CCudaTimeMeasure::~CCudaTimeMeasure() { cudaCheckError(cudaEventDestroy(m_ceStartEvent)); cudaCheckError(cudaEventDestroy(m_ceStopEvent)); }
ProfilerDPD::~ProfilerDPD() { CUDA_CHECK(cudaEventDestroy(evstart)); CUDA_CHECK(cudaEventDestroy(evforce)); }
cudaTimer::~cudaTimer(){ cudaError_t status; status = cudaEventDestroy(inicio); status = cudaEventDestroy(fin); }
/** * free allocated memory */ static void destroy(const CudaEvent& ev) { CUDA_CHECK(cudaEventSynchronize(ev.event)); CUDA_CHECK(cudaEventDestroy(ev.event)); }
// virtual libcgt::cuda::Event::~Event() { cudaEventDestroy( m_stop ); cudaEventDestroy( m_start ); }
TimerGPU::~TimerGPU() { cudaEventDestroy(start_); cudaEventDestroy(stop_); }
extern "C" magma_int_t magma_ssytrd_sy2sb( char uplo, magma_int_t n, magma_int_t nb, float *a, magma_int_t lda, float *tau, float *work, magma_int_t lwork, float *dT, magma_int_t threads, magma_int_t *info) { /* -- MAGMA (version 1.3.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver November 2012 Purpose ======= SSYTRD_HE2HB reduces a real symmetric matrix A to real symmetric band-diagonal form T by an orthogonal similarity transformation: Q**T * A * Q = T. This version stores the triangular matrices T used in the accumulated Householder transformations (I - V T V'). Arguments ========= UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) REAL array, dimension (LDA,N) On entry, the symmetric matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, 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 Upper band-diagonal of A is overwritten by the corresponding elements of the band-diagonal matrix T, and the elements above the band diagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = 'L', the the Lower band-diagonal of A is overwritten by the corresponding elements of the band-diagonal matrix T, and the elements below the band-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 >= max(1,N). TAU (output) REAL array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). WORK (workspace/output) REAL array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. LWORK >= 1. For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. dT (output) REAL array on the GPU, dimension N*NB, where NB is the optimal blocksize. On exit dT holds the upper triangular matrices T from the accumulated Householder transformations (I - V T V') used in the factorization. The nb x nb matrices T are ordered consecutively in memory one after another. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value Further Details =============== If UPLO = 'U', the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(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+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = 'L', the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-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(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = 'U': if UPLO = 'L': ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). ===================================================================== */ #define a_ref(a_1,a_2) ( a + ((a_2)-1)*( lda) + (a_1)-1) #define da_ref(a_1,a_2) (da + ((a_2)-1)*(ldda) + (a_1)-1) #define tau_ref(a_1) (tau + (a_1)-1) #define t_ref(a_1) (dT + ((a_1)-1)*(lddt)) char uplo_[2] = {uplo, 0}; int ldda = ((n+31)/32)*32; int lddt = nb; float c_neg_one = MAGMA_S_NEG_ONE; float c_neg_half = MAGMA_S_NEG_HALF; float c_one = MAGMA_S_ONE ; float c_zero = MAGMA_S_ZERO; float d_one = MAGMA_D_ONE; magma_int_t pm, pn, indi, indj, pk; magma_int_t pm_old=0, pn_old=0, indi_old=0, indj_old=0; int i; int lwkopt; int lquery; *info = 0; int upper = lapackf77_lsame(uplo_, "U"); lquery = lwork == -1; if (! upper && ! lapackf77_lsame(uplo_, "L")) { *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,n)) { *info = -4; } else if (lwork < 1 && ! lquery) { *info = -9; } if (*info == 0) { /* Determine the block size. */ lwkopt = n * nb; MAGMA_S_SET2REAL( work[0], lwkopt ); } if (*info != 0) return *info; else if (lquery) return *info; /* Quick return if possible */ if (n == 0) { work[0] = c_one; return *info; } float *da; if (MAGMA_SUCCESS != magma_smalloc( &da, (n + 2*nb)*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_int_t mklth = min(threads,12); #if defined(USEMKL) mkl_set_num_threads(mklth); #endif #if defined(USEACML) omp_set_num_threads(mklth); #endif /* Use the first panel of da as work space */ float *dwork = da+n*ldda; float *dW = dwork + nb*ldda; #ifdef TRACING char buf[80]; #endif cudaStream_t stream[3]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); stream[2] = 0; // default stream trace_init( 1, 1, 3, stream ); float *hT = work + lwork - nb*nb; lwork -= nb*nb; memset( hT, 0, nb*nb*sizeof(float)); magmablasSetKernelStream( stream[0] ); cudaEvent_t Pupdate_event; cudaEventCreateWithFlags(&Pupdate_event,cudaEventDisableTiming); //cudaEventCreate(&Pupdate_event); if (upper) { printf("SSYTRD_HE2HB is not yet implemented for upper matrix storage. Exit.\n"); exit(1); }else { /* Copy the matrix to the GPU */ if (1 <= n-nb){ trace_gpu_start( 0, 0, "set", "set A" ); magma_ssetmatrix_async( (n-nb), (n-nb), a_ref(nb+1, nb+1), lda, da_ref(nb+1, nb+1), ldda, stream[0] ); trace_gpu_end( 0, 0 ); } /* Reduce the lower triangle of A */ for (i = 1; i <= n-nb; i += nb) { indi = i+nb; indj = i; pm = n - i - nb + 1; //pn = min(i+nb-1, n-nb) -i + 1; pn = nb; /* Get the current panel (no need for the 1st iteration) */ if (i > 1 ){ // spanel_to_q copy the upper oof diagonal part of // the matrix to work to be restored later. acctually // the zero's and one's putted are not used this is only // because we don't have a function that copy only the // upper part of A to be restored after copying the // lookahead panel that has been computted from GPU to CPU. spanel_to_q(MagmaUpper, pn-1, a_ref(i, i+1), lda, work); trace_gpu_start( 0, 1, "get", "get panel" ); //magma_queue_sync( stream[0] ); cudaStreamWaitEvent(stream[1], Pupdate_event, 0); magma_sgetmatrix_async( (pm+pn), pn, da_ref( i, i), ldda, a_ref ( i, i), lda, stream[1] ); trace_gpu_end( 0, 1 ); trace_gpu_start( 0, 2, "syr2k", "syr2k" ); magma_ssyr2k(MagmaLower, MagmaNoTrans, pm_old-pn_old, pn_old, c_neg_one, da_ref(indi_old+pn_old, indj_old), ldda, dW + pn_old , pm_old, d_one, da_ref(indi_old+pn_old, indi_old+pn_old), ldda); trace_gpu_end( 0, 2 ); trace_cpu_start( 0, "sync", "sync on 1" ); magma_queue_sync( stream[1] ); trace_cpu_end( 0 ); sq_to_panel(MagmaUpper, pn-1, a_ref(i, i+1), lda, work); } /* ========================================================== QR factorization on a panel starting nb off of the diagonal. Prepare the V and T matrices. ========================================================== */ #ifdef TRACING snprintf( buf, sizeof(buf), "panel %d", i ); #endif trace_cpu_start( 0, "geqrf", buf ); lapackf77_sgeqrf(&pm, &pn, a_ref(indi, indj), &lda, tau_ref(i), work, &lwork, info); /* Form the matrix T */ pk=min(pm,pn); lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr, &pm, &pk, a_ref(indi, indj), &lda, tau_ref(i), hT, &nb); /* Prepare V - put 0s in the upper triangular part of the panel (and 1s on the diagonal), temporaly storing the original in work */ spanel_to_q(MagmaUpper, pk, a_ref(indi, indj), lda, work); trace_cpu_end( 0 ); /* Send V from the CPU to the GPU */ trace_gpu_start( 0, 0, "set", "set V and T" ); magma_ssetmatrix_async( pm, pk, a_ref(indi, indj), lda, da_ref(indi, indj), ldda, stream[0] ); /* Send the triangular factor T to the GPU */ magma_ssetmatrix_async( pk, pk, hT, nb, t_ref(i), lddt, stream[0] ); trace_gpu_end( 0, 0 ); /* ========================================================== Compute W: 1. X = A (V T) 2. W = X - 0.5* V * (T' * (V' * X)) ========================================================== */ /* dwork = V T */ trace_cpu_start( 0, "sync", "sync on 0" ); // this sync is done here to be sure that the copy has been finished // because below we made a restore sq_to_panel and this restore need // to ensure that the copy has been finished. we did it here to allow // overlapp of restore with next gemm and symm. magma_queue_sync( stream[0] ); trace_cpu_end( 0 ); trace_gpu_start( 0, 2, "gemm", "work = V*T" ); magma_sgemm(MagmaNoTrans, MagmaNoTrans, pm, pk, pk, c_one, da_ref(indi, indj), ldda, t_ref(i), lddt, c_zero, dwork, pm); trace_gpu_end( 0, 2 ); /* dW = X = A*V*T. dW = A*dwork */ trace_gpu_start( 0, 2, "symm", "X = A*work" ); magma_ssymm(MagmaLeft, uplo, pm, pk, c_one, da_ref(indi, indi), ldda, dwork, pm, c_zero, dW, pm); trace_gpu_end( 0, 2 ); /* restore the panel */ sq_to_panel(MagmaUpper, pk, a_ref(indi, indj), lda, work); /* dwork = V*T already ==> dwork' = T'*V' * compute T'*V'*X ==> dwork'*W ==> * dwork + pm*nb = ((T' * V') * X) = dwork' * X = dwork' * W */ trace_gpu_start( 0, 2, "gemm", "work = T'*V'*X" ); magma_sgemm(MagmaTrans, MagmaNoTrans, pk, pk, pm, c_one, dwork, pm, dW, pm, c_zero, dwork + pm*nb, nb); trace_gpu_end( 0, 2 ); /* W = X - 0.5 * V * T'*V'*X * = X - 0.5 * V * (dwork + pm*nb) = W - 0.5 * V * (dwork + pm*nb) */ trace_gpu_start( 0, 2, "gemm", "W = X - 0.5*V*(T'*V'*X)" ); magma_sgemm(MagmaNoTrans, MagmaNoTrans, pm, pk, pk, c_neg_half, da_ref(indi, indj), ldda, dwork + pm*nb, nb, c_one, dW, pm); trace_gpu_end( 0, 2 ); /* ========================================================== Update the unreduced submatrix A(i+ib:n,i+ib:n), using an update of the form: A := A - V*W' - W*V' ========================================================== */ if (i + nb <= n-nb){ /* There would be next iteration; do lookahead - update the next panel */ trace_gpu_start( 0, 2, "gemm", "gemm 4 next panel left" ); magma_sgemm(MagmaNoTrans, MagmaTrans, pm, pn, pn, c_neg_one, da_ref(indi, indj), ldda, dW , pm, c_one, da_ref(indi, indi), ldda); trace_gpu_end( 0, 2 ); trace_gpu_start( 0, 2, "gemm", "gemm 5 next panel right" ); magma_sgemm(MagmaNoTrans, MagmaTrans, pm, pn, pn, c_neg_one, dW , pm, da_ref(indi, indj), ldda, c_one, da_ref(indi, indi), ldda); trace_gpu_end( 0, 2 ); cudaEventRecord(Pupdate_event, stream[0]); } else { /* no look-ahead as this is last iteration */ trace_gpu_start( 0, 2, "syr2k", "syr2k last iteration" ); magma_ssyr2k(MagmaLower, MagmaNoTrans, pk, pk, c_neg_one, da_ref(indi, indj), ldda, dW , pm, d_one, da_ref(indi, indi), ldda); trace_gpu_end( 0, 2 ); } indi_old = indi; indj_old = indj; pm_old = pm; pn_old = pn; } // end loop for(i) /* Send the last block to the CPU */ pk = min(pm,pn); if (1 <= n-nb){ spanel_to_q(MagmaUpper, pk-1, a_ref(n-pk+1, n-pk+2), lda, work); trace_gpu_start( 0, 2, "get", "get last block" ); magma_sgetmatrix( pk, pk, da_ref(n-pk+1, n-pk+1), ldda, a_ref(n-pk+1, n-pk+1), lda ); trace_gpu_end( 0, 2 ); sq_to_panel(MagmaUpper, pk-1, a_ref(n-pk+1, n-pk+2), lda, work); } }// end of LOWER trace_finalize( "ssytrd_sy2sb.svg", "trace.css" ); cudaEventDestroy(Pupdate_event); magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free( da ); MAGMA_S_SET2REAL( work[0], lwkopt ); magmablasSetKernelStream( 0 ); #if defined(USEMKL) mkl_set_num_threads(1); #endif #if defined(USEACML) omp_set_num_threads(1); #endif return *info; } /* ssytrd_sy2sb_ */
/** Purpose ------- CHETRD_HE2HB reduces a complex Hermitian matrix A to real symmetric band-diagonal form T by an orthogonal similarity transformation: Q**H * A * Q = T. This version stores the triangular matrices T used in the accumulated Householder transformations (I - V T V'). Arguments --------- @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX array, dimension (LDA,N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, 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 Upper band-diagonal of A is overwritten by the corresponding elements of the band-diagonal matrix T, and the elements above the band diagonal, with the array TAU, represent the orthogonal matrix Q as a product of elementary reflectors; if UPLO = MagmaLower, the the Lower band-diagonal of A is overwritten by the corresponding elements of the band-diagonal matrix T, and the elements below the band-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 >= max(1,N). @param[out] tau COMPLEX array, dimension (N-1) The scalar factors of the elementary reflectors (see Further Details). @param[out] work (workspace) COMPLEX array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. @param[in] lwork INTEGER The dimension of the array WORK. LWORK >= 1. For optimum performance LWORK >= N*NB, where NB is the optimal blocksize. \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. @param[out] dT COMPLEX array on the GPU, dimension N*NB, where NB is the optimal blocksize. On exit dT holds the upper triangular matrices T from the accumulated Householder transformations (I - V T V') used in the factorization. The nb x nb matrices T are ordered consecutively in memory one after another. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value Further Details --------------- If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary reflectors Q = H(n-1) . . . H(2) H(1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in A(1:i-1,i+1), and tau in TAU(i). If UPLO = MagmaLower, the matrix Q is represented as a product of elementary reflectors Q = H(1) H(2) . . . H(n-1). Each H(i) has the form H(i) = I - tau * v * v' where tau is a complex scalar, and v is a complex vector with v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i), and tau in TAU(i). The contents of A on exit are illustrated by the following examples with n = 5: if UPLO = MagmaUpper: if UPLO = MagmaLower: ( d e v2 v3 v4 ) ( d ) ( d e v3 v4 ) ( e d ) ( d e v4 ) ( v1 e d ) ( d e ) ( v1 v2 e d ) ( d ) ( v1 v2 v3 e d ) where d and e denote diagonal and off-diagonal elements of T, and vi denotes an element of the vector defining H(i). @ingroup magma_cheev_2stage ********************************************************************/ extern "C" magma_int_t magma_chetrd_he2hb( magma_uplo_t uplo, magma_int_t n, magma_int_t nb, magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *tau, magmaFloatComplex *work, magma_int_t lwork, magmaFloatComplex *dT, magma_int_t *info) { #define A(a_1,a_2) ( A + ((a_2)-1)*( lda) + (a_1)-1) #define dA(a_1,a_2) (dA + ((a_2)-1)*(ldda) + (a_1)-1) #define tau_ref(a_1) (tau + (a_1)-1) #define dT(a_1) (dT + ((a_1)-1)*(lddt)) int ldda = ((n+31)/32)*32; int lddt = nb; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex c_neg_half = MAGMA_C_NEG_HALF; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; float d_one = MAGMA_D_ONE; magma_int_t pm, pn, indi, indj, pk; magma_int_t pm_old=0, pn_old=0, indi_old=0, indj_old=0; int i; int lwkopt; int lquery; *info = 0; int upper = (uplo == MagmaUpper); lquery = (lwork == -1); if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,n)) { *info = -4; } else if (lwork < 1 && ! lquery) { *info = -9; } /* Determine the block size. */ lwkopt = n * nb; if (*info == 0) { work[0] = MAGMA_C_MAKE( lwkopt, 0 ); } if (*info != 0) return *info; else if (lquery) return *info; /* Quick return if possible */ if (n == 0) { work[0] = c_one; return *info; } magmaFloatComplex *dA; if (MAGMA_SUCCESS != magma_cmalloc( &dA, (n + 2*nb)*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_int_t threads = magma_get_lapack_numthreads(); magma_int_t mklth = min(threads,16); magma_set_lapack_numthreads(mklth); /* Use the first panel of dA as work space */ magmaFloatComplex *dwork = dA + n*ldda; magmaFloatComplex *dW = dwork + nb*ldda; #ifdef TRACING char buf[80]; #endif magma_queue_t stream[3]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); stream[2] = 0; // default stream trace_init( 1, 1, 3, stream ); magmaFloatComplex *hT = work + lwork - nb*nb; lwork -= nb*nb; memset( hT, 0, nb*nb*sizeof(magmaFloatComplex)); magmablasSetKernelStream( stream[0] ); magma_event_t Pupdate_event; cudaEventCreateWithFlags(&Pupdate_event,cudaEventDisableTiming); //cudaEventCreate(&Pupdate_event); if (upper) { printf("CHETRD_HE2HB is not yet implemented for upper matrix storage. Exit.\n"); exit(1); } else { /* Copy the matrix to the GPU */ if (1 <= n-nb) { trace_gpu_start( 0, 0, "set", "set A" ); magma_csetmatrix_async( (n-nb), (n-nb), A(nb+1, nb+1), lda, dA(nb+1, nb+1), ldda, stream[0] ); trace_gpu_end( 0, 0 ); } /* Reduce the lower triangle of A */ for (i = 1; i <= n-nb; i += nb) { indi = i+nb; indj = i; pm = n - i - nb + 1; //pn = min(i+nb-1, n-nb) -i + 1; pn = nb; /* Get the current panel (no need for the 1st iteration) */ if (i > 1 ) { // cpanel_to_q copy the upper oof diagonal part of // the matrix to work to be restored later. acctually // the zero's and one's putted are not used this is only // because we don't have a function that copy only the // upper part of A to be restored after copying the // lookahead panel that has been computted from GPU to CPU. cpanel_to_q(MagmaUpper, pn-1, A(i, i+1), lda, work); trace_gpu_start( 0, 1, "get", "get panel" ); //magma_queue_sync( stream[0] ); cudaStreamWaitEvent(stream[1], Pupdate_event, 0); magma_cgetmatrix_async( (pm+pn), pn, dA( i, i), ldda, A ( i, i), lda, stream[1] ); trace_gpu_end( 0, 1 ); trace_gpu_start( 0, 2, "her2k", "her2k" ); magma_cher2k(MagmaLower, MagmaNoTrans, pm_old-pn_old, pn_old, c_neg_one, dA(indi_old+pn_old, indj_old), ldda, dW + pn_old, pm_old, d_one, dA(indi_old+pn_old, indi_old+pn_old), ldda); trace_gpu_end( 0, 2 ); trace_cpu_start( 0, "sync", "sync on 1" ); magma_queue_sync( stream[1] ); trace_cpu_end( 0 ); cq_to_panel(MagmaUpper, pn-1, A(i, i+1), lda, work); } /* ========================================================== QR factorization on a panel starting nb off of the diagonal. Prepare the V and T matrices. ========================================================== */ #ifdef TRACING snprintf( buf, sizeof(buf), "panel %d", i ); #endif trace_cpu_start( 0, "geqrf", buf ); lapackf77_cgeqrf(&pm, &pn, A(indi, indj), &lda, tau_ref(i), work, &lwork, info); /* Form the matrix T */ pk=min(pm,pn); lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr, &pm, &pk, A(indi, indj), &lda, tau_ref(i), hT, &nb); /* Prepare V - put 0s in the upper triangular part of the panel (and 1s on the diagonal), temporaly storing the original in work */ cpanel_to_q(MagmaUpper, pk, A(indi, indj), lda, work); trace_cpu_end( 0 ); /* Send V from the CPU to the GPU */ trace_gpu_start( 0, 0, "set", "set V and T" ); magma_csetmatrix_async( pm, pk, A(indi, indj), lda, dA(indi, indj), ldda, stream[0] ); /* Send the triangular factor T to the GPU */ magma_csetmatrix_async( pk, pk, hT, nb, dT(i), lddt, stream[0] ); trace_gpu_end( 0, 0 ); /* ========================================================== Compute W: 1. X = A (V T) 2. W = X - 0.5* V * (T' * (V' * X)) ========================================================== */ /* dwork = V T */ trace_cpu_start( 0, "sync", "sync on 0" ); // this sync is done here to be sure that the copy has been finished // because below we made a restore cq_to_panel and this restore need // to ensure that the copy has been finished. we did it here to allow // overlapp of restore with next gemm and symm. magma_queue_sync( stream[0] ); trace_cpu_end( 0 ); trace_gpu_start( 0, 2, "gemm", "work = V*T" ); magma_cgemm(MagmaNoTrans, MagmaNoTrans, pm, pk, pk, c_one, dA(indi, indj), ldda, dT(i), lddt, c_zero, dwork, pm); trace_gpu_end( 0, 2 ); /* dW = X = A*V*T. dW = A*dwork */ trace_gpu_start( 0, 2, "hemm", "X = A*work" ); magma_chemm(MagmaLeft, uplo, pm, pk, c_one, dA(indi, indi), ldda, dwork, pm, c_zero, dW, pm); trace_gpu_end( 0, 2 ); /* restore the panel */ cq_to_panel(MagmaUpper, pk, A(indi, indj), lda, work); /* dwork = V*T already ==> dwork' = T'*V' * compute T'*V'*X ==> dwork'*W ==> * dwork + pm*nb = ((T' * V') * X) = dwork' * X = dwork' * W */ trace_gpu_start( 0, 2, "gemm", "work = T'*V'*X" ); magma_cgemm(MagmaConjTrans, MagmaNoTrans, pk, pk, pm, c_one, dwork, pm, dW, pm, c_zero, dwork + pm*nb, nb); trace_gpu_end( 0, 2 ); /* W = X - 0.5 * V * T'*V'*X * = X - 0.5 * V * (dwork + pm*nb) = W - 0.5 * V * (dwork + pm*nb) */ trace_gpu_start( 0, 2, "gemm", "W = X - 0.5*V*(T'*V'*X)" ); magma_cgemm(MagmaNoTrans, MagmaNoTrans, pm, pk, pk, c_neg_half, dA(indi, indj), ldda, dwork + pm*nb, nb, c_one, dW, pm); trace_gpu_end( 0, 2 ); /* ========================================================== Update the unreduced submatrix A(i+ib:n,i+ib:n), using an update of the form: A := A - V*W' - W*V' ========================================================== */ if (i + nb <= n-nb) { /* There would be next iteration; do lookahead - update the next panel */ trace_gpu_start( 0, 2, "gemm", "gemm 4 next panel left" ); magma_cgemm(MagmaNoTrans, MagmaConjTrans, pm, pn, pn, c_neg_one, dA(indi, indj), ldda, dW, pm, c_one, dA(indi, indi), ldda); trace_gpu_end( 0, 2 ); trace_gpu_start( 0, 2, "gemm", "gemm 5 next panel right" ); magma_cgemm(MagmaNoTrans, MagmaConjTrans, pm, pn, pn, c_neg_one, dW, pm, dA(indi, indj), ldda, c_one, dA(indi, indi), ldda); trace_gpu_end( 0, 2 ); cudaEventRecord(Pupdate_event, stream[0]); } else { /* no look-ahead as this is last iteration */ trace_gpu_start( 0, 2, "her2k", "her2k last iteration" ); magma_cher2k(MagmaLower, MagmaNoTrans, pk, pk, c_neg_one, dA(indi, indj), ldda, dW, pm, d_one, dA(indi, indi), ldda); trace_gpu_end( 0, 2 ); } indi_old = indi; indj_old = indj; pm_old = pm; pn_old = pn; } // end loop for (i) /* Send the last block to the CPU */ pk = min(pm,pn); if (1 <= n-nb) { cpanel_to_q(MagmaUpper, pk-1, A(n-pk+1, n-pk+2), lda, work); trace_gpu_start( 0, 2, "get", "get last block" ); magma_cgetmatrix( pk, pk, dA(n-pk+1, n-pk+1), ldda, A(n-pk+1, n-pk+1), lda ); trace_gpu_end( 0, 2 ); cq_to_panel(MagmaUpper, pk-1, A(n-pk+1, n-pk+2), lda, work); } }// end of LOWER trace_finalize( "chetrd_he2hb.svg", "trace.css" ); cudaEventDestroy(Pupdate_event); magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free( dA ); work[0] = MAGMA_C_MAKE( lwkopt, 0 ); magmablasSetKernelStream( 0 ); magma_set_lapack_numthreads(threads); return *info; } /* magma_chetrd_he2hb */
extern "C" magma_int_t magma_cbulge_applyQ_v2(magma_side_t side, magma_int_t NE, magma_int_t N, magma_int_t NB, magma_int_t Vblksiz, magmaFloatComplex *dE, magma_int_t ldde, magmaFloatComplex *V, magma_int_t ldv, magmaFloatComplex *T, magma_int_t ldt, magma_int_t *info) { //%=========================== //% local variables //%=========================== magma_int_t Vm, Vn, mt, nt; magma_int_t myrow, mycol, blkj, blki; magma_int_t blkid,vpos,tpos; magma_int_t firstrow, nbcolinvolvd; magma_int_t versionL = 113; magma_int_t versionR = 92; magma_int_t Vchunksiz = 10; *info=0; /* Quick return */ if ( NE == 0 ) { return MAGMA_SUCCESS; } if ( N == 0 ) { return MAGMA_SUCCESS; } if ( NB == 0 ) { return MAGMA_SUCCESS; } /* ========================================== * some infos for developer * Initialisation and checking nb of cores * ==========================================*/ /* we have 2 algo for left (113 114) and 2 algo for right (91 92) * which correspond to versionL versionR. * They are very similar (detail explained in tech report and matlab code) * however version 114 and 92 improve locality. * while version 113 is used in case WNATZ=1 (construct Q2) which allow * the construction to be done in an optimized way taking into * consideration that the matrix is Identity so making less flops. * */ // Initialize streaming and events cudaDeviceSynchronize(); magma_queue_t cstream; magmablasGetKernelStream(&cstream); magma_queue_t stream[2]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); magma_event_t myevent[2]; cudaEventCreateWithFlags(&myevent[0],cudaEventDisableTiming); cudaEventCreateWithFlags(&myevent[1],cudaEventDisableTiming); // Azzam 21/11/2012 // NOTE THAT dwork was of size 2*NE*Vblksiz+... // but I am thinking why not modifing it to NE*Vblksiz+... // BUT NO because the 2* is used because of making 2 streams working and so // they might be using dwork in parallel magmaFloatComplex *dwork, *dwork0, *dwork1, *dwvt0, *dwvt1; magmaFloatComplex *dT0, *dV0, *dT1, *dV1; magma_int_t lddv = ldv; magma_int_t lddt = ldt; magma_int_t lddw = 0; magma_int_t lddwork = ((NE+31)/32)*32; magma_int_t dwVTsiz = lddv*Vblksiz; // lddv*lddv + lddv*lddwork; (v2) // lddv*Vblksiz; (v1,v3) magma_int_t dworksiz = lddwork*Vblksiz; // lddv*Vblksiz; (v2) // NE*Vblksiz=lddwork*Vblksiz; (v1,v3) if (MAGMA_SUCCESS != magma_cmalloc( &dwork, 2*dworksiz + 2*dwVTsiz + 2*Vchunksiz* (Vblksiz* (lddv+lddt)) )) { printf ("!!!! magma_cbulge_applyQ magma_alloc failed for: dwork\n" ); exit(-1); } dwork0 = dwork; // size = dworksiz; dwork1 = dwork0 + dworksiz; // size = dworksiz; dwvt0 = dwork + 2*dworksiz; // size = dwVTsiz; dwvt1 = dwvt0 + dwVTsiz; // size = dwVTsiz; dV0 = dwork + 2*dworksiz + 2*dwVTsiz; dT0 = dV0 + Vchunksiz*Vblksiz*lddv; dV1 = dT0 + Vchunksiz*Vblksiz*lddt; dT1 = dV1 + Vchunksiz*Vblksiz*lddv; // make overlapped copy magma_int_t ncpy = 0; magma_int_t copyed=0, copyst=0; magma_int_t blkcnt,nothing, mysiz, flip, vld,tld, locpos; findVTsiz(N, NB, Vblksiz, &blkcnt, ¬hing); flip = 0; // performance loss if the reflector are applied to a big number of eigenvectors (~10000) // => apply the reflectors to blocks of eigenvectors. //magma_int_t nr_bl = magma_ceildiv(NE,10000); //nr of blocks magma_int_t sz_bl = NE; //magma_ceildiv(NE,nr_bl*64)*64; //maximum size of blocks (to have blocks of around the same size and multiple of 64) magma_int_t ib; //size of current block /* SIDE LEFT meaning apply E = Q*E = (q_1*q_2*.....*q_n) * E ==> so traverse Vs in reverse order (forward) from q_n to q_1 * Also E is splitten by row meaning each apply consist in a block of row (horizontal block) */ /* SIDE RIGHT meaning apply E = E*Q = E * (q_1*q_2*.....*q_n) ==> so tarverse Vs in normal order (forward) from q_1 to q_n * Also E is splitten by col meaning each apply consist in a block of col (vertical block) */ #ifdef ENABLE_DEBUG printf(" APPLY Q_v22 GPU with N %d, NE %d, NB %d, Vblksiz %d, versionL %d versionR %d SIDE %c \n", N, NE, NB, Vblksiz, versionL, versionR, side); #endif /* * MagmamaLeft */ if (side == MagmaLeft) { /* * Version 113: * loop over the block_col (nt) and for each find the * number of tiles (mt) in this block_col. then loop over mt, find * the size of the V's(Vm,Vn) and apply it to the corresponding * portion of E. */ if ( versionL == 113 ) { nt = magma_ceildiv((N-1),Vblksiz); for (blkj=nt-1; blkj >= 0; blkj--) { /* the index of the first row on the top of block (blkj) */ firstrow = blkj * Vblksiz + 1; /*find the number of tile for this block */ if ( blkj == nt-1 ) mt = magma_ceildiv( N - firstrow, NB); else mt = magma_ceildiv( N - (firstrow+1), NB); /*loop over the tiles find the size of the Vs and apply it */ for (blki=mt; blki > 0; blki--) { /*calculate the size of each losange of Vs= (Vm,Vn)*/ myrow = firstrow + (mt-blki)*NB; mycol = blkj*Vblksiz; Vm = min( NB+Vblksiz-1, N-myrow); if ( ( blkj == nt-1 ) && ( blki == mt ) ) { Vn = min (Vblksiz, Vm); } else { Vn = min (Vblksiz, Vm-1); } /*calculate the pointer to the Vs and the Ts. * Note that Vs and Ts have special storage done * by the bulgechasing function*/ //printf("voici blkj %d blki %d Vm %d Vn %d mycol %d vpos %d \n",blkj,blki,Vm, Vn,mycol,vpos); magma_bulge_findpos113(N, NB, Vblksiz, mycol, myrow, &blkid); // COPY Vchunksiz Vs and Vchunksiz Ts to GPU and store it in dV0/dV1 and dT0/dT1 if (ncpy == 0) { // flip = 1 for this. copyst = 0; // meaning that copy will start copying from blkid =copyst copyed = min(copyst+Vchunksiz, blkcnt); // meaning that copy will end copying at blkid =copyed-1==> next copy had to start at copyed mysiz = copyed-copyst; // the size of the chunk to be copied if (mysiz > 0) { ncpy = 1; flip = 1; vpos = copyst*Vblksiz*ldv; tpos = copyst*Vblksiz*ldt; vld = mysiz * ldv; tld = mysiz * ldt; magmablasSetKernelStream(stream[1]); magma_csetmatrix_async(vld, Vblksiz, V(vpos), vld, dV1, vld, stream[1]); magma_csetmatrix_async(tld, Vblksiz, T(tpos), tld, dT1, tld, stream[1]); //printf("doing the first copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV1 dT1\n",mysiz,copyst,copyed,vpos,tpos); } } if (blkid == copyst) { flip = ncpy % 2; copyst = copyed; // meaning that copy will start copying from blkid =copyst copyed = min(copyst+Vchunksiz, blkcnt); // meaning that copy will end copying at blkid =copyed-1==> next copy had to start at copyed mysiz = copyed-copyst; // the size of the chunk to be copied //printf(" get to copy blkid %d blkid+(2*Vchunksiz) %d copyst %d copyed %d\n",blkid,blkid+(Vchunksiz),copyst,copyed); if (mysiz > 0) { ncpy = ncpy + 1; vpos = copyst*Vblksiz*ldv; tpos = copyst*Vblksiz*ldt; vld = mysiz * ldv; tld = mysiz * ldt; if (flip == 0) { // now I am working on dV0 so copy the next and put it on dV1 //printf("doing overlapping copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV1 dT1\n",mysiz,copyst,copyed,vpos,tpos); magmablasSetKernelStream(stream[1]); magma_csetmatrix_async(vld, Vblksiz, V(vpos), vld, dV1, vld, stream[1]); magma_csetmatrix_async(tld, Vblksiz, T(tpos), tld, dT1, tld, stream[1]); } else { // now I am working on dV1 so copy the next and put it on dV0 //printf("doing overlapping copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV0 dT0\n",mysiz,copyst,copyed,vpos,tpos); magmablasSetKernelStream(stream[0]); magma_csetmatrix_async(vld, Vblksiz, V(vpos), vld, dV0, vld, stream[0]); magma_csetmatrix_async(tld, Vblksiz, T(tpos), tld, dT0, tld, stream[0]); } } } if ((Vm > 0) && (Vn > 0)) { locpos = blkid%Vchunksiz; magma_int_t lcvpos = locpos*Vblksiz*lddv; magma_int_t lctpos = locpos*Vblksiz*lddt; //printf("voici blkj %d blki %d Vm %d Vn %d mycol %d locvpos %5d loctpos %5d blkid %2d using data in dV%1d dT%1d \n",blkj,blki,Vm, Vn,mycol,lcvpos,lctpos, blkid,flip,flip); if (flip == 0) { magmablasSetKernelStream(stream[0]); cudaStreamWaitEvent(stream[0], myevent[1], 0); for (magma_int_t i=0; i < NE; i += sz_bl) { ib = min(sz_bl, NE-i); lddw = min(lddwork,sz_bl); //magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0+lcvpos, lddv, dT0+lctpos, lddt, dE(myrow,i), ldde, dwork0, lddw); magma_clarfb_gpu_gemm( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0+lcvpos, lddv, dT0+lctpos, lddt, dE(myrow,i), ldde, dwork0, lddw, dwvt0, lddv); } cudaEventRecord(myevent[0], stream[0]); } else { magmablasSetKernelStream(stream[1]); cudaStreamWaitEvent(stream[1], myevent[0], 0); for (magma_int_t i=0; i < NE; i += sz_bl) { ib = min(sz_bl, NE-i); lddw = min(lddwork,sz_bl); //magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV1+lcvpos, lddv, dT1+lctpos, lddt, dE(myrow,i), ldde, dwork1, lddw); magma_clarfb_gpu_gemm( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV1+lcvpos, lddv, dT1+lctpos, lddt, dE(myrow,i), ldde, dwork1, lddw, dwvt1, lddv); } cudaEventRecord(myevent[1], stream[1]); } } // end for (Vm &Vn) > 0 } // end for blki } // end for blkj } // end if version=113 /* * Version 114: * loop over the block_row (mt) and for each find diagonally the * number of tiles (nt) in this block_row. then loop over nt, find * the size of the V's(Vm,Vn) and apply it to the corresponding * portion of E. */ else { mt = magma_ceildiv((N-1),NB); for (blki = mt; blki > 0; blki--) { /* nbcolinvolvd = number of column corresponding to this block_row (blki) */ nbcolinvolvd = min(N-1, blki*NB); /*find the number of tile for this block (diagonal row of tiles) */ nt = magma_ceildiv(nbcolinvolvd,Vblksiz); /*loop over the tiles find the size of the Vs and apply it */ for (blkj = nt-1; blkj >= 0; blkj--) { /* the index of the first row of the first col meaning * the block on the top left (blki) */ firstrow = (mt-blki)*NB+1; /*calculate the size of each losange of Vs= (Vm,Vn)*/ myrow = firstrow + blkj*Vblksiz; mycol = blkj*Vblksiz; Vm = min( NB+Vblksiz-1, N-myrow); if ( ( blkj == nt-1 ) && ( blki == mt ) ) { Vn = min (Vblksiz, Vm); } else { Vn = min (Vblksiz, Vm-1); } if ((Vm > 0) && (Vn > 0)) { /*calculate the pointer to the Vs and the Ts. * Note that Vs and Ts have special storage done * by the bulgechasing function*/ magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos); magma_csetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL); magma_csetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL); //printf("voici blki %d rownbm %d mycol %d coled %d blkid %d vpos %d tpos %d\n", blki, rownbm, mycol, coled, blkid, vpos, tpos); for (magma_int_t i=0; i < NE; i += sz_bl) { ib = min(sz_bl, NE-i); magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0, lddv, dT0, lddt, dE(myrow,i), ldde, dwork, NE); } } // end for (Vm &Vn) > 0 } // end for blkj } // end for blki } // end version 114 } // end LEFT /* * MagmaRight */ else { /* * Version 91: */ if ( versionR == 91 ) { nt = magma_ceildiv((N-1),Vblksiz); for (blkj=0; blkj < nt; blkj++) { /* the index of the first myrow on the top of block (blkj) */ firstrow = blkj * Vblksiz + 1; /*find the number of tile for this block */ if ( blkj == nt-1 ) mt = magma_ceildiv( N - firstrow, NB); else mt = magma_ceildiv( N - (firstrow+1), NB); /*loop over the tiles find the size of the Vs and apply it */ for (blki=1; blki <= mt; blki++) { /*calculate the size of each losange of Vs= (Vm,Vn)*/ myrow = firstrow + (mt-blki)*NB; Vm = min( NB+Vblksiz-1, N-myrow); if ( (blkj == nt-1) && (blki == mt) ) { Vn = min( Vblksiz, Vm ); } else { Vn = min( Vblksiz, Vm-1 ); } mycol = blkj*Vblksiz; if ((Vm > 0) && (Vn > 0)) { /*calculate the pointer to the Vs and the Ts. * Note that Vs and Ts have special storage done * by the bulgechasing function*/ magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos); magma_csetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL); magma_csetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL); magma_clarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, Vm, Vn, dV0, lddv, dT0, lddt, dE(0, myrow), ldde, dwork, NE); } // end for (Vm &Vn) > 0 } // end for blki } // end fo blkj } // end of version 91 /* * Version 92: */ else { mt = magma_ceildiv((N-1),NB); for (blki = 1; blki <= mt; blki++) { /* nbcolinvolvd = number of column corresponding to this block_row (blki) */ nbcolinvolvd = min(N-1, blki*NB); /*find the number of tile for this block (diagonal row of tiles) */ nt = magma_ceildiv(nbcolinvolvd,Vblksiz); /*loop over the tiles find the size of the Vs and apply it */ for (blkj = 0; blkj < nt; blkj++) { /* the index of the first row of the first col meaning * the block on the top left (blki) */ firstrow = (mt-blki)*NB+1; /*calculate the size of each losange of Vs= (Vm,Vn)*/ myrow = firstrow + blkj*Vblksiz; mycol = blkj*Vblksiz; Vm = min( NB+Vblksiz-1, N-myrow); if ( ( blkj == nt-1 ) && ( blki == mt ) ) { Vn = min (Vblksiz, Vm); } else { Vn = min (Vblksiz, Vm-1); } if ((Vm > 0) && (Vn > 0)) { /*calculate the pointer to the Vs and the Ts. * Note that Vs and Ts have special storage done * by the bulgechasing function*/ magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos); magma_csetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL); magma_csetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL); magma_clarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, Vm, Vn, dV0, lddv, dT0, lddt, dE(0, myrow), ldde, dwork, NE); } // end for (Vm &Vn) > 0 } //end for blkj } // end for blki } //end of version 92 } // end RIGHT cudaDeviceSynchronize(); magmablasSetKernelStream(cstream); cudaEventDestroy(myevent[0]); cudaEventDestroy(myevent[1]); magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free(dwork); return MAGMA_SUCCESS; }
//----------------------------------------------------------------------------- CUDA::Timer::~Timer () { cudaEventDestroy(mStart); cudaEventDestroy(mStop); }
Event::~Event() { cudaEventDestroy( m_stop ); cudaEventDestroy( m_start ); }