */ i__2 = m - i; i__3 = n - i + 1; // 1. Send the block reflector A(i,i+1:n) to the GPU ------ magma_ssetvector( i__3, A(i,i), lda, dA(i-1,i-1), ldda ); // 2. Multiply --------------------------------------------- //magma_scopy( i__3, dA(i-1,i-1), ldda, dY(1,1), 1 ); magma_sgemv( MagmaNoTrans, i__2, i__3, c_one, dA(i,i-1), ldda, dA(i-1,i-1), ldda, //dY(1,1), 1, c_zero, dX(i+1,i), ione ); // 3. Put the result back ---------------------------------- magma_sgetmatrix_async( i__2, 1, dX(i+1,i), lddx, X(i+1,i), ldx, stream ); i__2 = n - i + 1; i__3 = i - 1; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_one, Y(i,1), &ldy, A(i,i), &lda, &c_zero, X(1,i), &ione ); i__2 = m - i; i__3 = i - 1; blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, A(i+1,1), &lda, X(1,i), &ione, &c_zero, f, &ione ); i__2 = i - 1; i__3 = n - i + 1; blasf77_sgemv( "No transpose", &i__2, &i__3, &c_one, A(1,i), &lda, A(i,i), &lda, &c_zero, X(1,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( stream ); if (i__2 != 0) { i__3 = m - i; blasf77_saxpy( &i__3, &c_one, f, &ione, X(i+1,i), &ione ); } i__2 = m - i; i__3 = i - 1; blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, X(i+1,1), &ldx, X(1,i), &ione, &c_one, X(i+1,i), &ione ); i__2 = m - i; blasf77_sscal( &i__2, &taup[i], X(i+1,i), &ione ); i__2 = n - i + 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__2, A(i,i), &lda ); magma_ssetvector( i__2, A(i,i), lda, dA(i-1,i-1), ldda ); #endif /* Update A(i+1:m,i) */ i__2 = m - i; i__3 = i - 1; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__3, Y(i,1), &ldy ); #endif blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, A(i+1,1), &lda, Y(i,1), &ldy, &c_one, A(i+1,i), &ione ); i__2 = m - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv( &i__3, Y(i,1), &ldy ); #endif blasf77_sgemv( "No transpose", &i__2, &i, &c_neg_one, X(i+1,1), &ldx, A(1,i), &ione, &c_one, A(i+1,i), &ione ); /* Generate reflection Q(i) to annihilate A(i+2:m,i) */ i__2 = m - i; i__3 = i + 2; alpha = *A(i+1,i); lapackf77_slarfg( &i__2, &alpha, A(min(i__3,m),i), &ione, &tauq[i] ); e[i] = MAGMA_S_REAL( alpha ); *A(i+1,i) = c_one; /* Compute Y(i+1:n,i) */ i__2 = m - i; i__3 = n - i; // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( i__2, A(i+1,i), 1, dA(i,i-1), 1 ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaConjTrans, i__2, i__3, c_one, dA(i,i), ldda, dA(i,i-1), ione, c_zero, dY(i+1,i), ione ); // 3. Put the result back ---------------------------------- magma_sgetmatrix_async( i__3, 1, dY(i+1,i), lddy, Y(i+1,i), ldy, stream ); i__2 = m - i; i__3 = i - 1; blasf77_sgemv( MagmaConjTransStr, &i__2, &i__3, &c_one, A(i+1,1), &lda, A(i+1,i), &ione, &c_zero, Y(1,i), &ione ); i__2 = n - i; i__3 = i - 1; blasf77_sgemv( "No transpose", &i__2, &i__3, &c_neg_one, Y(i+1,1), &ldy, Y(1,i), &ione, &c_zero, f, &ione ); i__2 = m - i; blasf77_sgemv( MagmaConjTransStr, &i__2, &i, &c_one, X(i+1,1), &ldx, A(i+1,i), &ione, &c_zero, Y(1,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( stream ); if (i__3 != 0) { i__2 = n - i; blasf77_saxpy( &i__2, &c_one, f, &ione, Y(i+1,i), &ione ); } i__2 = n - i; blasf77_sgemv( MagmaConjTransStr, &i, &i__2, &c_neg_one, A(1,i+1), &lda, Y(1,i), &ione, &c_one, Y(i+1,i), &ione ); i__2 = n - i; blasf77_sscal( &i__2, &tauq[i], Y(i+1,i), &ione ); } #if defined(PRECISION_z) || defined(PRECISION_c) else { i__2 = n - i + 1; lapackf77_slacgv( &i__2, A(i,i), &lda ); magma_ssetvector( i__2, A(i,i), lda, dA(i-1,i-1), ldda ); } #endif } } magma_queue_destroy( stream ); magma_free_cpu( f ); return info; } /* magma_slabrd_gpu */
/***************************************************************************//** Purpose ------- SLABRD reduces the first NB rows and columns of a real general m by n matrix A to upper or lower bidiagonal form by an orthogonal transformation Q' * A * P, and returns the matrices X and Y which are needed to apply the transformation to the unreduced part of A. If m >= n, A is reduced to upper bidiagonal form; if m < n, to lower bidiagonal form. This is an auxiliary routine called by SGEBRD. Arguments --------- @param[in] m INTEGER The number of rows in the matrix A. @param[in] n INTEGER The number of columns in the matrix A. @param[in] nb INTEGER The number of leading rows and columns of A to be reduced. @param[in,out] A REAL array, dimension (LDA,N) On entry, the m by n general matrix to be reduced. On exit, the first NB rows and columns of the matrix are overwritten; the rest of the array is unchanged. If m >= n, elements on and below the diagonal in the first NB columns, with the array TAUQ, represent the orthogonal matrix Q as a product of elementary reflectors; and elements above the diagonal in the first NB rows, with the array TAUP, represent the orthogonal matrix P as a product of elementary reflectors. \n If m < n, elements below the diagonal in the first NB columns, with the array TAUQ, represent the orthogonal matrix Q as a product of elementary reflectors, and elements on and above the diagonal in the first NB rows, with the array TAUP, represent the orthogonal matrix P as a product of elementary reflectors. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,M). @param[in,out] dA REAL array, dimension (LDDA,N) Copy of A on GPU. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,M). @param[out] d REAL array, dimension (NB) The diagonal elements of the first NB rows and columns of the reduced matrix. D(i) = A(i,i). @param[out] e REAL array, dimension (NB) The off-diagonal elements of the first NB rows and columns of the reduced matrix. @param[out] tauq REAL array dimension (NB) The scalar factors of the elementary reflectors which represent the orthogonal matrix Q. See Further Details. @param[out] taup REAL array, dimension (NB) The scalar factors of the elementary reflectors which represent the orthogonal matrix P. See Further Details. @param[out] X REAL array, dimension (LDX,NB) The m-by-nb matrix X required to update the unreduced part of A. @param[in] ldx INTEGER The leading dimension of the array X. LDX >= M. @param[out] dX REAL array, dimension (LDDX,NB) Copy of X on GPU. @param[in] lddx INTEGER The leading dimension of the array dX. LDDX >= M. @param[out] Y REAL array, dimension (LDY,NB) The n-by-nb matrix Y required to update the unreduced part of A. @param[in] ldy INTEGER The leading dimension of the array Y. LDY >= N. @param[out] dY REAL array, dimension (LDDY,NB) Copy of Y on GPU. @param[in] lddy INTEGER The leading dimension of the array dY. LDDY >= N. @param work REAL array, dimension (LWORK) Workspace. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= max( M, N ). @param[in] queue magma_queue_t Queue to execute in. Further Details --------------- The matrices Q and P are represented as products of elementary reflectors: Q = H(1) H(2) . . . H(nb) and P = G(1) G(2) . . . G(nb) Each H(i) and G(i) has the form: H(i) = I - tauq * v * v' and G(i) = I - taup * u * u' where tauq and taup are real scalars, and v and u are real vectors. If m >= n, v(1:i-1) = 0, v(i) = 1, and v(i:m) is stored on exit in A(i:m,i); u(1:i) = 0, u(i+1) = 1, and u(i+1:n) is stored on exit in A(i,i+1:n); tauq is stored in TAUQ(i) and taup in TAUP(i). If m < n, v(1:i) = 0, v(i+1) = 1, and v(i+1:m) is stored on exit in A(i+2:m,i); u(1:i-1) = 0, u(i) = 1, and u(i:n) is stored on exit in A(i,i+1:n); tauq is stored in TAUQ(i) and taup in TAUP(i). The elements of the vectors v and u together form the m-by-nb matrix V and the nb-by-n matrix U' which are needed, with X and Y, to apply the transformation to the unreduced part of the matrix, using a block update of the form: A := A - V*Y' - X*U'. The contents of A on exit are illustrated by the following examples with nb = 2: @verbatim m = 6 and n = 5 (m > n): m = 5 and n = 6 (m < n): ( 1 1 u1 u1 u1 ) ( 1 u1 u1 u1 u1 u1 ) ( v1 1 1 u2 u2 ) ( 1 1 u2 u2 u2 u2 ) ( v1 v2 a a a ) ( v1 1 a a a a ) ( v1 v2 a a a ) ( v1 v2 a a a a ) ( v1 v2 a a a ) ( v1 v2 a a a a ) ( v1 v2 a a a ) @endverbatim where a denotes an element of the original matrix which is unchanged, vi denotes an element of the vector defining H(i), and ui an element of the vector defining G(i). @ingroup magma_labrd *******************************************************************************/ extern "C" magma_int_t magma_slabrd_gpu( magma_int_t m, magma_int_t n, magma_int_t nb, float *A, magma_int_t lda, magmaFloat_ptr dA, magma_int_t ldda, float *d, float *e, float *tauq, float *taup, float *X, magma_int_t ldx, magmaFloat_ptr dX, magma_int_t lddx, float *Y, magma_int_t ldy, magmaFloat_ptr dY, magma_int_t lddy, float *work, magma_int_t lwork, magma_queue_t queue ) { #define A(i_,j_) ( A + (i_) + (j_)*lda) #define X(i_,j_) ( X + (i_) + (j_)*ldx) #define Y(i_,j_) ( Y + (i_) + (j_)*ldy) #define dA(i_,j_) (dA + (i_) + (j_)*ldda) #define dY(i_,j_) (dY + (i_) + (j_)*lddy) #define dX(i_,j_) (dX + (i_) + (j_)*lddx) /* Constants */ const float c_neg_one = MAGMA_S_NEG_ONE; const float c_one = MAGMA_S_ONE; const float c_zero = MAGMA_S_ZERO; const magma_int_t ione = 1; /* Local variables */ magma_int_t i, i1, m_i, m_i1, n_i, n_i1; float alpha; /* Quick return if possible */ magma_int_t info = 0; if (m <= 0 || n <= 0) { return info; } if (m >= n) { /* Reduce to upper bidiagonal form */ for (i=0; i < nb; ++i) { /* Update A(i:m,i) */ i1 = i + 1; m_i = m - i; m_i1 = m - (i+1); n_i1 = n - (i+1); #ifdef COMPLEX lapackf77_slacgv( &i, Y(i,0), &ldy ); #endif blasf77_sgemv( "No transpose", &m_i, &i, &c_neg_one, A(i,0), &lda, Y(i,0), &ldy, &c_one, A(i,i), &ione ); #ifdef COMPLEX lapackf77_slacgv( &i, Y(i,0), &ldy ); #endif blasf77_sgemv( "No transpose", &m_i, &i, &c_neg_one, X(i,0), &ldx, A(0,i), &ione, &c_one, A(i,i), &ione ); /* Generate reflection Q(i) to annihilate A(i+1:m,i) */ alpha = *A(i,i); lapackf77_slarfg( &m_i, &alpha, A(min(i+1,m-1),i), &ione, &tauq[i] ); d[i] = MAGMA_S_REAL( alpha ); if (i+1 < n) { *A(i,i) = c_one; /* Compute Y(i+1:n,i) */ // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( m_i, A(i,i), 1, dA(i,i), 1, queue ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaConjTrans, m_i, n_i1, c_one, dA(i,i+1), ldda, dA(i,i), ione, c_zero, dY(i+1,i), ione, queue ); // 3. Get the result back ---------------------------------- magma_sgetmatrix_async( n_i1, 1, dY(i+1,i), lddy, Y(i+1,i), ldy, queue ); blasf77_sgemv( MagmaConjTransStr, &m_i, &i, &c_one, A(i,0), &lda, A(i,i), &ione, &c_zero, Y(0,i), &ione ); blasf77_sgemv( "N", &n_i1, &i, &c_neg_one, Y(i+1,0), &ldy, Y(0,i), &ione, &c_zero, work, &ione ); blasf77_sgemv( MagmaConjTransStr, &m_i, &i, &c_one, X(i,0), &ldx, A(i,i), &ione, &c_zero, Y(0,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( queue ); if (i != 0) { blasf77_saxpy( &n_i1, &c_one, work, &ione, Y(i+1,i), &ione ); } blasf77_sgemv( MagmaConjTransStr, &i, &n_i1, &c_neg_one, A(0,i+1), &lda, Y(0,i), &ione, &c_one, Y(i+1,i), &ione ); blasf77_sscal( &n_i1, &tauq[i], Y(i+1,i), &ione ); /* Update A(i,i+1:n) */ #ifdef COMPLEX lapackf77_slacgv( &n_i1, A(i,i+1), &lda ); lapackf77_slacgv( &i1, A(i,0), &lda ); #endif blasf77_sgemv( "No transpose", &n_i1, &i1, &c_neg_one, Y(i+1,0), &ldy, A(i,0), &lda, &c_one, A(i,i+1), &lda ); #ifdef COMPLEX lapackf77_slacgv( &i1, A(i,0), &lda ); lapackf77_slacgv( &i, X(i,0), &ldx ); #endif blasf77_sgemv( MagmaConjTransStr, &i, &n_i1, &c_neg_one, A(0,i+1), &lda, X(i,0), &ldx, &c_one, A(i,i+1), &lda ); #ifdef COMPLEX lapackf77_slacgv( &i, X(i,0), &ldx ); #endif /* Generate reflection P(i) to annihilate A(i,i+2:n) */ alpha = *A(i,i+1); lapackf77_slarfg( &n_i1, &alpha, A(i,min(i+2,n-1)), &lda, &taup[i] ); e[i] = MAGMA_S_REAL( alpha ); *A(i,i+1) = c_one; /* Compute X(i+1:m,i) */ // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( n_i1, A(i,i+1), lda, dA(i,i+1), ldda, queue ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaNoTrans, m_i1, n_i1, c_one, dA(i+1,i+1), ldda, dA(i,i+1), ldda, //dY(0,0), 1, c_zero, dX(i+1,i), ione, queue ); // 3. Get the result back ---------------------------------- magma_sgetmatrix_async( m_i1, 1, dX(i+1,i), lddx, X(i+1,i), ldx, queue ); blasf77_sgemv( MagmaConjTransStr, &n_i1, &i1, &c_one, Y(i+1,0), &ldy, A(i,i+1), &lda, &c_zero, X(0,i), &ione ); blasf77_sgemv( "N", &m_i1, &i1, &c_neg_one, A(i+1,0), &lda, X(0,i), &ione, &c_zero, work, &ione ); blasf77_sgemv( "N", &i, &n_i1, &c_one, A(0,i+1), &lda, A(i,i+1), &lda, &c_zero, X(0,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( queue ); if ((i+1) != 0) { blasf77_saxpy( &m_i1, &c_one, work, &ione, X(i+1,i), &ione ); } blasf77_sgemv( "No transpose", &m_i1, &i, &c_neg_one, X(i+1,0), &ldx, X(0,i), &ione, &c_one, X(i+1,i), &ione ); blasf77_sscal( &m_i1, &taup[i], X(i+1,i), &ione ); #ifdef COMPLEX lapackf77_slacgv( &n_i1, A(i,i+1), &lda ); // 4. Send the block reflector A(i+1:m,i) to the GPU after SLACGV() magma_ssetvector( n_i1, A(i,i+1), lda, dA(i,i+1), ldda, queue ); #endif } } } else { /* Reduce to lower bidiagonal form */ for (i=0; i < nb; ++i) { /* Update A(i,i:n) */ i1 = i + 1; m_i1 = m - (i+1); n_i = n - i; n_i1 = n - (i+1); #ifdef COMPLEX lapackf77_slacgv( &n_i, A(i,i), &lda ); lapackf77_slacgv( &i, A(i,0), &lda ); #endif blasf77_sgemv( "No transpose", &n_i, &i, &c_neg_one, Y(i,0), &ldy, A(i,0), &lda, &c_one, A(i,i), &lda ); #ifdef COMPLEX lapackf77_slacgv( &i, A(i,0), &lda ); lapackf77_slacgv( &i, X(i,0), &ldx ); #endif blasf77_sgemv( MagmaConjTransStr, &i, &n_i, &c_neg_one, A(0,i), &lda, X(i,0), &ldx, &c_one, A(i,i), &lda ); #ifdef COMPLEX lapackf77_slacgv( &i, X(i,0), &ldx ); #endif /* Generate reflection P(i) to annihilate A(i,i+1:n) */ alpha = *A(i,i); lapackf77_slarfg( &n_i, &alpha, A(i,min(i+1,n-1)), &lda, &taup[i] ); d[i] = MAGMA_S_REAL( alpha ); if (i+1 < m) { *A(i,i) = c_one; /* Compute X(i+1:m,i) */ // 1. Send the block reflector A(i,i+1:n) to the GPU ------ magma_ssetvector( n_i, A(i,i), lda, dA(i,i), ldda, queue ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaNoTrans, m_i1, n_i, c_one, dA(i+1,i), ldda, dA(i,i), ldda, //dY(0,0), 1, c_zero, dX(i+1,i), ione, queue ); // 3. Get the result back ---------------------------------- magma_sgetmatrix_async( m_i1, 1, dX(i+1,i), lddx, X(i+1,i), ldx, queue ); blasf77_sgemv( MagmaConjTransStr, &n_i, &i, &c_one, Y(i,0), &ldy, A(i,i), &lda, &c_zero, X(0,i), &ione ); blasf77_sgemv( "No transpose", &m_i1, &i, &c_neg_one, A(i+1,0), &lda, X(0,i), &ione, &c_zero, work, &ione ); blasf77_sgemv( "No transpose", &i, &n_i, &c_one, A(0,i), &lda, A(i,i), &lda, &c_zero, X(0,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( queue ); if (i != 0) { blasf77_saxpy( &m_i1, &c_one, work, &ione, X(i+1,i), &ione ); } blasf77_sgemv( "No transpose", &m_i1, &i, &c_neg_one, X(i+1,0), &ldx, X(0,i), &ione, &c_one, X(i+1,i), &ione ); blasf77_sscal( &m_i1, &taup[i], X(i+1,i), &ione ); #ifdef COMPLEX lapackf77_slacgv( &n_i, A(i,i), &lda ); magma_ssetvector( n_i, A(i,i), lda, dA(i,i), ldda, queue ); #endif /* Update A(i+1:m,i) */ #ifdef COMPLEX lapackf77_slacgv( &i, Y(i,0), &ldy ); #endif blasf77_sgemv( "No transpose", &m_i1, &i, &c_neg_one, A(i+1,0), &lda, Y(i,0), &ldy, &c_one, A(i+1,i), &ione ); #ifdef COMPLEX lapackf77_slacgv( &i, Y(i,0), &ldy ); #endif blasf77_sgemv( "No transpose", &m_i1, &i1, &c_neg_one, X(i+1,0), &ldx, A(0,i), &ione, &c_one, A(i+1,i), &ione ); /* Generate reflection Q(i) to annihilate A(i+2:m,i) */ alpha = *A(i+1,i); lapackf77_slarfg( &m_i1, &alpha, A(min(i+2,m-1),i), &ione, &tauq[i] ); e[i] = MAGMA_S_REAL( alpha ); *A(i+1,i) = c_one; /* Compute Y(i+1:n,i) */ // 1. Send the block reflector A(i+1:m,i) to the GPU ------ magma_ssetvector( m_i1, A(i+1,i), 1, dA(i+1,i), 1, queue ); // 2. Multiply --------------------------------------------- magma_sgemv( MagmaConjTrans, m_i1, n_i1, c_one, dA(i+1,i+1), ldda, dA(i+1,i), ione, c_zero, dY(i+1,i), ione, queue ); // 3. Get the result back ---------------------------------- magma_sgetmatrix_async( n_i1, 1, dY(i+1,i), lddy, Y(i+1,i), ldy, queue ); blasf77_sgemv( MagmaConjTransStr, &m_i1, &i, &c_one, A(i+1,0), &lda, A(i+1,i), &ione, &c_zero, Y(0,i), &ione ); blasf77_sgemv( "No transpose", &n_i1, &i, &c_neg_one, Y(i+1,0), &ldy, Y(0,i), &ione, &c_zero, work, &ione ); blasf77_sgemv( MagmaConjTransStr, &m_i1, &i1, &c_one, X(i+1,0), &ldx, A(i+1,i), &ione, &c_zero, Y(0,i), &ione ); // 4. Sync to make sure the result is back ---------------- magma_queue_sync( queue ); if (i != 0) { blasf77_saxpy( &n_i1, &c_one, work, &ione, Y(i+1,i), &ione ); } blasf77_sgemv( MagmaConjTransStr, &i1, &n_i1, &c_neg_one, A(0,i+1), &lda, Y(0,i), &ione, &c_one, Y(i+1,i), &ione ); blasf77_sscal( &n_i1, &tauq[i], Y(i+1,i), &ione ); } #ifdef COMPLEX else { lapackf77_slacgv( &n_i, A(i,i), &lda ); magma_ssetvector( n_i, A(i,i), lda, dA(i,i), ldda, queue ); } #endif } } return info; } /* magma_slabrd_gpu */
/** Purpose ------- SLAHR2 reduces the first NB columns of a real general n-BY-(n-k+1) matrix A so that elements below the k-th subdiagonal are zero. The reduction is performed by an orthogonal similarity transformation Q' * A * Q. The routine returns the matrices V and T which determine Q as a block reflector I - V*T*V', and also the matrix Y = A * V. (Note this is different than LAPACK, which computes Y = A * V * T.) This is an auxiliary routine called by SGEHRD. Arguments --------- @param[in] n INTEGER The order of the matrix A. @param[in] k INTEGER The offset for the reduction. Elements below the k-th subdiagonal in the first NB columns are reduced to zero. K < N. @param[in] nb INTEGER The number of columns to be reduced. @param[in,out] A REAL array, dimension (LDA,N-K+1) On entry, the n-by-(n-k+1) general matrix A. On exit, the elements on and above the k-th subdiagonal in the first NB columns are overwritten with the corresponding elements of the reduced matrix; the elements below the k-th subdiagonal, with the array TAU, represent the matrix Q as a product of elementary reflectors. The other columns of A are unchanged. See Further Details. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[out] tau REAL array, dimension (NB) The scalar factors of the elementary reflectors. See Further Details. @param[out] T REAL array, dimension (LDT,NB) The upper triangular matrix T. @param[in] ldt INTEGER The leading dimension of the array T. LDT >= NB. @param[out] Y REAL array, dimension (LDY,NB) The n-by-nb matrix Y. @param[in] ldy INTEGER The leading dimension of the array Y. LDY >= N. @param[in,out] data Structure with pointers to dA, dT, dV, dW, dY which are distributed across multiple GPUs. Further Details --------------- The matrix Q is represented as a product of nb elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i+k-1) = 0, v(i+k) = 1; v(i+k+1:n) is stored on exit in A(i+k+1:n,i), and tau in TAU(i). The elements of the vectors v together form the (n-k+1)-by-nb matrix V which is needed, with T and Y, to apply the transformation to the unreduced part of the matrix, using an update of the form: A := (I - V*T*V') * (A - Y*T*V'). The contents of A on exit are illustrated by the following example with n = 7, k = 3 and nb = 2: @verbatim ( a a a a a ) ( a a a a a ) ( a a a a a ) ( h h a a a ) ( v1 h a a a ) ( v1 v2 a a a ) ( v1 v2 a 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. @ingroup magma_sgeev_aux ********************************************************************/ extern "C" magma_int_t magma_slahr2_m( magma_int_t n, magma_int_t k, magma_int_t nb, float *A, magma_int_t lda, float *tau, float *T, magma_int_t ldt, float *Y, magma_int_t ldy, struct sgehrd_data *data ) { #define A( i, j ) ( A + (i) + (j)*lda) #define Y( i, j ) ( Y + (i) + (j)*ldy) #define T( i, j ) ( T + (i) + (j)*ldt) #define dA( d, i, j ) (data->A [d] + (i) + (j)*ldda) #define dTi( d ) (data->Ti[d]) #define dV( d, i, j ) (data->V [d] + (i) + (j)*ldv ) #define dVd( d, i, j ) (data->Vd[d] + (i) + (j)*ldvd) #define dY( d, i, j ) (data->Y [d] + (i) + (j)*ldda) float c_zero = MAGMA_S_ZERO; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; float tmp; magma_int_t ngpu = data->ngpu; magma_int_t ldda = data->ldda; magma_int_t ldv = data->ldv; magma_int_t ldvd = data->ldvd; magma_int_t ione = 1; magma_int_t d, dki1, dn, nblocks, gblock, lblock, lgid; magma_int_t n_k_i_1, n_k; float scale; magma_int_t i; float ei = MAGMA_S_ZERO; magma_int_t info_data = 0; magma_int_t *info = &info_data; if (n < 0) { *info = -1; } else if (k < 0 || k >= n) { *info = -2; } else if (nb < 1 || nb > n) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if (ldt < nb) { *info = -8; } else if (ldy < max(1,n)) { *info = -10; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } // adjust from 1-based indexing k -= 1; // Function Body if (n <= 1) return *info; magma_device_t orig_dev; magma_getdevice( &orig_dev ); // zero out current top block of V on all GPUs for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); magmablas_slaset( MagmaFull, nb, nb, c_zero, c_zero, dV(d,k,0), ldv, data->queues[d] ); } // set all Y=0 lapackf77_slaset( "Full", &n, &nb, &c_zero, &c_zero, Y, &ldy ); for (i = 0; i < nb; ++i) { n_k_i_1 = n - k - i - 1; n_k = n - k; if (i > 0) { // Finish applying I - V * T * V' on right tmp = MAGMA_S_NEGATE( tau[i-1] ); blasf77_saxpy( &n_k, &tmp, Y(k,i-1), &ione, A(k,i), &ione ); // Apply I - V * T' * V' to this column (call it b) from the // left, using the last column of T as workspace, w. // // Let V = ( V1 ) and b = ( b1 ) (first i-1 rows) // ( V2 ) ( b2 ) // where V1 is unit lower triangular // w := b1 = A(k+1:k+i, i) blasf77_scopy( &i, A(k+1,i), &ione, T(0,nb-1), &ione ); // w := V1' * b1 = VA(k+1:k+i, 0:i-1)' * w blasf77_strmv( "Lower", "Conj", "Unit", &i, A(k+1,0), &lda, T(0,nb-1), &ione ); // w := w + V2'*b2 = w + VA(k+i+1:n-1, 0:i-1)' * A(k+i+1:n-1, i) blasf77_sgemv( "Conj", &n_k_i_1, &i, &c_one, A(k+i+1,0), &lda, A(k+i+1,i), &ione, &c_one, T(0,nb-1), &ione ); // w := T'*w = T(0:i-1, 0:i-1)' * w blasf77_strmv( "Upper", "Conj", "Non-unit", &i, T(0,0), &ldt, T(0,nb-1), &ione ); // b2 := b2 - V2*w = A(k+i+1:n-1, i) - VA(k+i+1:n-1, 0:i-1) * w blasf77_sgemv( "No trans", &n_k_i_1, &i, &c_neg_one, A(k+i+1,0), &lda, T(0,nb-1), &ione, &c_one, A(k+i+1,i), &ione ); // w := V1*w = VA(k+1:k+i, 0:i-1) * w blasf77_strmv( "Lower", "No trans", "Unit", &i, A(k+1,0), &lda, T(0,nb-1), &ione ); // b1 := b1 - w = A(k+1:k+i-1, i) - w blasf77_saxpy( &i, &c_neg_one, T(0,nb-1), &ione, A(k+1,i), &ione ); // Restore diagonal element, saved below during previous iteration *A(k+i,i-1) = ei; } // Generate the elementary reflector H(i) to annihilate A(k+i+1:n-1,i) lapackf77_slarfg( &n_k_i_1, A(k+i+1,i), A(k+i+2,i), &ione, &tau[i] ); // Save diagonal element and set to one, to simplify multiplying by V ei = *A(k+i+1,i); *A(k+i+1,i) = c_one; // compute yi = A vi = sum_g A{d} vi{d} nblocks = (n-1) / nb / ngpu + 1; for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); // dV(k+i+1:n-1, i) = VA(k+i:n, i) magma_ssetvector_async( n_k_i_1, A(k+i+1,i), 1, dV(d, k+i+1, i), 1, data->queues[d] ); // copy column of dV -> dVd, using block cyclic distribution. // This assumes V and Vd have been padded so that // a 2D matrix copy doesn't access them out-of-bounds gblock = k / nb; lblock = gblock / ngpu; lgid = gblock % ngpu; if ( d < lgid ) { lblock += 1; } // treat V as (nb*ngpu) x nblock matrix, and Vd as nb x nblock matrix magmablas_slacpy( MagmaFull, nb, nblocks-lblock, dV (d, d*nb + lblock*nb*ngpu, i), nb*ngpu, dVd(d, 0 + lblock*nb, i), nb, data->queues[d] ); // convert global indices (k) to local indices (dk) magma_indices_1D_bcyclic( nb, ngpu, d, k+i+1, n, &dki1, &dn ); // dY(k:n, i) = dA(k:n, k+i+1:n) * dV(k+i+1:n, i) // skip if matrix is empty // each GPU copies to different temporary vector in Y, // which are summed in separate loop below if ( dn-dki1 > 0 ) { magma_sgemv( MagmaNoTrans, n-k, dn-dki1, c_one, dA (d, k, dki1), ldda, dVd(d, dki1, i), 1, c_zero, dY (d, k, i), 1, data->queues[d] ); // copy vector to host, storing in column nb+d of Y // as temporary space (Y has >= nb+ngpu columns) magma_sgetvector_async( n-k, dY(d, k, i), 1, Y(k, nb+d), 1, data->queues[d] ); } } // while GPU is doing above Ag*v... // Compute T(0:i,i) = [ -tau T V' vi ] // [ tau ] // T(0:i-1, i) = -tau VA(k+i+1:n-1, 0:i-1)' VA(k+i+1:n-1, i) scale = MAGMA_S_NEGATE( tau[i] ); blasf77_sgemv( "Conj", &n_k_i_1, &i, &scale, A(k+i+1,0), &lda, A(k+i+1,i), &ione, &c_zero, T(0,i), &ione ); // T(0:i-1, i) = T(0:i-1, 0:i-1) * T(0:i-1, i) blasf77_strmv( "Upper", "No trans", "Non-unit", &i, T(0,0), &ldt, T(0,i), &ione ); *T(i,i) = tau[i]; // apply reflectors to next column, A(i+1), on right only. // one axpy will be required to finish this, in the next iteration above if ( i > 0 && i+1 < nb ) { // Update next column, A(k:n,i+1), applying Q on right. // One axpy will be required to finish this, in the next iteration // above, after yi is computed. // This updates one more row than LAPACK does (row k), // making block above panel an even multiple of nb. // Use last column of T as workspace, w. magma_int_t i1 = i+1; // If real, conjugate row of V, and undo afterwards #ifdef COMPLEX lapackf77_slacgv( &i1, A(k+i1,0), &lda ); #endif // w = T(0:i, 0:i+1) * VA(k+i+1, 0:i+1)' // T is now rectangular, so we use gemv instead of trmv as in lapack. blasf77_sgemv( "No trans", &i, &i1, &c_one, T(0,0), &ldt, A(k+i1,0), &lda, &c_zero, T(0,nb-1), &ione ); #ifdef COMPLEX lapackf77_slacgv( &i1, A(k+i1,0), &lda ); #endif // A(k:n, i+1) -= Y(k:n, 0:i) * w blasf77_sgemv( "No trans", &n_k, &i, &c_neg_one, Y(k,0), &ldy, T(0,nb-1), &ione, &c_one, A(k,i1), &ione ); } // yi = sum_g yi{d} for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); magma_queue_sync( data->queues[d] ); magma_indices_1D_bcyclic( nb, ngpu, d, k+i+1, n, &dki1, &dn ); if ( dn-dki1 > 0 ) { // yi = yi + yi{d} blasf77_saxpy( &n_k, &c_one, Y(k,nb+d), &ione, Y(k,i), &ione ); } } } // Restore diagonal element *A(k+nb,nb-1) = ei; // compute Y = Am V = sum_g Am{d} V{d} --- top part, Y(0:k-1,:) for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); // convert global indices (k) to local indices (dk) magma_indices_1D_bcyclic( nb, ngpu, d, k+1, n, &dki1, &dn ); // dY(0:k, :) = dA(0:k, k+i+1:n-1) * dV(k+i+1:n-1, :) // skip if matrix is empty // each GPU copies to different temporary block in Y, // which are summed in separate loop below if ( dn-dki1 > 0 ) { magma_sgemm( MagmaNoTrans, MagmaNoTrans, k, nb, dn-dki1, c_one, dA (d, 0, dki1), ldda, dVd(d, dki1, 0), ldvd, c_zero, dY (d, 0, 0), ldda, data->queues[d] ); // copy result to host, storing in columns [nb + nb*d : nb + nb*(d+1)] of Y // as temporary space (Y has nb + nb*ngpu columns) magma_sgetmatrix_async( k, nb, dY(d, 0, 0), ldda, Y(0,nb+nb*d), ldy, data->queues[d] ); } } // Y = sum_g Y{d} for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); magma_queue_sync( 0 ); magma_indices_1D_bcyclic( nb, ngpu, d, k+1, n, &dki1, &dn ); if ( dn-dki1 > 0 ) { // Y = Y + Am V for( i = 0; i < nb; ++i ) { blasf77_saxpy( &k, &c_one, Y(0,nb+nb*d+i), &ione, Y(0,i), &ione ); } } } // copy Y and T matrices to GPUs for( d = 0; d < ngpu; ++d ) { magma_setdevice( d ); magma_ssetmatrix_async( n, nb, Y, ldy, dY(d, 0, 0), ldda, data->queues[d] ); magma_ssetmatrix_async( nb, nb, T, nb, dTi(d), nb, data->queues[d] ); } magma_setdevice( orig_dev ); return *info; } /* magma_slahr2 */
/***************************************************************************//** Purpose ------- SGELQF computes an LQ factorization of a REAL M-by-N matrix A: A = L * Q. Arguments --------- @param[in] m INTEGER The number of rows of the matrix A. M >= 0. @param[in] n INTEGER The number of columns of the matrix A. N >= 0. @param[in,out] A REAL array, dimension (LDA,N) On entry, the M-by-N matrix A. On exit, the elements on and below the diagonal of the array contain the m-by-min(m,n) lower trapezoidal matrix L (L is lower triangular if m <= n); the elements above the diagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors (see Further Details). \n Higher performance is achieved if A is in pinned memory, e.g. allocated using magma_malloc_pinned. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,M). @param[out] tau REAL array, dimension (min(M,N)) The scalar factors of the elementary reflectors (see Further Details). @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 >= max(1,M). For optimum performance LWORK >= M*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. \n TODO: work is currently unused. sgeqrf2 allocates its own work of (m + n)*nb. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value or another error occured, such as memory allocation failed. Further Details --------------- The matrix Q is represented as a product of elementary reflectors Q = H(k) . . . H(2) H(1), where k = min(m,n). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i-1) = 0 and v(i) = 1; v(i+1:n) is stored on exit in A(i,i+1:n), and tau in TAU(i). @ingroup magma_gelqf *******************************************************************************/ extern "C" magma_int_t magma_sgelqf( magma_int_t m, magma_int_t n, float *A, magma_int_t lda, float *tau, float *work, magma_int_t lwork, magma_int_t *info) { #define dA(i_, j_) (dA + (i_) + (j_)*ldda) #define dAT(i_, j_) (dAT + (i_) + (j_)*ldda) /* Constants */ const float c_one = MAGMA_S_ONE; const magma_int_t ione = 1; MAGMA_UNUSED( ione ); // used only for real /* Local variables */ magmaFloat_ptr dA=NULL, dAT=NULL; magma_int_t min_mn, maxm, maxn, maxdim, nb; magma_int_t iinfo, ldda, lddat; /* Function Body */ *info = 0; nb = magma_get_sgelqf_nb( m, n ); min_mn = min( m, n ); work[0] = magma_smake_lwork( m*nb ); bool lquery = (lwork == -1); if (m < 0) { *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,m)) { *info = -4; } else if (lwork < max(1,m) && ! lquery) { *info = -7; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (min_mn == 0) { work[0] = c_one; return *info; } maxm = magma_roundup( m, 32 ); maxn = magma_roundup( n, 32 ); maxdim = max( maxm, maxn ); magma_queue_t queue = NULL; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queue ); // copy to GPU and transpose if (maxdim*maxdim < 2*maxm*maxn) { // close to square, do everything in-place ldda = maxdim; lddat = maxdim; if (MAGMA_SUCCESS != magma_smalloc( &dA, maxdim*maxdim )) { *info = MAGMA_ERR_DEVICE_ALLOC; goto cleanup; } magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queue ); dAT = dA; magmablas_stranspose_inplace( lddat, dAT(0,0), lddat, queue ); } else { // rectangular, do everything out-of-place ldda = maxm; lddat = maxn; if (MAGMA_SUCCESS != magma_smalloc( &dA, 2*maxn*maxm )) { *info = MAGMA_ERR_DEVICE_ALLOC; goto cleanup; } magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queue ); dAT = dA + maxn * maxm; magmablas_stranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queue ); } // factor QR magma_sgeqrf2_gpu( n, m, dAT(0,0), lddat, tau, &iinfo ); assert( iinfo >= 0 ); if ( iinfo > 0 ) { *info = iinfo; } // conjugate tau #ifdef COMPLEX lapackf77_slacgv( &min_mn, tau, &ione ); #endif // undo transpose if (maxdim*maxdim < 2*maxm*maxn) { magmablas_stranspose_inplace( lddat, dAT(0,0), lddat, queue ); magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queue ); } else { magmablas_stranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queue ); magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queue ); } cleanup: magma_queue_destroy( queue ); magma_free( dA ); return *info; } /* magma_sgelqf */
/** Purpose ------- SLATRD reduces NB rows and columns of a real symmetric matrix A to symmetric tridiagonal form by an orthogonal similarity transformation Q' * A * Q, and returns the matrices V and W which are needed to apply the transformation to the unreduced part of A. If UPLO = MagmaUpper, SLATRD reduces the last NB rows and columns of a matrix, of which the upper triangle is supplied; if UPLO = MagmaLower, SLATRD reduces the first NB rows and columns of a matrix, of which the lower triangle is supplied. This is an auxiliary routine called by SSYTRD. Arguments --------- @param[in] uplo magma_uplo_t Specifies whether the upper or lower triangular part of the symmetric matrix A is stored: - = MagmaUpper: Upper triangular - = MagmaLower: Lower triangular @param[in] n INTEGER The order of the matrix A. @param[in] nb INTEGER The number of rows and columns to be reduced. @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, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading n-by-n lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. On exit: - if UPLO = MagmaUpper, the last NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements above the diagonal with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; - if UPLO = MagmaLower, the first NB columns have been reduced to tridiagonal form, with the diagonal elements overwriting the diagonal elements of A; the elements below the diagonal 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 >= (1,N). @param[out] e REAL array, dimension (N-1) If UPLO = MagmaUpper, E(n-nb:n-1) contains the superdiagonal elements of the last NB columns of the reduced matrix; if UPLO = MagmaLower, E(1:nb) contains the subdiagonal elements of the first NB columns of the reduced matrix. @param[out] tau REAL array, dimension (N-1) The scalar factors of the elementary reflectors, stored in TAU(n-nb:n-1) if UPLO = MagmaUpper, and in TAU(1:nb) if UPLO = MagmaLower. See Further Details. @param[out] W REAL array, dimension (LDW,NB) The n-by-nb matrix W required to update the unreduced part of A. @param[in] ldw INTEGER The leading dimension of the array W. LDW >= max(1,N). Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n) H(n-1) . . . H(n-nb+1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(i:n) = 0 and v(i-1) = 1; v(1:i-1) is stored on exit in A(1:i-1,i), and tau in TAU(i-1). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(nb). Each H(i) has the form H(i) = I - tau * v * v' where tau is a real scalar, and v is a real vector with v(1:i) = 0 and v(i+1) = 1; v(i+1:n) is stored on exit in A(i+1:n,i), and tau in TAU(i). The elements of the vectors v together form the n-by-nb matrix V which is needed, with W, to apply the transformation to the unreduced part of the matrix, using a symmetric rank-2k update of the form: A := A - V*W' - W*V'. The contents of A on exit are illustrated by the following examples with n = 5 and nb = 2: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( a a a v4 v5 ) ( d ) ( a a v4 v5 ) ( 1 d ) ( a 1 v5 ) ( v1 1 a ) ( d 1 ) ( v1 v2 a a ) ( d ) ( v1 v2 a a a ) where d denotes a diagonal element of the reduced matrix, a denotes an element of the original matrix that is unchanged, and vi denotes an element of the vector defining H(i). @ingroup magma_ssyev_aux ********************************************************************/ extern "C" magma_int_t magma_slatrd(magma_uplo_t uplo, magma_int_t n, magma_int_t nb, float *A, magma_int_t lda, float *e, float *tau, float *W, magma_int_t ldw, float *dA, magma_int_t ldda, float *dW, magma_int_t lddw) { #define A(i, j) (A + (j)*lda + (i)) #define W(i, j) (W + (j)*ldw + (i)) #define dA(i, j) (dA + (j)*ldda + (i)) #define dW(i, j) (dW + (j)*lddw + (i)) magma_int_t i; float c_neg_one = MAGMA_S_NEG_ONE; float c_one = MAGMA_S_ONE; float c_zero = MAGMA_S_ZERO; float value = MAGMA_S_ZERO; magma_int_t ione = 1; magma_int_t i_n, i_1, iw; float alpha; float *f; if (n <= 0) { return 0; } magma_queue_t stream; magma_queue_create( &stream ); magma_smalloc_cpu( &f, n ); assert( f != NULL ); // TODO return error, or allocate outside slatrd if (uplo == MagmaUpper) { /* Reduce last NB columns of upper triangle */ for (i = n-1; i >= n - nb; --i) { i_1 = i + 1; i_n = n - i - 1; iw = i - n + nb; if (i < n-1) { /* Update A(1:i,i) */ #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, W(i, iw+1), &ldw); #endif blasf77_sgemv("No transpose", &i_1, &i_n, &c_neg_one, A(0, i+1), &lda, W(i, iw+1), &ldw, &c_one, A(0, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, W(i, iw+1), &ldw); lapackf77_slacgv(&i_n, A(i, i+1), &lda); #endif blasf77_sgemv("No transpose", &i_1, &i_n, &c_neg_one, W(0, iw+1), &ldw, A(i, i+1), &lda, &c_one, A(0, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i_n, A(i, i+1), &lda); #endif } if (i > 0) { /* Generate elementary reflector H(i) to annihilate A(1:i-2,i) */ alpha = *A(i-1, i); lapackf77_slarfg(&i, &alpha, A(0, i), &ione, &tau[i - 1]); e[i-1] = MAGMA_S_REAL( alpha ); *A(i-1,i) = MAGMA_S_ONE; /* Compute W(1:i-1,i) */ // 1. Send the block reflector A(0:n-i-1,i) to the GPU magma_ssetvector( i, A(0, i), 1, dA(0, i), 1 ); magma_ssymv(MagmaUpper, i, c_one, dA(0, 0), ldda, dA(0, i), ione, c_zero, dW(0, iw), ione); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i, 1, dW(0, iw), lddw, W(0, iw) /*test*/, ldw, stream ); if (i < n-1) { blasf77_sgemv(MagmaTransStr, &i, &i_n, &c_one, W(0, iw+1), &ldw, A(0, i), &ione, &c_zero, W(i+1, iw), &ione); } // 3. Here is where we need it // TODO find the right place magma_queue_sync( stream ); if (i < n-1) { blasf77_sgemv("No transpose", &i, &i_n, &c_neg_one, A(0, i+1), &lda, W(i+1, iw), &ione, &c_one, W(0, iw), &ione); blasf77_sgemv(MagmaTransStr, &i, &i_n, &c_one, A(0, i+1), &lda, A(0, i), &ione, &c_zero, W(i+1, iw), &ione); blasf77_sgemv("No transpose", &i, &i_n, &c_neg_one, W(0, iw+1), &ldw, W(i+1, iw), &ione, &c_one, W(0, iw), &ione); } blasf77_sscal(&i, &tau[i - 1], W(0, iw), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) cblas_sdot_sub( i, W(0,iw), ione, A(0,i), ione, &value ); #else value = cblas_sdot( i, W(0,iw), ione, A(0,i), ione ); #endif alpha = tau[i - 1] * -0.5f * value; blasf77_saxpy(&i, &alpha, A(0, i), &ione, W(0, iw), &ione); } } } else { /* Reduce first NB columns of lower triangle */ for (i = 0; i < nb; ++i) { /* Update A(i:n,i) */ i_n = n - i; #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, W(i, 0), &ldw); #endif blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, A(i, 0), &lda, W(i, 0), &ldw, &c_one, A(i, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, W(i, 0), &ldw); lapackf77_slacgv(&i, A(i, 0), &lda); #endif blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, W(i, 0), &ldw, A(i, 0), &lda, &c_one, A(i, i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_slacgv(&i, A(i, 0), &lda); #endif if (i < n-1) { /* Generate elementary reflector H(i) to annihilate A(i+2:n,i) */ i_n = n - i - 1; alpha = *A(i+1, i); lapackf77_slarfg(&i_n, &alpha, A(min(i+2,n-1), i), &ione, &tau[i]); e[i] = MAGMA_S_REAL( alpha ); *A(i+1,i) = MAGMA_S_ONE; /* Compute W(i+1:n,i) */ // 1. Send the block reflector A(i+1:n,i) to the GPU magma_ssetvector( i_n, A(i+1, i), 1, dA(i+1, i), 1 ); magma_ssymv(MagmaLower, i_n, c_one, dA(i+1, i+1), ldda, dA(i+1, i), ione, c_zero, dW(i+1, i), ione); // 2. Start putting the result back (asynchronously) magma_sgetmatrix_async( i_n, 1, dW(i+1, i), lddw, W(i+1, i), ldw, stream ); blasf77_sgemv(MagmaTransStr, &i_n, &i, &c_one, W(i+1, 0), &ldw, A(i+1, i), &ione, &c_zero, W(0, i), &ione); blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, A(i+1, 0), &lda, W(0, i), &ione, &c_zero, f, &ione); blasf77_sgemv(MagmaTransStr, &i_n, &i, &c_one, A(i+1, 0), &lda, A(i+1, i), &ione, &c_zero, W(0, i), &ione); // 3. Here is where we need it magma_queue_sync( stream ); if (i != 0) blasf77_saxpy(&i_n, &c_one, f, &ione, W(i+1, i), &ione); blasf77_sgemv("No transpose", &i_n, &i, &c_neg_one, W(i+1, 0), &ldw, W(0, i), &ione, &c_one, W(i+1, i), &ione); blasf77_sscal(&i_n, &tau[i], W(i+1,i), &ione); #if defined(PRECISION_z) || defined(PRECISION_c) cblas_sdot_sub( i_n, W(i+1,i), ione, A(i+1,i), ione, &value ); #else value = cblas_sdot( i_n, W(i+1,i), ione, A(i+1,i), ione ); #endif alpha = tau[i] * -0.5f * value; blasf77_saxpy(&i_n, &alpha, A(i+1, i), &ione, W(i+1,i), &ione); } } } magma_free_cpu( f ); magma_queue_destroy( stream ); return 0; } /* magma_slatrd */