void get_QR_error(magma_int_t M, magma_int_t N, magma_int_t min_mn, magmaDoubleComplex *h_R, magmaDoubleComplex *h_A, magma_int_t lda, magmaDoubleComplex *tau, magmaDoubleComplex *Q, magma_int_t ldq, magmaDoubleComplex *R, magma_int_t ldr, magmaDoubleComplex *h_work, magma_int_t lwork, double *work, double *error, double *error2) { /* h_R:input the factorized matrix by lapack QR, h_A:input the original matrix copy tau: input */ const double d_neg_one = MAGMA_D_NEG_ONE; const double d_one = MAGMA_D_ONE; const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; const magmaDoubleComplex c_one = MAGMA_Z_ONE; const magmaDoubleComplex c_zero = MAGMA_Z_ZERO; double Anorm; magma_int_t info; // generate M by K matrix Q, where K = min(M,N) lapackf77_zlacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq ); lapackf77_zungqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info ); assert( info == 0 ); // copy K by N matrix R lapackf77_zlaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr ); lapackf77_zlacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr ); // error = || R - Q^H*A || / (N * ||A||) blasf77_zgemm( "Conj", "NoTrans", &min_mn, &N, &M, &c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr ); Anorm = lapackf77_zlange( "1", &M, &N, h_A, &lda, work ); *error = lapackf77_zlange( "1", &min_mn, &N, R, &ldr, work ); if ( N > 0 && Anorm > 0 ) *error /= (N*Anorm); // set R = I (K by K identity), then R = I - Q^H*Q // error = || I - Q^H*Q || / N lapackf77_zlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr ); blasf77_zherk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr ); *error2 = safe_lapackf77_zlanhe( "1", "Upper", &min_mn, R, &ldr, work ); if ( N > 0 ) *error2 /= N; }
// On input, LU and ipiv is LU factorization of A. On output, LU is overwritten. // Works for any m, n. // Uses init_matrix() to re-generate original A as needed. // Returns error in factorization, |PA - LU| / (n |A|) // This allocates 3 more matrices to store A, L, and U. double get_LU_error(magma_int_t M, magma_int_t N, magmaDoubleComplex *LU, magma_int_t lda, magma_int_t *ipiv) { magma_int_t min_mn = min(M,N); magma_int_t ione = 1; magma_int_t i, j; magmaDoubleComplex alpha = MAGMA_Z_ONE; magmaDoubleComplex beta = MAGMA_Z_ZERO; magmaDoubleComplex *A, *L, *U; double work[1], matnorm, residual; TESTING_MALLOC_CPU( A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( L, magmaDoubleComplex, M*min_mn ); TESTING_MALLOC_CPU( U, magmaDoubleComplex, min_mn*N ); memset( L, 0, M*min_mn*sizeof(magmaDoubleComplex) ); memset( U, 0, min_mn*N*sizeof(magmaDoubleComplex) ); // set to original A init_matrix( M, N, A, lda ); lapackf77_zlaswp( &N, A, &lda, &ione, &min_mn, ipiv, &ione); // copy LU to L and U, and set diagonal to 1 lapackf77_zlacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M ); lapackf77_zlacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn ); for(j=0; j<min_mn; j++) L[j+j*M] = MAGMA_Z_MAKE( 1., 0. ); matnorm = lapackf77_zlange("f", &M, &N, A, &lda, work); blasf77_zgemm("N", "N", &M, &N, &min_mn, &alpha, L, &M, U, &min_mn, &beta, LU, &lda); for( j = 0; j < N; j++ ) { for( i = 0; i < M; i++ ) { LU[i+j*lda] = MAGMA_Z_SUB( LU[i+j*lda], A[i+j*lda] ); } } residual = lapackf77_zlange("f", &M, &N, LU, &lda, work); TESTING_FREE_CPU( A ); TESTING_FREE_CPU( L ); TESTING_FREE_CPU( U ); return residual / (matnorm * N); }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgeqlf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_R, *tau, *h_work, tmp[1]; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, 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 ); double tol = 2. * opts.tolerance * lapackf77_dlamch("E"); printf(" M 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 ) { M = opts.msize[itest]; N = opts.nsize[itest]; min_mn = min(M, N); lda = M; n2 = lda*N; ldda = ((M+31)/32)*32; nb = magma_get_zgeqrf_nb(M); gflops = FLOPS_ZGEQLF( M, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_zgeqlf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_Z_REAL( tmp[0] ); lwork = max( lwork, N*nb ); lwork = max( lwork, 2*nb*nb); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zgeqlf( M, N, h_R, lda, tau, h_work, lwork, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zgeqlf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_zgeqlf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapack_zgeqlf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ error = lapackf77_zlange("f", &M, &N, h_A, &lda, work); blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_zlange("f", &M, &N, h_R, &lda, work) / error; printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed")); status += ! (error < tol); TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgehrd */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; magmaDoubleComplex *h_A, *h_R, *h_Q, *h_work, *tau, *twork; magmaDoubleComplex_ptr dT; #if defined(PRECISION_z) || defined(PRECISION_c) double *rwork; #endif double eps, result[2]; magma_int_t N, n2, lda, nb, lwork, ltwork, info; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; eps = lapackf77_dlamch( "E" ); magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) |A-QHQ'|/N|A| |I-QQ'|/N\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_zgehrd_nb(N); /* We suppose the magma nb is bigger than lapack nb */ lwork = N*nb; gflops = FLOPS_ZGEHRD( N ) / 1e9; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_DEV( dT, magmaDoubleComplex, nb*N ); /* Initialize the matrices */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zgehrd( N, ione, N, h_R, lda, tau, h_work, lwork, dT, 0, opts.queue, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zgehrd returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the factorization =================================================================== */ if ( opts.check ) { ltwork = 2*(N*N); TESTING_MALLOC_PIN( h_Q, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( twork, magmaDoubleComplex, ltwork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_MALLOC_CPU( rwork, double, N ); #endif lapackf77_zlacpy(MagmaUpperLowerStr, &N, &N, h_R, &lda, h_Q, &lda); for( int j = 0; j < N-1; ++j ) for( int i = j+2; i < N; ++i ) h_R[i+j*lda] = MAGMA_Z_ZERO; magma_zunghr( N, ione, N, h_Q, lda, tau, dT, 0, nb, opts.queue, &info ); if (info != 0) { printf("magma_zunghr returned error %d: %s.\n", (int) info, magma_strerror( info )); exit(1); } #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_zhst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, rwork, result); #else lapackf77_zhst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, result); #endif TESTING_FREE_PIN( h_Q ); TESTING_FREE_CPU( twork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_FREE_CPU( rwork ); #endif } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_zgehrd(&N, &ione, &N, h_R, &lda, tau, h_work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zgehrd returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Print performance and error. =================================================================== */ if ( opts.lapack ) { printf("%5d %7.2f (%7.2f) %7.2f (%7.2f)", (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time ); } else { printf("%5d --- ( --- ) %7.2f (%7.2f)", (int) N, gpu_perf, gpu_time ); } if ( opts.check ) { printf(" %8.2e %8.2e %s\n", result[0]*eps, result[1]*eps, ( ( (result[0]*eps < tol) && (result[1]*eps < tol) ) ? "ok" : "failed") ); status += ! (result[0]*eps < tol); status += ! (result[1]*eps < tol); } else { printf(" --- ---\n"); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( tau ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_work ); TESTING_FREE_DEV( dT ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhesv */ int main( int argc, char** argv) { TESTING_INIT(); magmaDoubleComplex *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_ZPOTRF( N ) + FLOPS_ZPOTRS( N, opts.nrhs ) ) / 1e9; TESTING_MALLOC_CPU( ipiv, magma_int_t, N ); TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_B, magmaDoubleComplex, sizeB ); TESTING_MALLOC_PIN( h_X, magmaDoubleComplex, sizeB ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { lwork = -1; lapackf77_zhesv(lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, ipiv, h_X, &ldb, &temp, &lwork, &info); lwork = (int)MAGMA_Z_REAL(temp); TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork ); init_matrix( N, N, h_A, lda ); lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B ); lapackf77_zlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb ); cpu_time = magma_wtime(); lapackf77_zhesv(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_zhesv 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_zlarnv( &ione, ISEED, &sizeB, h_B ); lapackf77_zlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb ); magma_setdevice(0); gpu_time = magma_wtime(); magma_zhesv( 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_zhesv 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; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgetri_batched */ int main( int argc, char** argv) { TESTING_INIT(); // constants const magmaDoubleComplex c_zero = MAGMA_Z_ZERO; const magmaDoubleComplex c_one = MAGMA_Z_ONE; const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; magmaDoubleComplex *h_A, *h_Ainv, *h_R, *work; magmaDoubleComplex_ptr d_A, d_invA; magmaDoubleComplex_ptr *dA_array; magmaDoubleComplex_ptr *dinvA_array; magma_int_t **dipiv_array; magma_int_t *dinfo_array; magma_int_t *ipiv, *cpu_info; magma_int_t *d_ipiv, *d_info; magma_int_t N, n2, lda, ldda, info, info1, info2, lwork; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magmaDoubleComplex tmp; double error, rwork[1]; magma_int_t columns; magma_int_t status = 0; magma_opts opts( MagmaOptsBatched ); opts.parse_opts( argc, argv ); magma_int_t batchCount = opts.batchcount; double tol = opts.tolerance * lapackf77_dlamch("E"); printf("%% batchCount N CPU Gflop/s (ms) GPU Gflop/s (ms) ||I - A*A^{-1}||_1 / (N*cond(A))\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 * batchCount; ldda = magma_roundup( N, opts.align ); // multiple of 32 by default // This is the correct flops but since this getri_batched is based on // 2 trsm = getrs and to know the real flops I am using the getrs one //gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRI( N ))/ 1e9 * batchCount; gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, N ))/ 1e9 * batchCount; // query for workspace size lwork = -1; lapackf77_zgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info ); if (info != 0) { printf("lapackf77_zgetri returned error %d: %s.\n", (int) info, magma_strerror( info )); } lwork = magma_int_t( MAGMA_Z_REAL( tmp )); TESTING_MALLOC_CPU( cpu_info, magma_int_t, batchCount ); TESTING_MALLOC_CPU( ipiv, magma_int_t, N * batchCount ); TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork*batchCount ); TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_Ainv, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, n2 ); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N * batchCount ); TESTING_MALLOC_DEV( d_invA, magmaDoubleComplex, ldda*N * batchCount ); TESTING_MALLOC_DEV( d_ipiv, magma_int_t, N * batchCount ); TESTING_MALLOC_DEV( d_info, magma_int_t, batchCount ); TESTING_MALLOC_DEV( dA_array, magmaDoubleComplex*, batchCount ); TESTING_MALLOC_DEV( dinvA_array, magmaDoubleComplex*, batchCount ); TESTING_MALLOC_DEV( dinfo_array, magma_int_t, batchCount ); TESTING_MALLOC_DEV( dipiv_array, magma_int_t*, batchCount ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); columns = N * batchCount; lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_Ainv, &lda ); magma_zsetmatrix( N, columns, h_R, lda, d_A, ldda, opts.queue ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_zset_pointer( dA_array, d_A, ldda, 0, 0, ldda * N, batchCount, opts.queue ); magma_zset_pointer( dinvA_array, d_invA, ldda, 0, 0, ldda * N, batchCount, opts.queue ); magma_iset_pointer( dipiv_array, d_ipiv, 1, 0, 0, N, batchCount, opts.queue ); gpu_time = magma_sync_wtime( opts.queue ); info1 = magma_zgetrf_batched( N, N, dA_array, ldda, dipiv_array, dinfo_array, batchCount, opts.queue); info2 = magma_zgetri_outofplace_batched( N, dA_array, ldda, dipiv_array, dinvA_array, ldda, dinfo_array, batchCount, opts.queue); gpu_time = magma_sync_wtime( opts.queue ) - gpu_time; gpu_perf = gflops / gpu_time; // check correctness of results throught "dinfo_magma" and correctness of argument throught "info" magma_getvector( batchCount, sizeof(magma_int_t), dinfo_array, 1, cpu_info, 1, opts.queue ); for (magma_int_t i=0; i < batchCount; i++) { if (cpu_info[i] != 0 ) { printf("magma_zgetrf_batched matrix %d returned error %d\n", (int) i, (int)cpu_info[i] ); } } if (info1 != 0) printf("magma_zgetrf_batched returned argument error %d: %s.\n", (int) info1, magma_strerror( info1 )); if (info2 != 0) printf("magma_zgetri_batched returned argument error %d: %s.\n", (int) info2, magma_strerror( info2 )); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); #if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP) magma_int_t nthreads = magma_get_lapack_numthreads(); magma_set_lapack_numthreads(1); magma_set_omp_numthreads(nthreads); #pragma omp parallel for schedule(dynamic) #endif for (int i=0; i < batchCount; i++) { magma_int_t locinfo; lapackf77_zgetrf(&N, &N, h_Ainv + i*lda*N, &lda, ipiv + i*N, &locinfo); if (locinfo != 0) { printf("lapackf77_zgetrf returned error %d: %s.\n", (int) locinfo, magma_strerror( locinfo )); } lapackf77_zgetri(&N, h_Ainv + i*lda*N, &lda, ipiv + i*N, work + i*lwork, &lwork, &locinfo ); if (locinfo != 0) { printf("lapackf77_zgetri returned error %d: %s.\n", (int) locinfo, magma_strerror( locinfo )); } } #if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP) magma_set_lapack_numthreads(nthreads); #endif cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; printf("%10d %5d %7.2f (%7.2f) %7.2f (%7.2f)", (int) batchCount, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000. ); } else { printf("%10d %5d --- ( --- ) %7.2f (%7.2f)", (int) batchCount, (int) N, gpu_perf, gpu_time*1000. ); } /* ===================================================================== Check the result =================================================================== */ if ( opts.check ) { magma_igetvector( N*batchCount, d_ipiv, 1, ipiv, 1, opts.queue ); magma_zgetmatrix( N, N*batchCount, d_invA, ldda, h_Ainv, lda, opts.queue ); error = 0; for (magma_int_t i=0; i < batchCount; i++) { for (magma_int_t k=0; k < N; k++) { if (ipiv[i*N+k] < 1 || ipiv[i*N+k] > N ) { printf("error for matrix %d ipiv @ %d = %d\n", (int) i, (int) k, (int) ipiv[i*N+k]); error = -1; } } if (error == -1) { break; } // compute 1-norm condition number estimate, following LAPACK's zget03 double normA, normAinv, rcond, err; normA = lapackf77_zlange( "1", &N, &N, h_A + i*lda*N, &lda, rwork ); normAinv = lapackf77_zlange( "1", &N, &N, h_Ainv + i*lda*N, &lda, rwork ); if ( normA <= 0 || normAinv <= 0 ) { rcond = 0; err = 1 / (tol/opts.tolerance); // == 1/eps } else { rcond = (1 / normA) / normAinv; // R = I // R -= A*A^{-1} // err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm lapackf77_zlaset( "full", &N, &N, &c_zero, &c_one, h_R + i*lda*N, &lda ); blasf77_zgemm( "no", "no", &N, &N, &N, &c_neg_one, h_A + i*lda*N, &lda, h_Ainv + i*lda*N, &lda, &c_one, h_R + i*lda*N, &lda ); err = lapackf77_zlange( "1", &N, &N, h_R + i*lda*N, &lda, rwork ); err = err * rcond / N; } if ( isnan(err) || isinf(err) ) { error = err; break; } error = max( err, error ); } bool okay = (error < tol); status += ! okay; printf(" %8.2e %s\n", error, (okay ? "ok" : "failed") ); } else { printf("\n"); } TESTING_FREE_CPU( cpu_info ); TESTING_FREE_CPU( ipiv ); TESTING_FREE_CPU( work ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_Ainv ); TESTING_FREE_CPU( h_R ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_invA ); TESTING_FREE_DEV( d_ipiv ); TESTING_FREE_DEV( d_info ); TESTING_FREE_DEV( dA_array ); TESTING_FREE_DEV( dinvA_array ); TESTING_FREE_DEV( dinfo_array ); TESTING_FREE_DEV( dipiv_array ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_zheevx_gpu(char jobz, char range, char uplo, magma_int_t n, magmaDoubleComplex *da, magma_int_t ldda, double vl, double vu, magma_int_t il, magma_int_t iu, double abstol, magma_int_t *m, double *w, magmaDoubleComplex *dz, magma_int_t lddz, magmaDoubleComplex *wa, magma_int_t ldwa, magmaDoubleComplex *wz, magma_int_t ldwz, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t *iwork, magma_int_t *ifail, magma_int_t *info) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= ZHEEVX computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. Arguments ========= JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. RANGE (input) CHARACTER*1 = 'A': all eigenvalues will be found. = 'V': all eigenvalues in the half-open interval (VL,VU] will be found. = 'I': the IL-th through IU-th eigenvalues will be found. UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. DA (device input/output) COMPLEX_16 array, dimension (LDDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, the lower triangle (if UPLO='L') or the upper triangle (if UPLO='U') of A, including the diagonal, is destroyed. LDDA (input) INTEGER The leading dimension of the array DA. LDDA >= max(1,N). VL (input) DOUBLE PRECISION VU (input) DOUBLE PRECISION If RANGE='V', the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = 'A' or 'I'. IL (input) INTEGER IU (input) INTEGER If RANGE='I', the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = 'A' or 'V'. ABSTOL (input) DOUBLE PRECISION The absolute error tolerance for the eigenvalues. An approximate eigenvalue is accepted as converged when it is determined to lie in an interval [a,b] of width less than or equal to ABSTOL + EPS * max( |a|,|b| ) , where EPS is the machine precision. If ABSTOL is less than or equal to zero, then EPS*|T| will be used in its place, where |T| is the 1-norm of the tridiagonal matrix obtained by reducing A to tridiagonal form. Eigenvalues will be computed most accurately when ABSTOL is set to twice the underflow threshold 2*DLAMCH('S'), not zero. If this routine returns with INFO>0, indicating that some eigenvectors did not converge, try setting ABSTOL to 2*DLAMCH('S'). See "Computing Small Singular Values of Bidiagonal Matrices with Guaranteed High Relative Accuracy," by Demmel and Kahan, LAPACK Working Note #3. M (output) INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1. W (output) DOUBLE PRECISION array, dimension (N) On normal exit, the first M elements contain the selected eigenvalues in ascending order. DZ (device output) COMPLEX_16 array, dimension (LDDZ, max(1,M)) If JOBZ = 'V', then if INFO = 0, the first M columns of Z contain the orthonormal eigenvectors of the matrix A corresponding to the selected eigenvalues, with the i-th column of Z holding the eigenvector associated with W(i). If an eigenvector fails to converge, then that column of Z contains the latest approximation to the eigenvector, and the index of the eigenvector is returned in IFAIL. If JOBZ = 'N', then Z is not referenced. Note: the user must ensure that at least max(1,M) columns are supplied in the array Z; if RANGE = 'V', the exact value of M is not known in advance and an upper bound must be used. ********* (workspace) If FAST_HEMV is defined DZ should be (LDDZ, max(1,N)) in both cases. LDDZ (input) INTEGER The leading dimension of the array DZ. LDDZ >= 1, and if JOBZ = 'V', LDDZ >= max(1,N). WA (workspace) COMPLEX_16 array, dimension (LDWA, N) LDWA (input) INTEGER The leading dimension of the array WA. LDWA >= max(1,N). WZ (workspace) COMPLEX_16 array, dimension (LDWZ, max(1,M)) LDWZ (input) INTEGER The leading dimension of the array DZ. LDWZ >= 1, and if JOBZ = 'V', LDWZ >= max(1,N). WORK (workspace/output) COMPLEX_16 array, dimension (LWORK) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. LWORK >= (NB+1)*N, where NB is the max of the blocksize for ZHETRD. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. RWORK (workspace) DOUBLE PRECISION array, dimension (7*N) IWORK (workspace) INTEGER array, dimension (5*N) IFAIL (output) INTEGER array, dimension (N) If JOBZ = 'V', then if INFO = 0, the first M elements of IFAIL are zero. If INFO > 0, then IFAIL contains the indices of the eigenvectors that failed to converge. If JOBZ = 'N', then IFAIL is not referenced. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i, then i eigenvectors failed to converge. Their indices are stored in array IFAIL. ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; char range_[2] = {range, 0}; magma_int_t ione = 1; char order[1]; magma_int_t indd, inde; magma_int_t imax; magma_int_t lopt, itmp1, indee; magma_int_t lower, wantz; magma_int_t i, j, jj, i__1; magma_int_t alleig, valeig, indeig; magma_int_t iscale, indibl; magma_int_t indiwk, indisp, indtau; magma_int_t indrwk, indwrk; magma_int_t llwork, nsplit; magma_int_t lquery; magma_int_t iinfo; double safmin; double bignum; double smlnum; double eps, tmp1; double anrm; double sigma, d__1; double rmin, rmax; double *dwork; /* Function Body */ lower = lapackf77_lsame(uplo_, MagmaLowerStr); wantz = lapackf77_lsame(jobz_, MagmaVecStr); alleig = lapackf77_lsame(range_, "A"); valeig = lapackf77_lsame(range_, "V"); indeig = lapackf77_lsame(range_, "I"); lquery = lwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -3; } else if (n < 0) { *info = -4; } else if (ldda < max(1,n)) { *info = -6; } else if (lddz < 1 || (wantz && lddz < n)) { *info = -15; } else if (ldwa < max(1,n)) { *info = -17; } else if (ldwz < 1 || (wantz && ldwz < n)) { *info = -19; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_zhetrd_nb(n); lopt = n * (nb + 1); work[0] = MAGMA_Z_MAKE( lopt, 0 ); if (lwork < lopt && ! lquery) { *info = -21; } if (*info != 0) { magma_xerbla( __func__, -(*info)); return *info; } else if (lquery) { return *info; } *m = 0; /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif magmaDoubleComplex *a = (magmaDoubleComplex *) malloc( n * n * sizeof(magmaDoubleComplex) ); magma_zgetmatrix(n, n, da, ldda, a, n); lapackf77_zheevx(jobz_, range_, uplo_, &n, a, &n, &vl, &vu, &il, &iu, &abstol, m, w, wz, &ldwz, work, &lwork, rwork, iwork, ifail, info); magma_zsetmatrix( n, n, a, n, da, ldda); magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz); free(a); return *info; } if (MAGMA_SUCCESS != magma_dmalloc( &dwork, n )) { fprintf (stderr, "!!!! device memory allocation error (magma_zheevx_gpu)\n"); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } --w; --work; --rwork; --iwork; --ifail; /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = magmablas_zlanhe('M', uplo, n, da, ldda, dwork); iscale = 0; sigma = 1; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { d__1 = 1.; magmablas_zlascl(uplo, 0, 0, 1., sigma, n, n, da, ldda, info); if (abstol > 0.) { abstol *= sigma; } if (valeig) { vl *= sigma; vu *= sigma; } } /* Call ZHETRD to reduce Hermitian matrix to tridiagonal form. */ indd = 1; inde = indd + n; indrwk = inde + n; indtau = 1; indwrk = indtau + n; llwork = lwork - indwrk + 1; #ifdef FAST_HEMV magma_zhetrd2_gpu(uplo, n, da, ldda, &rwork[indd], &rwork[inde], &work[indtau], wa, ldwa, &work[indwrk], llwork, dz, lddz*n, &iinfo); #else magma_zhetrd_gpu (uplo, n, da, ldda, &rwork[indd], &rwork[inde], &work[indtau], wa, ldwa, &work[indwrk], llwork, &iinfo); #endif lopt = n + (magma_int_t)MAGMA_Z_REAL(work[indwrk]); /* If all eigenvalues are desired and ABSTOL is less than or equal to zero, then call DSTERF or ZUNGTR and ZSTEQR. If this fails for some eigenvalue, then try DSTEBZ. */ if ((alleig || (indeig && il == 1 && iu == n)) && abstol <= 0.) { blasf77_dcopy(&n, &rwork[indd], &ione, &w[1], &ione); indee = indrwk + 2*n; if (! wantz) { i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_dsterf(&n, &w[1], &rwork[indee], info); } else { lapackf77_zlacpy("A", &n, &n, wa, &ldwa, wz, &ldwz); lapackf77_zungtr(uplo_, &n, wz, &ldwz, &work[indtau], &work[indwrk], &llwork, &iinfo); i__1 = n - 1; blasf77_dcopy(&i__1, &rwork[inde], &ione, &rwork[indee], &ione); lapackf77_zsteqr(jobz_, &n, &w[1], &rwork[indee], wz, &ldwz, &rwork[indrwk], info); if (*info == 0) { for (i = 1; i <= n; ++i) { ifail[i] = 0; } magma_zsetmatrix( n, n, wz, ldwz, dz, lddz ); } } if (*info == 0) { *m = n; } } /* Otherwise, call DSTEBZ and, if eigenvectors are desired, ZSTEIN. */ if (*m == 0) { *info = 0; if (wantz) { *(unsigned char *)order = 'B'; } else { *(unsigned char *)order = 'E'; } indibl = 1; indisp = indibl + n; indiwk = indisp + n; lapackf77_dstebz(range_, order, &n, &vl, &vu, &il, &iu, &abstol, &rwork[indd], &rwork[inde], m, &nsplit, &w[1], &iwork[indibl], &iwork[indisp], &rwork[indrwk], &iwork[indiwk], info); if (wantz) { lapackf77_zstein(&n, &rwork[indd], &rwork[inde], m, &w[1], &iwork[indibl], &iwork[indisp], wz, &ldwz, &rwork[indrwk], &iwork[indiwk], &ifail[1], info); magma_zsetmatrix( n, *m, wz, ldwz, dz, lddz ); /* Apply unitary matrix used in reduction to tridiagonal form to eigenvectors returned by ZSTEIN. */ magma_zunmtr_gpu(MagmaLeft, uplo, MagmaNoTrans, n, *m, da, ldda, &work[indtau], dz, lddz, wa, ldwa, &iinfo); } } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = *m; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_dscal(&imax, &d__1, &w[1], &ione); } /* If eigenvalues are not in order, then sort them, along with eigenvectors. */ if (wantz) { for (j = 1; j <= *m-1; ++j) { i = 0; tmp1 = w[j]; for (jj = j + 1; jj <= *m; ++jj) { if (w[jj] < tmp1) { i = jj; tmp1 = w[jj]; } } if (i != 0) { itmp1 = iwork[indibl + i - 1]; w[i] = w[j]; iwork[indibl + i - 1] = iwork[indibl + j - 1]; w[j] = tmp1; iwork[indibl + j - 1] = itmp1; magma_zswap(n, dz + (i-1)*lddz, ione, dz + (j-1)*lddz, ione); if (*info != 0) { itmp1 = ifail[i]; ifail[i] = ifail[j]; ifail[j] = itmp1; } } } } /* Set WORK(1) to optimal complex workspace size. */ work[1] = MAGMA_Z_MAKE( lopt, 0 ); return *info; } /* magma_zheevx_gpu */
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhegvdx */ int main( int argc, char** argv) { //#define USE_MGPU #ifdef USE_MGPU TESTING_CUDA_INIT_MGPU(); #else TESTING_CUDA_INIT(); #endif magma_int_t nrgpu =1; cuDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work; double *rwork, *w1, *w2; magma_int_t *iwork; double gpu_time, cpu_time; magma_timestr_t start, end; /* Matrix size */ magma_int_t N=0, n2; magma_int_t size[4] = {1024,2048,4100,6001}; magma_int_t i, itype, info; magma_int_t ione = 1, izero = 0; magma_int_t five = 5; cuDoubleComplex c_zero = MAGMA_Z_ZERO; cuDoubleComplex c_one = MAGMA_Z_ONE; cuDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; double d_one = 1.; double d_neg_one = -1.; double d_ten = 10.; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t il,iu,m1,m2; double vl,vu; double fraction_ev = 0; //const char *uplo = MagmaLowerStr; char *uplo = (char*)MagmaLowerStr; //char *uplo = (char*)MagmaUpperStr; char *jobz = (char*)MagmaVectorsStr; char range = 'A'; itype = 1; magma_int_t checkres; double result[2]; int flagN = 0; if (argc != 1){ for(i = 1; i<argc; i++){ if (strcmp("-N", argv[i])==0){ N = atoi(argv[++i]); if (N>0){ printf(" testing_zhegvdx -N %d\n\n", (int) N); flagN=1; } else { printf("\nUsage: \n"); printf(" testing_zhegvdx -N %d\n\n", (int) N); exit(1); } } if (strcmp("-ngpu", argv[i])==0){ nrgpu = atoi(argv[++i]); if (nrgpu>0){ printf(" testing_zhegvdx -ngpu %d\n\n", (int) nrgpu); } else { printf("\nUsage: \n"); printf(" testing_zhegvdx -ngpu %d\n\n", (int) nrgpu); exit(1); } } if (strcmp("-itype", argv[i])==0){ itype = atoi(argv[++i]); if (itype>0 && itype <= 3){ printf(" testing_zhegvdx -itype %d\n\n", (int) itype); } else { printf("\nUsage: \n"); printf(" testing_zhegvdx -itype %d\n\n", (int) itype); exit(1); } } if (strcmp("-FE", argv[i])==0){ fraction_ev = atof(argv[++i]); if (fraction_ev > 0 && fraction_ev <= 1){ printf(" testing_zhegvdx -FE %f\n\n", fraction_ev); } else { fraction_ev = 0; } } if (strcmp("-L", argv[i])==0){ uplo = (char*)MagmaLowerStr; printf(" testing_zhegvdx -L"); } if (strcmp("-U", argv[i])==0){ uplo = (char*)MagmaUpperStr; printf(" testing_zhegvdx -U"); } } } else { printf("\nUsage: \n"); printf(" testing_zhegvdx -L/U -N %d -itype %d\n\n", 1024, 1); } if(!flagN) N = size[3]; checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL; n2 = N * N; /* Allocate host memory for the matrix */ TESTING_MALLOC( h_A, cuDoubleComplex, n2); TESTING_MALLOC( h_B, cuDoubleComplex, n2); TESTING_MALLOC( w1, double , N); TESTING_MALLOC( w2, double , N); TESTING_HOSTALLOC(h_R, cuDoubleComplex, n2); TESTING_HOSTALLOC(h_S, cuDoubleComplex, n2); magma_int_t nb = magma_get_zhetrd_nb(N); magma_int_t lwork = magma_zbulge_get_lq2(N) + 2*N + N*N; magma_int_t lrwork = 1 + 5*N +2*N*N; magma_int_t liwork = 3 + 5*N; TESTING_HOSTALLOC(h_work, cuDoubleComplex, lwork); TESTING_HOSTALLOC( rwork, double, lrwork); TESTING_MALLOC( iwork, magma_int_t, liwork); printf(" N M GPU Time(s) \n"); printf("==========================\n"); for(i=0; i<4; i++){ if (!flagN){ N = size[i]; n2 = N*N; } if (fraction_ev == 0){ il = N / 10; iu = N / 5+il; } else { il = 1; iu = (int)(fraction_ev*N); if (iu < 1) iu = 1; } /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); //lapackf77_zlatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten, // &d_one, &N, &N, uplo, h_B, &N, h_work, &info); //lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_B, &N); lapackf77_zlarnv( &ione, ISEED, &n2, h_B ); /* increase the diagonal */ { magma_int_t i, j; for(i=0; i<N; i++) { MAGMA_Z_SET2REAL( h_B[i*N+i], ( MAGMA_Z_REAL(h_B[i*N+i]) + 1.*N ) ); } } lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); #ifdef USE_MGPU magma_zhegvdx_2stage_m(nrgpu, itype, jobz[0], range, uplo[0], N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); #else magma_zhegvdx_2stage(itype, jobz[0], range, uplo[0], N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); #endif lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ start = get_current_time(); #ifdef USE_MGPU magma_zhegvdx_2stage_m(nrgpu, itype, jobz[0], range, uplo[0], N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); #else magma_zhegvdx_2stage(itype, jobz[0], range, uplo[0], N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info); #endif end = get_current_time(); gpu_time = GetTimerValue(start,end)/1000.; if ( checkres ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvdx routine. A x = lambda B x is solved and the following 3 tests computed: (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) | A B Z - Z D | / ( |A||Z| N ) (itype = 2) | B A Z - Z D | / ( |A||Z| N ) (itype = 3) (2) | S(with V) - S(w/o V) | / | S | =================================================================== */ double temp1, temp2; cuDoubleComplex *tau; result[0] = 1.; result[0] /= lapackf77_zlanhe("1",uplo, &N, h_A, &N, rwork); result[0] /= lapackf77_zlange("1",&N , &m1, h_R, &N, rwork); if (itype == 1){ blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", uplo, &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N; } else if (itype == 2){ blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N; } else if (itype == 3){ blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N; } lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); magma_zhegvdx(itype, 'N', range, uplo[0], N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2, h_work, lwork, rwork, lrwork, iwork, liwork, &info); temp1 = temp2 = 0; for(int j=0; j<m2; j++){ temp1 = max(temp1, absv(w1[j])); temp1 = max(temp1, absv(w2[j])); temp2 = max(temp2, absv(w1[j]-w2[j])); } result[1] = temp2 / temp1; } /* ===================================================================== Print execution time =================================================================== */ printf("%5d %5d %6.2f\n", (int) N, (int) m1, gpu_time); if ( checkres ){ printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if(itype==1) printf("(1) | A Z - B Z D | / (|A| |Z| N) = %e\n", result[0]); else if(itype==2) printf("(1) | A B Z - Z D | / (|A| |Z| N) = %e\n", result[0]); else if(itype==3) printf("(1) | B A Z - Z D | / (|A| |Z| N) = %e\n", result[0]); printf("(2) | D(w/ Z)-D(w/o Z)|/ |D| = %e\n\n", result[1]); } if (flagN) break; } cudaSetDevice(0); /* Memory clean up */ TESTING_FREE( h_A); TESTING_FREE( h_B); TESTING_FREE( w1); TESTING_FREE( w2); TESTING_HOSTFREE( rwork); TESTING_FREE( iwork); TESTING_HOSTFREE(h_work); TESTING_HOSTFREE( h_R); TESTING_HOSTFREE( h_S); /* Shutdown */ #ifdef USE_MGPU TESTING_CUDA_FINALIZE_MGPU(); #else TESTING_CUDA_FINALIZE(); #endif }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgehrd2 */ int main( int argc, char** argv) { real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time; //*h_R1 is used for warm-up magmaDoubleComplex *h_A, *h_R, *h_Q, *h_work, *tau, *twork, *h_R1; magmaDoubleComplex_ptr dT; double *rwork; double result[2] = {0., 0.}; double eps; int checkres; checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL; /* Matrix size */ int N=0, n2, lda, nb, lwork, ltwork, once = 0; #if defined (PRECISION_z) magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7000,7000,7000,7000}; #else magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7040,8064,9088,9900}; #endif int i, info; int ione = 1; int ISEED[4] = {0,0,0,1}; if (argc != 1){ for(i = 1; i<argc; i++){ if (strcmp("-N", argv[i])==0) N = atoi(argv[++i]); } if ( N > 0 ){ printf(" testing_zgehrd -N %d\n\n", N); once = 1; } else { printf("\nUsage: \n"); printf(" testing_zgehrd -N %d\n\n", 1024); exit(1); } } else { printf("\nUsage: \n"); printf(" testing_zgehrd -N %d\n\n", 1024); N = size[9]; } /* Initialize */ magma_queue_t queue; magma_device_t device; int num = 0; magma_err_t err; magma_init(); err = magma_get_devices( &device, 1, &num ); if ( err != 0 || num < 1 ) { fprintf( stderr, "magma_get_devices failed: %d\n", err ); exit(-1); } err = magma_queue_create( device, &queue ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", err ); exit(-1); } eps = lapackf77_dlamch( "E" ); lda = N; n2 = N*lda; nb = magma_get_zgehrd_nb(N); /* We suppose the magma nb is bigger than lapack nb */ lwork = N*nb; TESTING_MALLOC_HOST( h_A , magmaDoubleComplex, n2 ); TESTING_MALLOC_HOST( tau , magmaDoubleComplex, N ); TESTING_MALLOC_HOST( h_R , magmaDoubleComplex, n2 ); TESTING_MALLOC_HOST( h_R1 , magmaDoubleComplex, n2 ); TESTING_MALLOC_HOST( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_DEV ( dT , magmaDoubleComplex, nb*N ); /* To avoid uninitialized variable warning */ h_Q = NULL; twork = NULL; rwork = NULL; if ( checkres ) { ltwork = 2*(N*N); TESTING_MALLOC_HOST( h_Q, magmaDoubleComplex, lda*N ); TESTING_MALLOC_HOST( twork, magmaDoubleComplex, ltwork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_MALLOC_HOST( rwork, double, N ); #endif } printf("\n\n"); printf(" N CPU GFlop/s GPU GFlop/s |A-QHQ'|/N|A| |I-QQ'|/N \n"); printf("=============================================================\n"); for(i=0; i<10; i++){ if ( !once ) { N = size[i]; } lda = N; n2 = lda*N; gflops = FLOPS( (double)N ) / 1e9; /* Initialize the matrices */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R1, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_zgehrd ( N, ione, N, h_R1, lda, tau, h_work, lwork, dT, 0, &info, queue); if ( info < 0 ) printf("Argument %d of magma_zgehrd had an illegal value\n", -info); clFinish(queue); gpu_time = get_time(); magma_zgehrd ( N, ione, N, h_R, lda, tau, h_work, lwork, dT, 0, &info, queue); gpu_time = get_time() - gpu_time; if ( info < 0 ) printf("Argument %d of magma_zgehrd had an illegal value\n", -info); gpu_perf = gflops / gpu_time; /* ===================================================================== Check the factorization =================================================================== */ if ( checkres ) { lapackf77_zlacpy(MagmaUpperLowerStr, &N, &N, h_R, &lda, h_Q, &lda); { int i, j; for(j=0; j<N-1; j++) for(i=j+2; i<lda; i++) h_R[i+j*lda] = MAGMA_Z_ZERO; } nb = magma_get_zgehrd_nb(N); magma_zunghr(N, ione, N, h_Q, lda, tau, dT, 0, nb, &info, queue); #if defined(PRECISION_z) || defined(PRECISION_c) lapackf77_zhst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, rwork, result); #else lapackf77_zhst01(&N, &ione, &N, h_A, &lda, h_R, &lda, h_Q, &lda, twork, <work, result); #endif } /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = get_time(); lapackf77_zgehrd(&N, &ione, &N, h_A, &lda, tau, h_work, &lwork, &info); cpu_time = get_time() - cpu_time; if (info < 0) printf("Argument %d of lapack_zgehrd had an illegal value.\n", -info); cpu_perf = gflops / cpu_time; /* ===================================================================== Print performance and error. =================================================================== */ if ( checkres ) { printf("%5d %6.2f %6.2f %e %e\n", N, cpu_perf, gpu_perf, result[0]*eps, result[1]*eps ); } else { printf("%5d %6.2f %6.2f\n", N, cpu_perf, gpu_perf ); } if ( once ) break; } /* Memory clean up */ TESTING_FREE ( h_A ); TESTING_FREE ( tau ); TESTING_FREE_HOST( h_work); TESTING_FREE_HOST( h_R ); TESTING_FREE_HOST( h_R1 ); TESTING_FREE_DEV ( dT ); if ( checkres ) { TESTING_FREE_HOST( h_Q ); TESTING_FREE( twork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_FREE( rwork ); #endif } /* Shutdown */ magma_queue_destroy( queue ); magma_finalize(); return EXIT_SUCCESS; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zungqr_gpu */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double Anorm, error, work[1]; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *hA, *hR, *tau, *h_work; magmaDoubleComplex_ptr dA, dT; magma_int_t m, n, k; magma_int_t n2, lda, ldda, lwork, min_mn, nb, info; 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 ); double tol = opts.tolerance * lapackf77_dlamch("E"); opts.lapack |= opts.check; // check (-c) implies lapack (-l) printf(" m n k CPU GFlop/s (sec) GPU GFlop/s (sec) ||R|| / ||A||\n"); printf("=========================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { m = opts.msize[itest]; n = opts.nsize[itest]; k = opts.ksize[itest]; if ( m < n || n < k ) { printf( "%5d %5d %5d skipping because m < n or n < k\n", (int) m, (int) n, (int) k ); continue; } lda = m; ldda = ((m + 31)/32)*32; n2 = lda*n; min_mn = min(m, n); nb = magma_get_zgeqrf_nb( m ); lwork = (m + 2*n+nb)*nb; gflops = FLOPS_ZUNGQR( m, n, k ) / 1e9; TESTING_MALLOC_PIN( hA, magmaDoubleComplex, lda*n ); TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_CPU( hR, magmaDoubleComplex, lda*n ); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*n ); TESTING_MALLOC_DEV( dT, magmaDoubleComplex, ( 2*min_mn + ((n + 31)/32)*32 )*nb ); lapackf77_zlarnv( &ione, ISEED, &n2, hA ); lapackf77_zlacpy( MagmaFullStr, &m, &n, hA, &lda, hR, &lda ); Anorm = lapackf77_zlange("f", &m, &n, hA, &lda, work ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ // first, get QR factors in both hA and dA // okay that magma_zgeqrf_gpu has special structure for R; R isn't used here. magma_zsetmatrix( m, n, hA, lda, dA, ldda ); magma_zgeqrf_gpu( m, n, dA, ldda, tau, dT, &info ); if (info != 0) printf("magma_zgeqrf_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_zgetmatrix( m, n, dA, ldda, hA, lda ); gpu_time = magma_wtime(); magma_zungqr_gpu( m, n, k, dA, ldda, tau, dT, nb, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zungqr_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); // Get dA back to the CPU to compare with the CPU result. magma_zgetmatrix( m, n, dA, ldda, hR, lda ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_zungqr( &m, &n, &k, hA, &lda, tau, h_work, &lwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zungqr returned error %d: %s.\n", (int) info, magma_strerror( info )); // compute relative error |R|/|A| := |Q_magma - Q_lapack|/|A| blasf77_zaxpy( &n2, &c_neg_one, hA, &ione, hR, &ione ); error = lapackf77_zlange("f", &m, &n, hR, &lda, work) / Anorm; bool okay = (error < tol); status += ! okay; printf("%5d %5d %5d %7.1f (%7.2f) %7.1f (%7.2f) %8.2e %s\n", (int) m, (int) n, (int) k, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (okay ? "ok" : "failed")); } else { printf("%5d %5d %5d --- ( --- ) %7.1f (%7.2f) --- \n", (int) m, (int) n, (int) k, gpu_perf, gpu_time ); } TESTING_FREE_PIN( hA ); TESTING_FREE_PIN( h_work ); TESTING_FREE_CPU( hR ); TESTING_FREE_CPU( tau ); TESTING_FREE_DEV( dA ); TESTING_FREE_DEV( dT ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_zheevd(char jobz, char uplo, magma_int_t n, magmaDoubleComplex *a, magma_int_t lda, double *w, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t lrwork, magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= ZHEEVD computes all eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments ========= JOBZ (input) CHARACTER*1 = 'N': Compute eigenvalues only; = 'V': Compute eigenvalues and eigenvectors. UPLO (input) CHARACTER*1 = 'U': Upper triangle of A is stored; = 'L': Lower triangle of A is stored. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX_16 array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = 'U', the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = 'L', the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = 'V', then if INFO = 0, A contains the orthonormal eigenvectors of the matrix A. If JOBZ = 'N', then on exit the lower triangle (if UPLO='L') or the upper triangle (if UPLO='U') of A, including the diagonal, is destroyed. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). W (output) DOUBLE PRECISION array, dimension (N) If INFO = 0, the eigenvalues in ascending order. WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. LWORK (input) INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = 'N' and N > 1, LWORK >= N + N*NB. If JOBZ = 'V' and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_zhetrd_nb(N). If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. RWORK (workspace/output) DOUBLE PRECISION array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. LRWORK (input) INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = 'N' and N > 1, LRWORK >= N. If JOBZ = 'V' and N > 1, LRWORK >= 1 + 5*N + 2*N**2. If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. LIWORK (input) INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = 'N' and N > 1, LIWORK >= 1. If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N. If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value > 0: if INFO = i and JOBZ = 'N', then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = 'V', then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details =============== Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. ===================================================================== */ char uplo_[2] = {uplo, 0}; char jobz_[2] = {jobz, 0}; magma_int_t ione = 1; magma_int_t izero = 0; double d_one = 1.; double d__1; double eps; magma_int_t inde; double anrm; magma_int_t imax; double rmin, rmax; double sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; double safmin; double bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; double smlnum; magma_int_t lquery; double* dwork; wantz = lapackf77_lsame(jobz_, MagmaVecStr); lower = lapackf77_lsame(uplo_, MagmaLowerStr); lquery = lwork == -1 || lrwork == -1 || liwork == -1; *info = 0; if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) { *info = -1; } else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } magma_int_t nb = magma_get_zhetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps to ensure length gets rounded up, // if it cannot be exactly represented in floating point. work[0] = MAGMA_Z_MAKE( lwmin * (1. + lapackf77_dlamch("Epsilon")), 0.); rwork[0] = lrwmin * (1. + lapackf77_dlamch("Epsilon")); iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -8; } else if ((lrwork < lrwmin) && ! lquery) { *info = -10; } else if ((liwork < liwmin) && ! lquery) { *info = -12; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_Z_REAL(a[0]); if (wantz) { a[0] = MAGMA_Z_ONE; } return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128){ #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_zheevd(jobz_, uplo_, &n, a, &lda, w, work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); return *info; } /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_zlanhe("M", uplo_, &n, a, &lda, rwork); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_zlascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a, &lda, info); } /* Call ZHETRD to reduce Hermitian matrix to tridiagonal form. */ // zhetrd rwork: e (n) // zstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // zhetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // zstedx work: tau (n) + z (n^2) // zunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; // #ifdef ENABLE_TIMER magma_timestr_t start, end; start = get_current_time(); #endif magma_zhetrd(uplo_[0], n, a, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo); #ifdef ENABLE_TIMER end = get_current_time(); printf("time zhetrd = %6.2f\n", GetTimerValue(start,end)/1000.); #endif /* For eigenvalues only, call DSTERF. For eigenvectors, first call ZSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call ZUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_dsterf(&n, w, &rwork[inde], info); } else { #ifdef ENABLE_TIMER start = get_current_time(); #endif if (MAGMA_SUCCESS != magma_dmalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_zstedx('A', n, 0., 0., 0, 0, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info); magma_free( dwork ); #ifdef ENABLE_TIMER end = get_current_time(); printf("time zstedx = %6.2f\n", GetTimerValue(start,end)/1000.); start = get_current_time(); #endif magma_zunmtr(MagmaLeft, uplo, MagmaNoTrans, n, n, a, lda, &work[indtau], &work[indwrk], n, &work[indwk2], llwrk2, &iinfo); lapackf77_zlacpy("A", &n, &n, &work[indwrk], &n, a, &lda); #ifdef ENABLE_TIMER end = get_current_time(); printf("time zunmtr + copy = %6.2f\n", GetTimerValue(start,end)/1000.); #endif } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_dscal(&imax, &d__1, w, &ione); } work[0] = MAGMA_Z_MAKE( lwmin * (1. + lapackf77_dlamch("Epsilon")), 0.); // round up rwork[0] = lrwmin * (1. + lapackf77_dlamch("Epsilon")); iwork[0] = liwmin; return *info; } /* magma_zheevd */
/** Purpose ------- ZHEEVDX computes selected eigenvalues and, optionally, eigenvectors of a complex Hermitian matrix A. Eigenvalues and eigenvectors can be selected by specifying either a range of values or a range of indices for the desired eigenvalues. If eigenvectors are desired, it uses a divide and conquer algorithm. The divide and conquer algorithm makes very mild assumptions about floating point arithmetic. It will work on machines with a guard digit in add/subtract, or on those binary machines without guard digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2. It could conceivably fail on hexadecimal or decimal machines without guard digits, but we know of none. Arguments --------- @param[in] jobz magma_vec_t - = MagmaNoVec: Compute eigenvalues only; - = MagmaVec: Compute eigenvalues and eigenvectors. @param[in] range magma_range_t - = MagmaRangeAll: all eigenvalues will be found. - = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU] will be found. - = MagmaRangeI: the IL-th through IU-th eigenvalues will be found. @param[in] uplo magma_uplo_t - = MagmaUpper: Upper triangle of A is stored; - = MagmaLower: Lower triangle of A is stored. @param[in] n INTEGER The order of the matrix A. N >= 0. @param[in,out] A COMPLEX_16 array, dimension (LDA, N) On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading N-by-N upper triangular part of A contains the upper triangular part of the matrix A. If UPLO = MagmaLower, the leading N-by-N lower triangular part of A contains the lower triangular part of the matrix A. On exit, if JOBZ = MagmaVec, then if INFO = 0, the first m columns of A contains the required orthonormal eigenvectors of the matrix A. If JOBZ = MagmaNoVec, then on exit the lower triangle (if UPLO=MagmaLower) or the upper triangle (if UPLO=MagmaUpper) of A, including the diagonal, is destroyed. @param[in] lda INTEGER The leading dimension of the array A. LDA >= max(1,N). @param[in] vl DOUBLE PRECISION @param[in] vu DOUBLE PRECISION If RANGE=MagmaRangeV, the lower and upper bounds of the interval to be searched for eigenvalues. VL < VU. Not referenced if RANGE = MagmaRangeAll or MagmaRangeI. @param[in] il INTEGER @param[in] iu INTEGER If RANGE=MagmaRangeI, the indices (in ascending order) of the smallest and largest eigenvalues to be returned. 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. Not referenced if RANGE = MagmaRangeAll or MagmaRangeV. @param[out] m INTEGER The total number of eigenvalues found. 0 <= M <= N. If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1. @param[out] w DOUBLE PRECISION array, dimension (N) If INFO = 0, the required m eigenvalues in ascending order. @param[out] work (workspace) COMPLEX_16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK[0] returns the optimal LWORK. @param[in] lwork INTEGER The length of the array WORK. If N <= 1, LWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_zhetrd_nb(N). \n If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] rwork (workspace) DOUBLE PRECISION array, dimension (LRWORK) On exit, if INFO = 0, RWORK[0] returns the optimal LRWORK. @param[in] lrwork INTEGER The dimension of the array RWORK. If N <= 1, LRWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LRWORK >= N. If JOBZ = MagmaVec and N > 1, LRWORK >= 1 + 5*N + 2*N**2. \n If LRWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK)) On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK. @param[in] liwork INTEGER The dimension of the array IWORK. If N <= 1, LIWORK >= 1. If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1. If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N. \n If LIWORK = -1, then a workspace query is assumed; the routine only calculates the optimal sizes of the WORK, RWORK and IWORK arrays, returns these values as the first entries of the WORK, RWORK and IWORK arrays, and no error message related to LWORK or LRWORK or LIWORK is issued by XERBLA. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument had an illegal value - > 0: if INFO = i and JOBZ = MagmaNoVec, then the algorithm failed to converge; i off-diagonal elements of an intermediate tridiagonal form did not converge to zero; if INFO = i and JOBZ = MagmaVec, then the algorithm failed to compute an eigenvalue while working on the submatrix lying in rows and columns INFO/(N+1) through mod(INFO,N+1). Further Details --------------- Based on contributions by Jeff Rutter, Computer Science Division, University of California at Berkeley, USA Modified description of INFO. Sven, 16 Feb 05. @ingroup magma_zheev_driver ********************************************************************/ extern "C" magma_int_t magma_zheevdx( magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n, magmaDoubleComplex *A, magma_int_t lda, double vl, double vu, magma_int_t il, magma_int_t iu, magma_int_t *m, double *w, magmaDoubleComplex *work, magma_int_t lwork, #ifdef COMPLEX double *rwork, magma_int_t lrwork, #endif magma_int_t *iwork, magma_int_t liwork, magma_int_t *info) { const char* uplo_ = lapack_uplo_const( uplo ); const char* jobz_ = lapack_vec_const( jobz ); magma_int_t ione = 1; magma_int_t izero = 0; double d_one = 1.; double d__1; double eps; magma_int_t inde; double anrm; magma_int_t imax; double rmin, rmax; double sigma; magma_int_t iinfo, lwmin; magma_int_t lower; magma_int_t llrwk; magma_int_t wantz; magma_int_t indwk2, llwrk2; magma_int_t iscale; double safmin; double bignum; magma_int_t indtau; magma_int_t indrwk, indwrk, liwmin; magma_int_t lrwmin, llwork; double smlnum; magma_int_t lquery; magma_int_t alleig, valeig, indeig; double* dwork; wantz = (jobz == MagmaVec); lower = (uplo == MagmaLower); alleig = (range == MagmaRangeAll); valeig = (range == MagmaRangeV); indeig = (range == MagmaRangeI); lquery = (lwork == -1 || lrwork == -1 || liwork == -1); *info = 0; if (! (wantz || (jobz == MagmaNoVec))) { *info = -1; } else if (! (alleig || valeig || indeig)) { *info = -2; } else if (! (lower || (uplo == MagmaUpper))) { *info = -3; } else if (n < 0) { *info = -4; } else if (lda < max(1,n)) { *info = -6; } else { if (valeig) { if (n > 0 && vu <= vl) { *info = -8; } } else if (indeig) { if (il < 1 || il > max(1,n)) { *info = -9; } else if (iu < min(n,il) || iu > n) { *info = -10; } } } magma_int_t nb = magma_get_zhetrd_nb( n ); if ( n <= 1 ) { lwmin = 1; lrwmin = 1; liwmin = 1; } else if ( wantz ) { lwmin = max( n + n*nb, 2*n + n*n ); lrwmin = 1 + 5*n + 2*n*n; liwmin = 3 + 5*n; } else { lwmin = n + n*nb; lrwmin = n; liwmin = 1; } // multiply by 1+eps (in Double!) to ensure length gets rounded up, // if it cannot be exactly represented in floating point. real_Double_t one_eps = 1. + lapackf77_dlamch("Epsilon"); work[0] = MAGMA_Z_MAKE( lwmin * one_eps, 0.); rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; if ((lwork < lwmin) && !lquery) { *info = -14; } else if ((lrwork < lrwmin) && ! lquery) { *info = -16; } else if ((liwork < liwmin) && ! lquery) { *info = -18; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } if (n == 1) { w[0] = MAGMA_Z_REAL(A[0]); if (wantz) { A[0] = MAGMA_Z_ONE; } return *info; } /* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */ if (n <= 128) { #ifdef ENABLE_DEBUG printf("--------------------------------------------------------------\n"); printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb); printf("--------------------------------------------------------------\n"); #endif lapackf77_zheevd(jobz_, uplo_, &n, A, &lda, w, work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, info); return *info; } /* Get machine constants. */ safmin = lapackf77_dlamch("Safe minimum"); eps = lapackf77_dlamch("Precision"); smlnum = safmin / eps; bignum = 1. / smlnum; rmin = magma_dsqrt(smlnum); rmax = magma_dsqrt(bignum); /* Scale matrix to allowable range, if necessary. */ anrm = lapackf77_zlanhe("M", uplo_, &n, A, &lda, rwork); iscale = 0; if (anrm > 0. && anrm < rmin) { iscale = 1; sigma = rmin / anrm; } else if (anrm > rmax) { iscale = 1; sigma = rmax / anrm; } if (iscale == 1) { lapackf77_zlascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, A, &lda, info); } /* Call ZHETRD to reduce Hermitian matrix to tridiagonal form. */ // zhetrd rwork: e (n) // zstedx rwork: e (n) + llrwk (1 + 4*N + 2*N**2) ==> 1 + 5n + 2n^2 inde = 0; indrwk = inde + n; llrwk = lrwork - indrwk; // zhetrd work: tau (n) + llwork (n*nb) ==> n + n*nb // zstedx work: tau (n) + z (n^2) // zunmtr work: tau (n) + z (n^2) + llwrk2 (n or n*nb) ==> 2n + n^2, or n + n*nb + n^2 indtau = 0; indwrk = indtau + n; indwk2 = indwrk + n*n; llwork = lwork - indwrk; llwrk2 = lwork - indwk2; magma_timer_t time=0; timer_start( time ); magma_zhetrd(uplo, n, A, lda, w, &rwork[inde], &work[indtau], &work[indwrk], llwork, &iinfo); timer_stop( time ); timer_printf( "time zhetrd = %6.2f\n", time ); /* For eigenvalues only, call DSTERF. For eigenvectors, first call ZSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the tridiagonal matrix, then call ZUNMTR to multiply it to the Householder transformations represented as Householder vectors in A. */ if (! wantz) { lapackf77_dsterf(&n, w, &rwork[inde], info); magma_dmove_eig(range, n, w, &il, &iu, vl, vu, m); } else { timer_start( time ); if (MAGMA_SUCCESS != magma_dmalloc( &dwork, 3*n*(n/2 + 1) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } magma_zstedx(range, n, vl, vu, il, iu, w, &rwork[inde], &work[indwrk], n, &rwork[indrwk], llrwk, iwork, liwork, dwork, info); magma_free( dwork ); timer_stop( time ); timer_printf( "time zstedx = %6.2f\n", time ); timer_start( time ); magma_dmove_eig(range, n, w, &il, &iu, vl, vu, m); magma_zunmtr(MagmaLeft, uplo, MagmaNoTrans, n, *m, A, lda, &work[indtau], &work[indwrk + n * (il-1) ], n, &work[indwk2], llwrk2, &iinfo); lapackf77_zlacpy("A", &n, m, &work[indwrk + n * (il-1)], &n, A, &lda); timer_stop( time ); timer_printf( "time zunmtr + copy = %6.2f\n", time ); } /* If matrix was scaled, then rescale eigenvalues appropriately. */ if (iscale == 1) { if (*info == 0) { imax = n; } else { imax = *info - 1; } d__1 = 1. / sigma; blasf77_dscal(&imax, &d__1, w, &ione); } work[0] = MAGMA_Z_MAKE( lwmin * one_eps, 0.); // round up rwork[0] = lrwmin * one_eps; iwork[0] = liwmin; return *info; } /* magma_zheevdx */
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhetrd_he2hb */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_time, gpu_perf; magmaDoubleComplex *h_A, *h_R, *h_work; magmaDoubleComplex *tau; double *D, *E; magma_int_t N, n2, lda, ldda, lwork, ldt, info, nstream; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; // TODO add these options to parse_opts magma_int_t NE = 0; magma_int_t distblk = 0; magma_opts opts; opts.parse_opts( argc, argv ); magma_int_t WANTZ = (opts.jobz == MagmaVec); double tol = opts.tolerance * lapackf77_dlamch("E"); if (opts.nb == 0) opts.nb = 64; //magma_get_zhetrd_he2hb_nb(N); if (NE < 1) NE = N; //64; //magma_get_zhetrd_he2hb_nb(N); nstream = max(3, opts.ngpu+2); magma_queue_t streams[MagmaMaxGPUs][20]; magmaDoubleComplex_ptr da[MagmaMaxGPUs], dT1[MagmaMaxGPUs]; if ((distblk == 0) || (distblk < opts.nb)) distblk = max(256, opts.nb); printf("%% ngpu %d, distblk %d, NB %d, nstream %d\n", (int) opts.ngpu, (int) distblk, (int) opts.nb, (int) nstream); for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) { magma_setdevice( dev ); for( int i = 0; i < nstream; ++i ) { magma_queue_create( &streams[dev][i] ); } } magma_setdevice( 0 ); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; ldt = N; ldda = magma_roundup( N, opts.align ); // multiple of 32 by default n2 = lda*N; /* We suppose the magma NB is bigger than lapack NB */ lwork = N*opts.nb; //gflops = ....? /* Allocate host memory for the matrix */ TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 ); TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N ); TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_PIN( D, double, N ); TESTING_MALLOC_PIN( E, double, N ); for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) { magma_int_t mlocal = ((N / distblk) / opts.ngpu + 1) * distblk; magma_setdevice( dev ); TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal ); TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*opts.nb ); } /* ==================================================================== Initialize the matrix =================================================================== */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); magma_zmake_hermitian( N, h_A, lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ /* Copy the matrix to the GPU */ magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, opts.ngpu, distblk); //magmaDoubleComplex_ptr dabis; //TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N ); //magma_zsetmatrix(N, N, h_R, lda, dabis, ldda); for (int count=0; count < 1; ++count) { magma_setdevice(0); gpu_time = magma_wtime(); if (opts.version == 30) { // see src/obsolete and magmablas/obsolete printf( "magma_zhetrd_he2hb_mgpu_spec not compiled\n" ); //magma_zhetrd_he2hb_mgpu_spec( // opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, // da, ldda, dT1, opts.nb, opts.ngpu, distblk, // streams, nstream, opts.nthread, &info); } else { nstream = 3; magma_zhetrd_he2hb_mgpu( opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, da, ldda, dT1, opts.nb, opts.ngpu, distblk, streams, nstream, opts.nthread, &info); } // magma_zhetrd_he2hb(opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, dT1[0], &info); gpu_time = magma_wtime() - gpu_time; printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n", N, opts.nb, distblk, opts.ngpu, opts.version, gpu_time); } magma_setdevice(0); for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) { magma_setdevice(dev); magma_device_sync(); } magma_setdevice(0); magmablasSetKernelStream( NULL ); // todo neither of these is declared in headers // magma_zhetrd_bhe2trc_v5(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt); // magma_zhetrd_bhe2trc(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt); // todo where is this timer started? // gpu_time = magma_wtime() - gpu_time; // todo what are the gflops? gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zhetrd_he2hb returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Print performance and error. =================================================================== */ #if defined(CHECKEIG) #if defined(PRECISION_z) || defined(PRECISION_d) if ( opts.check ) { printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time ); double nrmI=0.0, nrm1=0.0, nrm2=0.0; int lwork2 = 256*N; magmaDoubleComplex *work2, *AINIT; double *rwork2, *D2; // TODO free this memory ! magma_zmalloc_cpu( &work2, lwork2 ); magma_dmalloc_cpu( &rwork2, N ); magma_dmalloc_cpu( &D2, N ); magma_zmalloc_cpu( &AINIT, N*lda ); memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex)); /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); int nt = min(12, opts.nthread); magma_set_lapack_numthreads(nt); lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, #ifdef COMPLEX rwork2, #endif &info ); ///* call eigensolver for our resulting tridiag [D E] and for Q */ //dstedc_withZ('V', N, D, E, h_R, lda); ////dsterf_( &N, D, E, &info); cpu_time = magma_wtime() - cpu_time; printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, nt); /* compare result */ cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2); magmaDoubleComplex *WORKAJETER; double *RWORKAJETER, *RESU; // TODO free this memory ! magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) ); magma_dmalloc_cpu( &RWORKAJETER, N ); magma_dmalloc_cpu( &RESU, 10 ); int MATYPE; memset(RESU, 0, 10*sizeof(double)); MATYPE=3; double NOTHING=0.0; cpu_time = magma_wtime(); // check results zcheck_eig_( lapack_vec_const(opts.jobz), &MATYPE, &N, &opts.nb, AINIT, &lda, &NOTHING, &NOTHING, D2, D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU ); cpu_time = magma_wtime() - cpu_time; printf(" Finish CHECK - results timing= %f\n", cpu_time); magma_set_lapack_numthreads(1); printf("\n"); printf(" ================================================================================================================\n"); printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) opts.nthread, (int) N, (int) opts.nb, (int) WANTZ); printf(" ================================================================================================================\n"); printf(" DSBTRD : %15s \n", "STATblgv9withQ "); printf(" ================================================================================================================\n"); if (WANTZ > 0) printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]); if (WANTZ > 0) printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]); printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]); printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]); printf(" ================================================================================================================\n\n\n"); printf(" ****************************************************************************************************************\n"); printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2); printf(" ****************************************************************************************************************\n\n"); } #endif // PRECISION_z || PRECISION_d #endif // CHECKEIG printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time ); printf("%%===========================================================================\n\n\n"); TESTING_FREE_CPU( tau ); TESTING_FREE_PIN( h_A ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_work ); TESTING_FREE_PIN( D ); TESTING_FREE_PIN( E ); for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) { magma_setdevice( dev ); TESTING_FREE_DEV( da[dev] ); TESTING_FREE_DEV( dT1[dev] ); } magma_setdevice( 0 ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) { for( int i = 0; i < nstream; ++i ) { magma_queue_destroy( streams[dev][i] ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhegvdx */ int main( int argc, char** argv) { TESTING_INIT_MGPU(); real_Double_t mgpu_time; magmaDoubleComplex *h_A, *h_Ainit, *h_B, *h_Binit, *h_work; #if defined(PRECISION_z) || defined(PRECISION_c) double *rwork; magma_int_t lrwork; #endif double *w1, result; magma_int_t *iwork; magma_int_t N, n2, info, lwork, liwork; magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_timestr_t start, end; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); char jobz = opts.jobz; int checkres = opts.check; char range = 'A'; char uplo = opts.uplo; magma_int_t itype = opts.itype; double f = opts.fraction; if (f != 1) range='I'; if ( checkres && jobz == MagmaNoVec ) { fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" ); jobz = MagmaVec; } printf("using: nrgpu = %d, itype = %d, jobz = %c, range = %c, uplo = %c, checkres = %d, fraction = %6.4f\n", (int) opts.ngpu, (int) itype, jobz, range, uplo, (int) checkres, f); printf(" N M nr GPU MGPU Time(s) \n"); printf("====================================\n"); magma_int_t threads = magma_get_numthreads(); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[i]; n2 = N*N; #if defined(PRECISION_z) || defined(PRECISION_c) lwork = magma_zbulge_get_lq2(N, threads) + 2*N + N*N; lrwork = 1 + 5*N +2*N*N; #else lwork = magma_zbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N; #endif liwork = 3 + 5*N; //magma_int_t NB = 96;//magma_bulge_get_nb(N); //magma_int_t sizvblg = magma_zbulge_get_lq2(N, threads); //magma_int_t siz = max(sizvblg,n2)+2*(N*NB+N)+24*N; /* Allocate host memory for the matrix */ TESTING_HOSTALLOC( h_A, magmaDoubleComplex, n2); TESTING_HOSTALLOC( h_B, magmaDoubleComplex, n2); TESTING_MALLOC( w1, double , N); TESTING_HOSTALLOC(h_work, magmaDoubleComplex, lwork); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_HOSTALLOC( rwork, double, lrwork); #endif TESTING_MALLOC( iwork, magma_int_t, liwork); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlarnv( &ione, ISEED, &n2, h_B ); /* increase the diagonal */ { for(int i=0; i<N; i++) { MAGMA_Z_SET2REAL( h_B[i*N+i], ( MAGMA_Z_REAL(h_B[i*N+i]) + 1.*N ) ); MAGMA_Z_SET2REAL( h_A[i*N+i], MAGMA_Z_REAL(h_A[i*N+i]) ); } } if((opts.warmup)||( checkres )){ TESTING_MALLOC(h_Ainit, magmaDoubleComplex, n2); TESTING_MALLOC(h_Binit, magmaDoubleComplex, n2); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N ); } magma_int_t m1 = 0; double vl = 0; double vu = 0; magma_int_t il = 0; magma_int_t iu = 0; if (range == 'I'){ il = 1; iu = (int) (f*N); } if(opts.warmup){ // ================================================================== // Warmup using MAGMA. I prefer to use smalltest to warmup A- // ================================================================== magma_zhegvdx_2stage_m(opts.ngpu, itype, jobz, range, uplo, N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N ); } // =================================================================== // Performs operation using MAGMA // =================================================================== start = get_current_time(); magma_zhegvdx_2stage_m(opts.ngpu, itype, jobz, range, uplo, N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); end = get_current_time(); mgpu_time = GetTimerValue(start,end)/1000.; if ( checkres ) { // =================================================================== // Check the results following the LAPACK's [zc]hegvdx routine. // A x = lambda B x is solved // and the following 3 tests computed: // (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) // | A B Z - Z D | / ( |A||Z| N ) (itype = 2) // | B A Z - Z D | / ( |A||Z| N ) (itype = 3) // =================================================================== #if defined(PRECISION_d) || defined(PRECISION_s) double *rwork = h_work + N*N; #endif result = 1.; result /= lapackf77_zlanhe("1",&uplo, &N, h_Ainit, &N, rwork); result /= lapackf77_zlange("1",&N , &m1, h_A, &N, rwork); if (itype == 1){ blasf77_zhemm("L", &uplo, &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_zhemm("L", &uplo, &N, &m1, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N); result *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N; } else if (itype == 2){ blasf77_zhemm("L", &uplo, &N, &m1, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_zhemm("L", &uplo, &N, &m1, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N); result *= lapackf77_zlange("1", &N, &m1, h_A, &N, rwork)/N; } else if (itype == 3){ blasf77_zhemm("L", &uplo, &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_A[i*N], &ione); blasf77_zhemm("L", &uplo, &N, &m1, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N); result *= lapackf77_zlange("1", &N, &m1, h_A, &N, rwork)/N; } } // =================================================================== // Print execution time // =================================================================== printf("%5d %5d %2d %6.2f\n", (int) N, (int) m1, (int) opts.ngpu, mgpu_time); if ( checkres ){ printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if(itype==1) printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") ); else if(itype==2) printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") ); else if(itype==3) printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") ); } TESTING_HOSTFREE( h_A); TESTING_HOSTFREE( h_B); TESTING_FREE( w1); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_HOSTFREE( rwork); #endif TESTING_FREE( iwork); TESTING_HOSTFREE(h_work); if((opts.warmup)||( checkres )){ TESTING_FREE( h_Ainit); TESTING_FREE( h_Binit); } } if ( opts.niter > 1 ) { printf( "\n" ); } } /* Shutdown */ TESTING_FINALIZE_MGPU(); return 0; }
int main( int argc, char** argv) { real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time; magmaDoubleComplex *hA, *hR; magmaDoubleComplex_ptr dA; magma_int_t N = 0, n2, lda, ldda; magma_int_t size[10] = { 1024, 2048, 3072, 4032, 5184, 6048, 7200, 8064, 8928, 10560 }; magma_int_t i, info; magmaDoubleComplex mz_one = MAGMA_Z_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; double work[1], matnorm, diffnorm; 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; else exit(1); } else { printf("\nUsage: \n"); printf(" testing_zpotrf_gpu -N %d\n\n", 1024); } /* Initialize */ magma_queue_t queue; magma_device_t device; int num = 0; magma_err_t err; magma_init(); err = magma_get_devices( &device, 1, &num ); if ( err != 0 || num < 1 ) { fprintf( stderr, "magma_get_devices failed: %d\n", err ); exit(-1); } err = magma_queue_create( device, &queue ); if ( err != 0 ) { fprintf( stderr, "magma_queue_create failed: %d\n", err ); exit(-1); } /* Allocate memory for the largest matrix */ N = size[9]; n2 = N * N; ldda = ((N+31)/32) * 32; TESTING_MALLOC( hA, magmaDoubleComplex, n2 ); TESTING_MALLOC_HOST( hR, magmaDoubleComplex, n2 ); TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N ); printf("\n\n"); 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( (double)N ) * 1e-9; /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, hA ); /* Symmetrize and increase the diagonal */ for( int i = 0; i < N; ++i ) { MAGMA_Z_SET2REAL( hA(i,i), MAGMA_Z_REAL(hA(i,i)) + N ); for( int j = 0; j < i; ++j ) { hA(i, j) = MAGMA_Z_CNJG( hA(j,i) ); } } lapackf77_zlacpy( MagmaFullStr, &N, &N, hA, &lda, hR, &lda ); /* Warm up to measure the performance */ magma_zsetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue ); magma_zpotrf_gpu( MagmaUpper, N, dA, 0, ldda, &info, queue ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_zsetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue ); gpu_time = get_time(); magma_zpotrf_gpu( MagmaUpper, N, dA, 0, ldda, &info, queue ); gpu_time = get_time() - gpu_time; if (info != 0) printf( "magma_zpotrf had error %d.\n", info ); gpu_perf = gflops / gpu_time; /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = get_time(); lapackf77_zpotrf( MagmaUpperStr, &N, hA, &lda, &info ); cpu_time = get_time() - cpu_time; if (info != 0) printf( "lapackf77_zpotrf had error %d.\n", info ); cpu_perf = gflops / cpu_time; /* ===================================================================== Check the result compared to LAPACK |R_magma - R_lapack| / |R_lapack| =================================================================== */ magma_zgetmatrix( N, N, dA, 0, ldda, hR, 0, lda, queue ); matnorm = lapackf77_zlange("f", &N, &N, hA, &lda, work); blasf77_zaxpy(&n2, &mz_one, hA, &ione, hR, &ione); diffnorm = lapackf77_zlange("f", &N, &N, hR, &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 (argc != 1) break; } /* clean up */ TESTING_FREE( hA ); TESTING_FREE_HOST( hR ); TESTING_FREE_DEV( dA ); magma_queue_destroy( queue ); magma_finalize(); }
extern "C" magma_int_t magma_zgeev_m( char jobvl, char jobvr, magma_int_t n, magmaDoubleComplex *A, magma_int_t lda, magmaDoubleComplex *W, magmaDoubleComplex *vl, magma_int_t ldvl, magmaDoubleComplex *vr, magma_int_t ldvr, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t *info ) { /* -- MAGMA (version 1.4.1) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver December 2013 Purpose ======= ZGEEV computes for an N-by-N complex nonsymmetric matrix A, the eigenvalues and, optionally, the left and/or right eigenvectors. The right eigenvector v(j) of A satisfies A * v(j) = lambda(j) * v(j) where lambda(j) is its eigenvalue. The left eigenvector u(j) of A satisfies u(j)**H * A = lambda(j) * u(j)**H where u(j)**H denotes the conjugate transpose of u(j). The computed eigenvectors are normalized to have Euclidean norm equal to 1 and largest component real. Arguments ========= JOBVL (input) CHARACTER*1 = 'N': left eigenvectors of A are not computed; = 'V': left eigenvectors of are computed. JOBVR (input) CHARACTER*1 = 'N': right eigenvectors of A are not computed; = 'V': right eigenvectors of A are computed. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX*16 array, dimension (LDA,N) On entry, the N-by-N matrix A. On exit, A has been overwritten. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). W (output) COMPLEX*16 array, dimension (N) W contains the computed eigenvalues. VL (output) COMPLEX*16 array, dimension (LDVL,N) If JOBVL = 'V', the left eigenvectors u(j) are stored one after another in the columns of VL, in the same order as their eigenvalues. If JOBVL = 'N', VL is not referenced. u(j) = VL(:,j), the j-th column of VL. LDVL (input) INTEGER The leading dimension of the array VL. LDVL >= 1; if JOBVL = 'V', LDVL >= N. VR (output) COMPLEX*16 array, dimension (LDVR,N) If JOBVR = 'V', the right eigenvectors v(j) are stored one after another in the columns of VR, in the same order as their eigenvalues. If JOBVR = 'N', VR is not referenced. v(j) = VR(:,j), the j-th column of VR. LDVR (input) INTEGER The leading dimension of the array VR. LDVR >= 1; if JOBVR = 'V', LDVR >= N. WORK (workspace/output) COMPLEX*16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. LWORK >= (1+nb)*N. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. RWORK (workspace) DOUBLE PRECISION array, dimension (2*N) INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value. > 0: if INFO = i, the QR algorithm failed to compute all the eigenvalues, and no eigenvectors have been computed; elements and i+1:N of W contain eigenvalues which have converged. ===================================================================== */ #define vl(i,j) (vl + (i) + (j)*ldvl) #define vr(i,j) (vr + (i) + (j)*ldvr) magma_int_t c_one = 1; magma_int_t c_zero = 0; double d__1, d__2; magmaDoubleComplex z__1, z__2; magmaDoubleComplex tmp; double scl; double dum[1], eps; double anrm, cscale, bignum, smlnum; magma_int_t i, k, ilo, ihi; magma_int_t ibal, ierr, itau, iwrk, nout, liwrk, i__1, i__2, nb; magma_int_t scalea, minwrk, irwork, lquery, wantvl, wantvr, select[1]; char side[2] = {0, 0}; char jobvl_[2] = {jobvl, 0}; char jobvr_[2] = {jobvr, 0}; irwork = 0; *info = 0; lquery = lwork == -1; wantvl = lapackf77_lsame( jobvl_, "V" ); wantvr = lapackf77_lsame( jobvr_, "V" ); if (! wantvl && ! lapackf77_lsame( jobvl_, "N" )) { *info = -1; } else if (! wantvr && ! lapackf77_lsame( jobvr_, "N" )) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if ( (ldvl < 1) || (wantvl && (ldvl < n))) { *info = -8; } else if ( (ldvr < 1) || (wantvr && (ldvr < n))) { *info = -10; } /* Compute workspace */ nb = magma_get_zgehrd_nb( n ); if (*info == 0) { minwrk = (1+nb)*n; work[0] = MAGMA_Z_MAKE( minwrk, 0 ); if (lwork < minwrk && ! lquery) { *info = -12; } } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } #if defined(Version3) || defined(Version4) || defined(Version5) magmaDoubleComplex *dT; if (MAGMA_SUCCESS != magma_zmalloc( &dT, nb*n )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } #endif #if defined(Version4) || defined(Version5) magmaDoubleComplex *T; if (MAGMA_SUCCESS != magma_zmalloc_cpu( &T, nb*n )) { magma_free( dT ); *info = MAGMA_ERR_HOST_ALLOC; return *info; } #endif /* Get machine constants */ eps = lapackf77_dlamch( "P" ); smlnum = lapackf77_dlamch( "S" ); bignum = 1. / smlnum; lapackf77_dlabad( &smlnum, &bignum ); smlnum = magma_dsqrt( smlnum ) / eps; bignum = 1. / smlnum; /* Scale A if max element outside range [SMLNUM,BIGNUM] */ anrm = lapackf77_zlange( "M", &n, &n, A, &lda, dum ); scalea = 0; if (anrm > 0. && anrm < smlnum) { scalea = 1; cscale = smlnum; } else if (anrm > bignum) { scalea = 1; cscale = bignum; } if (scalea) { lapackf77_zlascl( "G", &c_zero, &c_zero, &anrm, &cscale, &n, &n, A, &lda, &ierr ); } /* Balance the matrix * (CWorkspace: none) * (RWorkspace: need N) */ ibal = 0; lapackf77_zgebal( "B", &n, A, &lda, &ilo, &ihi, &rwork[ibal], &ierr ); /* Reduce to upper Hessenberg form * (CWorkspace: need 2*N, prefer N + N*NB) * (RWorkspace: none) */ itau = 0; iwrk = itau + n; liwrk = lwork - iwrk; #if defined(Version1) // Version 1 - LAPACK lapackf77_zgehrd( &n, &ilo, &ihi, A, &lda, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version2) // Version 2 - LAPACK consistent HRD magma_zgehrd2( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) // Version 3 - LAPACK consistent MAGMA HRD + matrices T stored, magma_zgehrd( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, dT, &ierr ); #elif defined(Version4) || defined(Version5) // Version 4 - Multi-GPU, T on host magma_zgehrd_m( n, ilo, ihi, A, lda, &work[itau], &work[iwrk], liwrk, T, &ierr ); magma_zsetmatrix( nb, n, T, nb, dT, nb ); #endif if (wantvl) { /* Want left eigenvectors * Copy Householder vectors to VL */ side[0] = 'L'; lapackf77_zlacpy( MagmaLowerStr, &n, &n, A, &lda, vl, &ldvl ); /* Generate unitary matrix in VL * (CWorkspace: need 2*N-1, prefer N + (N-1)*NB) * (RWorkspace: none) */ #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_zunghr( &n, &ilo, &ihi, vl, &ldvl, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) || defined(Version4) // Version 3 - LAPACK consistent MAGMA HRD + matrices T stored magma_zunghr( n, ilo, ihi, vl, ldvl, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_zunghr_m( n, ilo, ihi, vl, ldvl, &work[itau], T, nb, &ierr ); #endif /* Perform QR iteration, accumulating Schur vectors in VL * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: none) */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_zhseqr( "S", "V", &n, &ilo, &ihi, A, &lda, W, vl, &ldvl, &work[iwrk], &liwrk, info ); if (wantvr) { /* Want left and right eigenvectors * Copy Schur vectors to VR */ side[0] = 'B'; lapackf77_zlacpy( "F", &n, &n, vl, &ldvl, vr, &ldvr ); } } else if (wantvr) { /* Want right eigenvectors * Copy Householder vectors to VR */ side[0] = 'R'; lapackf77_zlacpy( "L", &n, &n, A, &lda, vr, &ldvr ); /* Generate unitary matrix in VR * (CWorkspace: need 2*N-1, prefer N + (N-1)*NB) * (RWorkspace: none) */ #if defined(Version1) || defined(Version2) // Version 1 & 2 - LAPACK lapackf77_zunghr( &n, &ilo, &ihi, vr, &ldvr, &work[itau], &work[iwrk], &liwrk, &ierr ); #elif defined(Version3) || defined(Version4) // Version 3 - LAPACK consistent MAGMA HRD + matrices T stored magma_zunghr( n, ilo, ihi, vr, ldvr, &work[itau], dT, nb, &ierr ); #elif defined(Version5) // Version 5 - Multi-GPU, T on host magma_zunghr_m( n, ilo, ihi, vr, ldvr, &work[itau], T, nb, &ierr ); #endif /* Perform QR iteration, accumulating Schur vectors in VR * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: none) */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_zhseqr( "S", "V", &n, &ilo, &ihi, A, &lda, W, vr, &ldvr, &work[iwrk], &liwrk, info ); } else { /* Compute eigenvalues only * (CWorkspace: need 1, prefer HSWORK (see comments) ) * (RWorkspace: none) */ iwrk = itau; liwrk = lwork - iwrk; lapackf77_zhseqr( "E", "N", &n, &ilo, &ihi, A, &lda, W, vr, &ldvr, &work[iwrk], &liwrk, info ); } /* If INFO > 0 from ZHSEQR, then quit */ if (*info > 0) { goto CLEANUP; } if (wantvl || wantvr) { /* Compute left and/or right eigenvectors * (CWorkspace: need 2*N) * (RWorkspace: need 2*N) */ irwork = ibal + n; lapackf77_ztrevc( side, "B", select, &n, A, &lda, vl, &ldvl, vr, &ldvr, &n, &nout, &work[iwrk], &rwork[irwork], &ierr ); } if (wantvl) { /* Undo balancing of left eigenvectors * (CWorkspace: none) * (RWorkspace: need N) */ lapackf77_zgebak( "B", "L", &n, &ilo, &ihi, &rwork[ibal], &n, vl, &ldvl, &ierr ); /* Normalize left eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { scl = 1. / cblas_dznrm2( n, vl(0,i), 1 ); cblas_zdscal( n, scl, vl(0,i), 1 ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = MAGMA_Z_REAL( *vl(k,i) ); d__2 = MAGMA_Z_IMAG( *vl(k,i) ); rwork[irwork + k] = d__1*d__1 + d__2*d__2; } k = cblas_idamax( n, &rwork[irwork], 1 ); z__2 = MAGMA_Z_CNJG( *vl(k,i) ); d__1 = magma_dsqrt( rwork[irwork + k] ); MAGMA_Z_DSCALE( z__1, z__2, d__1 ); tmp = z__1; cblas_zscal( n, CBLAS_SADDR(tmp), vl(0,i), 1 ); d__1 = MAGMA_Z_REAL( *vl(k,i) ); z__1 = MAGMA_Z_MAKE( d__1, 0 ); *vl(k,i) = z__1; } } if (wantvr) { /* Undo balancing of right eigenvectors * (CWorkspace: none) * (RWorkspace: need N) */ lapackf77_zgebak( "B", "R", &n, &ilo, &ihi, &rwork[ibal], &n, vr, &ldvr, &ierr ); /* Normalize right eigenvectors and make largest component real */ for (i = 0; i < n; ++i) { scl = 1. / cblas_dznrm2( n, vr(0,i), 1 ); cblas_zdscal( n, scl, vr(0,i), 1 ); for (k = 0; k < n; ++k) { /* Computing 2nd power */ d__1 = MAGMA_Z_REAL( *vr(k,i) ); d__2 = MAGMA_Z_IMAG( *vr(k,i) ); rwork[irwork + k] = d__1*d__1 + d__2*d__2; } k = cblas_idamax( n, &rwork[irwork], 1 ); z__2 = MAGMA_Z_CNJG( *vr(k,i) ); d__1 = magma_dsqrt( rwork[irwork + k] ); MAGMA_Z_DSCALE( z__1, z__2, d__1 ); tmp = z__1; cblas_zscal( n, CBLAS_SADDR(tmp), vr(0,i), 1 ); d__1 = MAGMA_Z_REAL( *vr(k,i) ); z__1 = MAGMA_Z_MAKE( d__1, 0 ); *vr(k,i) = z__1; } } CLEANUP: /* Undo scaling if necessary */ if (scalea) { i__1 = n - (*info); i__2 = max( n - (*info), 1 ); lapackf77_zlascl( "G", &c_zero, &c_zero, &cscale, &anrm, &i__1, &c_one, W + (*info), &i__2, &ierr ); if (*info > 0) { i__1 = ilo - 1; lapackf77_zlascl( "G", &c_zero, &c_zero, &cscale, &anrm, &i__1, &c_one, W, &n, &ierr ); } } #if defined(Version3) || defined(Version4) || defined(Version5) magma_free( dT ); #endif #if defined(Version4) || defined(Version5) magma_free_cpu( T ); #endif return *info; } /* magma_zgeev */
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgeqrf */ int main( magma_int_t argc, char** argv) { magma_int_t nquarkthreads=2; magma_int_t nthreads=2; magma_int_t num_gpus = 1; TRACE = 0; //magma_qr_params mp; cuDoubleComplex *h_A, *h_R, *h_work, *tau; double gpu_perf, cpu_perf, flops; magma_timestr_t start, end; magma_qr_params *mp = (magma_qr_params*)malloc(sizeof(magma_qr_params)); /* Matrix size */ magma_int_t M=0, N=0, n2; magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7040,8064,9088,10112}; cublasStatus status; magma_int_t i, j, info; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; mp->nb=-1; mp->ob=-1; mp->fb=-1; mp->ib=32; magma_int_t loop = argc; magma_int_t accuracyflag = 1; char precision; magma_int_t nc = -1; magma_int_t ncps = -1; if (argc != 1) { for(i = 1; i<argc; i++){ if (strcmp("-N", argv[i])==0) N = atoi(argv[++i]); else if (strcmp("-M", argv[i])==0) M = atoi(argv[++i]); else if (strcmp("-F", argv[i])==0) mp->fb = atoi(argv[++i]); else if (strcmp("-O", argv[i])==0) mp->ob = atoi(argv[++i]); else if (strcmp("-B", argv[i])==0) mp->nb = atoi(argv[++i]); else if (strcmp("-b", argv[i])==0) mp->ib = atoi(argv[++i]); else if (strcmp("-A", argv[i])==0) accuracyflag = atoi(argv[++i]); else if (strcmp("-P", argv[i])==0) nthreads = atoi(argv[++i]); else if (strcmp("-Q", argv[i])==0) nquarkthreads = atoi(argv[++i]); else if (strcmp("-nc", argv[i])==0) nc = atoi(argv[++i]); else if (strcmp("-ncps", argv[i])==0) ncps = atoi(argv[++i]); } if ((M>0 && N>0) || (M==0 && N==0)) { printf(" testing_zgeqrf-v2 -M %d -N %d\n\n", M, N); if (M==0 && N==0) { M = N = size[9]; loop = 1; } } else { printf("\nUsage: \n"); printf(" Make sure you set the number of BLAS threads to 1, e.g.,\n"); printf(" > setenv MKL_NUM_THREADS 1\n"); printf(" > testing_zgeqrf-v2 -M %d -N %d -B 128 -T 1\n\n", 1024, 1024); exit(1); } } else { printf("\nUsage: \n"); printf(" Make sure you set the number of BLAS threads to 1, e.g.,\n"); printf(" > setenv MKL_NUM_THREADS 1\n"); printf(" Set number of cores per socket and number of cores.\n"); printf(" > testing_zgeqrf-v2 -M %d -N %d -ncps 6 -nc 12\n\n", 1024, 1024); printf(" Alternatively, set:\n"); printf(" Q: Number of threads for panel factorization.\n"); printf(" P: Number of threads for trailing matrix update (CPU).\n"); printf(" B: Block size.\n"); printf(" b: Inner block size.\n"); printf(" O: Block size for trailing matrix update (CPU).\n"); printf(" > testing_zgeqrf-v2 -M %d -N %d -Q 4 -P 4 -B 128 -b 32 -O 200\n\n", 10112, 10112); M = N = size[9]; } /* Auto tune based on number of cores and number of cores per socket if provided */ if ((nc > 0) && (ncps > 0)) { precision = 's'; #if (defined(PRECISION_d)) precision = 'd'; #endif #if (defined(PRECISION_c)) precision = 'c'; #endif #if (defined(PRECISION_z)) precision = 'z'; #endif auto_tune('q', precision, nc, ncps, M, N, &(mp->nb), &(mp->ob), &(mp->ib), &nthreads, &nquarkthreads); fprintf(stderr,"%d %d %d %d %d\n",mp->nb,mp->ob,mp->ib,nquarkthreads,nthreads); } /* Initialize MAGMA hardware context, seeting how many CPU cores and how many GPUs to be used in the consequent computations */ mp->sync0 = 0; magma_context *context; context = magma_init((void*)(mp),cpu_thread, nthreads, nquarkthreads, num_gpus, argc, argv); context->params = (void *)(mp); mp->sync1 = (volatile magma_int_t *) malloc (sizeof(int)*nthreads); for (i = 0; i < nthreads; i++) mp->sync1[i] = 0; n2 = M * N; magma_int_t min_mn = min(M, N); magma_int_t nb = magma_get_zgeqrf_nb(min_mn); magma_int_t lwork = N*nb; /* Allocate host memory for the matrix */ TESTING_MALLOC ( h_A , cuDoubleComplex, n2 ); TESTING_MALLOC ( tau , cuDoubleComplex, min_mn); TESTING_HOSTALLOC( h_R , cuDoubleComplex, n2 ); TESTING_HOSTALLOC(h_work, cuDoubleComplex, lwork ); printf("\n\n"); printf(" M N CPU GFlop/s GPU GFlop/s ||R||_F / ||A||_F\n"); printf("==========================================================\n"); for(i=0; i<10; i++){ if (loop==1){ M = N = min_mn = size[i]; n2 = M*N; } flops = FLOPS( (double)M, (double)N ) / 1000000; /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &M, h_R, &M ); //magma_zgeqrf(M, N, h_R, M, tau, h_work, lwork, &info); for(j=0; j<n2; j++) h_R[j] = h_A[j]; /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_qr_init(mp, M, N, h_R, nthreads); start = get_current_time(); magma_zgeqrf3(context, M, N, h_R, M, tau, h_work, lwork, &info); end = get_current_time(); gpu_perf = flops / GetTimerValue(start, end); /* ===================================================================== Performs operation using LAPACK =================================================================== */ start = get_current_time(); if (accuracyflag == 1) lapackf77_zgeqrf(&M, &N, h_A, &M, tau, h_work, &lwork, &info); end = get_current_time(); if (info < 0) printf("Argument %d of zgeqrf had an illegal value.\n", -info); cpu_perf = 4.*M*N*min_mn/(3.*1000000*GetTimerValue(start,end)); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ double work[1], matnorm = 1.; cuDoubleComplex mone = MAGMA_Z_NEG_ONE; magma_int_t one = 1; if (accuracyflag == 1){ matnorm = lapackf77_zlange("f", &M, &N, h_A, &M, work); blasf77_zaxpy(&n2, &mone, h_A, &one, h_R, &one); } if (accuracyflag == 1){ printf("%5d %5d %6.2f %6.2f %e\n", M, N, cpu_perf, gpu_perf, lapackf77_zlange("f", &M, &N, h_R, &M, work) / matnorm); } else { printf("%5d %5d %6.2f \n", M, N, gpu_perf); } if (loop != 1) break; } /* Memory clean up */ TESTING_FREE ( h_A ); TESTING_FREE ( tau ); TESTING_HOSTFREE(h_work); TESTING_HOSTFREE( h_R ); /* Shut down the MAGMA context */ magma_finalize(context); }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhegvdx */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gpu_time; magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work; #if defined(PRECISION_z) || defined(PRECISION_c) double *rwork; magma_int_t lrwork; #endif /* Matrix size */ double *w1, *w2, result[2]={0,0}; magma_int_t *iwork; magma_int_t N, n2, info, lwork, liwork; magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; 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 ); double tol = opts.tolerance * lapackf77_dlamch("E"); double tolulp = opts.tolerance * lapackf77_dlamch("P"); magma_range_t range = MagmaRangeAll; if (opts.fraction != 1) range = MagmaRangeI; if ( opts.check && opts.jobz == MagmaNoVec ) { fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" ); opts.jobz = MagmaVec; } printf("using: itype = %d, jobz = %s, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n", (int) opts.itype, lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo), (int) opts.check, opts.fraction); printf(" N M GPU Time (sec)\n"); printf("============================\n"); magma_int_t threads = magma_get_parallel_numthreads(); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; n2 = N*N; #if defined(PRECISION_z) || defined(PRECISION_c) lwork = magma_zbulge_get_lq2(N, threads) + 2*N + N*N; lrwork = 1 + 5*N +2*N*N; #else lwork = magma_zbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N; #endif liwork = 3 + 5*N; /* Allocate host memory for the matrix */ TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( w1, double, N ); TESTING_MALLOC_CPU( w2, double, N ); TESTING_MALLOC_CPU( iwork, magma_int_t, liwork ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_S, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_MALLOC_PIN( rwork, double, lrwork); #endif /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlarnv( &ione, ISEED, &n2, h_B ); magma_zmake_hpd( N, h_B, N ); magma_zmake_hermitian( N, h_A, N ); magma_int_t m1 = 0; double vl = 0; double vu = 0; magma_int_t il = 0; magma_int_t iu = 0; if (range == MagmaRangeI) { il = 1; iu = (int) (opts.fraction*N); } // ================================================================== // Warmup using MAGMA // ================================================================== if (opts.warmup) { lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo, N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); } // =================================================================== // Performs operation using MAGMA // =================================================================== lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); gpu_time = magma_wtime(); magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo, N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info); gpu_time = magma_wtime() - gpu_time; if ( opts.check && opts.jobz != MagmaNoVec ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvdx routine. A x = lambda B x is solved and the following 3 tests computed: (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1) | A B Z - Z D | / ( |A||Z| N ) (itype = 2) | B A Z - Z D | / ( |A||Z| N ) (itype = 3) (2) | S(with V) - S(w/o V) | / | S | =================================================================== */ #if defined(PRECISION_d) || defined(PRECISION_s) double *rwork = h_work + N*N; #endif result[0] = 1.; result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork); result[0] /= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork); if (opts.itype == 1) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N; } else if (opts.itype == 2) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N; } else if (opts.itype == 3) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N); for(int i=0; i<m1; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N); result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N; } lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N ); magma_int_t m2 = m1; lapackf77_zhegvd(&opts.itype, "N", lapack_uplo_const(opts.uplo), &N, h_R, &N, h_S, &N, w2, h_work, &lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, &lrwork, #endif iwork, &liwork, &info); double maxw=0, diff=0; for(int j=0; j<m2; j++) { maxw = max(maxw, fabs(w1[j])); maxw = max(maxw, fabs(w2[j])); diff = max(diff, fabs(w1[j] - w2[j])); } result[1] = diff / (m2*maxw); } /* ===================================================================== Print execution time =================================================================== */ printf("%5d %5d %7.2f\n", (int) N, (int) m1, gpu_time); if ( opts.check && opts.jobz != MagmaNoVec ) { printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if (opts.itype==1) { printf(" | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } else if (opts.itype==2) { printf(" | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } else if (opts.itype==3) { printf(" | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } printf( " | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed")); status += ! (result[0] < tol && result[1] < tolulp); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( w1 ); TESTING_FREE_CPU( w2 ); TESTING_FREE_CPU( iwork ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_S ); TESTING_FREE_PIN( h_work ); #if defined(PRECISION_z) || defined(PRECISION_c) TESTING_FREE_PIN( rwork ); #endif fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } /* Shutdown */ TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing magma_zhemm_mgpu */ int main( int argc, char** argv) { TESTING_INIT(); magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex calpha = MAGMA_Z_MAKE( 3.456, 5.678 ); magmaDoubleComplex cbeta = MAGMA_Z_MAKE( 1.234, 2.456 ); real_Double_t gflops, gpu_perf=0., cpu_perf=0., gpu_time=0., cpu_time=0.; real_Double_t gpu_perf2=0., gpu_time2=0.; double error=0., errorbis=0., work[1]; magmaDoubleComplex *hA, *hX, *hB, *hR; magmaDoubleComplex_ptr dA[MagmaMaxGPUs], dX[MagmaMaxGPUs], dB[MagmaMaxGPUs], dwork[MagmaMaxGPUs], hwork[MagmaMaxGPUs+1]; magmaDoubleComplex_ptr dA2; magma_int_t M, N, size, lda, ldda, msize, nb, nstream; 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 ); double tol = opts.tolerance * lapackf77_dlamch("E"); // default values nb = (opts.nb > 0 ? opts.nb : 64); nstream = (opts.nstream > 0 ? opts.nstream : 2); magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2]; magma_int_t nbcmplx = 0; magma_buildconnection_mgpu(gnode, &nbcmplx, opts.ngpu); printf("Initializing communication pattern... GPU-ncmplx %d\n\n", (int) nbcmplx); for (int i=0; i < nbcmplx; ++i) { int myngpu = gnode[i][MagmaMaxGPUs]; printf("cmplx %d has %d gpu ", i, myngpu); for(int j=0; j < myngpu; ++j) printf(" %d", (int) gnode[i][j]); printf("\n"); } magma_int_t nbevents = 2; magma_queue_t streams[MagmaMaxGPUs][20]; magma_event_t redevents[MagmaMaxGPUs][20]; magma_event_t redevents2[MagmaMaxGPUs][MagmaMaxGPUs*MagmaMaxGPUs+10]; for( int d = 0; d < opts.ngpu; ++d ) { for( magma_int_t i = 0; i < nstream; ++i ) { magma_queue_create( &streams[d][i] ); } for( magma_int_t i = 0; i < nbevents; ++i ) { cudaEventCreateWithFlags(&redevents[d][i], cudaEventDisableTiming); cudaEventCreateWithFlags(&redevents2[d][i], cudaEventDisableTiming); } } printf( "nb %d, ngpu %d, nstream %d version %d\n", (int) nb, (int) opts.ngpu, (int) nstream, (int) opts.version ); printf(" M N nb offset CPU GFlop/s (sec) GPU GFlop/s (sec) CUBLAS hemm (sec) ||R|| / ||A||*||X||\n"); printf("=========================================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { M = opts.msize[itest]; N = opts.nsize[itest]; for( int offset = 0; offset < N; offset += min(N,nb) ) { for( int iter = 0; iter < opts.niter; ++iter ) { msize = M - offset; lda = M; ldda = ((M + 31)/32)*32; size = lda*M; gflops = FLOPS_ZHEMM( MagmaLeft, (double)msize, (double)N ) / 1e9; magma_int_t dworksiz = ldda*N*3; magma_int_t hworksiz = lda*N; TESTING_MALLOC_CPU( hA, magmaDoubleComplex, lda*M ); TESTING_MALLOC_CPU( hX, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( hB, magmaDoubleComplex, lda*N ); TESTING_MALLOC_PIN( hR, magmaDoubleComplex, lda*N ); for( int d = 0; d < opts.ngpu; ++d ) { magma_int_t mlocal = ((M / nb) / opts.ngpu + 1) * nb; magma_setdevice( d ); TESTING_MALLOC_DEV( dA[d], magmaDoubleComplex, ldda*mlocal ); TESTING_MALLOC_DEV( dX[d], magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( dB[d], magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( dwork[d], magmaDoubleComplex, dworksiz ); TESTING_MALLOC_PIN( hwork[d], magmaDoubleComplex, hworksiz ); } TESTING_MALLOC_PIN( hwork[opts.ngpu], magmaDoubleComplex, lda*N ); if ( opts.check ) { magma_setdevice( 0 ); TESTING_MALLOC_DEV( dA2, magmaDoubleComplex, ldda*M ); } lapackf77_zlarnv( &ione, iseed, &size, hA ); magma_zmake_hermitian( M, hA, lda ); size = lda*N; lapackf77_zlarnv( &ione, iseed, &size, hX ); lapackf77_zlarnv( &ione, iseed, &size, hB ); lapackf77_zlacpy( "Full", &M, &N, hB, &lda, hR, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_zsetmatrix_1D_col_bcyclic( M, M, hA, lda, dA, ldda, opts.ngpu, nb ); for( int d = 0; d < opts.ngpu; ++d ) { magma_setdevice( d ); //magmablasSetKernelStream( streams[ d ][ 0 ] ); magma_zsetmatrix( M, N, hX, lda, dX[d], ldda ); //if (d == 0) magma_zsetmatrix( M, N, hB, lda, dB[d], ldda ); // this is wrong coz when offset != 0 the gpu who do the beta*C may be not 0 so this should be related to stdev(starting device who own i=0 first col) magma_zsetmatrix( M, N, hB, lda, dB[d], ldda ); } //memset(hR, 0, lda*N*sizeof(magmaDoubleComplex)); trace_init( 1, opts.ngpu, nstream, (magma_queue_t*) streams ); //magma_int_t offset = 0; //nb; gpu_time = magma_sync_wtime(0); magmablas_zhemm_mgpu_com( MagmaLeft, MagmaLower, msize, N, calpha, dA, ldda, offset, dX, ldda, cbeta, dB, ldda, dwork, dworksiz, hR, lda, hwork, hworksiz, opts.ngpu, nb, streams, nstream, redevents2, nbevents, gnode, nbcmplx); gpu_time = magma_sync_wtime(0) - gpu_time; gpu_perf = gflops / gpu_time; #ifdef TRACING char buf[80]; snprintf( buf, sizeof(buf), "zhemm-m%d-n%d-nb%d-stream%d-ngpu%d-run%d.svg", (int) M, (int) N, (int) nb, (int) nstream, (int) opts.ngpu, (int) iter ); trace_finalize( buf, "trace.css" ); #endif /* ==================================================================== Performs operation using CUBLAS =================================================================== */ if ( opts.check && iter == 0 ) { magma_setdevice( 0 ); magmablasSetKernelStream( 0 ); magma_zsetmatrix( M, M, hA, lda, dA2, ldda ); magma_zsetmatrix( M, N, hX, lda, dX[0], ldda ); magma_zsetmatrix( M, N, hB, lda, dwork[0], ldda ); gpu_time2 = magma_sync_wtime(0); magma_zhemm( MagmaLeft, MagmaLower, msize, N, calpha, dA2+offset*ldda+offset, ldda, dX[0], ldda, cbeta, dwork[0], ldda ); gpu_time2 = magma_sync_wtime(0) - gpu_time2; gpu_perf2 = gflops / gpu_time2; } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.check ) { // store ||A||*||X|| errorbis = lapackf77_zlange("fro", &msize, &msize, hA+offset*lda+offset, &lda, work ); errorbis *= lapackf77_zlange("fro", &msize, &N, hX, &lda, work ); //printf( "A =" ); magma_zprint( M, M, hA, lda ); //printf( "X =" ); magma_zprint( M, N, hX, lda ); //printf( "B =" ); magma_zprint( M, N, hB, lda ); cpu_time = magma_wtime(); blasf77_zhemm( "Left", "Lower", &msize, &N, &calpha, hA+offset*lda+offset, &lda, hX, &lda, &cbeta, hB, &lda ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; /* trace_file = fopen("AJETE/C", "w"); for (int j = 0; j < N; j++) for (int i = 0; i < siz; i++) fprintf(trace_file, "%10d%10d%40.30e\n", i+1, j+1, hB[j*lda+i]); fclose(trace_file); */ magma_int_t firstprint=0; for(magma_int_t dev=0; dev < opts.ngpu; ++dev) { magma_setdevice( dev ); magma_zgetmatrix( M, N, dB[dev], ldda, hR, lda ); // compute relative error ||R||/||A||*||X||, where R := B_magma - B_lapack = R - B size = lda*N; blasf77_zaxpy( &size, &c_neg_one, hB, &ione, hR, &ione ); error = lapackf77_zlange("fro", &msize, &N, hR, &lda, work) / errorbis; //printf( "R =" ); magma_zprint( M, N, hR, lda ); if (firstprint == 0) { printf( "%5d %5d %5d %5d %7.1f (%7.4f) %7.1f (%7.4f) %7.1f (%7.4f) %8.2e %s\n", (int) M, (int) N, (int) nb, (int) offset, cpu_perf, cpu_time, gpu_perf, gpu_time, gpu_perf2, gpu_time2, error, (error < tol ? "ok" : "failed") ); } else { printf( "%89s %8.2e %s\n", " ", error, (error < tol ? "ok" : "failed") ); } status += ! (error < tol); firstprint =1; } } else { printf( "%5d %5d %5d %5d --- ( --- ) %7.1f (%7.4f) --- ( --- ) ---\n", (int) M, (int) N, (int) nb, (int) offset, gpu_perf, gpu_time ); } TESTING_FREE_CPU( hA ); TESTING_FREE_CPU( hX ); TESTING_FREE_CPU( hB ); TESTING_FREE_PIN( hR ); for( int d = 0; d < opts.ngpu; ++d ) { magma_setdevice( d ); TESTING_FREE_DEV( dA[d] ); TESTING_FREE_DEV( dX[d] ); TESTING_FREE_DEV( dB[d] ); TESTING_FREE_DEV( dwork[d] ); TESTING_FREE_PIN( hwork[d] ); } TESTING_FREE_PIN( hwork[opts.ngpu] ); if ( opts.check ) { magma_setdevice( 0 ); TESTING_FREE_DEV( dA2 ); } fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } // offset printf( "\n" ); } for( int d = 0; d < opts.ngpu; ++d ) { magma_setdevice( d ); for( magma_int_t i = 0; i < nstream; ++i ) { magma_queue_destroy( streams[d][i] ); } for( magma_int_t i = 0; i < nbevents; ++i ) { magma_event_destroy( redevents[d][i] ); magma_event_destroy( redevents2[d][i] ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zposv */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time; double error, Rnorm, Anorm, Xnorm, *work; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_R, *h_B, *h_X; magma_int_t N, lda, ldb, info, sizeA, sizeB; 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 ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf("ngpu = %d, uplo = %s\n", (int) opts.ngpu, lapack_uplo_const(opts.uplo) ); printf(" N NRHS CPU Gflop/s (sec) GPU GFlop/s (sec) ||B - AX|| / 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]; lda = ldb = N; gflops = ( FLOPS_ZPOTRF( N ) + FLOPS_ZPOTRS( N, opts.nrhs ) ) / 1e9; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*opts.nrhs ); TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*opts.nrhs ); TESTING_MALLOC_CPU( work, double, N ); /* ==================================================================== Initialize the matrix =================================================================== */ sizeA = lda*N; sizeB = ldb*opts.nrhs; lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A ); lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B ); magma_zmake_hpd( N, h_A, lda ); // copy A to R and B to X; save A and B for residual lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zposv( opts.uplo, N, opts.nrhs, h_R, lda, h_X, ldb, opts.queues2, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Residual =================================================================== */ Anorm = lapackf77_zlange("I", &N, &N, h_A, &lda, work); Xnorm = lapackf77_zlange("I", &N, &opts.nrhs, h_X, &ldb, work); blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &opts.nrhs, &N, &c_one, h_A, &lda, h_X, &ldb, &c_neg_one, h_B, &ldb ); Rnorm = lapackf77_zlange("I", &N, &opts.nrhs, h_B, &ldb, work); error = Rnorm/(N*Anorm*Xnorm); status += ! (error < tol); /* ==================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_zposv( lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, h_B, &ldb, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zposv returned error %d: %s.\n", (int) info, magma_strerror( info )); printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) N, (int) opts.nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed")); } else { printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n", (int) N, (int) opts.nrhs, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed")); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_R ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( h_X ); TESTING_FREE_CPU( work ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
extern "C" magma_int_t magma_zgeev(magma_vec_t jobvl, magma_vec_t jobvr, magma_int_t n, magmaDoubleComplex *a, magma_int_t lda, magmaDoubleComplex *geev_w_array, magmaDoubleComplex *vl, magma_int_t ldvl, magmaDoubleComplex *vr, magma_int_t ldvr, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t *info, magma_queue_t queue) { /* -- clMAGMA (version 1.0.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver September 2012 Purpose ======= ZGEEV computes for an N-by-N complex nonsymmetric matrix A, the eigenvalues and, optionally, the left and/or right eigenvectors. The right eigenvector v(j) of A satisfies A * v(j) = lambda(j) * v(j) where lambda(j) is its eigenvalue. The left eigenvector u(j) of A satisfies u(j)**H * A = lambda(j) * u(j)**H where u(j)**H denotes the conjugate transpose of u(j). The computed eigenvectors are normalized to have Euclidean norm equal to 1 and largest component real. Arguments ========= JOBVL (input) CHARACTER*1 = 'N': left eigenvectors of A are not computed; = 'V': left eigenvectors of are computed. JOBVR (input) CHARACTER*1 = 'N': right eigenvectors of A are not computed; = 'V': right eigenvectors of A are computed. N (input) INTEGER The order of the matrix A. N >= 0. A (input/output) COMPLEX*16 array, dimension (LDA,N) On entry, the N-by-N matrix A. On exit, A has been overwritten. LDA (input) INTEGER The leading dimension of the array A. LDA >= max(1,N). W (output) COMPLEX*16 array, dimension (N) W contains the computed eigenvalues. VL (output) COMPLEX*16 array, dimension (LDVL,N) If JOBVL = 'V', the left eigenvectors u(j) are stored one after another in the columns of VL, in the same order as their eigenvalues. If JOBVL = 'N', VL is not referenced. u(j) = VL(:,j), the j-th column of VL. LDVL (input) INTEGER The leading dimension of the array VL. LDVL >= 1; if JOBVL = 'V', LDVL >= N. VR (output) COMPLEX*16 array, dimension (LDVR,N) If JOBVR = 'V', the right eigenvectors v(j) are stored one after another in the columns of VR, in the same order as their eigenvalues. If JOBVR = 'N', VR is not referenced. v(j) = VR(:,j), the j-th column of VR. LDVR (input) INTEGER The leading dimension of the array VR. LDVR >= 1; if JOBVR = 'V', LDVR >= N. WORK (workspace/output) COMPLEX*16 array, dimension (MAX(1,LWORK)) On exit, if INFO = 0, WORK(1) returns the optimal LWORK. LWORK (input) INTEGER The dimension of the array WORK. LWORK >= (1+nb)*N. If LWORK = -1, then a workspace query is assumed; the routine only calculates the optimal size of the WORK array, returns this value as the first entry of the WORK array, and no error message related to LWORK is issued by XERBLA. RWORK (workspace) DOUBLE PRECISION array, dimension (2*N) INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument had an illegal value. > 0: if INFO = i, the QR algorithm failed to compute all the eigenvalues, and no eigenvectors have been computed; elements and i+1:N of W contain eigenvalues which have converged. ===================================================================== */ magma_int_t c__1 = 1; magma_int_t c__0 = 0; magma_int_t a_dim1, a_offset, vl_dim1, vl_offset, vr_dim1, vr_offset, i__1, i__2, i__3; double d__1, d__2; magmaDoubleComplex z__1, z__2; magma_int_t i__, k, ihi; double scl; magma_int_t ilo; double dum[1], eps; magmaDoubleComplex tmp; magma_int_t ibal; double anrm; magma_int_t ierr, itau, iwrk, nout; magma_int_t scalea; double cscale; magma_int_t select[1]; double bignum; magma_int_t minwrk; magma_int_t wantvl; double smlnum; magma_int_t irwork; magma_int_t lquery, wantvr; magma_int_t nb = 0; magmaDoubleComplex_ptr dT; //magma_timestr_t start, end; char side[2] = {0, 0}; magma_vec_t jobvl_ = jobvl; magma_vec_t jobvr_ = jobvr; *info = 0; lquery = lwork == -1; wantvl = lapackf77_lsame(lapack_const(jobvl_), "V"); wantvr = lapackf77_lsame(lapack_const(jobvr_), "V"); if (! wantvl && ! lapackf77_lsame(lapack_const(jobvl_), "N")) { *info = -1; } else if (! wantvr && ! lapackf77_lsame(lapack_const(jobvr_), "N")) { *info = -2; } else if (n < 0) { *info = -3; } else if (lda < max(1,n)) { *info = -5; } else if ( (ldvl < 1) || (wantvl && (ldvl < n))) { *info = -8; } else if ( (ldvr < 1) || (wantvr && (ldvr < n))) { *info = -10; } /* Compute workspace */ if (*info == 0) { nb = magma_get_zgehrd_nb(n); minwrk = (1+nb)*n; work[0] = MAGMA_Z_MAKE((double) minwrk, 0.); if (lwork < minwrk && ! lquery) { *info = -12; } } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } else if (lquery) { return *info; } /* Quick return if possible */ if (n == 0) { return *info; } // if eigenvectors are needed #if defined(VERSION3) if (MAGMA_SUCCESS != magma_malloc(&dT, nb*n*sizeof(magmaDoubleComplex) )) { *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } #endif a_dim1 = lda; a_offset = 1 + a_dim1; a -= a_offset; vl_dim1 = ldvl; vl_offset = 1 + vl_dim1; vl -= vl_offset; vr_dim1 = ldvr; vr_offset = 1 + vr_dim1; vr -= vr_offset; --work; --rwork; /* Get machine constants */ eps = lapackf77_dlamch("P"); smlnum = lapackf77_dlamch("S"); bignum = 1. / smlnum; lapackf77_dlabad(&smlnum, &bignum); smlnum = magma_dsqrt(smlnum) / eps; bignum = 1. / smlnum; /* Scale A if max element outside range [SMLNUM,BIGNUM] */ anrm = lapackf77_zlange("M", &n, &n, &a[a_offset], &lda, dum); scalea = 0; if (anrm > 0. && anrm < smlnum) { scalea = 1; cscale = smlnum; } else if (anrm > bignum) { scalea = 1; cscale = bignum; } if (scalea) { lapackf77_zlascl("G", &c__0, &c__0, &anrm, &cscale, &n, &n, &a[a_offset], &lda, & ierr); } /* Balance the matrix (CWorkspace: none) (RWorkspace: need N) */ ibal = 1; lapackf77_zgebal("B", &n, &a[a_offset], &lda, &ilo, &ihi, &rwork[ibal], &ierr); /* Reduce to upper Hessenberg form (CWorkspace: need 2*N, prefer N+N*NB) (RWorkspace: none) */ itau = 1; iwrk = itau + n; i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) /* * Version 1 - LAPACK */ lapackf77_zgehrd(&n, &ilo, &ihi, &a[a_offset], &lda, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION2) /* * Version 2 - LAPACK consistent HRD */ magma_zgehrd2(n, ilo, ihi, &a[a_offset], lda, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored, */ magma_zgehrd(n, ilo, ihi, &a[a_offset], lda, &work[itau], &work[iwrk], i__1, dT, 0, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for zgehrd = %5.2f sec\n", GetTimerValue(start,end)/1000.); if (wantvl) { /* Want left eigenvectors Copy Householder vectors to VL */ side[0] = 'L'; lapackf77_zlacpy(MagmaLowerStr, &n, &n, &a[a_offset], &lda, &vl[vl_offset], &ldvl); /* Generate unitary matrix in VL (CWorkspace: need 2*N-1, prefer N+(N-1)*NB) (RWorkspace: none) */ i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) || defined(VERSION2) /* * Version 1 & 2 - LAPACK */ lapackf77_zunghr(&n, &ilo, &ihi, &vl[vl_offset], &ldvl, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored */ magma_zunghr(n, ilo, ihi, &vl[vl_offset], ldvl, &work[itau], dT, 0, nb, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for zunghr = %5.2f sec\n", GetTimerValue(start,end)/1000.); /* Perform QR iteration, accumulating Schur vectors in VL (CWorkspace: need 1, prefer HSWORK (see comments) ) (RWorkspace: none) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_zhseqr("S", "V", &n, &ilo, &ihi, &a[a_offset], &lda, geev_w_array, &vl[vl_offset], &ldvl, &work[iwrk], &i__1, info); if (wantvr) { /* Want left and right eigenvectors Copy Schur vectors to VR */ side[0] = 'B'; lapackf77_zlacpy("F", &n, &n, &vl[vl_offset], &ldvl, &vr[vr_offset], &ldvr); } } else if (wantvr) { /* Want right eigenvectors Copy Householder vectors to VR */ side[0] = 'R'; lapackf77_zlacpy("L", &n, &n, &a[a_offset], &lda, &vr[vr_offset], &ldvr); /* Generate unitary matrix in VR (CWorkspace: need 2*N-1, prefer N+(N-1)*NB) (RWorkspace: none) */ i__1 = lwork - iwrk + 1; //start = get_current_time(); #if defined(VERSION1) || defined(VERSION2) /* * Version 1 & 2 - LAPACK */ lapackf77_zunghr(&n, &ilo, &ihi, &vr[vr_offset], &ldvr, &work[itau], &work[iwrk], &i__1, &ierr); #elif defined(VERSION3) /* * Version 3 - LAPACK consistent MAGMA HRD + matrices T stored */ magma_zunghr(n, ilo, ihi, &vr[vr_offset], ldvr, &work[itau], dT, 0, nb, &ierr, queue); #endif //end = get_current_time(); //printf(" Time for zunghr = %5.2f sec\n", GetTimerValue(start,end)/1000.); /* Perform QR iteration, accumulating Schur vectors in VR (CWorkspace: need 1, prefer HSWORK (see comments) ) (RWorkspace: none) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_zhseqr("S", "V", &n, &ilo, &ihi, &a[a_offset], &lda, geev_w_array, &vr[vr_offset], &ldvr, &work[iwrk], &i__1, info); } else { /* Compute eigenvalues only (CWorkspace: need 1, prefer HSWORK (see comments) ) (RWorkspace: none) */ iwrk = itau; i__1 = lwork - iwrk + 1; lapackf77_zhseqr("E", "N", &n, &ilo, &ihi, &a[a_offset], &lda, geev_w_array, &vr[vr_offset], &ldvr, &work[iwrk], &i__1, info); } /* If INFO > 0 from ZHSEQR, then quit */ if (*info > 0) { goto L50; } if (wantvl || wantvr) { /* Compute left and/or right eigenvectors (CWorkspace: need 2*N) (RWorkspace: need 2*N) */ irwork = ibal + n; lapackf77_ztrevc(side, "B", select, &n, &a[a_offset], &lda, &vl[vl_offset], &ldvl, &vr[vr_offset], &ldvr, &n, &nout, &work[iwrk], &rwork[irwork], &ierr); } if (wantvl) { /* Undo balancing of left eigenvectors (CWorkspace: none) (RWorkspace: need N) */ lapackf77_zgebak("B", "L", &n, &ilo, &ihi, &rwork[ibal], &n, &vl[vl_offset], &ldvl, &ierr); /* Normalize left eigenvectors and make largest component real */ for (i__ = 1; i__ <= n; ++i__) { scl = 1. / cblas_dznrm2(n, &vl[i__ * vl_dim1 + 1], 1); cblas_zdscal(n, scl, &vl[i__ * vl_dim1 + 1], 1); i__2 = n; for (k = 1; k <= i__2; ++k) { i__3 = k + i__ * vl_dim1; /* Computing 2nd power */ d__1 = MAGMA_Z_REAL(vl[i__3]); /* Computing 2nd power */ d__2 = MAGMA_Z_IMAG(vl[k + i__ * vl_dim1]); rwork[irwork + k - 1] = d__1 * d__1 + d__2 * d__2; } /* Comment: Fortran BLAS does not have to add 1 C BLAS must add one to cblas_idamax */ k = cblas_idamax(n, &rwork[irwork], 1)+1; z__2 = MAGMA_Z_CNJG(vl[k + i__ * vl_dim1]); d__1 = magma_dsqrt(rwork[irwork + k - 1]); MAGMA_Z_DSCALE(z__1, z__2, d__1); tmp = z__1; cblas_zscal(n, CBLAS_SADDR(tmp), &vl[i__ * vl_dim1 + 1], 1); i__2 = k + i__ * vl_dim1; i__3 = k + i__ * vl_dim1; d__1 = MAGMA_Z_REAL(vl[i__3]); MAGMA_Z_SET2REAL(z__1, d__1); vl[i__2] = z__1; } } if (wantvr) { /* Undo balancing of right eigenvectors (CWorkspace: none) (RWorkspace: need N) */ lapackf77_zgebak("B", "R", &n, &ilo, &ihi, &rwork[ibal], &n, &vr[vr_offset], &ldvr, &ierr); /* Normalize right eigenvectors and make largest component real */ for (i__ = 1; i__ <= n; ++i__) { scl = 1. / cblas_dznrm2(n, &vr[i__ * vr_dim1 + 1], 1); cblas_zdscal(n, scl, &vr[i__ * vr_dim1 + 1], 1); i__2 = n; for (k = 1; k <= i__2; ++k) { i__3 = k + i__ * vr_dim1; /* Computing 2nd power */ d__1 = MAGMA_Z_REAL(vr[i__3]); /* Computing 2nd power */ d__2 = MAGMA_Z_IMAG(vr[k + i__ * vr_dim1]); rwork[irwork + k - 1] = d__1 * d__1 + d__2 * d__2; } /* Comment: Fortran BLAS does not have to add 1 C BLAS must add one to cblas_idamax */ k = cblas_idamax(n, &rwork[irwork], 1)+1; z__2 = MAGMA_Z_CNJG(vr[k + i__ * vr_dim1]); d__1 = magma_dsqrt(rwork[irwork + k - 1]); MAGMA_Z_DSCALE(z__1, z__2, d__1); tmp = z__1; cblas_zscal(n, CBLAS_SADDR(tmp), &vr[i__ * vr_dim1 + 1], 1); i__2 = k + i__ * vr_dim1; i__3 = k + i__ * vr_dim1; d__1 = MAGMA_Z_REAL(vr[i__3]); MAGMA_Z_SET2REAL(z__1, d__1); vr[i__2] = z__1; } } /* Undo scaling if necessary */ L50: if (scalea) { i__1 = n - *info; /* Computing MAX */ i__3 = n - *info; i__2 = max(i__3,1); lapackf77_zlascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, geev_w_array + *info, &i__2, &ierr); if (*info > 0) { i__1 = ilo - 1; lapackf77_zlascl("G", &c__0, &c__0, &cscale, &anrm, &i__1, &c__1, geev_w_array, &n, &ierr); } } #if defined(VERSION3) magma_free( dT ); #endif return *info; } /* magma_zgeev */
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgeqlf */ int main( int argc, char** argv) { TESTING_INIT(); const double d_neg_one = MAGMA_D_NEG_ONE; const double d_one = MAGMA_D_ONE; const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; const magmaDoubleComplex c_one = MAGMA_Z_ONE; const magmaDoubleComplex c_zero = MAGMA_Z_ZERO; const magma_int_t ione = 1; real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0; double Anorm, error=0, error2=0; magmaDoubleComplex *h_A, *h_R, *tau, *h_work, tmp[1]; magma_int_t M, N, n2, lda, lwork, info, min_mn, nb; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |L - Q^H*A| |I - Q^H*Q|\n"); printf("===============================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; min_mn = min(M, N); lda = M; n2 = lda*N; nb = magma_get_zgeqlf_nb(M); gflops = FLOPS_ZGEQLF( M, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_zgeqlf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_Z_REAL( tmp[0] ); lwork = max( lwork, N*nb ); lwork = max( lwork, 2*nb*nb); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_CPU( h_work, magmaDoubleComplex, lwork ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zgeqlf( M, N, h_R, lda, tau, h_work, lwork, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zgeqlf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result, following zqlt01 except using the reduced Q. This works for any M,N (square, tall, wide). =================================================================== */ if ( opts.check ) { magma_int_t ldq = M; magma_int_t ldl = min_mn; magmaDoubleComplex *Q, *L; double *work; TESTING_MALLOC_CPU( Q, magmaDoubleComplex, ldq*min_mn ); // M by K TESTING_MALLOC_CPU( L, magmaDoubleComplex, ldl*N ); // K by N TESTING_MALLOC_CPU( work, double, min_mn ); // copy M by K matrix V to Q (copying diagonal, which isn't needed) and // copy K by N matrix L lapackf77_zlaset( "Full", &min_mn, &N, &c_zero, &c_zero, L, &ldl ); if ( M >= N ) { // for M=5, N=3: A = [ V V V ] <= V full block (M-N by K) // K=N [ V V V ] // [ ----- ] // [ L V V ] <= V triangle (N by K, copying diagonal too) // [ L L V ] <= L triangle (K by N) // [ L L L ] magma_int_t M_N = M - N; lapackf77_zlacpy( "Full", &M_N, &min_mn, h_R, &lda, Q, &ldq ); lapackf77_zlacpy( "Upper", &N, &min_mn, &h_R[M_N], &lda, &Q[M_N], &ldq ); lapackf77_zlacpy( "Lower", &min_mn, &N, &h_R[M_N], &lda, L, &ldl ); } else { // for M=3, N=5: A = [ L L | L V V ] <= V triangle (K by K) // K=M [ L L | L L V ] <= L triangle (K by M) // [ L L | L L L ] // ^^^============= L full block (K by N-M) magma_int_t N_M = N - M; lapackf77_zlacpy( "Upper", &M, &min_mn, &h_R[N_M*lda], &lda, Q, &ldq ); lapackf77_zlacpy( "Full", &min_mn, &N_M, h_R, &lda, L, &ldl ); lapackf77_zlacpy( "Lower", &min_mn, &M, &h_R[N_M*lda], &lda, &L[N_M*ldl], &ldl ); } // generate M by K matrix Q, where K = min(M,N) lapackf77_zungql( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info ); assert( info == 0 ); // error = || L - Q^H*A || / (N * ||A||) blasf77_zgemm( "Conj", "NoTrans", &min_mn, &N, &M, &c_neg_one, Q, &ldq, h_A, &lda, &c_one, L, &ldl ); Anorm = lapackf77_zlange( "1", &M, &N, h_A, &lda, work ); error = lapackf77_zlange( "1", &min_mn, &N, L, &ldl, work ); if ( N > 0 && Anorm > 0 ) error /= (N*Anorm); // set L = I (K by K identity), then L = I - Q^H*Q // error = || I - Q^H*Q || / N lapackf77_zlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, L, &ldl ); blasf77_zherk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, L, &ldl ); error2 = lapackf77_zlanhe( "1", "Upper", &min_mn, L, &ldl, work ); if ( N > 0 ) error2 /= N; TESTING_FREE_CPU( Q ); Q = NULL; TESTING_FREE_CPU( L ); L = NULL; TESTING_FREE_CPU( work ); work = NULL; } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_zgeqlf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapack_zgeqlf returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Print performance and error. =================================================================== */ printf("%5d %5d ", (int) M, (int) N ); if ( opts.lapack ) { printf( "%7.2f (%7.2f)", cpu_perf, cpu_time ); } else { printf(" --- ( --- )" ); } printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time ); if ( opts.check ) { bool okay = (error < tol && error2 < tol); status += ! okay; printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") ); } else { printf( " ---\n" ); } TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zcgeqrsv */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time, gpu_perfd, gpu_perfs; double error, gpu_error, cpu_error, Anorm, work[1]; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_A2, *h_B, *h_X, *h_R; magmaDoubleComplex_ptr d_A, d_B, d_X, d_T; magmaFloatComplex *d_SA, *d_SB; magmaDoubleComplex *h_workd, *tau, tmp[1]; magmaFloatComplex *h_works; magma_int_t lda, ldb, lhwork, lworkgpu; magma_int_t ldda, lddb, lddx; magma_int_t M, N, nrhs, qrsv_iters, info, size, min_mn, max_mn, nb; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; printf("Epsilon(double): %8.6e\n" "Epsilon(single): %8.6e\n\n", lapackf77_dlamch("Epsilon"), lapackf77_slamch("Epsilon") ); magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); nrhs = opts.nrhs; printf(" CPU Gflop/s GPU Gflop/s |b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n"); printf(" M N NRHS double double single mixed Iter CPU GPU \n"); printf("=============================================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; if ( M < N ) { printf( "%5d %5d %5d skipping because M < N is not yet supported.\n", (int) M, (int) N, (int) nrhs ); continue; } min_mn = min(M, N); max_mn = max(M, N); lda = M; ldb = max_mn; ldda = ((M+31)/32) * 32; lddb = ((max_mn+31)/32)*32; lddx = ((N+31)/32) * 32; nb = max( magma_get_zgeqrf_nb( M ), magma_get_cgeqrf_nb( M ) ); gflops = (FLOPS_ZGEQRF( M, N ) + FLOPS_ZGEQRS( M, N, nrhs )) / 1e9; lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb; // query for workspace size lhwork = -1; lapackf77_zgels( MagmaNoTransStr, &M, &N, &nrhs, NULL, &lda, NULL, &ldb, tmp, &lhwork, &info ); lhwork = (magma_int_t) MAGMA_Z_REAL( tmp[0] ); lhwork = max( lhwork, lworkgpu ); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( h_A2, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_workd, magmaDoubleComplex, lhwork ); h_works = (magmaFloatComplex*)h_workd; TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*nrhs ); TESTING_MALLOC_DEV( d_X, magmaDoubleComplex, lddx*nrhs ); TESTING_MALLOC_DEV( d_T, magmaDoubleComplex, ( 2*min_mn + (N+31)/32*32 )*nb ); /* Initialize the matrices */ size = lda*N; lapackf77_zlarnv( &ione, ISEED, &size, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda ); // make random RHS size = ldb*nrhs; lapackf77_zlarnv( &ione, ISEED, &size, h_B ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb ); magma_zsetmatrix( M, N, h_A, lda, d_A, ldda ); magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb ); //===================================================================== // Mixed Precision Iterative Refinement - GPU //===================================================================== gpu_time = magma_wtime(); magma_zcgeqrsv_gpu( M, N, nrhs, d_A, ldda, d_B, lddb, d_X, lddx, &qrsv_iters, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zcgeqrsv returned error %d: %s.\n", (int) info, magma_strerror( info )); // compute the residual magma_zgetmatrix( N, nrhs, d_X, lddx, h_X, ldb ); blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A, &lda, h_X, &ldb, &c_one, h_R, &ldb); Anorm = lapackf77_zlange("f", &M, &N, h_A, &lda, work); //===================================================================== // Double Precision Solve //===================================================================== magma_zsetmatrix( M, N, h_A, lda, d_A, ldda ); magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb ); gpu_time = magma_wtime(); magma_zgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda, d_B, lddb, h_workd, lworkgpu, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfd = gflops / gpu_time; //===================================================================== // Single Precision Solve //===================================================================== magma_zsetmatrix( M, N, h_A, lda, d_A, ldda ); magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb ); /* The allocation of d_SA and d_SB is done here to avoid * to double the memory used on GPU with zcgeqrsv */ TESTING_MALLOC_DEV( d_SA, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( d_SB, magmaFloatComplex, lddb*nrhs ); magmablas_zlag2c( M, N, d_A, ldda, d_SA, ldda, &info ); magmablas_zlag2c( N, nrhs, d_B, lddb, d_SB, lddb, &info ); gpu_time = magma_wtime(); magma_cgels_gpu( MagmaNoTrans, M, N, nrhs, d_SA, ldda, d_SB, lddb, h_works, lhwork, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfs = gflops / gpu_time; TESTING_FREE_DEV( d_SA ); TESTING_FREE_DEV( d_SB ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb ); cpu_time = magma_wtime(); lapackf77_zgels( MagmaNoTransStr, &M, &N, &nrhs, h_A, &lda, h_X, &ldb, h_workd, &lhwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zgels returned error %d: %s.\n", (int) info, magma_strerror( info )); blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A2, &lda, h_X, &ldb, &c_one, h_B, &ldb ); cpu_error = lapackf77_zlange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm); gpu_error = lapackf77_zlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm); // error relative to LAPACK size = M*nrhs; blasf77_zaxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione ); error = lapackf77_zlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm); printf("%5d %5d %5d %7.2f %7.2f %7.2f %7.2f %4d %8.2e %8.2e %8.2e %s\n", (int) M, (int) N, (int) nrhs, cpu_perf, gpu_perfd, gpu_perfs, gpu_perf, (int) qrsv_iters, cpu_error, gpu_error, error, (error < tol ? "ok" : "failed")); status += ! (error < tol); TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_A2 ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( h_X ); TESTING_FREE_CPU( h_R ); TESTING_FREE_CPU( h_workd ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); TESTING_FREE_DEV( d_X ); TESTING_FREE_DEV( d_T ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zunmqr */ int main( int argc, char** argv ) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t ione = 1; magma_int_t mm, m, n, k, size, info; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t nb, ldc, lda, lwork, lwork_max; magmaDoubleComplex *C, *R, *A, *W, *tau; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); // need slightly looser bound (60*eps instead of 30*eps) for some tests opts.tolerance = max( 60., opts.tolerance ); double tol = opts.tolerance * lapackf77_dlamch("E"); // test all combinations of input parameters magma_side_t side [] = { MagmaLeft, MagmaRight }; magma_trans_t trans[] = { MagmaConjTrans, MagmaNoTrans }; printf(" M N K side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n"); printf("===============================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iside = 0; iside < 2; ++iside ) { for( int itran = 0; itran < 2; ++itran ) { for( int iter = 0; iter < opts.niter; ++iter ) { m = opts.msize[itest]; n = opts.nsize[itest]; k = opts.ksize[itest]; nb = magma_get_zgeqrf_nb( m ); ldc = m; // A is m x k (left) or n x k (right) mm = (side[iside] == MagmaLeft ? m : n); lda = mm; gflops = FLOPS_ZUNMQR( m, n, k, side[iside] ) / 1e9; if ( side[iside] == MagmaLeft && m < k ) { printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n", (int) m, (int) n, (int) k, lapacke_side_const( side[iside] ), lapacke_trans_const( trans[itran] ) ); continue; } if ( side[iside] == MagmaRight && n < k ) { printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n", (int) m, (int) n, (int) k, lapacke_side_const( side[iside] ), lapacke_trans_const( trans[itran] ) ); continue; } // need at least 2*nb*nb for geqrf lwork_max = max( max( m*nb, n*nb ), 2*nb*nb ); TESTING_MALLOC_CPU( C, magmaDoubleComplex, ldc*n ); TESTING_MALLOC_CPU( R, magmaDoubleComplex, ldc*n ); TESTING_MALLOC_CPU( A, magmaDoubleComplex, lda*k ); TESTING_MALLOC_CPU( W, magmaDoubleComplex, lwork_max ); TESTING_MALLOC_CPU( tau, magmaDoubleComplex, k ); // C is full, m x n size = ldc*n; lapackf77_zlarnv( &ione, ISEED, &size, C ); lapackf77_zlacpy( "Full", &m, &n, C, &ldc, R, &ldc ); size = lda*k; lapackf77_zlarnv( &ione, ISEED, &size, A ); // compute QR factorization to get Householder vectors in A, tau magma_zgeqrf( mm, k, A, lda, tau, W, lwork_max, &info ); if (info != 0) printf("magma_zgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_zunmqr( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ), &m, &n, &k, A, &lda, tau, C, &ldc, W, &lwork_max, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zunmqr returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ==================================================================== Performs operation using MAGMA =================================================================== */ // query for workspace size lwork = -1; magma_zunmqr( side[iside], trans[itran], m, n, k, A, lda, tau, R, ldc, W, lwork, &info ); if (info != 0) printf("magma_zunmqr (lwork query) returned error %d: %s.\n", (int) info, magma_strerror( info )); lwork = (magma_int_t) MAGMA_Z_REAL( W[0] ); if ( lwork < 0 || lwork > lwork_max ) { printf("optimal lwork %d > lwork_max %d\n", (int) lwork, (int) lwork_max ); lwork = lwork_max; } gpu_time = magma_wtime(); magma_zunmqr( side[iside], trans[itran], m, n, k, A, lda, tau, R, ldc, W, lwork, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zunmqr returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== compute relative error |QC_magma - QC_lapack| / |QC_lapack| =================================================================== */ error = lapackf77_zlange( "Fro", &m, &n, C, &ldc, work ); size = ldc*n; blasf77_zaxpy( &size, &c_neg_one, C, &ione, R, &ione ); error = lapackf77_zlange( "Fro", &m, &n, R, &ldc, work ) / error; printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) m, (int) n, (int) k, lapacke_side_const( side[iside] ), lapacke_trans_const( trans[itran] ), cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed") ); status += ! (error < tol); TESTING_FREE_CPU( C ); TESTING_FREE_CPU( R ); TESTING_FREE_CPU( A ); TESTING_FREE_CPU( W ); TESTING_FREE_CPU( tau ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } }} // end iside, itran printf( "\n" ); } TESTING_FINALIZE(); return status; }
magma_int_t magma_ztrevc3_mt( magma_side_t side, magma_vec_t howmany, magma_int_t *select, // logical in Fortran magma_int_t n, magmaDoubleComplex *T, magma_int_t ldt, magmaDoubleComplex *VL, magma_int_t ldvl, magmaDoubleComplex *VR, magma_int_t ldvr, magma_int_t mm, magma_int_t *mout, magmaDoubleComplex *work, magma_int_t lwork, #ifdef COMPLEX double *rwork, #endif magma_int_t *info ) { #define T(i,j) ( T + (i) + (j)*ldt ) #define VL(i,j) (VL + (i) + (j)*ldvl) #define VR(i,j) (VR + (i) + (j)*ldvr) #define work(i,j) (work + (i) + (j)*n) // .. Parameters .. const magmaDoubleComplex c_zero = MAGMA_Z_ZERO; const magmaDoubleComplex c_one = MAGMA_Z_ONE; const magma_int_t nbmin = 16, nbmax = 128; const magma_int_t ione = 1; // .. Local Scalars .. magma_int_t allv, bothv, leftv, over, rightv, somev; magma_int_t i, ii, is, j, k, ki, iv, n2, nb, nb2, version; double ovfl, remax, unfl; //smlnum, smin, ulp // Decode and test the input parameters bothv = (side == MagmaBothSides); rightv = (side == MagmaRight) || bothv; leftv = (side == MagmaLeft ) || bothv; allv = (howmany == MagmaAllVec); over = (howmany == MagmaBacktransVec); somev = (howmany == MagmaSomeVec); // Set mout to the number of columns required to store the selected // eigenvectors. if ( somev ) { *mout = 0; for( j=0; j < n; ++j ) { if ( select[j] ) { *mout += 1; } } } else { *mout = n; } *info = 0; if ( ! rightv && ! leftv ) *info = -1; else if ( ! allv && ! over && ! somev ) *info = -2; else if ( n < 0 ) *info = -4; else if ( ldt < max( 1, n ) ) *info = -6; else if ( ldvl < 1 || ( leftv && ldvl < n ) ) *info = -8; else if ( ldvr < 1 || ( rightv && ldvr < n ) ) *info = -10; else if ( mm < *mout ) *info = -11; else if ( lwork < max( 1, 2*n ) ) *info = -14; if ( *info != 0 ) { magma_xerbla( __func__, -(*info) ); return *info; } // Quick return if possible. if ( n == 0 ) { return *info; } // Use blocked version (2) if sufficient workspace. // Requires 1 vector to save diagonal elements, and 2*nb vectors for x and Q*x. // (Compared to dtrevc3, rwork stores 1-norms.) // Zero-out the workspace to avoid potential NaN propagation. nb = 2; if ( lwork >= n + 2*n*nbmin ) { version = 2; nb = (lwork - n) / (2*n); nb = min( nb, nbmax ); nb2 = 1 + 2*nb; lapackf77_zlaset( "F", &n, &nb2, &c_zero, &c_zero, work, &n ); } else { version = 1; } // Set the constants to control overflow. unfl = lapackf77_dlamch( "Safe minimum" ); ovfl = 1. / unfl; lapackf77_dlabad( &unfl, &ovfl ); //ulp = lapackf77_dlamch( "Precision" ); //smlnum = unfl*( n / ulp ); // Store the diagonal elements of T in working array work. for( i=0; i < n; ++i ) { *work(i,0) = *T(i,i); } // Compute 1-norm of each column of strictly upper triangular // part of T to control overflow in triangular solver. rwork[0] = 0.; for( j=1; j < n; ++j ) { rwork[j] = magma_cblas_dzasum( j, T(0,j), ione ); } // launch threads -- each single-threaded MKL magma_int_t nthread = magma_get_parallel_numthreads(); magma_int_t lapack_nthread = magma_get_lapack_numthreads(); magma_set_lapack_numthreads( 1 ); magma_thread_queue queue; queue.launch( nthread ); //printf( "nthread %d, %d\n", nthread, lapack_nthread ); // gemm_nb = N/thread, rounded up to multiple of 16, // but avoid multiples of page size, e.g., 512*8 bytes = 4096. magma_int_t gemm_nb = magma_int_t( ceil( ceil( ((double)n) / nthread ) / 16. ) * 16. ); if ( gemm_nb % 512 == 0 ) { gemm_nb += 32; } magma_timer_t time_total=0, time_trsv=0, time_gemm=0, time_gemv=0, time_trsv_sum=0, time_gemm_sum=0, time_gemv_sum=0; timer_start( time_total ); if ( rightv ) { // ============================================================ // Compute right eigenvectors. // iv is index of column in current block. // Non-blocked version always uses iv=1; // blocked version starts with iv=nb, goes down to 1. // (Note the "0-th" column is used to store the original diagonal.) iv = 1; if ( version == 2 ) { iv = nb; } timer_start( time_trsv ); is = *mout - 1; for( ki=n-1; ki >= 0; --ki ) { if ( somev ) { if ( ! select[ki] ) { continue; } } //smin = max( ulp*MAGMA_Z_ABS1( *T(ki,ki) ), smlnum ); // -------------------------------------------------------- // Complex right eigenvector *work(ki,iv) = c_one; // Form right-hand side. for( k=0; k < ki; ++k ) { *work(k,iv) = -(*T(k,ki)); } // Solve upper triangular system: // [ T(1:ki-1,1:ki-1) - T(ki,ki) ]*X = scale*work. if ( ki > 0 ) { queue.push_task( new magma_zlatrsd_task( MagmaUpper, MagmaNoTrans, MagmaNonUnit, MagmaTrue, ki, T, ldt, *T(ki,ki), work(0,iv), work(ki,iv), rwork )); } // Copy the vector x or Q*x to VR and normalize. if ( ! over ) { // ------------------------------ // no back-transform: copy x to VR and normalize queue.sync(); n2 = ki+1; blasf77_zcopy( &n2, work(0,iv), &ione, VR(0,is), &ione ); ii = blasf77_izamax( &n2, VR(0,is), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VR(ii,is) ); blasf77_zdscal( &n2, &remax, VR(0,is), &ione ); for( k=ki+1; k < n; ++k ) { *VR(k,is) = c_zero; } } else if ( version == 1 ) { // ------------------------------ // version 1: back-transform each vector with GEMV, Q*x. queue.sync(); time_trsv_sum += timer_stop( time_trsv ); timer_start( time_gemv ); if ( ki > 0 ) { blasf77_zgemv( "n", &n, &ki, &c_one, VR, &ldvr, work(0, iv), &ione, work(ki,iv), VR(0,ki), &ione ); } time_gemv_sum += timer_stop( time_gemv ); ii = blasf77_izamax( &n, VR(0,ki), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VR(ii,ki) ); blasf77_zdscal( &n, &remax, VR(0,ki), &ione ); timer_start( time_trsv ); } else if ( version == 2 ) { // ------------------------------ // version 2: back-transform block of vectors with GEMM // zero out below vector for( k=ki+1; k < n; ++k ) { *work(k,iv) = c_zero; } // Columns iv:nb of work are valid vectors. // When the number of vectors stored reaches nb, // or if this was last vector, do the GEMM if ( (iv == 1) || (ki == 0) ) { queue.sync(); time_trsv_sum += timer_stop( time_trsv ); timer_start( time_gemm ); nb2 = nb-iv+1; n2 = ki+nb-iv+1; // split gemm into multiple tasks, each doing one block row for( i=0; i < n; i += gemm_nb ) { magma_int_t ib = min( gemm_nb, n-i ); queue.push_task( new zgemm_task( MagmaNoTrans, MagmaNoTrans, ib, nb2, n2, c_one, VR(i,0), ldvr, work(0,iv ), n, c_zero, work(i,nb+iv), n )); } queue.sync(); time_gemm_sum += timer_stop( time_gemm ); // normalize vectors // TODO if somev, should copy vectors individually to correct location. for( k = iv; k <= nb; ++k ) { ii = blasf77_izamax( &n, work(0,nb+k), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *work(ii,nb+k) ); blasf77_zdscal( &n, &remax, work(0,nb+k), &ione ); } lapackf77_zlacpy( "F", &n, &nb2, work(0,nb+iv), &n, VR(0,ki), &ldvr ); iv = nb; timer_start( time_trsv ); } else { iv -= 1; } } // blocked back-transform is -= 1; } } timer_stop( time_trsv ); timer_stop( time_total ); timer_printf( "trevc trsv %.4f, gemm %.4f, gemv %.4f, total %.4f\n", time_trsv_sum, time_gemm_sum, time_gemv_sum, time_total ); if ( leftv ) { // ============================================================ // Compute left eigenvectors. // iv is index of column in current block. // Non-blocked version always uses iv=1; // blocked version starts with iv=1, goes up to nb. // (Note the "0-th" column is used to store the original diagonal.) iv = 1; is = 0; for( ki=0; ki < n; ++ki ) { if ( somev ) { if ( ! select[ki] ) { continue; } } //smin = max( ulp*MAGMA_Z_ABS1( *T(ki,ki) ), smlnum ); // -------------------------------------------------------- // Complex left eigenvector *work(ki,iv) = c_one; // Form right-hand side. for( k = ki + 1; k < n; ++k ) { *work(k,iv) = -MAGMA_Z_CONJ( *T(ki,k) ); } // Solve conjugate-transposed triangular system: // [ T(ki+1:n,ki+1:n) - T(ki,ki) ]**H * X = scale*work. // TODO what happens with T(k,k) - lambda is small? Used to have < smin test. if ( ki < n-1 ) { n2 = n-ki-1; queue.push_task( new magma_zlatrsd_task( MagmaUpper, MagmaConjTrans, MagmaNonUnit, MagmaTrue, n2, T(ki+1,ki+1), ldt, *T(ki,ki), work(ki+1,iv), work(ki,iv), rwork )); } // Copy the vector x or Q*x to VL and normalize. if ( ! over ) { // ------------------------------ // no back-transform: copy x to VL and normalize queue.sync(); n2 = n-ki; blasf77_zcopy( &n2, work(ki,iv), &ione, VL(ki,is), &ione ); ii = blasf77_izamax( &n2, VL(ki,is), &ione ) + ki - 1; remax = 1. / MAGMA_Z_ABS1( *VL(ii,is) ); blasf77_zdscal( &n2, &remax, VL(ki,is), &ione ); for( k=0; k < ki; ++k ) { *VL(k,is) = c_zero; } } else if ( version == 1 ) { // ------------------------------ // version 1: back-transform each vector with GEMV, Q*x. queue.sync(); if ( ki < n-1 ) { n2 = n-ki-1; blasf77_zgemv( "n", &n, &n2, &c_one, VL(0,ki+1), &ldvl, work(ki+1,iv), &ione, work(ki, iv), VL(0,ki), &ione ); } ii = blasf77_izamax( &n, VL(0,ki), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VL(ii,ki) ); blasf77_zdscal( &n, &remax, VL(0,ki), &ione ); } else if ( version == 2 ) { // ------------------------------ // version 2: back-transform block of vectors with GEMM // zero out above vector // could go from (ki+1)-NV+1 to ki for( k=0; k < ki; ++k ) { *work(k,iv) = c_zero; } // Columns 1:iv of work are valid vectors. // When the number of vectors stored reaches nb, // or if this was last vector, do the GEMM if ( (iv == nb) || (ki == n-1) ) { queue.sync(); n2 = n-(ki+1)+iv; // split gemm into multiple tasks, each doing one block row for( i=0; i < n; i += gemm_nb ) { magma_int_t ib = min( gemm_nb, n-i ); queue.push_task( new zgemm_task( MagmaNoTrans, MagmaNoTrans, ib, iv, n2, c_one, VL(i,ki-iv+1), ldvl, work(ki-iv+1,1), n, c_zero, work(i,nb+1), n )); } queue.sync(); // normalize vectors for( k=1; k <= iv; ++k ) { ii = blasf77_izamax( &n, work(0,nb+k), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *work(ii,nb+k) ); blasf77_zdscal( &n, &remax, work(0,nb+k), &ione ); } lapackf77_zlacpy( "F", &n, &iv, work(0,nb+1), &n, VL(0,ki-iv+1), &ldvl ); iv = 1; } else { iv += 1; } } // blocked back-transform is += 1; } } // close down threads queue.quit(); magma_set_lapack_numthreads( lapack_nthread ); return *info; } // End of ZTREVC
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgeqrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double error, work[1]; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_R, *tau, *dtau, *h_work, tmp[1]; magmaDoubleComplex *d_A; double *dwork; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R||_F / ||A||_F\n"); printf("=======================================================================\n"); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[i]; N = opts.nsize[i]; min_mn = min(M, N); lda = M; n2 = lda*N; ldda = ((M+31)/32)*32; gflops = FLOPS_ZGEQRF( M, N ) / 1e9; lwork = -1; lapackf77_zgeqrf(&M, &N, h_A, &M, tau, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_Z_REAL( tmp[0] ); TESTING_MALLOC( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC( h_A, magmaDoubleComplex, n2 ); TESTING_HOSTALLOC( h_R, magmaDoubleComplex, n2 ); TESTING_DEVALLOC( d_A, magmaDoubleComplex, ldda*N ); TESTING_DEVALLOC( dtau, magmaDoubleComplex, min_mn ); TESTING_DEVALLOC(dwork, double, min_mn ); TESTING_MALLOC( h_work, magmaDoubleComplex, lwork ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_zsetmatrix( M, N, h_R, lda, d_A, ldda ); // warmup magma_zgeqr2_gpu( M, N, d_A, ldda, dtau, dwork, &info ); magma_zsetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( 0 ); magma_zgeqr2_gpu( M, N, d_A, ldda, dtau, dwork, &info ); gpu_time = magma_sync_wtime( 0 ) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_zgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_zgetmatrix( M, N, d_A, ldda, h_R, M ); error = lapackf77_zlange("f", &M, &N, h_A, &lda, work); blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_zlange("f", &M, &N, h_R, &lda, work) / error; printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e\n", (int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, error ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, 1000.*gpu_time ); } TESTING_FREE( tau ); TESTING_FREE( h_A ); TESTING_FREE( h_work ); TESTING_HOSTFREE( h_R ); TESTING_DEVFREE( d_A ); TESTING_DEVFREE( dtau ); TESTING_DEVFREE( dwork ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return 0; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zgeqrs */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; double gpu_error, cpu_error, matnorm, work[1]; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1]; magmaDoubleComplex *d_A, *d_B; magma_int_t M, N, n2, nrhs, lda, ldb, ldda, lddb, min_mn, max_mn, nb, info; magma_int_t lworkgpu, lhwork, lhwork2; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_opts opts; parse_opts( argc, argv, &opts ); magma_int_t status = 0; double tol = opts.tolerance * lapackf77_dlamch("E"); nrhs = opts.nrhs; printf(" ||b-Ax|| / (N||A||)\n"); printf(" M N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) CPU GPU \n"); printf("===============================================================================\n"); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[i]; N = opts.nsize[i]; if ( M < N ) { printf( "skipping M=%d, N=%d because M < N is not yet supported.\n", (int) M, (int) N ); continue; } min_mn = min(M, N); max_mn = max(M, N); lda = M; ldb = max_mn; n2 = lda*N; ldda = ((M+31)/32)*32; lddb = ((max_mn+31)/32)*32; nb = magma_get_zgeqrf_nb(M); gflops = (FLOPS_ZGEQRF( M, N ) + FLOPS_ZGEQRS( M, N, nrhs )) / 1e9; // query for workspace size lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb; lhwork = -1; lapackf77_zgeqrf(&M, &N, h_A, &M, tau, tmp, &lhwork, &info); lhwork2 = (magma_int_t) MAGMA_Z_REAL( tmp[0] ); lhwork = -1; lapackf77_zunmqr( MagmaLeftStr, MagmaConjTransStr, &M, &nrhs, &min_mn, h_A, &lda, tau, h_X, &ldb, tmp, &lhwork, &info); lhwork = (magma_int_t) MAGMA_Z_REAL( tmp[0] ); lhwork = max( max( lhwork, lhwork2 ), lworkgpu ); TESTING_MALLOC( tau, magmaDoubleComplex, min_mn ); TESTING_MALLOC( h_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC( h_A2, magmaDoubleComplex, lda*N ); TESTING_MALLOC( h_B, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC( h_X, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC( h_R, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC( h_work, magmaDoubleComplex, lhwork ); TESTING_DEVALLOC( d_A, magmaDoubleComplex, ldda*N ); TESTING_DEVALLOC( d_B, magmaDoubleComplex, lddb*nrhs ); /* Initialize the matrices */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda ); // make random RHS n2 = M*nrhs; lapackf77_zlarnv( &ione, ISEED, &n2, h_B ); lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb ); // make consistent RHS //n2 = N*nrhs; //lapackf77_zlarnv( &ione, ISEED, &n2, h_X ); //blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, // &c_one, h_A, &lda, // h_X, &ldb, // &c_zero, h_B, &ldb ); //lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_zsetmatrix( M, N, h_A, lda, d_A, ldda ); magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb ); gpu_time = magma_wtime(); magma_zgels3_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda, d_B, lddb, h_work, lworkgpu, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zgels returned error %d: %s.\n", (int) info, magma_strerror( info )); // Get the solution in h_X magma_zgetmatrix( N, nrhs, d_B, lddb, h_X, ldb ); // compute the residual blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A, &lda, h_X, &ldb, &c_one, h_R, &ldb); matnorm = lapackf77_zlange("f", &M, &N, h_A, &lda, work); /* ===================================================================== Performs operation using LAPACK =================================================================== */ lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb ); cpu_time = magma_wtime(); lapackf77_zgels( MagmaNoTransStr, &M, &N, &nrhs, h_A, &lda, h_X, &ldb, h_work, &lhwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zgels returned error %d: %s.\n", (int) info, magma_strerror( info )); blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A2, &lda, h_X, &ldb, &c_one, h_B, &ldb); cpu_error = lapackf77_zlange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*matnorm); gpu_error = lapackf77_zlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*matnorm); printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e", (int) M, (int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time, cpu_error, gpu_error ); printf("%s\n", (gpu_error < tol ? "" : " failed")); status |= ! (gpu_error < tol); TESTING_FREE( tau ); TESTING_FREE( h_A ); TESTING_FREE( h_A2 ); TESTING_FREE( h_B ); TESTING_FREE( h_X ); TESTING_FREE( h_R ); TESTING_FREE( h_work ); TESTING_DEVFREE( d_A ); TESTING_DEVFREE( d_B ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
magma_int_t magma_ztrevc3( magma_side_t side, magma_vec_t howmany, magma_int_t *select, // logical in Fortran magma_int_t n, magmaDoubleComplex *T, magma_int_t ldt, magmaDoubleComplex *VL, magma_int_t ldvl, magmaDoubleComplex *VR, magma_int_t ldvr, magma_int_t mm, magma_int_t *mout, magmaDoubleComplex *work, magma_int_t lwork, double *rwork, magma_int_t *info ) { #define T(i,j) ( T + (i) + (j)*ldt ) #define VL(i,j) (VL + (i) + (j)*ldvl) #define VR(i,j) (VR + (i) + (j)*ldvr) #define work(i,j) (work + (i) + (j)*n) // .. Parameters .. const magmaDoubleComplex c_zero = MAGMA_Z_ZERO; const magmaDoubleComplex c_one = MAGMA_Z_ONE; const magma_int_t nbmin = 16, nbmax = 128; const magma_int_t ione = 1; // .. Local Scalars .. magma_int_t allv, bothv, leftv, over, rightv, somev; magma_int_t i, ii, is, j, k, ki, iv, n2, nb, nb2, version; double ovfl, remax, scale, smin, smlnum, ulp, unfl; // Decode and test the input parameters bothv = (side == MagmaBothSides); rightv = (side == MagmaRight) || bothv; leftv = (side == MagmaLeft ) || bothv; allv = (howmany == MagmaAllVec); over = (howmany == MagmaBacktransVec); somev = (howmany == MagmaSomeVec); // Set mout to the number of columns required to store the selected // eigenvectors. if ( somev ) { *mout = 0; for( j=0; j < n; ++j ) { if ( select[j] ) { *mout += 1; } } } else { *mout = n; } *info = 0; if ( ! rightv && ! leftv ) *info = -1; else if ( ! allv && ! over && ! somev ) *info = -2; else if ( n < 0 ) *info = -4; else if ( ldt < max( 1, n ) ) *info = -6; else if ( ldvl < 1 || ( leftv && ldvl < n ) ) *info = -8; else if ( ldvr < 1 || ( rightv && ldvr < n ) ) *info = -10; else if ( mm < *mout ) *info = -11; else if ( lwork < max( 1, 2*n ) ) *info = -14; if ( *info != 0 ) { magma_xerbla( __func__, -(*info) ); return *info; } // Quick return if possible. if ( n == 0 ) { return *info; } // Use blocked version (2) if sufficient workspace. // Requires 1 vector to save diagonal elements, and 2*nb vectors for x and Q*x. // (Compared to dtrevc3, rwork stores 1-norms.) // Zero-out the workspace to avoid potential NaN propagation. nb = 2; if ( lwork >= n + 2*n*nbmin ) { version = 2; nb = (lwork - n) / (2*n); nb = min( nb, nbmax ); nb2 = 1 + 2*nb; lapackf77_zlaset( "F", &n, &nb2, &c_zero, &c_zero, work, &n ); } else { version = 1; } // Set the constants to control overflow. unfl = lapackf77_dlamch( "Safe minimum" ); ovfl = 1. / unfl; lapackf77_dlabad( &unfl, &ovfl ); ulp = lapackf77_dlamch( "Precision" ); smlnum = unfl*( n / ulp ); // Store the diagonal elements of T in working array work. for( i=0; i < n; ++i ) { *work(i,0) = *T(i,i); } // Compute 1-norm of each column of strictly upper triangular // part of T to control overflow in triangular solver. rwork[0] = 0.; for( j=1; j < n; ++j ) { rwork[j] = cblas_dzasum( j, T(0,j), ione ); } magma_timer_t time_total=0, time_trsv=0, time_gemm=0, time_gemv=0, time_trsv_sum=0, time_gemm_sum=0, time_gemv_sum=0; timer_start( time_total ); if ( rightv ) { // ============================================================ // Compute right eigenvectors. // iv is index of column in current block. // Non-blocked version always uses iv=1; // blocked version starts with iv=nb, goes down to 1. // (Note the "0-th" column is used to store the original diagonal.) iv = 1; if ( version == 2 ) { iv = nb; } timer_start( time_trsv ); is = *mout - 1; for( ki=n-1; ki >= 0; --ki ) { if ( somev ) { if ( ! select[ki] ) { continue; } } smin = max( ulp*( MAGMA_Z_ABS1( *T(ki,ki) ) ), smlnum ); // -------------------------------------------------------- // Complex right eigenvector *work(ki,iv) = c_one; // Form right-hand side. for( k=0; k < ki; ++k ) { *work(k,iv) = -(*T(k,ki)); } // Solve upper triangular system: // [ T(1:ki-1,1:ki-1) - T(ki,ki) ]*X = scale*work. for( k=0; k < ki; ++k ) { *T(k,k) -= *T(ki,ki); if ( MAGMA_Z_ABS1( *T(k,k) ) < smin ) { *T(k,k) = MAGMA_Z_MAKE( smin, 0. ); } } if ( ki > 0 ) { lapackf77_zlatrs( "Upper", "No transpose", "Non-unit", "Y", &ki, T, &ldt, work(0,iv), &scale, rwork, info ); *work(ki,iv) = MAGMA_Z_MAKE( scale, 0. ); } // Copy the vector x or Q*x to VR and normalize. if ( ! over ) { // ------------------------------ // no back-transform: copy x to VR and normalize n2 = ki+1; blasf77_zcopy( &n2, work(0,iv), &ione, VR(0,is), &ione ); ii = blasf77_izamax( &n2, VR(0,is), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VR(ii,is) ); blasf77_zdscal( &n2, &remax, VR(0,is), &ione ); for( k=ki+1; k < n; ++k ) { *VR(k,is) = c_zero; } } else if ( version == 1 ) { // ------------------------------ // version 1: back-transform each vector with GEMV, Q*x. time_trsv_sum += timer_stop( time_trsv ); timer_start( time_gemv ); if ( ki > 0 ) { blasf77_zgemv( "n", &n, &ki, &c_one, VR, &ldvr, work(0, iv), &ione, work(ki,iv), VR(0,ki), &ione ); } time_gemv_sum += timer_stop( time_gemv ); ii = blasf77_izamax( &n, VR(0,ki), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VR(ii,ki) ); blasf77_zdscal( &n, &remax, VR(0,ki), &ione ); timer_start( time_trsv ); } else if ( version == 2 ) { // ------------------------------ // version 2: back-transform block of vectors with GEMM // zero out below vector for( k=ki+1; k < n; ++k ) { *work(k,iv) = c_zero; } // Columns iv:nb of work are valid vectors. // When the number of vectors stored reaches nb, // or if this was last vector, do the GEMM if ( (iv == 1) || (ki == 0) ) { time_trsv_sum += timer_stop( time_trsv ); timer_start( time_gemm ); nb2 = nb-iv+1; n2 = ki+nb-iv+1; blasf77_zgemm( "n", "n", &n, &nb2, &n2, &c_one, VR, &ldvr, work(0,iv ), &n, &c_zero, work(0,nb+iv), &n ); time_gemm_sum += timer_stop( time_gemm ); // normalize vectors // TODO if somev, should copy vectors individually to correct location. for( k = iv; k <= nb; ++k ) { ii = blasf77_izamax( &n, work(0,nb+k), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *work(ii,nb+k) ); blasf77_zdscal( &n, &remax, work(0,nb+k), &ione ); } lapackf77_zlacpy( "F", &n, &nb2, work(0,nb+iv), &n, VR(0,ki), &ldvr ); iv = nb; timer_start( time_trsv ); } else { iv -= 1; } } // blocked back-transform // Restore the original diagonal elements of T. for( k=0; k <= ki - 1; ++k ) { *T(k,k) = *work(k,0); } is -= 1; } } timer_stop( time_trsv ); timer_stop( time_total ); timer_printf( "trevc trsv %.4f, gemm %.4f, gemv %.4f, total %.4f\n", time_trsv_sum, time_gemm_sum, time_gemv_sum, time_total ); if ( leftv ) { // ============================================================ // Compute left eigenvectors. // iv is index of column in current block. // Non-blocked version always uses iv=1; // blocked version starts with iv=1, goes up to nb. // (Note the "0-th" column is used to store the original diagonal.) iv = 1; is = 0; for( ki=0; ki < n; ++ki ) { if ( somev ) { if ( ! select[ki] ) { continue; } } smin = max( ulp*MAGMA_Z_ABS1( *T(ki,ki) ), smlnum ); // -------------------------------------------------------- // Complex left eigenvector *work(ki,iv) = c_one; // Form right-hand side. for( k = ki + 1; k < n; ++k ) { *work(k,iv) = -MAGMA_Z_CNJG( *T(ki,k) ); } // Solve conjugate-transposed triangular system: // [ T(ki+1:n,ki+1:n) - T(ki,ki) ]**H * X = scale*work. for( k = ki + 1; k < n; ++k ) { *T(k,k) -= *T(ki,ki); if ( MAGMA_Z_ABS1( *T(k,k) ) < smin ) { *T(k,k) = MAGMA_Z_MAKE( smin, 0. ); } } if ( ki < n-1 ) { n2 = n-ki-1; lapackf77_zlatrs( "Upper", "Conjugate transpose", "Non-unit", "Y", &n2, T(ki+1,ki+1), &ldt, work(ki+1,iv), &scale, rwork, info ); *work(ki,iv) = MAGMA_Z_MAKE( scale, 0. ); } // Copy the vector x or Q*x to VL and normalize. if ( ! over ) { // ------------------------------ // no back-transform: copy x to VL and normalize n2 = n-ki; blasf77_zcopy( &n2, work(ki,iv), &ione, VL(ki,is), &ione ); ii = blasf77_izamax( &n2, VL(ki,is), &ione ) + ki - 1; remax = 1. / MAGMA_Z_ABS1( *VL(ii,is) ); blasf77_zdscal( &n2, &remax, VL(ki,is), &ione ); for( k=0; k < ki; ++k ) { *VL(k,is) = c_zero; } } else if ( version == 1 ) { // ------------------------------ // version 1: back-transform each vector with GEMV, Q*x. if ( ki < n-1 ) { n2 = n-ki-1; blasf77_zgemv( "n", &n, &n2, &c_one, VL(0,ki+1), &ldvl, work(ki+1,iv), &ione, work(ki, iv), VL(0,ki), &ione ); } ii = blasf77_izamax( &n, VL(0,ki), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *VL(ii,ki) ); blasf77_zdscal( &n, &remax, VL(0,ki), &ione ); } else if ( version == 2 ) { // ------------------------------ // version 2: back-transform block of vectors with GEMM // zero out above vector // could go from (ki+1)-NV+1 to ki for( k=0; k < ki; ++k ) { *work(k,iv) = c_zero; } // Columns 1:iv of work are valid vectors. // When the number of vectors stored reaches nb, // or if this was last vector, do the GEMM if ( (iv == nb) || (ki == n-1) ) { n2 = n-(ki+1)+iv; blasf77_zgemm( "n", "n", &n, &iv, &n2, &c_one, VL(0,ki-iv+1), &ldvl, work(ki-iv+1,1 ), &n, &c_zero, work(0, nb+1), &n ); // normalize vectors for( k=1; k <= iv; ++k ) { ii = blasf77_izamax( &n, work(0,nb+k), &ione ) - 1; remax = 1. / MAGMA_Z_ABS1( *work(ii,nb+k) ); blasf77_zdscal( &n, &remax, work(0,nb+k), &ione ); } lapackf77_zlacpy( "F", &n, &iv, work(0,nb+1), &n, VL(0,ki-iv+1), &ldvl ); iv = 1; } else { iv += 1; } } // blocked back-transform // Restore the original diagonal elements of T. for( k = ki + 1; k < n; ++k ) { *T(k,k) = *work(k,0); } is += 1; } } return *info; } // End of ZTREVC
/* //////////////////////////////////////////////////////////////////////////// -- Testing zpotrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; magmaDoubleComplex *h_A, *h_R; magma_int_t N, n2, lda, info; magmaDoubleComplex c_neg_one = MAGMA_Z_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_ZPOTRF( N ) / 1e9; TESTING_MALLOC( h_A, magmaDoubleComplex, n2 ); TESTING_HOSTALLOC( h_R, magmaDoubleComplex, n2 ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); magma_zmake_hpd( N, h_A, lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zpotrf( opts.uplo, N, h_R, lda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_zpotrf( &opts.uplo, &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ error = lapackf77_zlange("f", &N, &N, h_A, &lda, work); blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_zlange("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; }
int main( int argc, char** argv ) { TESTING_INIT(); real_Double_t gflops, t1, t2; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t ione = 1; magma_trans_t trans[] = { MagmaNoTrans, MagmaConjTrans, MagmaTrans }; magma_uplo_t uplo [] = { MagmaLower, MagmaUpper }; magma_diag_t diag [] = { MagmaUnit, MagmaNonUnit }; magma_side_t side [] = { MagmaLeft, MagmaRight }; magmaDoubleComplex *A, *B, *C, *C2, *LU; magmaDoubleComplex *dA, *dB, *dC1, *dC2; magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.5, 0.1 ); magmaDoubleComplex beta = MAGMA_Z_MAKE( 0.7, 0.2 ); double dalpha = 0.6; double dbeta = 0.8; double work[1], error, total_error; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t m, n, k, size, maxn, ld, info; magma_int_t *piv; magma_int_t err; magma_opts opts; parse_opts( argc, argv, &opts ); printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" ); total_error = 0.; for( int itest = 0; itest < opts.ntest; ++itest ) { m = opts.msize[itest]; n = opts.nsize[itest]; k = opts.ksize[itest]; printf("=========================================================================\n"); printf( "m=%d, n=%d, k=%d\n", (int) m, (int) n, (int) k ); // allocate matrices // over-allocate so they can be any combination of {m,n,k} x {m,n,k}. maxn = max( max( m, n ), k ); ld = max( 1, maxn ); size = ld*maxn; err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 ); err = magma_zmalloc_pinned( &A, size ); assert( err == 0 ); err = magma_zmalloc_pinned( &B, size ); assert( err == 0 ); err = magma_zmalloc_pinned( &C, size ); assert( err == 0 ); err = magma_zmalloc_pinned( &C2, size ); assert( err == 0 ); err = magma_zmalloc_pinned( &LU, size ); assert( err == 0 ); err = magma_zmalloc( &dA, size ); assert( err == 0 ); err = magma_zmalloc( &dB, size ); assert( err == 0 ); err = magma_zmalloc( &dC1, size ); assert( err == 0 ); err = magma_zmalloc( &dC2, size ); assert( err == 0 ); // initialize matrices size = maxn*maxn; lapackf77_zlarnv( &ione, ISEED, &size, A ); lapackf77_zlarnv( &ione, ISEED, &size, B ); lapackf77_zlarnv( &ione, ISEED, &size, C ); printf( "========== Level 1 BLAS ==========\n" ); // ----- test ZSWAP // swap columns 2 and 3 of dA, then copy to C2 and compare with A if ( n >= 3 ) { magma_zsetmatrix( m, n, A, ld, dA, ld ); magma_zsetmatrix( m, n, A, ld, dB, ld ); magma_zswap( m, dA(0,1), 1, dA(0,2), 1 ); magma_zswap( m, dB(0,1), 1, dB(0,2), 1 ); // check results, storing diff between magma and cuda calls in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dA, 1, dB, 1 ); magma_zgetmatrix( m, n, dB, ld, C2, ld ); error = lapackf77_zlange( "F", &m, &k, C2, &ld, work ); total_error += error; printf( "zswap diff %.2g\n", error ); } else { printf( "zswap skipped for n < 3\n" ); } // ----- test IZAMAX // get argmax of column of A magma_zsetmatrix( m, k, A, ld, dA, ld ); error = 0; for( int j = 0; j < k; ++j ) { magma_int_t i1 = magma_izamax( m, dA(0,j), 1 ); int i2; // NOT magma_int_t, for cublas cublasIzamax( handle, m, dA(0,j), 1, &i2 ); // todo need sync here? assert( i1 == i2 ); error += abs( i1 - i2 ); } total_error += error; gflops = (double)m * k / 1e9; printf( "izamax diff %.2g\n", error ); printf( "\n" ); printf( "========== Level 2 BLAS ==========\n" ); // ----- test ZGEMV // c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors // try no-trans/trans for( int ia = 0; ia < 3; ++ia ) { magma_zsetmatrix( m, n, A, ld, dA, ld ); magma_zsetvector( maxn, B, 1, dB, 1 ); magma_zsetvector( maxn, C, 1, dC1, 1 ); magma_zsetvector( maxn, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_zgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZgemv( handle, cublas_trans_const(trans[ia]), m, n, &alpha, dA, ld, dB, 1, &beta, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 size = (trans[ia] == MagmaNoTrans ? m : n); cublasZaxpy( handle, size, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetvector( size, dC2, 1, C2, 1 ); error = lapackf77_zlange( "F", &size, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_ZGEMV( m, n ) / 1e9; printf( "zgemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_trans_const(trans[ia]), error, gflops/t1, gflops/t2 ); } printf( "\n" ); // ----- test ZHEMV // c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors // try upper/lower for( int iu = 0; iu < 2; ++iu ) { magma_zsetmatrix( m, m, A, ld, dA, ld ); magma_zsetvector( m, B, 1, dB, 1 ); magma_zsetvector( m, C, 1, dC1, 1 ); magma_zsetvector( m, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_zhemv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZhemv( handle, cublas_uplo_const(uplo[iu]), m, &alpha, dA, ld, dB, 1, &beta, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetvector( m, dC2, 1, C2, 1 ); error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_ZHEMV( m ) / 1e9; printf( "zhemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 ); } printf( "\n" ); // ----- test ZTRSV // solve A*c = c, with A m*m triangular; c m-vector // try upper/lower, no-trans/trans, unit/non-unit diag // Factor A into LU to get well-conditioned triangles, else solve yields garbage. // Still can give garbage if solves aren't consistent with LU factors, // e.g., using unit diag for U, so copy lower triangle to upper triangle. // Also used for trsm later. lapackf77_zlacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld ); lapackf77_zgetrf( &maxn, &maxn, LU, &ld, piv, &info ); for( int j = 0; j < maxn; ++j ) { for( int i = 0; i < j; ++i ) { *LU(i,j) = *LU(j,i); } } for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { magma_zsetmatrix( m, m, LU, ld, dA, ld ); magma_zsetvector( m, C, 1, dC1, 1 ); magma_zsetvector( m, C, 1, dC2, 1 ); t1 = magma_sync_wtime( 0 ); magma_ztrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZtrsv( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]), cublas_diag_const(diag[id]), m, dA, ld, dC2, 1 ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetvector( m, dC2, 1, C2, 1 ); error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work ); total_error += error; gflops = FLOPS_ZTRSM( MagmaLeft, m, 1 ) / 1e9; printf( "ztrsv( %c, %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), lapacke_diag_const(diag[id]), error, gflops/t1, gflops/t2 ); }}} printf( "\n" ); printf( "========== Level 3 BLAS ==========\n" ); // ----- test ZGEMM // C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n // try combinations of no-trans/trans for( int ia = 0; ia < 3; ++ia ) { for( int ib = 0; ib < 3; ++ib ) { bool nta = (trans[ia] == MagmaNoTrans); bool ntb = (trans[ib] == MagmaNoTrans); magma_zsetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld ); magma_zsetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld ); magma_zsetmatrix( m, n, C, ld, dC1, ld ); magma_zsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_zgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZgemm( handle, cublas_trans_const(trans[ia]), cublas_trans_const(trans[ib]), m, n, k, &alpha, dA, ld, dB, ld, &beta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &m, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZGEMM( m, n, k ) / 1e9; printf( "zgemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_trans_const(trans[ia]), lapacke_trans_const(trans[ib]), error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test ZHEMM // C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or // C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n // try left/right, upper/lower for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { magma_zsetmatrix( m, m, A, ld, dA, ld ); magma_zsetmatrix( m, n, B, ld, dB, ld ); magma_zsetmatrix( m, n, C, ld, dC1, ld ); magma_zsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_zhemm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZhemm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]), m, n, &alpha, dA, ld, dB, ld, &beta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &m, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZHEMM( side[is], m, n ) / 1e9; printf( "zhemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_side_const(side[is]), lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test ZHERK // C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or // C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric // try upper/lower, no-trans/trans for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { magma_zsetmatrix( n, k, A, ld, dA, ld ); magma_zsetmatrix( n, n, C, ld, dC1, ld ); magma_zsetmatrix( n, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_zherk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZherk( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]), n, k, &dalpha, dA, ld, &dbeta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( n, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZHERK( k, n ) / 1e9; printf( "zherk( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test ZHER2K // C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or // C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric // try upper/lower, no-trans/trans for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { bool nt = (trans[it] == MagmaNoTrans); magma_zsetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld ); magma_zsetmatrix( n, n, C, ld, dC1, ld ); magma_zsetmatrix( n, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_zher2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZher2k( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]), n, k, &alpha, dA, ld, dB, ld, &dbeta, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( n, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZHER2K( k, n ) / 1e9; printf( "zher2k( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), error, gflops/t1, gflops/t2 ); }} printf( "\n" ); // ----- test ZTRMM // C = alpha*A*C (left) with A m*m triangular; C m*n; or // C = alpha*C*A (right) with A n*n triangular; C m*n // try left/right, upper/lower, no-trans/trans, unit/non-unit for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { bool left = (side[is] == MagmaLeft); magma_zsetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld ); magma_zsetmatrix( m, n, C, ld, dC1, ld ); magma_zsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_ztrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; // note cublas does trmm out-of-place (i.e., adds output matrix C), // but allows C=B to do in-place. t2 = magma_sync_wtime( 0 ); cublasZtrmm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]), cublas_diag_const(diag[id]), m, n, &alpha, dA, ld, dC2, ld, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZTRMM( side[is], m, n ) / 1e9; printf( "ztrmm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), error, gflops/t1, gflops/t2 ); }}}} printf( "\n" ); // ----- test ZTRSM // solve A*X = alpha*B (left) with A m*m triangular; B m*n; or // solve X*A = alpha*B (right) with A n*n triangular; B m*n // try left/right, upper/lower, no-trans/trans, unit/non-unit for( int is = 0; is < 2; ++is ) { for( int iu = 0; iu < 2; ++iu ) { for( int it = 0; it < 3; ++it ) { for( int id = 0; id < 2; ++id ) { bool left = (side[is] == MagmaLeft); magma_zsetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld ); magma_zsetmatrix( m, n, C, ld, dC1, ld ); magma_zsetmatrix( m, n, C, ld, dC2, ld ); t1 = magma_sync_wtime( 0 ); magma_ztrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld ); t1 = magma_sync_wtime( 0 ) - t1; t2 = magma_sync_wtime( 0 ); cublasZtrsm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]), cublas_diag_const(diag[id]), m, n, &alpha, dA, ld, dC2, ld ); t2 = magma_sync_wtime( 0 ) - t2; // check results, storing diff between magma and cuda call in C2 cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 ); magma_zgetmatrix( m, n, dC2, ld, C2, ld ); error = lapackf77_zlange( "F", &n, &n, C2, &ld, work ); total_error += error; gflops = FLOPS_ZTRSM( side[is], m, n ) / 1e9; printf( "ztrsm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n", lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), error, gflops/t1, gflops/t2 ); }}}} printf( "\n" ); // cleanup magma_free_cpu( piv ); magma_free_pinned( A ); magma_free_pinned( B ); magma_free_pinned( C ); magma_free_pinned( C2 ); magma_free_pinned( LU ); magma_free( dA ); magma_free( dB ); magma_free( dC1 ); magma_free( dC2 ); fflush( stdout ); } if ( total_error != 0. ) { printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n", total_error ); } else { printf( "all tests passed\n" ); } TESTING_FINALIZE(); int status = (total_error != 0.); return status; }