/** Purpose ------- SPOTRF computes the Cholesky factorization of a real symmetric positive definite matrix dA. The factorization has the form dA = U**H * U, if UPLO = MagmaUpper, or dA = 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] uplo magma_uplo_t - = MagmaUpper: Upper triangle of dA is stored; - = MagmaLower: Lower triangle of dA is stored. @param[in] n INTEGER The order of the matrix dA. N >= 0. @param[in,out] d_lA REAL array of pointers on the GPU, dimension (ngpu) On entry, the symmetric matrix dA distributed over GPUs (dl_A[d] points to the local matrix on the d-th GPU). It is distributed in 1D block column or row cyclic (with the block size of nb) if UPLO = MagmaUpper or MagmaLower, respectively. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of dA contains the upper triangular part of the matrix dA, and the strictly lower triangular part of dA is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of dA contains the lower triangular part of the matrix dA, and the strictly upper triangular part of dA is not referenced. \n On exit, if INFO = 0, the factor U or L from the Cholesky factorization dA = U**H * U or dA = L * L**H. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,N). To benefit from coalescent memory accesses LDDA must be divisible by 16. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i, the leading minor of order i is not positive definite, and the factorization could not be completed. @ingroup magma_sposv_comp ********************************************************************/ extern "C" magma_int_t magma_spotrf_mgpu_right( magma_int_t ngpu, magma_uplo_t uplo, magma_int_t n, magmaFloat_ptr d_lA[], magma_int_t ldda, magma_int_t *info ) { #define dlA(id, i, j) (d_lA[(id)] + (j) * ldda + (i)) #define dlP(id, i, j) (d_lP[(id)] + (j) * ldda + (i)) #define panel(j) (panel + (j)) #define tmppanel(j) (tmppanel + (j)) #define tmpprevpanel(j) (tmpprevpanel + (j)) #define STREAM_ID(i) (nqueue > 1 ? 1+((i)/nb)%(nqueue-1) : 0) float z_one = MAGMA_S_MAKE( 1.0, 0.0 ); float mz_one = MAGMA_S_MAKE( -1.0, 0.0 ); float one = 1.0; float m_one = -1.0; const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t j, nb, d, id, j_local, blkid, crosspoint, prevtrsmrows=0, nqueue = 5; float *panel, *tmppanel0, *tmppanel1, *tmppanel, *tmpprevpanel; float *d_lP[MagmaMaxGPUs], *dlpanel, *dlpanels[MagmaMaxGPUs]; magma_int_t rows, trsmrows, igpu, n_local[MagmaMaxGPUs], ldpanel; magma_queue_t queues[MagmaMaxGPUs][10]; *info = 0; if ( uplo != MagmaUpper && uplo != MagmaLower ) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } magma_device_t orig_dev; magma_getdevice( &orig_dev ); magma_queue_t orig_stream; magmablasGetKernelStream( &orig_stream ); nb = magma_get_spotrf_nb(n); ldpanel = ldda; magma_setdevice(0); if (MAGMA_SUCCESS != magma_smalloc_pinned( &panel, 2 * nb * ldpanel )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } tmppanel0 = panel; tmppanel1 = tmppanel0 + nb * ldpanel; if ((nb <= 1) || (nb >= n)) { // Use unblocked code. magma_sgetmatrix( n, n, dlA(0, 0, 0), ldda, panel, ldpanel); lapackf77_spotrf( uplo_, &n, panel, &ldpanel, info); magma_ssetmatrix( n, n, panel, ldpanel, dlA(0, 0, 0), ldda ); } else { for( d = 0; d < ngpu; d++ ) { // local-n and local-ld n_local[d] = ((n / nb) / ngpu) * nb; if (d < (n / nb) % ngpu) n_local[d] += nb; else if (d == (n / nb) % ngpu) n_local[d] += n % nb; magma_setdevice(d); magma_device_sync(); if (MAGMA_SUCCESS != magma_smalloc( &d_lP[d], nb * ldda )) { for( j = 0; j < d; j++ ) { magma_setdevice(j); magma_free( d_lP[d] ); } *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } for( j=0; j < nqueue; j++ ) { magma_queue_create( &queues[d][j] ); } } //#define ENABLE_TIMER #if defined (ENABLE_TIMER) real_Double_t therk[4], tmtc, tcchol, tctrsm, tctm, tmnp, tcnp; real_Double_t ttot_herk[4] = {0,0,0,0}, ttot_mtc = 0, ttot_cchol = 0, ttot_ctrsm = 0, ttot_ctm = 0, ttot_mnp = 0, ttot_cnp = 0; printf("\n\n %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s\n", "j", "nb", "row", "mtc", "CPU_np", "panel", "ctrsm", "CH+TRSM", "CPU", "dsyrk[0]", "dsyrk[1]", "dsyrk[2]", "dsyrk[3]", "ctm P", "gpu_np"); printf(" ====================================================================================================\n"); #endif // Use blocked code. if (uplo == MagmaUpper) { printf( " === not supported, yet ===\n" ); } else { blkid = -1; if (ngpu == 4) crosspoint = n; else if (ngpu == 3) crosspoint = n; else if (ngpu == 2) crosspoint = 20160; else crosspoint = 0; crosspoint = 0; //n; //n -- > gpu always does next panel, 0 --> cpu always does next panel crosspoint = n; #if defined (ENABLE_TIMER) real_Double_t tget = magma_wtime(), tset = 0.0, ttot = 0.0; #endif if ( n > nb ) { // send first panel to cpu magma_setdevice(0); tmppanel = tmppanel0; magma_sgetmatrix_async(n, nb, dlA(0, 0, 0), ldda, tmppanel(0), ldpanel, queues[0][0] ); } #if defined (ENABLE_TIMER) for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_device_sync(); } tget = magma_wtime()-tget; #endif // Compute the Cholesky factorization A = L*L' for (j = 0; (j + nb) < n; j += nb) { #if defined (ENABLE_TIMER) therk[0] = therk[1] = therk[2] = therk[3] = tmtc = tcchol = tctrsm = tctm = tmnp = tcnp = 0.0; #endif blkid += 1; tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1; // Set the gpu number that holds the current panel id = (j / nb) % ngpu; magma_setdevice(id); // Set the local index where the current panel is j_local = j / (nb * ngpu) * nb; rows = n - j; // Wait for the panel on cpu magma_queue_sync( queues[id][0] ); if (j > 0 && prevtrsmrows > crosspoint) { #if defined (ENABLE_TIMER) tcnp = magma_wtime(); #endif tmpprevpanel = ((blkid - 1) % 2) == 0 ? tmppanel0 : tmppanel1; blasf77_sgemm( MagmaNoTransStr, MagmaConjTransStr, &rows, &nb, &nb, &mz_one, tmpprevpanel(j), &ldpanel, tmpprevpanel(j), &ldpanel, &z_one, tmppanel(j), &ldpanel ); #if defined (ENABLE_TIMER) tcnp = magma_wtime() - tcnp; ttot_cnp += tcnp; #endif } #if defined (ENABLE_TIMER) tcchol = magma_wtime(); #endif lapackf77_spotrf(MagmaLowerStr, &nb, tmppanel(j), &ldpanel, info); if (*info != 0) { *info = *info + j; break; } #if defined (ENABLE_TIMER) tcchol = magma_wtime() - tcchol; ttot_cchol += tcchol; tctrsm = magma_wtime(); #endif trsmrows = rows - nb; if (trsmrows > 0) { blasf77_strsm(MagmaRightStr, MagmaLowerStr, MagmaConjTransStr, MagmaNonUnitStr, &trsmrows, &nb, &z_one, tmppanel(j), &ldpanel, tmppanel(j + nb), &ldpanel); } #if defined (ENABLE_TIMER) tctrsm = magma_wtime() - tctrsm; ttot_ctrsm += tctrsm; tctm = magma_wtime(); #endif d = (id + 1) % ngpu; // send current panel to gpus for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) { magma_int_t myrows = 0; magma_int_t row_offset = 0; if ( d == id ) { dlpanel = dlA(d, j, j_local); myrows = rows; row_offset = 0; } else { dlpanel = dlP(d, 0, 0); myrows = trsmrows; row_offset = nb; } if (myrows > 0) { magma_setdevice(d); magma_ssetmatrix_async(myrows, nb, tmppanel(j + row_offset), ldpanel, dlpanel, ldda, queues[d][0] ); } } /* make sure panel is on GPUs */ d = (id + 1) % ngpu; for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) { magma_setdevice(d); magma_queue_sync( queues[d][0] ); } #if defined (ENABLE_TIMER) tctm = magma_wtime() - tctm; ttot_ctm += tctm; #endif if ( (j + nb) < n) { magma_int_t offset = 0; magma_int_t row_offset = 0; if (j + nb + nb < n) { d = (id + 1) % ngpu; magma_setdevice(d); magma_int_t j_local2 = (j + nb) / (nb * ngpu) * nb; if (trsmrows <= crosspoint) { #if defined (ENABLE_TIMER) tmnp = magma_wtime(); #endif // do gemm on look ahead panel if ( d == id ) { dlpanel = dlA(d, j + nb, j_local); } else { dlpanel = dlP(d, 0, 0); } magmablasSetKernelStream( queues[d][STREAM_ID(j_local2)] ); #define SSYRK_ON_DIAG #ifdef SSYRK_ON_DIAG magma_ssyrk( MagmaLower, MagmaNoTrans, nb, nb, m_one, dlpanel, ldda, one, dlA(d, j + nb, j_local2), ldda); magma_sgemm( MagmaNoTrans, MagmaConjTrans, trsmrows-nb, nb, nb, mz_one, dlpanel+nb, ldda, dlpanel, ldda, z_one, dlA(d, j + nb +nb, j_local2), ldda); #else magma_sgemm( MagmaNoTrans, MagmaConjTrans, trsmrows, nb, nb, mz_one, dlpanel, ldda, dlpanel, ldda, z_one, dlA(d, j + nb, j_local2), ldda); #endif #if defined (ENABLE_TIMER) magma_device_sync(); tmnp = magma_wtime() - tmnp; ttot_mnp += tmnp; #endif } // send next panel to cpu magma_queue_sync( queues[d][STREAM_ID(j_local2)] ); // make sure lookahead is done tmppanel = ((blkid+1) % 2 == 0) ? tmppanel0 : tmppanel1; magma_sgetmatrix_async(rows-nb, nb, dlA(d, j+nb, j_local2), ldda, tmppanel(j+nb), ldpanel, queues[d][0] ); tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1; offset = j + nb + nb; row_offset = nb; } else { offset = j + nb; row_offset = 0; } if (n - offset > 0) { // syrk on multiple gpu for (d = 0; d < ngpu; d++ ) { if ( d == id ) { dlpanels[d] = dlA(d, j + nb + row_offset, j_local); } else { dlpanels[d] = dlP(d, row_offset, 0); } } #if defined (ENABLE_TIMER) for( d=0; d < ngpu; d++ ) therk[d] = magma_wtime(); #endif //magmablasSetKernelStream( queues[d] ); //magma_ssyrk(MagmaLower, MagmaNoTrans, n - offset, nb, // m_one, dlpanel, ldda, // one, &d_lA[d][offset + offset*ldda], ldda ); #ifdef SSYRK_ON_DIAG magma_ssyrk_mgpu #else magma_ssyrk_mgpu2 #endif (ngpu, MagmaLower, MagmaNoTrans, nb, n - offset, nb, m_one, dlpanels, ldda, 0, one, d_lA, ldda, offset, nqueue, queues ); #if defined (ENABLE_TIMER) for( d=0; d < ngpu; d++ ) { magma_setdevice(d); magma_device_sync(); therk[d] = magma_wtime() - therk[d]; ttot_herk[d] += therk[d]; } #endif } prevtrsmrows = trsmrows; #if defined (ENABLE_TIMER) ttot += (tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp); printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(%d) %10.3lf\n", j, nb, rows, tmtc, tcnp, // gemm tcchol, // potrf tctrsm, // trsm (tcchol + tctrsm), (tmtc+tcnp+tcchol+tctrsm), therk[0], therk[1], therk[2], therk[3], // syrk tctm, // copy panel to GPU tmnp, // lookahead on GPU (id + 1) % ngpu, (tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp)); fflush(0); #endif } } for( d = 0; d < ngpu; d++ ) { magma_setdevice(d); for( id=0; id < nqueue; id++ ) { magma_queue_sync( queues[d][id] ); } } #if defined (ENABLE_TIMER) printf("\n%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf\n", n, n, 0, ttot_mtc, ttot_cnp, // gemm ttot_cchol, // potrf ttot_ctrsm, // trsm (ttot_cchol + ttot_ctrsm), (ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm), ttot_herk[0], ttot_herk[1], ttot_herk[2], ttot_herk[3], // syrk ttot_ctm, // copy panel to GPU ttot_mnp, // lookahead on GPU (ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp)); printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf (ratio)\n", n, n, 0, ttot_mtc/ttot, ttot_cnp/ttot, // gemm ttot_cchol/ttot, // potrf ttot_ctrsm/ttot, // trsm (ttot_cchol + ttot_ctrsm)/ttot, (ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm)/ttot, ttot_herk[0]/ttot, ttot_herk[1]/ttot, ttot_herk[2]/ttot, ttot_herk[3]/ttot, // syrk ttot_ctm/ttot, // copy panel to GPU ttot_mnp/ttot, // lookahead on GPU (ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp)/ttot); #endif // cholesky for the last block if (j < n && *info == 0) { rows = n - j; id = (j / nb) % ngpu; // Set the local index where the current panel is j_local = j / (nb * ngpu) * nb; magma_setdevice(id); #if defined (ENABLE_TIMER) tset = magma_wtime(); #endif magma_sgetmatrix(rows, rows, dlA(id, j, j_local), ldda, panel(j), ldpanel); lapackf77_spotrf(MagmaLowerStr, &rows, panel(j), &ldpanel, info); magma_ssetmatrix(rows, rows, panel(j), ldpanel, dlA(id, j, j_local), ldda); #if defined (ENABLE_TIMER) tset = magma_wtime() - tset; #endif } #if defined (ENABLE_TIMER) printf( " matrix_get,set: %10.3lf %10.3lf -> %10.3lf\n",tget,tset,ttot+tget+tset ); #endif } // end of else not upper // clean up for( d = 0; d < ngpu; d++ ) { magma_setdevice(d); for( j=0; j < nqueue; j++ ) { magma_queue_destroy( queues[d][j] ); } magma_free( d_lP[d] ); } } // end of not lapack // free workspace magma_free_pinned( panel ); magma_setdevice( orig_dev ); magmablasSetKernelStream( orig_stream ); return *info; } /* magma_spotrf_mgpu_right */
/** Purpose ------- SLAUUM computes the product U * U' or L' * L, where the triangular factor U or L is stored in the upper or lower triangular part of the array A. If UPLO = MagmaUpper then the upper triangle of the result is stored, overwriting the factor U in A. If UPLO = MagmaLower then the lower triangle of the result is stored, overwriting the factor L in A. This is the blocked form of the algorithm, calling Level 3 BLAS. Arguments --------- @param[in] uplo magma_uplo_t Specifies whether the triangular factor stored in the array A is upper or lower triangular: - = MagmaUpper: Upper triangular - = MagmaLower: Lower triangular @param[in] n INTEGER The order of the triangular factor U or L. N >= 0. @param[in,out] A COPLEX_16 array, dimension (LDA,N) On entry, the triangular factor U or L. On exit, if UPLO = MagmaUpper, the upper triangle of A is overwritten with the upper triangle of the product U * U'; if UPLO = MagmaLower, the lower triangle of A is overwritten with the lower triangle of the product L' * L. @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 = -k, the k-th argument had an illegal value @ingroup magma_sposv_aux ***************************************************************************/ extern "C" magma_int_t magma_slauum( magma_uplo_t uplo, magma_int_t n, float *A, magma_int_t lda, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define dA(i, j) (dA + (j)*ldda + (i)) /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t ldda, nb; magma_int_t i, ib; float c_one = MAGMA_S_ONE; float d_one = MAGMA_D_ONE; float *dA; int upper = (uplo == MagmaUpper); *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; ldda = ((n+31)/32)*32; if (MAGMA_SUCCESS != magma_smalloc( &dA, (n)*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_queue_t stream[2]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); nb = magma_get_spotrf_nb(n); if (nb <= 1 || nb >= n) lapackf77_slauum(uplo_, &n, A, &lda, info); else { if (upper) { /* Compute the product U * U'. */ for (i=0; i < n; i += nb) { ib=min(nb,n-i); magma_ssetmatrix_async( ib, ib, A(i,i), lda, dA(i, i), ldda, stream[1] ); magma_ssetmatrix_async( ib, (n-i-ib), A(i,i+ib), lda, dA(i,i+ib), ldda, stream[0] ); magma_queue_sync( stream[1] ); magma_strmm( MagmaRight, MagmaUpper, MagmaConjTrans, MagmaNonUnit, i, ib, c_one, dA(i,i), ldda, dA(0, i),ldda); lapackf77_slauum(MagmaUpperStr, &ib, A(i,i), &lda, info); magma_ssetmatrix_async( ib, ib, A(i, i), lda, dA(i, i), ldda, stream[0] ); if (i+ib < n) { magma_sgemm( MagmaNoTrans, MagmaConjTrans, i, ib, (n-i-ib), c_one, dA(0,i+ib), ldda, dA(i, i+ib),ldda, c_one, dA(0,i), ldda); magma_queue_sync( stream[0] ); magma_ssyrk( MagmaUpper, MagmaNoTrans, ib,(n-i-ib), d_one, dA(i, i+ib), ldda, d_one, dA(i, i), ldda); } magma_sgetmatrix( i+ib, ib, dA(0, i), ldda, A(0, i), lda ); } } else { /* Compute the product L' * L. */ for (i=0; i < n; i += nb) { ib=min(nb,n-i); magma_ssetmatrix_async( ib, ib, A(i,i), lda, dA(i, i), ldda, stream[1] ); magma_ssetmatrix_async( (n-i-ib), ib, A(i+ib, i), lda, dA(i+ib, i), ldda, stream[0] ); magma_queue_sync( stream[1] ); magma_strmm( MagmaLeft, MagmaLower, MagmaConjTrans, MagmaNonUnit, ib, i, c_one, dA(i,i), ldda, dA(i, 0),ldda); lapackf77_slauum(MagmaLowerStr, &ib, A(i,i), &lda, info); magma_ssetmatrix_async( ib, ib, A(i, i), lda, dA(i, i), ldda, stream[0] ); if (i+ib < n) { magma_sgemm(MagmaConjTrans, MagmaNoTrans, ib, i, (n-i-ib), c_one, dA( i+ib,i), ldda, dA(i+ib, 0),ldda, c_one, dA(i,0), ldda); magma_queue_sync( stream[0] ); magma_ssyrk(MagmaLower, MagmaConjTrans, ib, (n-i-ib), d_one, dA(i+ib, i), ldda, d_one, dA(i, i), ldda); } magma_sgetmatrix( ib, i+ib, dA(i, 0), ldda, A(i, 0), lda ); } } } magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free( dA ); return *info; }
int main( int argc, char** argv) { real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time; float *h_R = NULL, *h_P = NULL; magmaFloat_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; float mz_one = MAGMA_S_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_spotrf_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_SPOTRF( N ) / 1e9;; nb = magma_get_spotrf_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(float), NULL, NULL); cl_mem buffer2 = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, lda*nb*sizeof(float), NULL, NULL); for (k=0; k<num_gpus; k++) { h_R = (float*)clEnqueueMapBuffer(queues[2*k], buffer1, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, n2*sizeof(float), 0, NULL, NULL, NULL); h_P = (float*)clEnqueueMapBuffer(queues[2*k], buffer2, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, lda*nb*sizeof(float), 0, NULL, NULL, NULL); } #else TESTING_MALLOC_PIN( h_P, float, lda*nb ); TESTING_MALLOC_PIN( h_R, float, 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], float, 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_ssetmatrix( 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_slacpy( 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_ssetmatrix( 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_ssetmatrix( nk, j+nk, &h_R[j], lda, d_lA[k], j/(nb*tot_subs)*nb, ldda, queues[2*(k%num_gpus)]); }*/ } magma_spotrf_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_ssetmatrix( 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_slacpy( 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_ssetmatrix( 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_ssetmatrix( 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_spotrf_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_spotrf 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_sgetmatrix( 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_sgetmatrix( 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_slacpy( 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_sgetmatrix( 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) { float work[1], matnorm, diffnorm; float *h_A; TESTING_MALLOC_PIN( h_A, float, n2 ); init_matrix( N, h_A, lda ); cpu_time = magma_wtime(); if (uplo == MagmaLower) { lapackf77_spotrf( MagmaLowerStr, &N, h_A, &lda, &info ); } else { lapackf77_spotrf( MagmaUpperStr, &N, h_A, &lda, &info ); } cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf( "lapackf77_spotrf had error %d.\n", info ); /* ===================================================================== Check the result compared to LAPACK |R_magma - R_lapack| / |R_lapack| =================================================================== */ matnorm = lapackf77_slange("f", &N, &N, h_A, &lda, work); blasf77_saxpy(&n2, &mz_one, h_A, &ione, h_R, &ione); diffnorm = lapackf77_slange("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; }
extern "C" magma_int_t magma_slauum_gpu(char uplo, magma_int_t n, float *dA, magma_int_t ldda, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= SLAUUM computes the product U * U' or L' * L, where the triangular factor U or L is stored in the upper or lower triangular part of the array dA. If UPLO = 'U' or 'u' then the upper triangle of the result is stored, overwriting the factor U in dA. If UPLO = 'L' or 'l' then the lower triangle of the result is stored, overwriting the factor L in dA. This is the blocked form of the algorithm, calling Level 3 BLAS. Arguments ========= UPLO (input) CHARACTER*1 Specifies whether the triangular factor stored in the array dA is upper or lower triangular: = 'U': Upper triangular = 'L': Lower triangular N (input) INTEGER The order of the triangular factor U or L. N >= 0. dA (input/output) DOUBLE PRECISION array on the GPU, dimension (LDDA,N) On entry, the triangular factor U or L. On exit, if UPLO = 'U', the upper triangle of dA is overwritten with the upper triangle of the product U * U'; if UPLO = 'L', the lower triangle of dA is overwritten with the lower triangle of the product L' * L. LDDA (input) INTEGER The leading dimension of the array A. LDDA >= max(1,N). INFO (output) INTEGER = 0: successful exit < 0: if INFO = -k, the k-th argument had an illegal value ===================================================================== */ /* Local variables */ char uplo_[2] = {uplo, 0}; magma_int_t nb, i, ib; float d_one = MAGMA_D_ONE; float c_one = MAGMA_S_ONE; float *work; int upper = lapackf77_lsame(uplo_, "U"); *info = 0; if ((! upper) && (! lapackf77_lsame(uplo_, "L"))) *info = -1; else if (n < 0) *info = -2; else if (ldda < max(1,n)) *info = -4; if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } nb = magma_get_spotrf_nb(n); if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, nb*nb )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } magma_queue_t stream[2]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); if (nb <= 1 || nb >= n) { magma_sgetmatrix( n, n, dA, ldda, work, n ); lapackf77_slauum(uplo_, &n, work, &n, info); magma_ssetmatrix( n, n, work, n, dA, ldda ); } else { if (upper) { /* Compute inverse of upper triangular matrix */ for (i=0; i < n; i += nb) { ib = min(nb, (n-i)); /* Compute the product U * U'. */ magma_strmm( MagmaRight, MagmaUpper, MagmaTrans, MagmaNonUnit, i, ib, c_one, dA(i,i), ldda, dA(0, i),ldda); magma_sgetmatrix( ib, ib, dA(i, i), ldda, work, ib ); lapackf77_slauum(MagmaUpperStr, &ib, work, &ib, info); magma_ssetmatrix( ib, ib, work, ib, dA(i, i), ldda ); if(i+ib < n) { magma_sgemm( MagmaNoTrans, MagmaTrans, i, ib, (n-i-ib), c_one, dA(0,i+ib), ldda, dA(i, i+ib), ldda, c_one, dA(0,i), ldda); magma_ssyrk( MagmaUpper, MagmaNoTrans, ib,(n-i-ib), d_one, dA(i, i+ib), ldda, d_one, dA(i, i), ldda); } } } else { /* Compute the product L' * L. */ for(i=0; i<n; i=i+nb) { ib=min(nb,(n-i)); magma_strmm( MagmaLeft, MagmaLower, MagmaTrans, MagmaNonUnit, ib, i, c_one, dA(i,i), ldda, dA(i, 0),ldda); magma_sgetmatrix( ib, ib, dA(i, i), ldda, work, ib ); lapackf77_slauum(MagmaLowerStr, &ib, work, &ib, info); magma_ssetmatrix( ib, ib, work, ib, dA(i, i), ldda ); if((i+ib) < n) { magma_sgemm( MagmaTrans, MagmaNoTrans, ib, i, (n-i-ib), c_one, dA( i+ib,i), ldda, dA(i+ib, 0),ldda, c_one, dA(i,0), ldda); magma_ssyrk( MagmaLower, MagmaTrans, ib, (n-i-ib), d_one, dA(i+ib, i), ldda, d_one, dA(i, i), ldda); } } } } magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free_pinned( work ); return *info; }
int main( int argc, char** argv) { real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time; float *h_A, *h_R; magmaFloat_ptr d_lA[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, info; float mz_one = MAGMA_S_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; float work[1], matnorm, diffnorm; magma_int_t num_gpus0 = 1, num_gpus, flag = 0; int nb, mb, 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; }else exit(1); } if(strcmp("-NGPU", argv[i])==0) num_gpus0 = atoi(argv[++i]); if(strcmp("-UPLO", argv[i])==0){ if(strcmp("L", argv[++i])==0){ uplo = MagmaLower; }else{ uplo = MagmaUpper; } } } } else { printf("\nUsage: \n"); printf(" testing_spotrf_mgpu -N %d -NGPU %d -UPLO -L\n\n", 1024, num_gpus0); } /* looking for max. ldda */ ldda = 0; n2 = 0; for(i=0;i<10;i++){ N = size[i]; nb = magma_get_spotrf_nb(N); mb = nb; if(num_gpus0 > N/nb){ num_gpus = N/nb; if(N%nb != 0) num_gpus ++; }else{ num_gpus = num_gpus0; } n_local = nb*(1+N/(nb*num_gpus))*mb*((N+mb-1)/mb); if(n_local > ldda) ldda = n_local; if(n2 < N*N) n2 = N*N; if(flag != 0) break; } /* Allocate host memory for the matrix */ TESTING_MALLOC_PIN( h_A, float, n2 ); TESTING_MALLOC_PIN( h_R, float, n2 ); /* Initialize */ magma_queue_t queues[MagmaMaxGPUs * 2]; //magma_queue_t queues[MagmaMaxGPUs]; magma_device_t devices[ MagmaMaxGPUs ]; int num = 0; magma_err_t err; magma_init(); err = magma_get_devices( devices, MagmaMaxGPUs, &num ); if ( err != 0 || num < 1 ) { fprintf( stderr, "magma_get_devices failed: %d\n", err ); exit(-1); } for(i=0;i<num_gpus;i++){ err = magma_queue_create( devices[i], &queues[2*i] ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", err ); exit(-1); } err = magma_queue_create( devices[i], &queues[2*i+1] ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", err ); exit(-1); } } printf("each buffer size: %d\n", ldda); /* allocate local matrix on Buffers */ for(i=0; i<num_gpus0; i++){ TESTING_MALLOC_DEV( d_lA[i], float, ldda ); } printf("\n\n"); printf("Using GPUs: %d\n", num_gpus0); if(uplo == MagmaUpper){ printf("\n testing_spotrf_mgpu -N %d -NGPU %d -UPLO U\n\n", N, num_gpus0); }else{ printf("\n testing_spotrf_mgpu -N %d -NGPU %d -UPLO L\n\n", N, num_gpus0); } 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; ldda = ((N+31)/32)*32; gflops = FLOPS( (float)N ) * 1e-9; /* Initialize the matrix */ lapackf77_slarnv( &ione, ISEED, &n2, h_A ); /* Symmetrize and increase the diagonal */ for( int i = 0; i < N; ++i ) { MAGMA_S_SET2REAL( h_A(i,i), MAGMA_S_REAL(h_A(i,i)) + N ); for( int j = 0; j < i; ++j ) { h_A(i, j) = MAGMA_S_CNJG( h_A(j,i) ); } } lapackf77_slacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda ); /* Warm up to measure the performance */ nb = magma_get_spotrf_nb(N); if(num_gpus0 > N/nb){ num_gpus = N/nb; 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; } /* distribute matrix to gpus */ if(uplo == MagmaUpper){ // Upper ldda = ((N+mb-1)/mb)*mb; for(j=0;j<N;j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_ssetmatrix(N, nk, &h_A[j*lda], 0, lda, d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda, queues[2*k]); } }else{ // Lower ldda = (1+N/(nb*num_gpus))*nb; for(j=0;j<N;j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_ssetmatrix(nk, N, &h_A[j], 0, lda, d_lA[k], (j/(nb*num_gpus)*nb), ldda, queues[2*k]); } } magma_spotrf_mgpu( num_gpus, uplo, N, d_lA, 0, ldda, &info, queues ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ /* distribute matrix to gpus */ if(uplo == MagmaUpper){ // Upper ldda = ((N+mb-1)/mb)*mb; for(j=0;j<N;j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_ssetmatrix(N, nk, &h_A[j*lda], 0, lda, d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda, queues[2*k]); } }else{ // Lower ldda = (1+N/(nb*num_gpus))*nb; for(j=0;j<N;j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_ssetmatrix(nk, N, &h_A[j], 0, lda, d_lA[k], (j/(nb*num_gpus)*nb), ldda, queues[2*k]); } } gpu_time = magma_wtime(); magma_spotrf_mgpu( num_gpus, uplo, N, d_lA, 0, ldda, &info, queues ); gpu_time = magma_wtime() - gpu_time; if (info != 0) printf( "magma_spotrf had error %d.\n", info ); gpu_perf = gflops / gpu_time; /* gather matrix from gpus */ if(uplo==MagmaUpper){ // Upper for(j=0;j<N;j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_sgetmatrix(N, nk, d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda, &h_R[j*lda], 0, lda, queues[2*k]); } }else{ // Lower for(j=0; j<N; j+=nb){ k = (j/nb)%num_gpus; nk = min(nb, N-j); magma_sgetmatrix( nk, N, d_lA[k], (j/(nb*num_gpus)*nb), ldda, &h_R[j], 0, lda, queues[2*k] ); } } /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); if(uplo == MagmaLower){ lapackf77_spotrf( MagmaLowerStr, &N, h_A, &lda, &info ); }else{ lapackf77_spotrf( MagmaUpperStr, &N, h_A, &lda, &info ); } cpu_time = magma_wtime() - cpu_time; if (info != 0) printf( "lapackf77_spotrf had error %d.\n", info ); cpu_perf = gflops / cpu_time; /* ===================================================================== Check the result compared to LAPACK |R_magma - R_lapack| / |R_lapack| =================================================================== */ matnorm = lapackf77_slange("f", &N, &N, h_A, &lda, work); blasf77_saxpy(&n2, &mz_one, h_A, &ione, h_R, &ione); diffnorm = lapackf77_slange("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 ); if (flag != 0) break; } /* clean up */ TESTING_FREE_PIN( h_A ); TESTING_FREE_PIN( h_R ); for(i=0;i<num_gpus;i++){ TESTING_FREE_DEV( d_lA[i] ); magma_queue_destroy( queues[2*i] ); magma_queue_destroy( queues[2*i+1] ); } magma_finalize(); }
/** Purpose ------- SPOTRF 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 factorization has the form A = U**T * U, if uplo = MagmaUpper, or A = L * L**T, 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. If the current stream is NULL, this version replaces it with user defined stream to overlap computation with communication. 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 REAL array, dimension (LDA,N) On entry, the symmetric matrix A. If uplo = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If uplo = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. \n On exit, if INFO = 0, the factor U or L from the Cholesky factorization A = U**T * U or A = L * L**T. \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_sposv_comp ********************************************************************/ extern "C" magma_int_t magma_spotrf(magma_uplo_t uplo, magma_int_t n, float *A, magma_int_t lda, magma_int_t *info) { #define A(i, j) (A + (j)*lda + (i)) #define dA(i, j) (work + (j)*ldda + (i)) /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); magma_int_t ldda, nb; magma_int_t j, jb; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; float *work; float d_one = 1.0; float d_neg_one = -1.0; int upper = (uplo == MagmaUpper); *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_int_t num_gpus = magma_num_gpus(); if ( num_gpus > 1 ) { /* call multiple-GPU interface */ return magma_spotrf_m(num_gpus, uplo, n, A, lda, info); } ldda = ((n+31)/32)*32; if (MAGMA_SUCCESS != magma_smalloc( &work, (n)*ldda )) { /* alloc failed so call the non-GPU-resident version */ return magma_spotrf_m(num_gpus, uplo, n, A, lda, info); } /* Define user stream if current stream is NULL */ cudaStream_t stream[3], current_stream; magmablasGetKernelStream(¤t_stream); magma_queue_create( &stream[0] ); magma_queue_create( &stream[2] ); if (current_stream == NULL) { magma_queue_create( &stream[1] ); magmablasSetKernelStream(stream[1]); } else stream[1] = current_stream; nb = magma_get_spotrf_nb(n); if (nb <= 1 || nb >= n) { lapackf77_spotrf(uplo_, &n, A, &lda, info); } else { /* Use hybrid blocked code. */ if (upper) { /* Compute the Cholesky factorization A = U'*U. */ for (j=0; j < n; j += nb) { /* Update and factorize the current diagonal block and test for non-positive-definiteness. Computing MIN */ jb = min(nb, (n-j)); magma_ssetmatrix_async( jb, (n-j), A(j, j), lda, dA(j, j), ldda, stream[1]); magma_ssyrk(MagmaUpper, MagmaTrans, jb, j, d_neg_one, dA(0, j), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, A(j, j), lda, stream[0] ); if ( (j+jb) < n) { magma_sgemm(MagmaTrans, MagmaNoTrans, jb, (n-j-jb), j, c_neg_one, dA(0, j ), ldda, dA(0, j+jb), ldda, c_one, dA(j, j+jb), ldda); } magma_sgetmatrix_async( j, jb, dA(0, j), ldda, A (0, j), lda, stream[2] ); magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaUpperStr, &jb, A(j, j), &lda, info); if (*info != 0) { *info = *info + j; break; } magma_ssetmatrix_async( jb, jb, A(j, j), lda, dA(j, j), ldda, stream[0] ); magma_queue_sync( stream[0] ); if ( (j+jb) < n ) { magma_strsm(MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit, jb, (n-j-jb), c_one, dA(j, j ), ldda, dA(j, j+jb), ldda); } } } else { //========================================================= // Compute the Cholesky factorization A = L*L'. for (j=0; j < n; j += nb) { // Update and factorize the current diagonal block and test // for non-positive-definiteness. Computing MIN jb = min(nb, (n-j)); magma_ssetmatrix_async( (n-j), jb, A(j, j), lda, dA(j, j), ldda, stream[1]); magma_ssyrk(MagmaLower, MagmaNoTrans, jb, j, d_neg_one, dA(j, 0), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j,j), ldda, A(j,j), lda, stream[0] ); if ( (j+jb) < n) { magma_sgemm( MagmaNoTrans, MagmaTrans, (n-j-jb), jb, j, c_neg_one, dA(j+jb, 0), ldda, dA(j, 0), ldda, c_one, dA(j+jb, j), ldda); } magma_sgetmatrix_async( jb, j, dA(j, 0), ldda, A(j, 0), lda, stream[2] ); magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaLowerStr, &jb, A(j, j), &lda, info); if (*info != 0) { *info = *info + j; break; } magma_ssetmatrix_async( jb, jb, A(j, j), lda, dA(j, j), ldda, stream[0] ); magma_queue_sync( stream[0] ); if ( (j+jb) < n) { magma_strsm(MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit, (n-j-jb), jb, c_one, dA(j, j), ldda, dA(j+jb, j), ldda); } } } } magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[2] ); if (current_stream == NULL) { magma_queue_destroy( stream[1] ); magmablasSetKernelStream(NULL); } magma_free( work ); return *info; } /* magma_spotrf */
magma_int_t magmaf_get_spotrf_nb( magma_int_t *m ) { return magma_get_spotrf_nb( *m ); }
/***************************************************************************//** Purpose ------- SPOTRF computes the Cholesky factorization of a real symmetric positive definite matrix dA. The factorization has the form dA = U**H * U, if UPLO = MagmaUpper, or dA = 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] uplo magma_uplo_t - = MagmaUpper: Upper triangle of dA is stored; - = MagmaLower: Lower triangle of dA is stored. @param[in] n INTEGER The order of the matrix dA. N >= 0. @param[in,out] dA REAL array on the GPU, dimension (LDDA,N) On entry, the symmetric matrix dA. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of dA contains the upper triangular part of the matrix dA, and the strictly lower triangular part of dA is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of dA contains the lower triangular part of the matrix dA, and the strictly upper triangular part of dA is not referenced. \n On exit, if INFO = 0, the factor U or L from the Cholesky factorization dA = U**H * U or dA = L * L**H. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,N). To benefit from coalescent memory accesses LDDA must be divisible by 16. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i, the leading minor of order i is not positive definite, and the factorization could not be completed. @ingroup magma_potrf *******************************************************************************/ extern "C" magma_int_t magma_spotrf_gpu( magma_uplo_t uplo, magma_int_t n, magmaFloat_ptr dA, magma_int_t ldda, magma_int_t *info ) { #ifdef HAVE_clBLAS #define dA(i_, j_) dA, ((i_) + (j_)*ldda + dA_offset) #else #define dA(i_, j_) (dA + (i_) + (j_)*ldda) #endif /* Constants */ const float c_one = MAGMA_S_ONE; const float c_neg_one = MAGMA_S_NEG_ONE; const float d_one = 1.0; const float d_neg_one = -1.0; /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); bool upper = (uplo == MagmaUpper); magma_int_t j, jb, nb; float *work; *info = 0; if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } nb = magma_get_spotrf_nb( n ); if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, nb*nb )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } magma_queue_t queues[2]; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queues[0] ); magma_queue_create( cdev, &queues[1] ); if (nb <= 1 || nb >= n) { /* Use unblocked code. */ magma_sgetmatrix( n, n, dA(0,0), ldda, work, n, queues[0] ); lapackf77_spotrf( uplo_, &n, work, &n, info ); magma_ssetmatrix( n, n, work, n, dA(0,0), ldda, queues[0] ); } else { /* Use blocked code. */ if (upper) { //========================================================= /* Compute the Cholesky factorization A = U'*U. */ for (j=0; j < n; j += nb) { // apply all previous updates to diagonal block, // then transfer it to CPU jb = min( nb, n-j ); magma_ssyrk( MagmaUpper, MagmaConjTrans, jb, j, d_neg_one, dA(0, j), ldda, d_one, dA(j, j), ldda, queues[1] ); magma_queue_sync( queues[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, queues[0] ); // apply all previous updates to block row right of diagonal block if (j+jb < n) { magma_sgemm( MagmaConjTrans, MagmaNoTrans, jb, n-j-jb, j, c_neg_one, dA(0, j ), ldda, dA(0, j+jb), ldda, c_one, dA(j, j+jb), ldda, queues[1] ); } // simultaneous with above sgemm, transfer diagonal block, // factor it on CPU, and test for positive definiteness magma_queue_sync( queues[0] ); lapackf77_spotrf( MagmaUpperStr, &jb, work, &jb, info ); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, queues[1] ); if (*info != 0) { *info = *info + j; break; } // apply diagonal block to block row right of diagonal block if (j+jb < n) { magma_strsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit, jb, n-j-jb, c_one, dA(j, j), ldda, dA(j, j+jb), ldda, queues[1] ); } } } else { //========================================================= // Compute the Cholesky factorization A = L*L'. for (j=0; j < n; j += nb) { // apply all previous updates to diagonal block, // then transfer it to CPU jb = min( nb, n-j ); magma_ssyrk( MagmaLower, MagmaNoTrans, jb, j, d_neg_one, dA(j, 0), ldda, d_one, dA(j, j), ldda, queues[1] ); magma_queue_sync( queues[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, queues[0] ); // apply all previous updates to block column below diagonal block if (j+jb < n) { magma_sgemm( MagmaNoTrans, MagmaConjTrans, n-j-jb, jb, j, c_neg_one, dA(j+jb, 0), ldda, dA(j, 0), ldda, c_one, dA(j+jb, j), ldda, queues[1] ); } // simultaneous with above sgemm, transfer diagonal block, // factor it on CPU, and test for positive definiteness magma_queue_sync( queues[0] ); lapackf77_spotrf( MagmaLowerStr, &jb, work, &jb, info ); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, queues[1] ); if (*info != 0) { *info = *info + j; break; } // apply diagonal block to block column below diagonal if (j+jb < n) { magma_strsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit, n-j-jb, jb, c_one, dA(j, j), ldda, dA(j+jb, j), ldda, queues[1] ); } } } } magma_queue_destroy( queues[0] ); magma_queue_destroy( queues[1] ); magma_free_pinned( work ); return *info; } /* magma_spotrf_gpu */
/***************************************************************************//** Purpose ------- SPOTRF 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 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. This uses multiple queues to overlap communication and computation. 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 REAL array, dimension (LDA,N) On entry, the symmetric matrix A. If uplo = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A, and the strictly lower triangular part of A is not referenced. If uplo = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A, and the strictly upper triangular part of A is not referenced. \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_potrf *******************************************************************************/ extern "C" magma_int_t magma_spotrf( magma_uplo_t uplo, magma_int_t n, float *A, magma_int_t lda, magma_int_t *info ) { #define A(i_, j_) (A + (i_) + (j_)*lda) #ifdef HAVE_clBLAS #define dA(i_, j_) dA, ((i_) + (j_)*ldda) #else #define dA(i_, j_) (dA + (i_) + (j_)*ldda) #endif /* Constants */ const float c_one = MAGMA_S_ONE; const float c_neg_one = MAGMA_S_NEG_ONE; const float d_one = 1.0; const float d_neg_one = -1.0; /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); bool upper = (uplo == MagmaUpper); magma_int_t j, jb, ldda, nb; magmaFloat_ptr dA = NULL; /* Check arguments */ *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; nb = magma_get_spotrf_nb( n ); if (nb <= 1 || nb >= n) { lapackf77_spotrf( uplo_, &n, A, &lda, info ); } else { /* Use hybrid blocked code. */ ldda = magma_roundup( n, 32 ); magma_int_t ngpu = magma_num_gpus(); if ( ngpu > 1 ) { /* call multi-GPU non-GPU-resident interface */ return magma_spotrf_m( ngpu, uplo, n, A, lda, info ); } if (MAGMA_SUCCESS != magma_smalloc( &dA, n*ldda )) { /* alloc failed so call the non-GPU-resident version */ return magma_spotrf_m( ngpu, uplo, n, A, lda, info ); } magma_queue_t queues[2] = { NULL, NULL }; magma_device_t cdev; magma_getdevice( &cdev ); magma_queue_create( cdev, &queues[0] ); magma_queue_create( cdev, &queues[1] ); if (upper) { /* Compute the Cholesky factorization A = U'*U. */ for (j=0; j < n; j += nb) { /* Update and factorize the current diagonal block and test for non-positive-definiteness. */ jb = min( nb, n-j ); magma_ssetmatrix_async( jb, n-j, A(j, j), lda, dA(j, j), ldda, queues[1] ); magma_ssyrk( MagmaUpper, MagmaConjTrans, jb, j, d_neg_one, dA(0, j), ldda, d_one, dA(j, j), ldda, queues[1] ); magma_queue_sync( queues[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, A(j, j), lda, queues[0] ); if (j+jb < n) { magma_sgemm( MagmaConjTrans, MagmaNoTrans, jb, n-j-jb, j, c_neg_one, dA(0, j ), ldda, dA(0, j+jb), ldda, c_one, dA(j, j+jb), ldda, queues[1] ); } magma_queue_sync( queues[0] ); // this could be on any queue; it isn't needed until exit. magma_sgetmatrix_async( j, jb, dA(0, j), ldda, A(0, j), lda, queues[0] ); lapackf77_spotrf( MagmaUpperStr, &jb, A(j, j), &lda, info ); if (*info != 0) { *info = *info + j; break; } magma_ssetmatrix_async( jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0] ); magma_queue_sync( queues[0] ); if (j+jb < n) { magma_strsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit, jb, n-j-jb, c_one, dA(j, j ), ldda, dA(j, j+jb), ldda, queues[1] ); } } } else { //========================================================= // Compute the Cholesky factorization A = L*L'. for (j=0; j < n; j += nb) { // Update and factorize the current diagonal block and test // for non-positive-definiteness. jb = min( nb, n-j ); magma_ssetmatrix_async( n-j, jb, A(j, j), lda, dA(j, j), ldda, queues[1] ); magma_ssyrk( MagmaLower, MagmaNoTrans, jb, j, d_neg_one, dA(j, 0), ldda, d_one, dA(j, j), ldda, queues[1] ); magma_queue_sync( queues[1] ); magma_sgetmatrix_async( jb, jb, dA(j,j), ldda, A(j,j), lda, queues[0] ); if (j+jb < n) { magma_sgemm( MagmaNoTrans, MagmaConjTrans, n-j-jb, jb, j, c_neg_one, dA(j+jb, 0), ldda, dA(j, 0), ldda, c_one, dA(j+jb, j), ldda, queues[1] ); } magma_queue_sync( queues[0] ); // this could be on any queue; it isn't needed until exit. magma_sgetmatrix_async( jb, j, dA(j, 0), ldda, A(j, 0), lda, queues[0] ); lapackf77_spotrf( MagmaLowerStr, &jb, A(j, j), &lda, info ); if (*info != 0) { *info = *info + j; break; } magma_ssetmatrix_async( jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0] ); magma_queue_sync( queues[0] ); if (j+jb < n) { magma_strsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit, n-j-jb, jb, c_one, dA(j, j), ldda, dA(j+jb, j), ldda, queues[1] ); } } } magma_queue_destroy( queues[0] ); magma_queue_destroy( queues[1] ); magma_free( dA ); } return *info; } /* magma_spotrf */
/** Purpose ------- SPOTRF computes the Cholesky factorization of a real symmetric positive definite matrix dA. The factorization has the form dA = U**H * U, if UPLO = MagmaUpper, or dA = 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. If the current stream is NULL, this version replaces it with a new stream to overlap computation with communication. Arguments --------- @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of dA is stored; - = MagmaLower: Lower triangle of dA is stored. @param[in] n INTEGER The order of the matrix dA. N >= 0. @param[in,out] dA REAL array on the GPU, dimension (LDDA,N) On entry, the symmetric matrix dA. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of dA contains the upper triangular part of the matrix dA, and the strictly lower triangular part of dA is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of dA contains the lower triangular part of the matrix dA, and the strictly upper triangular part of dA is not referenced. \n On exit, if INFO = 0, the factor U or L from the Cholesky factorization dA = U**H * U or dA = L * L**H. @param[in] ldda INTEGER The leading dimension of the array dA. LDDA >= max(1,N). To benefit from coalescent memory accesses LDDA must be divisible by 16. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i, the leading minor of order i is not positive definite, and the factorization could not be completed. @ingroup magma_sposv_comp ********************************************************************/ extern "C" magma_int_t magma_spotrf_gpu( magma_uplo_t uplo, magma_int_t n, magmaFloat_ptr dA, magma_int_t ldda, magma_int_t *info) { #define dA(i, j) (dA + (j)*ldda + (i)) magma_int_t j, jb, nb; const char* uplo_ = lapack_uplo_const( uplo ); float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; float *work; float d_one = 1.0; float d_neg_one = -1.0; int upper = (uplo == MagmaUpper); *info = 0; if (! upper && uplo != MagmaLower) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } nb = magma_get_spotrf_nb(n); if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, nb*nb )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } /* Define user stream if current stream is NULL */ magma_queue_t stream[2]; magma_queue_t orig_stream; magmablasGetKernelStream( &orig_stream ); magma_queue_create( &stream[0] ); if (orig_stream == NULL) { magma_queue_create( &stream[1] ); magmablasSetKernelStream(stream[1]); } else { stream[1] = orig_stream; } if ((nb <= 1) || (nb >= n)) { /* Use unblocked code. */ magma_sgetmatrix_async( n, n, dA, ldda, work, n, stream[1] ); magma_queue_sync( stream[1] ); lapackf77_spotrf(uplo_, &n, work, &n, info); magma_ssetmatrix_async( n, n, work, n, dA, ldda, stream[1] ); } else { /* Use blocked code. */ if (upper) { /* Compute the Cholesky factorization A = U'*U. */ for (j=0; j < n; j += nb) { /* Update and factorize the current diagonal block and test for non-positive-definiteness. Computing MIN */ jb = min(nb, (n-j)); magma_ssyrk(MagmaUpper, MagmaConjTrans, jb, j, d_neg_one, dA(0, j), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, stream[0] ); if ( (j+jb) < n) { /* Compute the current block row. */ magma_sgemm(MagmaConjTrans, MagmaNoTrans, jb, (n-j-jb), j, c_neg_one, dA(0, j ), ldda, dA(0, j+jb), ldda, c_one, dA(j, j+jb), ldda); } magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaUpperStr, &jb, work, &jb, info); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, stream[1] ); if (*info != 0) { *info = *info + j; break; } if ( (j+jb) < n) { magma_strsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit, jb, (n-j-jb), c_one, dA(j, j ), ldda, dA(j, j+jb), ldda); } } } else { //========================================================= // Compute the Cholesky factorization A = L*L'. for (j=0; j < n; j += nb) { // Update and factorize the current diagonal block and test // for non-positive-definiteness. Computing MIN jb = min(nb, (n-j)); magma_ssyrk(MagmaLower, MagmaNoTrans, jb, j, d_neg_one, dA(j, 0), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, stream[0] ); if ( (j+jb) < n) { magma_sgemm( MagmaNoTrans, MagmaConjTrans, (n-j-jb), jb, j, c_neg_one, dA(j+jb, 0), ldda, dA(j, 0), ldda, c_one, dA(j+jb, j), ldda); } magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaLowerStr, &jb, work, &jb, info); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, stream[1] ); if (*info != 0) { *info = *info + j; break; } if ( (j+jb) < n) { magma_strsm(MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit, (n-j-jb), jb, c_one, dA(j, j), ldda, dA(j+jb, j), ldda); } } } } magma_free_pinned( work ); magma_queue_destroy( stream[0] ); if (orig_stream == NULL) { magma_queue_destroy( stream[1] ); } magmablasSetKernelStream( orig_stream ); return *info; } /* magma_spotrf_gpu */
/** Purpose ------- STRTRI computes the inverse of a real upper or lower triangular matrix A. This is the Level 3 BLAS version of the algorithm. Arguments --------- @param[in] uplo magma_uplo_t - = MagmaUpper: A is upper triangular; - = MagmaLower: A is lower triangular. @param[in] diag magma_diag_t - = MagmaNonUnit: A is non-unit triangular; - = MagmaUnit: A is unit triangular. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A REAL array, dimension (LDA,N) On entry, the triangular matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of the array A contains the upper triangular matrix, and the strictly lower triangular part of A is not referenced. If UPLO = MagmaLower, the leading N-by-N lower triangular part of the array A contains the lower triangular matrix, and the strictly upper triangular part of A is not referenced. If DIAG = MagmaUnit, the diagonal elements of A are also not referenced and are assumed to be 1. On exit, the (triangular) inverse of the original matrix, in the same storage format. @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 - > 0: if INFO = i, A(i,i) is exactly zero. The triangular matrix is singular and its inverse cannot be computed. @ingroup magma_sgesv_aux ********************************************************************/ extern "C" magma_int_t magma_strtri( magma_uplo_t uplo, magma_diag_t diag, magma_int_t n, float *A, magma_int_t lda, magma_int_t *info) { #define A(i, j) ( A + (i) + (j)*lda ) #define dA(i, j) (dA + (i) + (j)*ldda) /* Local variables */ const char* uplo_ = lapack_uplo_const( uplo ); const char* diag_ = lapack_diag_const( diag ); magma_int_t ldda, nb, nn, j, jb; float c_zero = MAGMA_S_ZERO; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; float *dA; int upper = (uplo == MagmaUpper); int nounit = (diag == MagmaNonUnit); *info = 0; if (! upper && uplo != MagmaLower) *info = -1; else if (! nounit && diag != MagmaUnit) *info = -2; else if (n < 0) *info = -3; else if (lda < max(1,n)) *info = -5; if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } /* Quick return */ if ( n == 0 ) return *info; /* Check for singularity if non-unit */ if (nounit) { for (j=0; j < n; ++j) { if ( MAGMA_S_EQUAL( *A(j,j), c_zero )) { *info = j+1; // Fortran index return *info; } } } /* Determine the block size for this environment */ nb = magma_get_spotrf_nb(n); ldda = ((n+31)/32)*32; if (MAGMA_SUCCESS != magma_smalloc( &dA, (n)*ldda )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_queue_t stream[2]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); if (nb <= 1 || nb >= n) lapackf77_strtri(uplo_, diag_, &n, A, &lda, info); else { if (upper) { /* Compute inverse of upper triangular matrix */ for (j=0; j < n; j += nb) { jb = min(nb, (n-j)); magma_ssetmatrix( jb, (n-j), A(j, j), lda, dA(j, j), ldda ); /* Compute rows 1:j-1 of current block column */ magma_strmm( MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit, j, jb, c_one, dA(0,0), ldda, dA(0, j),ldda); magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, j, jb, c_neg_one, dA(j,j), ldda, dA(0, j),ldda); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, A(j, j), lda, stream[1] ); magma_sgetmatrix_async( j, jb, dA(0, j), ldda, A(0, j), lda, stream[0] ); magma_queue_sync( stream[1] ); /* Compute inverse of current diagonal block */ lapackf77_strtri(MagmaUpperStr, diag_, &jb, A(j,j), &lda, info); magma_ssetmatrix( jb, jb, A(j, j), lda, dA(j, j), ldda ); } } else { /* Compute inverse of lower triangular matrix */ nn=((n-1)/nb)*nb+1; for (j=nn-1; j >= 0; j -= nb) { jb=min(nb,(n-j)); if ((j+jb) < n) { magma_ssetmatrix( (n-j), jb, A(j, j), lda, dA(j, j), ldda ); /* Compute rows j+jb:n of current block column */ magma_strmm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaNonUnit, (n-j-jb), jb, c_one, dA(j+jb,j+jb), ldda, dA(j+jb, j), ldda ); magma_strsm( MagmaRight, MagmaLower, MagmaNoTrans, MagmaNonUnit, (n-j-jb), jb, c_neg_one, dA(j,j), ldda, dA(j+jb, j), ldda ); magma_sgetmatrix_async( n-j-jb, jb, dA(j+jb, j), ldda, A(j+jb, j), lda, stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j,j), ldda, A(j,j), lda, stream[0] ); magma_queue_sync( stream[0] ); } /* Compute inverse of current diagonal block */ lapackf77_strtri(MagmaLowerStr, diag_, &jb, A(j,j), &lda, info); magma_ssetmatrix( jb, jb, A(j, j), lda, dA(j, j), ldda ); } } } magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); magma_free( dA ); return *info; }
extern "C" magma_int_t magma_spotrf_gpu(char uplo, magma_int_t n, float *dA, magma_int_t ldda, magma_int_t *info) { /* -- MAGMA (version 1.4.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver August 2013 Purpose ======= SPOTRF computes the Cholesky factorization of a real symmetric positive definite matrix dA. The factorization has the form dA = U**T * U, if UPLO = 'U', or dA = L * L**T, if UPLO = 'L', where U is an upper triangular matrix and L is lower triangular. This is the block version of the algorithm, calling Level 3 BLAS. If the current stream is NULL, this version replaces it with user defined stream to overlap computation with communication. Arguments ========= UPLO (input) CHARACTER*1 = 'U': Upper triangle of dA is stored; = 'L': Lower triangle of dA is stored. N (input) INTEGER The order of the matrix dA. N >= 0. dA (input/output) REAL array on the GPU, dimension (LDDA,N) On entry, the symmetric matrix dA. If UPLO = 'U', the leading N-by-N upper triangular part of dA contains the upper triangular part of the matrix dA, and the strictly lower triangular part of dA is not referenced. If UPLO = 'L', the leading N-by-N lower triangular part of dA contains the lower triangular part of the matrix dA, and the strictly upper triangular part of dA is not referenced. On exit, if INFO = 0, the factor U or L from the Cholesky factorization dA = U**T * U or dA = L * L**T. LDDA (input) INTEGER The leading dimension of the array dA. LDDA >= max(1,N). To benefit from coalescent memory accesses LDDA must be dividable by 16. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i, the leading minor of order i is not positive definite, and the factorization could not be completed. ===================================================================== */ magma_int_t j, jb, nb; char uplo_[2] = {uplo, 0}; float c_one = MAGMA_S_ONE; float c_neg_one = MAGMA_S_NEG_ONE; float *work; float d_one = 1.0; float d_neg_one = -1.0; int upper = lapackf77_lsame(uplo_, "U"); *info = 0; if ( (! upper) && (! lapackf77_lsame(uplo_, "L")) ) { *info = -1; } else if (n < 0) { *info = -2; } else if (ldda < max(1,n)) { *info = -4; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } nb = magma_get_spotrf_nb(n); if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, nb*nb )) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } /* Define user stream if current stream is NULL */ cudaStream_t stream[2], current_stream; magmablasGetKernelStream(¤t_stream); magma_queue_create( &stream[0] ); if (current_stream == NULL) { magma_queue_create( &stream[1] ); magmablasSetKernelStream(stream[1]); } else stream[1] = current_stream; if ((nb <= 1) || (nb >= n)) { /* Use unblocked code. */ magma_sgetmatrix_async( n, n, dA, ldda, work, n, stream[1] ); magma_queue_sync( stream[1] ); lapackf77_spotrf(uplo_, &n, work, &n, info); magma_ssetmatrix_async( n, n, work, n, dA, ldda, stream[1] ); } else { /* Use blocked code. */ if (upper) { /* Compute the Cholesky factorization A = U'*U. */ for (j=0; j<n; j+=nb) { /* Update and factorize the current diagonal block and test for non-positive-definiteness. Computing MIN */ jb = min(nb, (n-j)); magma_ssyrk(MagmaUpper, MagmaTrans, jb, j, d_neg_one, dA(0, j), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, stream[0] ); if ( (j+jb) < n) { /* Compute the current block row. */ magma_sgemm(MagmaTrans, MagmaNoTrans, jb, (n-j-jb), j, c_neg_one, dA(0, j ), ldda, dA(0, j+jb), ldda, c_one, dA(j, j+jb), ldda); } magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaUpperStr, &jb, work, &jb, info); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, stream[1] ); if (*info != 0) { *info = *info + j; break; } if ( (j+jb) < n) { magma_strsm( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit, jb, (n-j-jb), c_one, dA(j, j ), ldda, dA(j, j+jb), ldda); } } } else { //========================================================= // Compute the Cholesky factorization A = L*L'. for (j=0; j<n; j+=nb) { // Update and factorize the current diagonal block and test // for non-positive-definiteness. Computing MIN jb = min(nb, (n-j)); magma_ssyrk(MagmaLower, MagmaNoTrans, jb, j, d_neg_one, dA(j, 0), ldda, d_one, dA(j, j), ldda); magma_queue_sync( stream[1] ); magma_sgetmatrix_async( jb, jb, dA(j, j), ldda, work, jb, stream[0] ); if ( (j+jb) < n) { magma_sgemm( MagmaNoTrans, MagmaTrans, (n-j-jb), jb, j, c_neg_one, dA(j+jb, 0), ldda, dA(j, 0), ldda, c_one, dA(j+jb, j), ldda); } magma_queue_sync( stream[0] ); lapackf77_spotrf(MagmaLowerStr, &jb, work, &jb, info); magma_ssetmatrix_async( jb, jb, work, jb, dA(j, j), ldda, stream[1] ); if (*info != 0) { *info = *info + j; break; } if ( (j+jb) < n) { magma_strsm(MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit, (n-j-jb), jb, c_one, dA(j, j), ldda, dA(j+jb, j), ldda); } } } } magma_free_pinned( work ); magma_queue_destroy( stream[0] ); if (current_stream == NULL) { magma_queue_destroy( stream[1] ); magmablasSetKernelStream(NULL); } return *info; } /* magma_spotrf_gpu */