/** Purpose ------- DORMQR overwrites the general real M-by-N matrix C with @verbatim SIDE = MagmaLeft SIDE = MagmaRight TRANS = MagmaNoTrans: Q * C C * Q TRANS = MagmaTrans: Q**H * C C * Q**H @endverbatim where Q is a real unitary matrix defined as the product of k elementary reflectors Q = H(1) H(2) . . . H(k) as returned by DGEQRF. Q is of order M if SIDE = MagmaLeft and of order N if SIDE = MagmaRight. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @param[in] side magma_side_t - = MagmaLeft: apply Q or Q**H from the Left; - = MagmaRight: apply Q or Q**H from the Right. @param[in] trans magma_trans_t - = MagmaNoTrans: No transpose, apply Q; - = MagmaTrans: Conjugate transpose, apply Q**H. @param[in] m INTEGER The number of rows of the matrix C. M >= 0. @param[in] n INTEGER The number of columns of the matrix C. N >= 0. @param[in] k INTEGER The number of elementary reflectors whose product defines the matrix Q. If SIDE = MagmaLeft, M >= K >= 0; if SIDE = MagmaRight, N >= K >= 0. @param[in] A DOUBLE_PRECISION array, dimension (LDA,K) The i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by DGEQRF in the first k columns of its array argument A. @param[in] lda INTEGER The leading dimension of the array A. If SIDE = MagmaLeft, LDA >= max(1,M); if SIDE = MagmaRight, LDA >= max(1,N). @param[in] tau DOUBLE_PRECISION array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by DGEQRF. @param[in,out] C DOUBLE_PRECISION array, dimension (LDC,N) On entry, the M-by-N matrix C. On exit, C is overwritten by Q*C or Q**H*C or C*Q**H or C*Q. @param[in] ldc INTEGER The leading dimension of the array C. LDC >= max(1,M). @param[out] work (workspace) DOUBLE_PRECISION 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. If SIDE = MagmaLeft, LWORK >= max(1,N); if SIDE = MagmaRight, LWORK >= max(1,M). For optimum performance LWORK >= N*NB if SIDE = MagmaLeft, and LWORK >= M*NB if SIDE = MagmaRight, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value @ingroup magma_dgeqrf_comp ********************************************************************/ extern "C" magma_int_t magma_dormqr_m( magma_int_t ngpu, magma_side_t side, magma_trans_t trans, magma_int_t m, magma_int_t n, magma_int_t k, double *A, magma_int_t lda, double *tau, double *C, magma_int_t ldc, double *work, magma_int_t lwork, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define C(i, j) (C + (j)*ldc + (i)) #define dC(gpui, i, j) (dw[gpui] + (j)*lddc + (i)) #define dA_c(gpui, ind, i, j) (dw[gpui] + maxnlocal*lddc + (ind)*lddar*lddac + (i) + (j)*lddac) #define dA_r(gpui, ind, i, j) (dw[gpui] + maxnlocal*lddc + (ind)*lddar*lddac + (i) + (j)*lddar) #define dT(gpui, ind) (dw[gpui] + maxnlocal*lddc + 2*lddac*lddar + (ind)*((nb+1)*nb)) #define dwork(gpui, ind) (dw[gpui] + maxnlocal*lddc + 2*lddac*lddar + 2*((nb+1)*nb) + (ind)*(lddwork*nb)) double c_zero = MAGMA_D_ZERO; double c_one = MAGMA_D_ONE; const char* side_ = lapack_side_const( side ); const char* trans_ = lapack_trans_const( trans ); // TODO fix memory leak (alloc after argument checks) magma_int_t nb = 128; double *T; magma_dmalloc_pinned(&T, nb*nb); //printf("calling dormqr_m with nb=%d\n", (int) nb); double* dw[MagmaMaxGPUs]; magma_queue_t stream [MagmaMaxGPUs][2]; magma_event_t event [MagmaMaxGPUs][2]; magma_int_t ind_c; magma_device_t igpu; magma_device_t orig_dev; magma_getdevice( &orig_dev ); magma_queue_t orig_stream; magmablasGetKernelStream( &orig_stream ); *info = 0; magma_int_t left = (side == MagmaLeft); magma_int_t notran = (trans == MagmaNoTrans); magma_int_t lquery = (lwork == -1); /* NQ is the order of Q and NW is the minimum dimension of WORK */ magma_int_t nq, nw; if (left) { nq = m; nw = n; } else { nq = n; nw = m; } if (! left && side != MagmaRight) { *info = -1; } else if (! notran && trans != MagmaTrans) { *info = -2; } else if (m < 0) { *info = -3; } else if (n < 0) { *info = -4; } else if (k < 0 || k > nq) { *info = -5; } else if (lda < max(1,nq)) { *info = -7; } else if (ldc < max(1,m)) { *info = -10; } else if (lwork < max(1,nw) && ! lquery) { *info = -12; } magma_int_t lwkopt = max(1,nw) * nb; if (*info == 0) { work[0] = MAGMA_D_MAKE( lwkopt, 0 ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (m == 0 || n == 0 || k == 0) { work[0] = c_one; return *info; } if (nb >= k) { /* Use CPU code */ lapackf77_dormqr(side_, trans_, &m, &n, &k, A, &lda, tau, C, &ldc, work, &lwork, info); return *info; } magma_int_t lddc = (m+63)/64*64; magma_int_t lddac = nq; magma_int_t lddar = nb; magma_int_t lddwork = nw; magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 }; magma_int_t nb_l=256; magma_int_t nbl = (n-1)/nb_l+1; // number of blocks magma_int_t maxnlocal = (nbl+ngpu-1)/ngpu*nb_l; ngpu = min(ngpu, (n+nb_l-1)/nb_l); // Don't use GPU that will not have data. magma_int_t ldw = maxnlocal*lddc // dC + 2*lddac*lddar // 2*dA + 2*(nb + 1 + lddwork)*nb; // 2*(dT and dwork) for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); if (MAGMA_SUCCESS != magma_dmalloc( &dw[igpu], ldw )) { *info = MAGMA_ERR_DEVICE_ALLOC; magma_xerbla( __func__, -(*info) ); return *info; } magma_queue_create( &stream[igpu][0] ); magma_queue_create( &stream[igpu][1] ); magma_event_create( &event[igpu][0] ); magma_event_create( &event[igpu][1] ); } /* Use hybrid CPU-MGPU code */ if (left) { //copy C to mgpus for (magma_int_t i = 0; i < nbl; ++i) { magma_int_t igpu = i%ngpu; magma_setdevice(igpu); magma_int_t kb = min(nb_l, n-i*nb_l); magma_dsetmatrix_async( m, kb, C(0, i*nb_l), ldc, dC(igpu, 0, i/ngpu*nb_l), lddc, stream[igpu][0] ); nlocal[igpu] += kb; } magma_int_t i1, i2, i3; if ( !notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } ind_c = 0; for (magma_int_t i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { // start the copy of A panel magma_int_t kb = min(nb, k - i); for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); magma_event_sync(event[igpu][ind_c]); // check if the new data can be copied magma_dsetmatrix_async(nq-i, kb, A(i, i), lda, dA_c(igpu, ind_c, i, 0), lddac, stream[igpu][0] ); // set upper triangular part of dA to identity magmablas_dlaset_band_q( MagmaUpper, kb, kb, kb, c_zero, c_one, dA_c(igpu, ind_c, i, 0), lddac, stream[igpu][0] ); } /* Form the triangular factor of the block reflector H = H(i) H(i+1) . . . H(i+ib-1) */ magma_int_t nqi = nq - i; lapackf77_dlarft("F", "C", &nqi, &kb, A(i, i), &lda, &tau[i], T, &kb); /* H or H' is applied to C(1:m,i:n) */ /* Apply H or H'; First copy T to the GPU */ for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); magma_dsetmatrix_async(kb, kb, T, kb, dT(igpu, ind_c), kb, stream[igpu][0] ); } for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); magma_queue_sync( stream[igpu][0] ); // check if the data was copied magmablasSetKernelStream(stream[igpu][1]); magma_dlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, m-i, nlocal[igpu], kb, dA_c(igpu, ind_c, i, 0), lddac, dT(igpu, ind_c), kb, dC(igpu, i, 0), lddc, dwork(igpu, ind_c), lddwork); magma_event_record(event[igpu][ind_c], stream[igpu][1] ); } ind_c = (ind_c+1)%2; } for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); magma_queue_sync( stream[igpu][1] ); } //copy C from mgpus for (magma_int_t i = 0; i < nbl; ++i) { magma_int_t igpu = i%ngpu; magma_setdevice(igpu); magma_int_t kb = min(nb_l, n-i*nb_l); magma_dgetmatrix( m, kb, dC(igpu, 0, i/ngpu*nb_l), lddc, C(0, i*nb_l), ldc ); // magma_dgetmatrix_async( m, kb, // dC(igpu, 0, i/ngpu*nb_l), lddc, // C(0, i*nb_l), ldc, stream[igpu][0] ); } } else { // TODO fix memory leak T, dw, event, stream fprintf(stderr, "The case (side == right) is not implemented\n"); *info = MAGMA_ERR_NOT_IMPLEMENTED; magma_xerbla( __func__, -(*info) ); return *info; /* if ( notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } mi = m; ic = 0; for (i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { ib = min(nb, k - i); // Form the triangular factor of the block reflector // H = H(i) H(i+1) . . . H(i+ib-1) i__4 = nq - i; lapackf77_dlarft("F", "C", &i__4, &ib, A(i, i), &lda, &tau[i], T, &ib); // 1) copy the panel from A to the GPU, and // 2) set upper triangular part of dA to identity magma_dsetmatrix( i__4, ib, A(i, i), lda, dA(i, 0), ldda ); magmablas_dlaset_band( MagmaUpper, ib, ib, ib, c_zero, c_one, dA(i, 0), ldda ); // H or H' is applied to C(1:m,i:n) ni = n - i; jc = i; // Apply H or H'; First copy T to the GPU magma_dsetmatrix( ib, ib, T, ib, dT, ib ); magma_dlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, mi, ni, ib, dA(i, 0), ldda, dT, ib, dC(ic, jc), lddc, dwork, lddwork); } */ } work[0] = MAGMA_D_MAKE( lwkopt, 0 ); for (igpu = 0; igpu < ngpu; ++igpu) { magma_setdevice(igpu); magma_event_destroy( event[igpu][0] ); magma_event_destroy( event[igpu][1] ); magma_queue_destroy( stream[igpu][0] ); magma_queue_destroy( stream[igpu][1] ); magma_free( dw[igpu] ); } magma_setdevice( orig_dev ); magmablasSetKernelStream( orig_stream ); return *info; } /* magma_dormqr */
/** Purpose ------- ZUNMQR overwrites the general complex M-by-N matrix C with @verbatim SIDE = MagmaLeft SIDE = MagmaRight TRANS = MagmaNoTrans: Q * C C * Q TRANS = Magma_ConjTrans: Q**H * C C * Q**H @endverbatim where Q is a complex unitary matrix defined as the product of k elementary reflectors Q = H(1) H(2) . . . H(k) as returned by ZGEQRF. Q is of order M if SIDE = MagmaLeft and of order N if SIDE = MagmaRight. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @param[in] side magma_side_t - = MagmaLeft: apply Q or Q**H from the Left; - = MagmaRight: apply Q or Q**H from the Right. @param[in] trans magma_trans_t - = MagmaNoTrans: No transpose, apply Q; - = Magma_ConjTrans: Conjugate transpose, apply Q**H. @param[in] m INTEGER The number of rows of the matrix C. M >= 0. @param[in] n INTEGER The number of columns of the matrix C. N >= 0. @param[in] k INTEGER The number of elementary reflectors whose product defines the matrix Q. If SIDE = MagmaLeft, M >= K >= 0; if SIDE = MagmaRight, N >= K >= 0. @param[in] A COMPLEX_16 array, dimension (LDA,K) The i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by ZGEQRF in the first k columns of its array argument A. @param[in] lda INTEGER The leading dimension of the array A. If SIDE = MagmaLeft, LDA >= max(1,M); if SIDE = MagmaRight, LDA >= max(1,N). @param[in] tau COMPLEX_16 array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by ZGEQRF. @param[in,out] C COMPLEX_16 array, dimension (LDC,N) On entry, the M-by-N matrix C. On exit, C is overwritten by Q*C or Q**H*C or C*Q**H or C*Q. @param[in] ldc INTEGER The leading dimension of the array C. LDC >= max(1,M). @param[out] work (workspace) COMPLEX_16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. If SIDE = MagmaLeft, LWORK >= max(1,N); if SIDE = MagmaRight, LWORK >= max(1,M). For optimum performance LWORK >= N*NB if SIDE = MagmaLeft, and LWORK >= M*NB if SIDE = MagmaRight, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value @ingroup magma_zgeqrf_comp ********************************************************************/ extern "C" magma_int_t magma_zunmqr_m( magma_int_t ngpu, magma_side_t side, magma_trans_t trans, magma_int_t m, magma_int_t n, magma_int_t k, magmaDoubleComplex *A, magma_int_t lda, magmaDoubleComplex *tau, magmaDoubleComplex *C, magma_int_t ldc, magmaDoubleComplex *work, magma_int_t lwork, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define C(i, j) (C + (j)*ldc + (i)) #define dC(gpui, i, j) (dw[gpui] + (j)*lddc + (i)) #define dA_c(gpui, ind, i, j) (dw[gpui] + maxnlocal*lddc + (ind)*lddar*lddac + (i) + (j)*lddac) #define dA_r(gpui, ind, i, j) (dw[gpui] + maxnlocal*lddc + (ind)*lddar*lddac + (i) + (j)*lddar) #define dT(gpui, ind) (dw[gpui] + maxnlocal*lddc + 2*lddac*lddar + (ind)*((nb+1)*nb)) #define dwork(gpui, ind) (dw[gpui] + maxnlocal*lddc + 2*lddac*lddar + 2*((nb+1)*nb) + (ind)*(lddwork*nb)) /* Constants */ magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; /* Local variables */ const char* side_ = lapack_side_const( side ); const char* trans_ = lapack_trans_const( trans ); magma_int_t nb = 128; magmaDoubleComplex *T = NULL; magmaDoubleComplex_ptr dw[MagmaMaxGPUs] = { NULL }; magma_queue_t queues[MagmaMaxGPUs][2] = {{ NULL }}; magma_event_t events[MagmaMaxGPUs][2] = {{ NULL }}; magma_int_t ind_c; magma_device_t dev; magma_device_t orig_dev; magma_getdevice( &orig_dev ); *info = 0; magma_int_t left = (side == MagmaLeft); magma_int_t notran = (trans == MagmaNoTrans); magma_int_t lquery = (lwork == -1); /* NQ is the order of Q and NW is the minimum dimension of WORK */ magma_int_t nq, nw; if (left) { nq = m; nw = n; } else { nq = n; nw = m; } if (! left && side != MagmaRight) { *info = -1; } else if (! notran && trans != Magma_ConjTrans) { *info = -2; } else if (m < 0) { *info = -3; } else if (n < 0) { *info = -4; } else if (k < 0 || k > nq) { *info = -5; } else if (lda < max(1,nq)) { *info = -7; } else if (ldc < max(1,m)) { *info = -10; } else if (lwork < max(1,nw) && ! lquery) { *info = -12; } magma_int_t lwkopt = max(1,nw) * nb; if (*info == 0) { work[0] = magma_zmake_lwork( lwkopt ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (m == 0 || n == 0 || k == 0) { work[0] = c_one; return *info; } if (nb >= k) { /* Use CPU code */ lapackf77_zunmqr(side_, trans_, &m, &n, &k, A, &lda, tau, C, &ldc, work, &lwork, info); return *info; } magma_int_t lddc = magma_roundup( m, 64 ); // TODO why 64 instead of 32 ? magma_int_t lddac = nq; magma_int_t lddar = nb; magma_int_t lddwork = nw; magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 }; magma_int_t nb_l=256; magma_int_t nbl = magma_ceildiv( n, nb_l ); // number of blocks magma_int_t maxnlocal = magma_ceildiv( nbl, ngpu )*nb_l; ngpu = min( ngpu, magma_ceildiv( n, nb_l )); // Don't use GPU that will not have data. magma_int_t ldw = maxnlocal*lddc // dC + 2*lddac*lddar // 2*dA + 2*(nb + 1 + lddwork)*nb; // 2*(dT and dwork) if (MAGMA_SUCCESS != magma_zmalloc_pinned( &T, nb*nb )) { *info = MAGMA_ERR_HOST_ALLOC; goto cleanup; } for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); if (MAGMA_SUCCESS != magma_zmalloc( &dw[dev], ldw )) { *info = MAGMA_ERR_DEVICE_ALLOC; goto cleanup; } magma_queue_create( dev, &queues[dev][0] ); magma_queue_create( dev, &queues[dev][1] ); magma_event_create( &events[dev][0] ); magma_event_create( &events[dev][1] ); } /* Use hybrid CPU-MGPU code */ if (left) { //copy C to mgpus for (magma_int_t i = 0; i < nbl; ++i) { dev = i % ngpu; magma_setdevice( dev ); magma_int_t kb = min(nb_l, n-i*nb_l); magma_zsetmatrix_async( m, kb, C(0, i*nb_l), ldc, dC(dev, 0, i/ngpu*nb_l), lddc, queues[dev][0] ); nlocal[dev] += kb; } magma_int_t i1, i2, i3; if ( !notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } ind_c = 0; for (magma_int_t i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { // start the copy of A panel magma_int_t kb = min(nb, k - i); for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); magma_event_sync( events[dev][ind_c] ); // check if the new data can be copied magma_zsetmatrix_async(nq-i, kb, A(i, i), lda, dA_c(dev, ind_c, i, 0), lddac, queues[dev][0] ); // set upper triangular part of dA to identity magmablas_zlaset_band( MagmaUpper, kb, kb, kb, c_zero, c_one, dA_c(dev, ind_c, i, 0), lddac, queues[dev][0] ); } /* Form the triangular factor of the block reflector H = H(i) H(i+1) . . . H(i+ib-1) */ magma_int_t nqi = nq - i; lapackf77_zlarft("F", "C", &nqi, &kb, A(i, i), &lda, &tau[i], T, &kb); /* H or H' is applied to C(1:m,i:n) */ /* Apply H or H'; First copy T to the GPU */ for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); magma_zsetmatrix_async(kb, kb, T, kb, dT(dev, ind_c), kb, queues[dev][0] ); } for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); magma_queue_sync( queues[dev][0] ); // check if the data was copied magma_zlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, m-i, nlocal[dev], kb, dA_c(dev, ind_c, i, 0), lddac, dT(dev, ind_c), kb, dC(dev, i, 0), lddc, dwork(dev, ind_c), lddwork, queues[dev][1] ); magma_event_record(events[dev][ind_c], queues[dev][1] ); } ind_c = (ind_c+1)%2; } for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); magma_queue_sync( queues[dev][1] ); } //copy C from mgpus for (magma_int_t i = 0; i < nbl; ++i) { dev = i % ngpu; magma_setdevice( dev ); magma_int_t kb = min(nb_l, n-i*nb_l); magma_zgetmatrix( m, kb, dC(dev, 0, i/ngpu*nb_l), lddc, C(0, i*nb_l), ldc, queues[dev][1] ); // magma_zgetmatrix_async( m, kb, // dC(dev, 0, i/ngpu*nb_l), lddc, // C(0, i*nb_l), ldc, queues[dev][0] ); } } else { *info = MAGMA_ERR_NOT_IMPLEMENTED; magma_xerbla( __func__, -(*info) ); goto cleanup; /* if ( notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } mi = m; ic = 0; for (i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { ib = min(nb, k - i); // Form the triangular factor of the block reflector // H = H(i) H(i+1) . . . H(i+ib-1) i__4 = nq - i; lapackf77_zlarft("F", "C", &i__4, &ib, A(i, i), &lda, &tau[i], T, &ib); // 1) copy the panel from A to the GPU, and // 2) set upper triangular part of dA to identity magma_zsetmatrix( i__4, ib, A(i, i), lda, dA(i, 0), ldda, queues[dev][1] ); magmablas_zlaset_band( MagmaUpper, ib, ib, ib, c_zero, c_one, dA(i, 0), ldda, queues[dev][1] ); // H or H' is applied to C(1:m,i:n) ni = n - i; jc = i; // Apply H or H'; First copy T to the GPU magma_zsetmatrix( ib, ib, T, ib, dT, ib, queues[dev][1] ); magma_zlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, mi, ni, ib, dA(i, 0), ldda, dT, ib, dC(ic, jc), lddc, dwork, lddwork, queues[dev][1] ); } */ } cleanup: work[0] = magma_zmake_lwork( lwkopt ); for (dev = 0; dev < ngpu; ++dev) { magma_setdevice( dev ); magma_event_destroy( events[dev][0] ); magma_event_destroy( events[dev][1] ); magma_queue_destroy( queues[dev][0] ); magma_queue_destroy( queues[dev][1] ); magma_free( dw[dev] ); } magma_setdevice( orig_dev ); magma_free_pinned( T ); return *info; } /* magma_zunmqr */
extern "C" magma_int_t magma_dormqr_m(magma_int_t nrgpu, char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, double *a, magma_int_t lda, double *tau, double *c, magma_int_t ldc, double *work, magma_int_t lwork, magma_int_t *info) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= DORMQR overwrites the general real M-by-N matrix C with SIDE = 'L' SIDE = 'R' TRANS = 'N': Q * C C * Q TRANS = 'T': Q**T * C C * Q**T where Q is a real orthogonal matrix defined as the product of k elementary reflectors Q = H(1) H(2) . . . H(k) as returned by DGEQRF. Q is of order M if SIDE = 'L' and of order N if SIDE = 'R'. Arguments ========= SIDE (input) CHARACTER*1 = 'L': apply Q or Q**T from the Left; = 'R': apply Q or Q**T from the Right. TRANS (input) CHARACTER*1 = 'N': No transpose, apply Q; = 'T': Transpose, apply Q**T. M (input) INTEGER The number of rows of the matrix C. M >= 0. N (input) INTEGER The number of columns of the matrix C. N >= 0. K (input) INTEGER The number of elementary reflectors whose product defines the matrix Q. If SIDE = 'L', M >= K >= 0; if SIDE = 'R', N >= K >= 0. A (input) DOUBLE_PRECISION array, dimension (LDA,K) The i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by DGEQRF in the first k columns of its array argument A. LDA (input) INTEGER The leading dimension of the array A. If SIDE = 'L', LDA >= max(1,M); if SIDE = 'R', LDA >= max(1,N). TAU (input) DOUBLE_PRECISION array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by DGEQRF. C (input/output) DOUBLE_PRECISION array, dimension (LDC,N) On entry, the M-by-N matrix C. On exit, C is overwritten by Q*C or Q**T*C or C*Q**T or C*Q. LDC (input) INTEGER The leading dimension of the array C. LDC >= max(1,M). WORK (workspace/output) DOUBLE_PRECISION array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(0) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. If SIDE = 'L', LWORK >= max(1,N); if SIDE = 'R', LWORK >= max(1,M). For optimum performance LWORK >= N*NB if SIDE = 'L', and LWORK >= M*NB if SIDE = 'R', 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 ===================================================================== */ double c_one = MAGMA_D_ONE; char side_[2] = {side, 0}; char trans_[2] = {trans, 0}; magma_int_t nb = 128; double *t ; magma_dmalloc_pinned (&t, nb*nb); //printf("calling dormqr_m with nb=%d\n", (int) nb); double* dw[MagmaMaxGPUs]; magma_queue_t stream [MagmaMaxGPUs][2]; magma_event_t event [MagmaMaxGPUs][2]; magma_int_t ind_c; magma_int_t igpu = 0; int gpu_b; magma_getdevice(&gpu_b); *info = 0; magma_int_t left = lapackf77_lsame(side_, "L"); magma_int_t notran = lapackf77_lsame(trans_, "N"); magma_int_t lquery = (lwork == -1); /* NQ is the order of Q and NW is the minimum dimension of WORK */ magma_int_t nq, nw; if (left) { nq = m; nw = n; } else { nq = n; nw = m; } if (! left && ! lapackf77_lsame(side_, "R")) { *info = -1; } else if (! notran && ! lapackf77_lsame(trans_, "T")) { *info = -2; } else if (m < 0) { *info = -3; } else if (n < 0) { *info = -4; } else if (k < 0 || k > nq) { *info = -5; } else if (lda < max(1,nq)) { *info = -7; } else if (ldc < max(1,m)) { *info = -10; } else if (lwork < max(1,nw) && ! lquery) { *info = -12; } magma_int_t lwkopt = max(1,nw) * nb; if (*info == 0) { work[0] = MAGMA_D_MAKE( lwkopt, 0 ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (m == 0 || n == 0 || k == 0) { work[0] = c_one; return *info; } if (nb >= k) { /* Use CPU code */ lapackf77_dormqr(side_, trans_, &m, &n, &k, a, &lda, tau, c, &ldc, work, &lwork, info); return *info; } magma_int_t lddc = (m+63)/64*64; magma_int_t lddac = nq; magma_int_t lddar = nb; magma_int_t lddwork = nw; magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 }; magma_int_t nb_l=256; magma_int_t nbl = (n-1)/nb_l+1; // number of blocks magma_int_t maxnlocal = (nbl+nrgpu-1)/nrgpu*nb_l; nrgpu = min(nrgpu, (n+nb_l-1)/nb_l); // Don't use GPU that will not have data. magma_int_t ldw = maxnlocal*lddc // dC + 2*lddac*lddar // 2*dA + 2*(nb + 1 + lddwork)*nb; // 2*(dT and dwork) for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); if (MAGMA_SUCCESS != magma_dmalloc( &dw[igpu], ldw)) { magma_xerbla( __func__, -(*info) ); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_queue_create( &stream[igpu][0] ); magma_queue_create( &stream[igpu][1] ); magma_event_create( &event[igpu][0] ); magma_event_create( &event[igpu][1] ); } /* Use hybrid CPU-MGPU code */ if (left) { //copy C to mgpus for (magma_int_t i = 0; i < nbl; ++i){ magma_int_t igpu = i%nrgpu; magma_setdevice(igpu); magma_int_t kb = min(nb_l, n-i*nb_l); magma_dsetmatrix_async( m, kb, C(0, i*nb_l), ldc, dC(igpu, 0, i/nrgpu*nb_l), lddc, stream[igpu][0] ); nlocal[igpu] += kb; } magma_int_t i1, i2, i3; if ( !notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } ind_c = 0; for (magma_int_t i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { // start the copy of A panel magma_int_t kb = min(nb, k - i); for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_event_sync(event[igpu][ind_c]); // check if the new data can be copied magma_dsetmatrix_async(nq-i, kb, A(i, i), lda, dA_c(igpu, ind_c, i, 0), lddac, stream[igpu][0] ); // Put 0s in the upper triangular part of dA; magmablas_dsetdiag1subdiag0_stream('L', kb, kb, dA_c(igpu, ind_c, i, 0), lddac, stream[igpu][0]); } /* Form the triangular factor of the block reflector H = H(i) H(i+1) . . . H(i+ib-1) */ magma_int_t nqi = nq - i; lapackf77_dlarft("F", "C", &nqi, &kb, A(i, i), &lda, &tau[i], t, &kb); /* H or H' is applied to C(1:m,i:n) */ /* Apply H or H'; First copy T to the GPU */ for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_dsetmatrix_async(kb, kb, t, kb, dt(igpu, ind_c), kb, stream[igpu][0] ); } for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_queue_sync( stream[igpu][0] ); // check if the data was copied magmablasSetKernelStream(stream[igpu][1]); magma_dlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, m-i, nlocal[igpu], kb, dA_c(igpu, ind_c, i, 0), lddac, dt(igpu, ind_c), kb, dC(igpu, i, 0), lddc, dwork(igpu, ind_c), lddwork); magma_event_record(event[igpu][ind_c], stream[igpu][1] ); } ind_c = (ind_c+1)%2; } for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_queue_sync( stream[igpu][1] ); } //copy C from mgpus for (magma_int_t i = 0; i < nbl; ++i){ magma_int_t igpu = i%nrgpu; magma_setdevice(igpu); magma_int_t kb = min(nb_l, n-i*nb_l); magma_dgetmatrix( m, kb, dC(igpu, 0, i/nrgpu*nb_l), lddc, C(0, i*nb_l), ldc ); // magma_dgetmatrix_async( m, kb, // dC(igpu, 0, i/nrgpu*nb_l), lddc, // C(0, i*nb_l), ldc, stream[igpu][0] ); } } else { fprintf(stderr, "The case (side == right) is not implemented\n"); magma_xerbla( __func__, 1 ); return *info; /* if ( notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } mi = m; ic = 0; for (i = i1; (i3 < 0 ? i >= i2 : i < i2); i += i3) { ib = min(nb, k - i); // Form the triangular factor of the block reflector // H = H(i) H(i+1) . . . H(i+ib-1) i__4 = nq - i; lapackf77_dlarft("F", "C", &i__4, &ib, A(i, i), &lda, &tau[i], t, &ib); // 1) copy the panel from A to the GPU, and // 2) Put 0s in the upper triangular part of dA; magma_dsetmatrix( i__4, ib, A(i, i), lda, dA(i, 0), ldda ); magmablas_dsetdiag1subdiag0('L', ib, ib, dA(i, 0), ldda); // H or H' is applied to C(1:m,i:n) ni = n - i; jc = i; // Apply H or H'; First copy T to the GPU magma_dsetmatrix( ib, ib, t, ib, dt, ib ); magma_dlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, mi, ni, ib, dA(i, 0), ldda, dt, ib, dC(ic, jc), lddc, dwork, lddwork); } */ } work[0] = MAGMA_D_MAKE( lwkopt, 0 ); for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magmablasSetKernelStream(NULL); magma_event_destroy( event[igpu][0] ); magma_event_destroy( event[igpu][1] ); magma_queue_destroy( stream[igpu][0] ); magma_queue_destroy( stream[igpu][1] ); magma_free( dw[igpu] ); } magma_setdevice(gpu_b); return *info; } /* magma_dormqr */
extern "C" magma_int_t magma_zunmqr_m(magma_int_t nrgpu, char side, char trans, magma_int_t m, magma_int_t n, magma_int_t k, cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *tau, cuDoubleComplex *c, magma_int_t ldc, cuDoubleComplex *work, magma_int_t lwork, magma_int_t *info) { /* -- MAGMA (version 1.3.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver November 2012 Purpose ======= ZUNMQR overwrites the general complex M-by-N matrix C with SIDE = 'L' SIDE = 'R' TRANS = 'N': Q * C C * Q TRANS = 'T': Q**H * C C * Q**H where Q is a complex orthogonal matrix defined as the product of k elementary reflectors Q = H(1) H(2) . . . H(k) as returned by ZGEQRF. Q is of order M if SIDE = 'L' and of order N if SIDE = 'R'. Arguments ========= SIDE (input) CHARACTER*1 = 'L': apply Q or Q**H from the Left; = 'R': apply Q or Q**H from the Right. TRANS (input) CHARACTER*1 = 'N': No transpose, apply Q; = 'T': Transpose, apply Q**H. M (input) INTEGER The number of rows of the matrix C. M >= 0. N (input) INTEGER The number of columns of the matrix C. N >= 0. K (input) INTEGER The number of elementary reflectors whose product defines the matrix Q. If SIDE = 'L', M >= K >= 0; if SIDE = 'R', N >= K >= 0. A (input) COMPLEX_16 array, dimension (LDA,K) The i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by ZGEQRF in the first k columns of its array argument A. LDA (input) INTEGER The leading dimension of the array A. If SIDE = 'L', LDA >= max(1,M); if SIDE = 'R', LDA >= max(1,N). TAU (input) COMPLEX_16 array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by ZGEQRF. C (input/output) COMPLEX_16 array, dimension (LDC,N) On entry, the M-by-N matrix C. On exit, C is overwritten by Q*C or Q**H*C or C*Q**H or C*Q. LDC (input) INTEGER The leading dimension of the array C. LDC >= max(1,M). WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(0) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. If SIDE = 'L', LWORK >= max(1,N); if SIDE = 'R', LWORK >= max(1,M). For optimum performance LWORK >= N*NB if SIDE = 'L', and LWORK >= M*NB if SIDE = 'R', 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 ===================================================================== */ cuDoubleComplex c_one = MAGMA_Z_ONE; char side_[2] = {side, 0}; char trans_[2] = {trans, 0}; cuDoubleComplex* dw[MagmaMaxGPUs]; cudaStream_t stream [MagmaMaxGPUs][2]; magma_int_t ind_c, kb; magma_int_t i__4; magma_int_t i; cuDoubleComplex t[4160]; /* was [65][64] */ magma_int_t i1, i2, i3, ib, nb, nq, nw; magma_int_t left, notran, lquery; magma_int_t iinfo, lwkopt; magma_int_t igpu = 0; int gpu_b; magma_getdevice(&gpu_b); *info = 0; left = lapackf77_lsame(side_, "L"); notran = lapackf77_lsame(trans_, "N"); lquery = (lwork == -1); /* NQ is the order of Q and NW is the minimum dimension of WORK */ if (left) { nq = m; nw = n; } else { nq = n; nw = m; } if (! left && ! lapackf77_lsame(side_, "R")) { *info = -1; } else if (! notran && ! lapackf77_lsame(trans_, "T")) { *info = -2; } else if (m < 0) { *info = -3; } else if (n < 0) { *info = -4; } else if (k < 0 || k > nq) { *info = -5; } else if (lda < max(1,nq)) { *info = -7; } else if (ldc < max(1,m)) { *info = -10; } else if (lwork < max(1,nw) && ! lquery) { *info = -12; } if (*info == 0) { /* Determine the block size. NB may be at most NBMAX, where NBMAX is used to define the local array T. */ nb = 64; lwkopt = max(1,nw) * nb; MAGMA_Z_SET2REAL( work[0], lwkopt ); } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (m == 0 || n == 0 || k == 0) { work[0] = c_one; return *info; } magma_int_t lddc = (m+63)/64*64; magma_int_t lddac = nq; magma_int_t lddar =nb; magma_int_t lddwork = nw; magma_int_t n_l = (n+nrgpu-1)/nrgpu; // local n n_l = ((n_l+63)/64)*64; if (n_l<256) n_l=256; nrgpu = min(nrgpu, (n+n_l-1)/n_l); // Don't use GPU that will not have data. for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magmablasSetKernelStream(NULL); if (MAGMA_SUCCESS != magma_zmalloc( &dw[igpu], (n_l*lddc + 2*lddac*lddar + 2*(nb + 1 + lddwork)*nb))) { magma_xerbla( __func__, -(*info) ); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_queue_create( &stream[igpu][0] ); magma_queue_create( &stream[igpu][1] ); } if (nb >= k) { /* Use CPU code */ lapackf77_zunmqr(side_, trans_, &m, &n, &k, a, &lda, tau, c, &ldc, work, &lwork, &iinfo); } else { /* Use hybrid CPU-MGPU code */ if (left) { //copy C to mgpus for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); kb = min(n_l, n-igpu*n_l); magma_zsetmatrix_async( m, kb, C(0, igpu*n_l), ldc, dC(igpu, 0, 0), lddc, stream[igpu][0] ); } if ( !notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } kb = min(nb, k-i1); for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_zsetmatrix_async( nq-i1, kb, A(i1, i1), lda, dA_c(igpu, 0, i1, 0), lddac, stream[igpu][0] ); } ind_c = 0; for (i = i1; i3 < 0 ? i >= i2 : i < i2; i += i3) { ib = min(nb, k - i); /* Form the triangular factor of the block reflector H = H(i) H(i+1) . . . H(i+ib-1) */ i__4 = nq - i; lapackf77_zlarft("F", "C", &i__4, &ib, A(i, i), &lda, &tau[i], t, &ib); /* H or H' is applied to C(1:m,i:n) */ /* Apply H or H'; First copy T to the GPU */ for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_zsetmatrix_async( ib, ib, t, ib, dt(igpu, ind_c), ib, stream[igpu][ind_c] ); magma_queue_sync( stream[igpu][ind_c] ); // Makes sure that we can change t next iteration. } // start the copy of next A panel kb = min(nb, k - i - i3); if (kb > 0 && i+i3 >= 0){ for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_zsetmatrix_async( nq-(i+i3), kb, A(i+i3, i+i3), lda, dA_c(igpu, (ind_c+1)%2, i+i3, 0), lddac, stream[igpu][(ind_c+1)%2] ); } } for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); // Put 0s in the upper triangular part of dA; magmablas_zsetdiag1subdiag0_stream('L', ib, ib, dA_c(igpu, ind_c, i, 0), lddac, stream[igpu][ind_c]); kb = min(n_l, n-igpu*n_l); magmablasSetKernelStream(stream[igpu][ind_c]); magma_zlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, m-i, kb, ib, dA_c(igpu, ind_c, i, 0), lddac, dt(igpu, ind_c), ib, dC(igpu, i, 0), lddc, dwork(igpu, ind_c), lddwork); } ind_c = (ind_c+1)%2; } //copy C from mgpus for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_queue_sync( stream[igpu][0] ); magma_queue_sync( stream[igpu][1] ); kb = min(n_l, n-igpu*n_l); //asynchronous copy gives problems sometimes... // magma_zgetmatrix_async( m, kb, // dC(igpu, 0, 0), lddc, // C(0, igpu*n_l), ldc, stream[igpu][0] ); magma_zgetmatrix( m, kb, dC(igpu, 0, 0), lddc, C(0, igpu*n_l), ldc ); } } else { fprintf(stderr, "The case (side == right) is not implemented\n"); magma_xerbla( __func__, 1 ); return *info; /*if ( notran ) { i1 = 0; i2 = k; i3 = nb; } else { i1 = (k - 1) / nb * nb; i2 = 0; i3 = -nb; } mi = m; ic = 0; for (i = i1; i3 < 0 ? i >= i2 : i < i2; i += i3) { ib = min(nb, k - i); // Form the triangular factor of the block reflector // H = H(i) H(i+1) . . . H(i+ib-1) i__4 = nq - i; lapackf77_zlarft("F", "C", &i__4, &ib, A(i, i), &lda, &tau[i], t, &ib); // 1) copy the panel from A to the GPU, and // 2) Put 0s in the upper triangular part of dA; magma_zsetmatrix( i__4, ib, A(i, i), lda, dA(i, 0), ldda ); magmablas_zsetdiag1subdiag0('L', ib, ib, dA(i, 0), ldda); // H or H' is applied to C(1:m,i:n) ni = n - i; jc = i; // Apply H or H'; First copy T to the GPU magma_zsetmatrix( ib, ib, t, ib, dt, ib ); magma_zlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise, mi, ni, ib, dA(i, 0), ldda, dt, ib, dC(ic, jc), lddc, dwork, lddwork); } */ } } MAGMA_Z_SET2REAL( work[0], lwkopt ); for (igpu = 0; igpu < nrgpu; ++igpu){ magma_setdevice(igpu); magma_queue_sync( stream[igpu][0] ); magmablasSetKernelStream(NULL); magma_queue_destroy( stream[igpu][0] ); magma_queue_destroy( stream[igpu][1] ); magma_free( dw[igpu] ); } magma_setdevice(gpu_b); return *info; } /* magma_zunmqr */