/* //////////////////////////////////////////////////////////////////////////// -- Testing dsysv */ int main( int argc, char** argv) { TESTING_INIT(); double *h_A, *h_B, *h_X, *work, temp; real_Double_t gflops, gpu_perf, gpu_time = 0.0, cpu_perf=0, cpu_time=0; double error, error_lapack = 0.0; magma_int_t *ipiv; magma_int_t N, n2, lda, ldb, sizeB, lwork, info; magma_int_t status = 0, ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_opts opts; opts.parse_opts( argc, argv ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |Ax-b|/(N*|A|*|x|)\n"); printf("%%========================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; ldb = N; lda = N; n2 = lda*N; sizeB = ldb*opts.nrhs; gflops = ( FLOPS_DPOTRF( N ) + FLOPS_DPOTRS( N, opts.nrhs ) ) / 1e9; TESTING_MALLOC_CPU( ipiv, magma_int_t, N ); TESTING_MALLOC_PIN( h_A, double, n2 ); TESTING_MALLOC_PIN( h_B, double, sizeB ); TESTING_MALLOC_PIN( h_X, double, sizeB ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { lwork = -1; lapackf77_dsysv(lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, ipiv, h_X, &ldb, &temp, &lwork, &info); lwork = (int)MAGMA_D_REAL(temp); TESTING_MALLOC_CPU( work, double, lwork ); init_matrix( N, N, h_A, lda ); lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B ); lapackf77_dlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb ); cpu_time = magma_wtime(); lapackf77_dsysv(lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, ipiv, h_X, &ldb, work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) { printf("lapackf77_dsysv returned error %d: %s.\n", (int) info, magma_strerror( info )); } error_lapack = get_residual( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, h_B, ldb ); TESTING_FREE_CPU( work ); } /* ==================================================================== Performs operation using MAGMA =================================================================== */ init_matrix( N, N, h_A, lda ); lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B ); lapackf77_dlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb ); magma_setdevice(0); gpu_time = magma_wtime(); magma_dsysv( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) { printf("magma_dsysv returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Check the factorization =================================================================== */ if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)", (int) N, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f)", (int) N, (int) N, gpu_perf, gpu_time ); } if ( opts.check == 0 ) { printf(" --- \n"); } else { error = get_residual( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, h_B, ldb ); printf(" %8.2e %s", error, (error < tol ? "ok" : "failed")); if (opts.lapack) printf(" (lapack rel.res. = %8.2e)", error_lapack); printf("\n"); status += ! (error < tol); } TESTING_FREE_CPU( ipiv ); TESTING_FREE_PIN( h_X ); TESTING_FREE_PIN( h_B ); TESTING_FREE_PIN( h_A ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
/** Purpose ------- DPOTRF_OOC computes the Cholesky factorization of a real symmetric positive definite matrix A. This version does not require work space on the GPU passed as input. GPU memory is allocated in the routine. The matrix A may exceed the GPU memory. The factorization has the form A = U**H * U, if UPLO = MagmaUpper, or A = L * L**H, if UPLO = MagmaLower, where U is an upper triangular matrix and L is lower triangular. This is the block version of the algorithm, calling Level 3 BLAS. Arguments --------- @param[in] ngpu INTEGER Number of GPUs to use. ngpu > 0. @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 DOUBLE_PRECISION 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. \n On exit, if INFO = 0, the factor U or L from the Cholesky factorization A = U**H * U or A = L * L**H. \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,N). @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. - > 0: if INFO = i, the leading minor of order i is not positive definite, and the factorization could not be completed. @ingroup magma_dposv_comp ********************************************************************/ extern "C" magma_int_t rr_dpotrf_m( magma_int_t ngpu, magma_uplo_t uplo, magma_int_t n, double *A, magma_int_t lda, magma_int_t *info) { #define A(i, j) ( A + (j)*lda + (i)) #define dA(d, i, j) (dwork[(d)] + (j)*lddla + (i)) #define dT(d, i, j) ( dt[(d)] + (j)*ldda + (i)) #define dAup(d, i, j) (dwork[(d)] + (j)*NB + (i)) #define dTup(d, i, j) ( dt[(d)] + (j)*nb + (i)) /* Local variables */ double d_one = 1.0; double d_neg_one = -1.0; double c_one = MAGMA_D_ONE; double c_neg_one = MAGMA_D_NEG_ONE; const char* uplo_ = lapack_uplo_const( uplo ); int upper = (uplo == MagmaUpper); double *dwork[MagmaMaxGPUs], *dt[MagmaMaxGPUs]; magma_int_t ldda, lddla, nb, iinfo, n_local[MagmaMaxGPUs], J2, d, ngpu0 = ngpu; magma_int_t j, jj, jb, J, JB, NB, h; magma_queue_t stream[MagmaMaxGPUs][3]; magma_event_t event[MagmaMaxGPUs][5]; magma_timer_t time_total=0, time_sum=0, time=0; *info = 0; if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (lda < max(1,n)) { *info = -4; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } /* Quick return */ if ( n == 0 ) return *info; magma_device_t orig_dev; magma_getdevice( &orig_dev ); magma_queue_t orig_stream; magmablasGetKernelStream( &orig_stream ); nb = magma_get_dpotrf_nb(n); if ( ngpu0 > n/nb ) { ngpu = n/nb; if ( n%nb != 0 ) ngpu ++; } else { ngpu = ngpu0; } //ldda = ((n+31)/32)*32; ldda = ((n+nb-1)/nb)*nb; lddla = ((nb*((n+nb*ngpu-1)/(nb*ngpu))+31)/32)*32; /* figure out NB */ size_t freeMem, totalMem; cudaMemGetInfo( &freeMem, &totalMem ); freeMem /= sizeof(double); //MB = n; /* number of rows in the big panel */ NB = (magma_int_t)((0.8*freeMem - max(2,ngpu)*nb*ldda - (n+nb)*nb)/lddla); /* number of columns in the big panel */ //NB = min(5*nb,n); if ( NB >= n ) { #ifdef CHECK_DPOTRF_OOC printf( " * still fits in GPU memory.\n" ); #endif NB = n; } else { #ifdef CHECK_DPOTRF_OOC printf( " * doesn't fit in GPU memory.\n" ); #endif NB = (NB/nb) * nb; /* making sure it's devisable by nb */ } #ifdef CHECK_DPOTRF_OOC if ( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d, lddla=%d, freeMem=%.2e).\n", n, NB, nb, lddla, (double)freeMem ); else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d, lddla=%d, freeMem=%.2e).\n", n, NB, nb, lddla, (double)freeMem ); fflush(stdout); #endif for (d=0; d < ngpu; d++ ) { magma_setdevice(d); if (MAGMA_SUCCESS != magma_dmalloc( &dt[d], NB*lddla + max(2,ngpu)*nb*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } dwork[d] = &dt[d][max(2,ngpu)*nb*ldda]; for( j=0; j < 3; j++ ) magma_queue_create( &stream[d][j] ); for( j=0; j < 5; j++ ) magma_event_create( &event[d][j] ); magma_device_sync(); // synch the device } magma_setdevice(0); timer_start( time_total ); if (nb <= 1 || nb >= n) { lapackf77_dpotrf(uplo_, &n, A, &lda, info); } else { /* Use hybrid blocked code. */ if (upper) { /* =========================================================== * * Compute the Cholesky factorization A = U'*U. * * big panel is divided by block-row and distributed in block * * column cyclic format */ /* for each big-panel */ for( J=0; J < n; J += NB ) { JB = min(NB,n-J); if ( ngpu0 > (n-J)/nb ) { ngpu = (n-J)/nb; if ( (n-J)%nb != 0 ) ngpu ++; } else { ngpu = ngpu0; } /* load the new big-panel by block-rows */ magma_dhtodpo( ngpu, uplo, JB, n, J, J, nb, A, lda, dwork, NB, stream, &iinfo); /* update with the previous big-panels */ timer_start( time ); for( j=0; j < J; j += nb ) { /* upload the diagonal of the block column (broadcast to all GPUs) */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_dsetmatrix_async( nb, JB, A(j, J), lda, dTup(d, 0, J), nb, stream[d][0] ); n_local[d] = 0; } /* distribute off-diagonal blocks to GPUs */ for( jj=J+JB; jj < n; jj += nb ) { d = ((jj-J)/nb)%ngpu; magma_setdevice(d); jb = min(nb, n-jj); magma_dsetmatrix_async( nb, jb, A(j, jj), lda, dTup(d, 0, J+JB+n_local[d]), nb, stream[d][0] ); n_local[d] += jb; } /* wait for the communication */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_queue_sync( stream[d][0] ); } /* update the current big-panel using the previous block-row */ /* -- process the big diagonal block of the big panel */ for( jj=0; jj < JB; jj += nb ) { // jj is 'local' column index within the big panel d = (jj/nb)%ngpu; J2 = jj/(nb*ngpu); magma_setdevice(d); magmablasSetKernelStream(stream[d][J2%2]); // the last stream (2) used to process off-diagonal J2 = nb*J2; jb = min(nb,JB-jj); // number of columns in this current block-row magma_dgemm( MagmaConjTrans, MagmaNoTrans, jj, jb, nb, c_neg_one, dTup(d, 0, J ), nb, dTup(d, 0, J+jj), nb, c_one, dAup(d, 0, J2), NB); magma_dsyrk(MagmaUpper, MagmaConjTrans, jb, nb, d_neg_one, dTup(d, 0, J+jj), nb, d_one, dAup(d, jj, J2), NB); } /* -- process the remaining big off-diagonal block of the big panel */ if ( n > J+JB ) { for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magmablasSetKernelStream(stream[d][2]); /* local number of columns in the big panel */ n_local[d] = ((n-J)/(nb*ngpu))*nb; if (d < ((n-J)/nb)%ngpu) n_local[d] += nb; else if (d == ((n-J)/nb)%ngpu) n_local[d] += (n-J)%nb; /* subtracting the local number of columns in the diagonal */ J2 = nb*(JB/(nb*ngpu)); if ( d < (JB/nb)%ngpu ) J2 += nb; n_local[d] -= J2; magma_dgemm( MagmaConjTrans, MagmaNoTrans, JB, n_local[d], nb, c_neg_one, dTup(d, 0, J ), nb, dTup(d, 0, J+JB), nb, c_one, dAup(d, 0, J2), NB); } } /* wait for the previous updates */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); for( jj=0; jj < 3; jj++ ) magma_queue_sync( stream[d][jj] ); magmablasSetKernelStream(NULL); } magma_setdevice(0); } /* end of updates with previous rows */ /* factor the big panel */ h = (JB+nb-1)/nb; // big diagonal of big panel will be on CPU // using three streams rr_dpotrf3_mgpu(ngpu, uplo, JB, n-J, J, J, nb, dwork, NB, dt, ldda, A, lda, h, stream, event, &iinfo); if ( iinfo != 0 ) { *info = J+iinfo; break; } time_sum += timer_stop( time ); /* upload the off-diagonal (and diagonal!!!) big panel */ magma_ddtohpo(ngpu, uplo, JB, n, J, J, nb, NB, A, lda, dwork, NB, stream, &iinfo); //magma_ddtohpo(ngpu, uplo, JB, n, J, J, nb, 0, A, lda, dwork, NB, stream, &iinfo); } } else { /* ========================================================= * * Compute the Cholesky factorization A = L*L'. */ /* for each big-panel */ for( J=0; J < n; J += NB ) { JB = min(NB,n-J); if ( ngpu0 > (n-J)/nb ) { ngpu = (n-J)/nb; if ( (n-J)%nb != 0 ) ngpu ++; } else { ngpu = ngpu0; } /* load the new big-panel by block-columns */ magma_dhtodpo( ngpu, uplo, n, JB, J, J, nb, A, lda, dwork, lddla, stream, &iinfo); /* update with the previous big-panels */ timer_start( time ); for( j=0; j < J; j += nb ) { /* upload the diagonal of big panel */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_dsetmatrix_async( JB, nb, A(J, j), lda, dT(d, J, 0), ldda, stream[d][0] ); n_local[d] = 0; } /* upload off-diagonals */ for( jj=J+JB; jj < n; jj += nb ) { d = ((jj-J)/nb)%ngpu; magma_setdevice(d); jb = min(nb, n-jj); magma_dsetmatrix_async( jb, nb, A(jj, j), lda, dT(d, J+JB+n_local[d], 0), ldda, stream[d][0] ); n_local[d] += jb; } /* wait for the communication */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_queue_sync( stream[d][0] ); } /* update the current big-panel using the previous block-row */ for( jj=0; jj < JB; jj += nb ) { /* diagonal */ d = (jj/nb)%ngpu; J2 = jj/(nb*ngpu); magma_setdevice(d); magmablasSetKernelStream(stream[d][J2%2]); J2 = nb*J2; jb = min(nb,JB-jj); magma_dgemm( MagmaNoTrans, MagmaConjTrans, jb, jj, nb, c_neg_one, dT(d, J+jj, 0), ldda, dT(d, J, 0), ldda, c_one, dA(d, J2, 0), lddla); magma_dsyrk(MagmaLower, MagmaNoTrans, jb, nb, d_neg_one, dT(d, J+jj, 0), ldda, d_one, dA(d, J2, jj), lddla); } if ( n > J+JB ) { /* off-diagonal */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magmablasSetKernelStream(stream[d][2]); /* local number of columns in the big panel */ n_local[d] = (((n-J)/nb)/ngpu)*nb; if (d < ((n-J)/nb)%ngpu) n_local[d] += nb; else if (d == ((n-J)/nb)%ngpu) n_local[d] += (n-J)%nb; /* subtracting local number of columns in diagonal */ J2 = nb*(JB/(nb*ngpu)); if ( d < (JB/nb)%ngpu ) J2 += nb; n_local[d] -= J2; magma_dgemm( MagmaNoTrans, MagmaConjTrans, n_local[d], JB, nb, c_neg_one, dT(d, J+JB, 0), ldda, dT(d, J, 0), ldda, c_one, dA(d, J2, 0), lddla); } } /* wait for the previous updates */ for( d=0; d < ngpu; d++ ) { magma_setdevice(d); for( jj=0; jj < 3; jj++ ) magma_queue_sync( stream[d][jj] ); magmablasSetKernelStream(NULL); } magma_setdevice(0); } /* factor the big panel */ h = (JB+nb-1)/nb; // big diagonal of big panel will be on CPU // using three streams magma_dpotrf3_mgpu(ngpu, uplo, n-J, JB, J, J, nb, dwork, lddla, dt, ldda, A, lda, h, stream, event, &iinfo); if ( iinfo != 0 ) { *info = J+iinfo; break; } time_sum += timer_stop( time ); /* upload the off-diagonal big panel */ magma_ddtohpo( ngpu, uplo, n, JB, J, J, nb, JB, A, lda, dwork, lddla, stream, &iinfo); } /* end of for J */ } /* if upper */ } /* if nb */ timer_stop( time_total ); if ( ngpu0 > n/nb ) { ngpu = n/nb; if ( n%nb != 0 ) ngpu ++; } else { ngpu = ngpu0; } for (d=0; d < ngpu; d++ ) { magma_setdevice(d); for( j=0; j < 3; j++ ) { magma_queue_destroy( stream[d][j] ); } magma_free( dt[d] ); for( j=0; j < 5; j++ ) { magma_event_destroy( event[d][j] ); } } magma_setdevice( orig_dev ); magmablasSetKernelStream( orig_stream ); timer_printf( "\n n=%d NB=%d nb=%d\n", (int) n, (int) NB, (int) nb ); timer_printf( " Without memory allocation: %f / %f = %f GFlop/s\n", FLOPS_DPOTRF(n) / 1e9, time_total, FLOPS_DPOTRF(n) / 1e9 / time_total ); timer_printf( " Performance %f / %f = %f GFlop/s\n", FLOPS_DPOTRF(n) / 1e9, time_sum, FLOPS_DPOTRF(n) / 1e9 / time_sum ); return *info; } /* magma_dpotrf_ooc */
int main( int argc, char** argv) { real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time; double *h_R = NULL, *h_P = NULL; magmaDouble_ptr d_lA[MagmaMaxSubs * MagmaMaxGPUs]; magma_int_t N = 0, n2, lda, ldda; magma_int_t size[10] = { 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000 }; magma_int_t i, j, k, check = 0, info; double mz_one = MAGMA_D_NEG_ONE; magma_int_t ione = 1; magma_int_t num_gpus0 = 1, num_gpus, num_subs0 = 1, num_subs, tot_subs, flag = 0; magma_int_t nb, n_local, nk; magma_uplo_t uplo = MagmaLower; if (argc != 1){ for(i = 1; i<argc; i++){ if (strcmp("-N", argv[i]) == 0){ N = atoi(argv[++i]); if (N > 0) { size[0] = size[9] = N; flag = 1; } } if(strcmp("-NGPU", argv[i]) == 0) num_gpus0 = atoi(argv[++i]); if(strcmp("-NSUB", argv[i]) == 0) num_subs0 = atoi(argv[++i]); if(strcmp("-UPLO", argv[i]) == 0) uplo = (strcmp("L", argv[++i]) == 0 ? MagmaLower : MagmaUpper); if(strcmp("-check", argv[i]) == 0) check = 1; } } /* Initialize */ magma_queue_t queues[2*MagmaMaxGPUs]; magma_device_t devices[ MagmaMaxGPUs ]; magma_int_t num = 0; magma_int_t err; magma_init(); err = magma_getdevices( devices, MagmaMaxGPUs, &num ); if ( err != 0 || num < 1 ) { fprintf( stderr, "magma_getdevices failed: %d\n", (int) err ); exit(-1); } for(i=0;i<num_gpus0;i++){ err = magma_queue_create( devices[i], &queues[2*i] ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", (int) err ); exit(-1); } err = magma_queue_create( devices[i], &queues[2*i+1] ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", (int) err ); exit(-1); } } printf("\nUsing %d GPUs:\n", num_gpus0); printf(" testing_dpotrf_msub -N %d -NGPU %d -NSUB %d -UPLO %c %s\n\n", size[0], num_gpus0,num_subs0, (uplo == MagmaLower ? 'L' : 'U'),(check == 1 ? "-check" : " ")); printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma-R_lapack||_F / ||R_lapack||_F\n"); printf("========================================================================================\n"); for(i=0; i<10; i++){ N = size[i]; lda = N; n2 = lda*N; gflops = FLOPS_DPOTRF( N ) / 1e9;; nb = magma_get_dpotrf_nb(N); if (num_subs0*num_gpus0 > N/nb) { num_gpus = N/nb; num_subs = 1; if(N%nb != 0) num_gpus ++; printf("too many GPUs for the matrix size, using %d GPUs\n", (int)num_gpus); } else { num_gpus = num_gpus0; num_subs = num_subs0; } tot_subs = num_subs * num_gpus; /* Allocate host memory for the matrix */ #ifdef USE_PINNED_CLMEMORY cl_mem buffer1 = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, n2*sizeof(double), NULL, NULL); cl_mem buffer2 = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, lda*nb*sizeof(double), NULL, NULL); for (k=0; k<num_gpus; k++) { h_R = (double*)clEnqueueMapBuffer(queues[2*k], buffer1, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, n2*sizeof(double), 0, NULL, NULL, NULL); h_P = (double*)clEnqueueMapBuffer(queues[2*k], buffer2, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, lda*nb*sizeof(double), 0, NULL, NULL, NULL); } #else TESTING_MALLOC_PIN( h_P, double, lda*nb ); TESTING_MALLOC_PIN( h_R, double, n2 ); #endif /* Initialize the matrix */ init_matrix( N, h_R, lda ); /* Allocate GPU memory */ if (uplo == MagmaUpper) { ldda = ((N+nb-1)/nb)*nb; n_local = ((N+nb*tot_subs-1)/(nb*tot_subs))*nb; } else { ldda = ((N+nb*tot_subs-1)/(nb*tot_subs))*nb; n_local = ((N+nb-1)/nb)*nb; } for (j=0; j<tot_subs; j++) { TESTING_MALLOC_DEV( d_lA[j], double, n_local*ldda ); } /* Warm up to measure the performance */ /* distribute matrix to gpus */ if (uplo == MagmaUpper) { for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dsetmatrix( j+nk, nk, &h_R[j*lda], lda, d_lA[k], j/(nb*tot_subs)*nb*ldda, ldda, queues[2*(k%num_gpus)]); } } else { for (j=0; j<N; j+=nb) { nk = min(nb, N-j); for (magma_int_t kk = 0; kk<tot_subs; kk++) { magma_int_t mk = 0; for (magma_int_t ii=j+kk*nb; ii<N; ii+=nb*tot_subs) { magma_int_t mii = min(nb, N-ii); lapackf77_dlacpy( MagmaFullStr, &mii, &nk, &h_R[ii+j*lda], &lda, &h_P[mk], &lda ); mk += mii; } k = ((j+kk*nb)/nb)%tot_subs; if (mk > 0 && nk > 0) { magma_dsetmatrix( mk, nk, h_P, lda, d_lA[k], j*ldda+(j+kk*nb)/(nb*tot_subs)*nb, ldda, queues[2*(k%num_gpus)]); } } } /*for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dsetmatrix( nk, j+nk, &h_R[j], lda, d_lA[k], j/(nb*tot_subs)*nb, ldda, queues[2*(k%num_gpus)]); }*/ } magma_dpotrf_msub( num_subs, num_gpus, uplo, N, d_lA, 0, ldda, queues, &info ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ /* distribute matrix to gpus */ if (uplo == MagmaUpper) { for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dsetmatrix( j+nk, nk, &h_R[j*lda], lda, d_lA[k], j/(nb*tot_subs)*nb*ldda, ldda, queues[2*(k%num_gpus)]); } } else { for (j=0; j<N; j+=nb) { nk = min(nb, N-j); for (magma_int_t kk = 0; kk<tot_subs; kk++) { magma_int_t mk = 0; for (magma_int_t ii=j+kk*nb; ii<N; ii+=nb*tot_subs) { magma_int_t mii = min(nb, N-ii); lapackf77_dlacpy( MagmaFullStr, &mii, &nk, &h_R[ii+j*lda], &lda, &h_P[mk], &lda ); mk += mii; } k = ((j+kk*nb)/nb)%tot_subs; if (mk > 0 && nk > 0) { magma_dsetmatrix( mk, nk, h_P, lda, d_lA[k], j*ldda+(j+kk*nb)/(nb*tot_subs)*nb, ldda, queues[2*(k%num_gpus)]); } } } /*for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dsetmatrix( nk, j+nk, &h_R[j], lda, d_lA[k], (j/(nb*tot_subs)*nb), ldda, queues[2*(k%num_gpus)]); }*/ } gpu_time = magma_wtime(); magma_dpotrf_msub( num_subs, num_gpus, uplo, N, d_lA, 0, ldda, queues, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf( "magma_dpotrf had error %d.\n", info ); /* gather matrix from gpus */ if (uplo==MagmaUpper) { for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dgetmatrix( j+nk, nk, d_lA[k], j/(nb*tot_subs)*nb*ldda, ldda, &h_R[j*lda], lda, queues[2*(k%num_gpus)]); } } else { for (j=0; j<N; j+=nb) { nk = min(nb, N-j); for (magma_int_t kk = 0; kk<tot_subs; kk++) { k = ((j+kk*nb)/nb)%tot_subs; magma_int_t mk = 0; mk = 0; for (magma_int_t ii=j+kk*nb; ii<N; ii+=nb*tot_subs) { mk += min(nb, N-ii); } if (mk > 0 && nk > 0) { magma_dgetmatrix( mk, nk, d_lA[k], j*ldda+(j+kk*nb)/(nb*tot_subs)*nb, ldda, h_P, lda, queues[2*(k%num_gpus)]); } mk = 0; for (magma_int_t ii=j+kk*nb; ii<N; ii+=nb*tot_subs) { magma_int_t mii = min(nb, N-ii); lapackf77_dlacpy( MagmaFullStr, &mii, &nk, &h_P[mk], &lda, &h_R[ii+j*lda], &lda ); mk += mii; } } } /*for (j=0; j<N; j+=nb) { k = (j/nb)%tot_subs; nk = min(nb, N-j); magma_dgetmatrix( nk, j+nk, d_lA[k], (j/(nb*tot_subs)*nb), ldda, &h_R[j], lda, queues[2*(k%num_gpus)] ); }*/ } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if (check == 1) { double work[1], matnorm, diffnorm; double *h_A; TESTING_MALLOC_PIN( h_A, double, n2 ); init_matrix( N, h_A, lda ); cpu_time = magma_wtime(); if (uplo == MagmaLower) { lapackf77_dpotrf( MagmaLowerStr, &N, h_A, &lda, &info ); } else { lapackf77_dpotrf( MagmaUpperStr, &N, h_A, &lda, &info ); } cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf( "lapackf77_dpotrf had error %d.\n", info ); /* ===================================================================== Check the result compared to LAPACK |R_magma - R_lapack| / |R_lapack| =================================================================== */ matnorm = lapackf77_dlange("f", &N, &N, h_A, &lda, work); blasf77_daxpy(&n2, &mz_one, h_A, &ione, h_R, &ione); diffnorm = lapackf77_dlange("f", &N, &N, h_R, &lda, work); printf( "%5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n", N, cpu_perf, cpu_time, gpu_perf, gpu_time, diffnorm / matnorm ); TESTING_FREE_PIN( h_A ); } else { printf( "%5d - - (- -) %6.2f (%6.2f) - -\n", N, gpu_perf, gpu_time ); } // free memory #ifdef USE_PINNED_CLMEMORY for (k=0; k<num_gpus; k++) { clEnqueueUnmapMemObject(queues[2*k], buffer1, h_R, 0, NULL, NULL); clEnqueueUnmapMemObject(queues[2*k], buffer2, h_P, 0, NULL, NULL); } clReleaseMemObject(buffer1); clReleaseMemObject(buffer2); #else TESTING_FREE_PIN( h_P ); TESTING_FREE_PIN( h_R ); #endif for (j=0; j<tot_subs; j++) { TESTING_FREE_DEV( d_lA[j] ); } if (flag != 0) break; } /* clean up */ for (i=0; i<num_gpus; i++) { magma_queue_destroy( queues[2*i] ); magma_queue_destroy( queues[2*i+1] ); } magma_finalize(); return 0; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dpotf2_gpu */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double *h_A, *h_R; magmaDouble_ptr d_A; magma_int_t N, n2, lda, ldda, info; double c_neg_one = MAGMA_D_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; double work[1], error; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); opts.lapack |= opts.check; // check (-c) implies lapack (-l) printf("uplo = %s\n", lapack_uplo_const(opts.uplo) ); printf(" N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R_magma - R_lapack||_F / ||R_lapack||_F\n"); printf("========================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; ldda = ((N+31)/32)*32; gflops = FLOPS_DPOTRF( N ) / 1e9; if ( N > 512 ) { printf( "%5d skipping because dpotf2 does not support N > 512\n", (int) N ); continue; } TESTING_MALLOC_CPU( h_A, double, n2 ); TESTING_MALLOC_PIN( h_R, double, n2 ); TESTING_MALLOC_DEV( d_A, double, ldda*N ); /* Initialize the matrix */ lapackf77_dlarnv( &ione, ISEED, &n2, h_A ); magma_dmake_hpd( N, h_A, lda ); lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); magma_dsetmatrix( N, N, h_A, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_dpotf2_gpu( opts.uplo, N, d_A, ldda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_dpotf2_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_dpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_dpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_dgetmatrix( N, N, d_A, ldda, h_R, lda ); error = lapackf77_dlange("f", &N, &N, h_A, &lda, work); blasf77_daxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_dlange("f", &N, &N, h_R, &lda, work) / error; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., error, (error < tol ? "ok" : "failed")); status += ! (error < tol); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) N, gpu_perf, gpu_time*1000. ); } TESTING_FREE_CPU( h_A ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dpotrf_mgpu */ int main( int argc, char** argv ) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0; double error, work[1]; double c_neg_one = MAGMA_D_NEG_ONE; double *h_A, *h_R; double *d_lA[ MagmaMaxGPUs ]; magma_int_t N, n2, lda, ldda, max_size, ngpu; magma_int_t info, nb; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) double tol = opts.tolerance * lapackf77_dlamch("E"); printf("ngpu = %d, uplo = %s\n", (int) opts.ngpu, lapack_uplo_const(opts.uplo) ); printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n"); printf("=================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; nb = magma_get_dpotrf_nb( N ); gflops = FLOPS_DPOTRF( N ) / 1e9; // ngpu must be at least the number of blocks ngpu = min( opts.ngpu, int((N+nb-1)/nb) ); if ( ngpu < opts.ngpu ) { printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu ); } // Allocate host memory for the matrix TESTING_MALLOC_CPU( h_A, double, n2 ); TESTING_MALLOC_PIN( h_R, double, n2 ); // Allocate device memory // matrix is distributed by block-rows or block-columns // this is maximum size that any GPU stores; // size is rounded up to full blocks in both rows and columns max_size = nb*(1+N/(nb*ngpu)) * nb*((N+nb-1)/nb); for( int dev=0; dev < ngpu; dev++ ) { magma_setdevice( dev ); TESTING_MALLOC_DEV( d_lA[dev], double, max_size ); } /* Initialize the matrix */ lapackf77_dlarnv( &ione, ISEED, &n2, h_A ); magma_dmake_hpd( N, h_A, lda ); lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_dpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_dpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ==================================================================== Performs operation using MAGMA =================================================================== */ if ( opts.uplo == MagmaUpper ) { ldda = ((N+nb-1)/nb)*nb; magma_dsetmatrix_1D_col_bcyclic( N, N, h_R, lda, d_lA, ldda, ngpu, nb ); } else { ldda = (1+N/(nb*ngpu))*nb; magma_dsetmatrix_1D_row_bcyclic( N, N, h_R, lda, d_lA, ldda, ngpu, nb ); } gpu_time = magma_wtime(); magma_dpotrf_mgpu( ngpu, opts.uplo, N, d_lA, ldda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_dpotrf_mgpu returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.uplo == MagmaUpper ) { magma_dgetmatrix_1D_col_bcyclic( N, N, d_lA, ldda, h_R, lda, ngpu, nb ); } else { magma_dgetmatrix_1D_row_bcyclic( N, N, d_lA, ldda, h_R, lda, ngpu, nb ); } /* ===================================================================== Check the result compared to LAPACK =================================================================== */ for( int dev=0; dev < ngpu; dev++ ){ magma_setdevice( dev ); magma_device_sync(); } if ( opts.lapack ) { error = lapackf77_dlange("f", &N, &N, h_A, &lda, work ); blasf77_daxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione ); error = lapackf77_dlange("f", &N, &N, h_R, &lda, work ) / error; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed") ); status += ! (error < tol); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) ---\n", (int) N, gpu_perf, gpu_time ); } TESTING_FREE_CPU( h_A ); TESTING_FREE_PIN( h_R ); for( int dev=0; dev < ngpu; dev++ ){ magma_setdevice( dev ); TESTING_FREE_DEV( d_lA[dev] ); } fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing dpotrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double *h_A, *h_R; magma_int_t N, n2, lda, info; double c_neg_one = MAGMA_D_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; double work[1], error; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) double tol = opts.tolerance * lapackf77_dlamch("E"); printf("ngpu %d, uplo %c\n", (int) opts.ngpu, opts.uplo ); printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma - R_lapack||_F / ||R_lapack||_F\n"); printf("========================================================\n"); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[i]; lda = N; n2 = lda*N; gflops = FLOPS_DPOTRF( N ) / 1e9; TESTING_MALLOC( h_A, double, n2 ); TESTING_HOSTALLOC( h_R, double, n2 ); /* Initialize the matrix */ lapackf77_dlarnv( &ione, ISEED, &n2, h_A ); magma_dmake_hpd( N, h_A, lda ); lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_dpotrf( opts.uplo, N, h_R, lda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_dpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_dpotrf( &opts.uplo, &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_dpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ error = lapackf77_dlange("f", &N, &N, h_A, &lda, work); blasf77_daxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_dlange("f", &N, &N, h_R, &lda, work) / error; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e%s\n", (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "" : " failed") ); status |= ! (error < tol); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) N, gpu_perf, gpu_time ); } TESTING_FREE( h_A ); TESTING_HOSTFREE( h_R ); } } TESTING_FINALIZE(); return status; }