/* //////////////////////////////////////////////////////////////////////////// -- 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 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; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zposv_gpu */ 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_B, *h_X; magmaDoubleComplex *d_A, *d_B; magma_int_t N, lda, ldb, ldda, lddb, 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("uplo = %s\n", 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; ldda = ((N+31)/32)*32; lddb = ldda; gflops = ( FLOPS_ZPOTRF( N ) + FLOPS_ZPOTRS( N, opts.nrhs ) ) / 1e9; TESTING_MALLOC_CPU( h_A, 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 ); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*opts.nrhs ); /* ==================================================================== 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 ); magma_zsetmatrix( N, N, h_A, N, d_A, ldda ); magma_zsetmatrix( N, opts.nrhs, h_B, N, d_B, lddb ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zposv_gpu( opts.uplo, N, opts.nrhs, d_A, ldda, d_B, lddb, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zpotrf_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Residual =================================================================== */ magma_zgetmatrix( N, opts.nrhs, d_B, lddb, h_X, ldb ); 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_B ); TESTING_FREE_CPU( h_X ); TESTING_FREE_CPU( work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zposv_batched */ int main(int argc, char **argv) { TESTING_INIT(); real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time; double err = 0.0, Rnorm, Anorm, Xnorm, *work; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_B, *h_X; magmaDoubleComplex_ptr d_A, d_B; magma_int_t *dinfo_array; magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_int_t batchCount = 1; magmaDoubleComplex **dA_array = NULL; magmaDoubleComplex **dB_array = NULL; magma_queue_t queue = magma_stream; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); nrhs = opts.nrhs; batchCount = opts.batchcount ; printf("uplo = %s\n", lapack_uplo_const(opts.uplo) ); printf("BatchCount 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 = N; ldb = lda; ldda = ((N+31)/32)*32; lddb = ldda; gflops = ( FLOPS_ZPOTRF( N) + FLOPS_ZPOTRS( N, nrhs ) ) / 1e9 * batchCount; sizeA = lda*N*batchCount; sizeB = ldb*nrhs*batchCount; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, sizeA ); TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, sizeB ); TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, sizeB ); TESTING_MALLOC_CPU( work, double, N); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N*batchCount ); TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*nrhs*batchCount ); TESTING_MALLOC_DEV( dinfo_array, magma_int_t, batchCount ); magma_malloc((void**)&dA_array, batchCount * sizeof(*dA_array)); magma_malloc((void**)&dB_array, batchCount * sizeof(*dB_array)); /* Initialize the matrices */ lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A ); lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B ); for(int i=0; i<batchCount; i++) { magma_zmake_hpd( N, h_A + i * lda * N, lda );// need modification } magma_zsetmatrix( N, N*batchCount, h_A, lda, d_A, ldda ); magma_zsetmatrix( N, nrhs*batchCount, h_B, ldb, d_B, lddb ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ zset_pointer(dA_array, d_A, ldda, 0, 0, ldda*N, batchCount, queue); zset_pointer(dB_array, d_B, lddb, 0, 0, lddb*nrhs, batchCount, queue); gpu_time = magma_wtime(); info = magma_zposv_batched(opts.uplo, N, nrhs, dA_array, ldda, dB_array, lddb, dinfo_array, batchCount, queue); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; // check correctness of results throught "dinfo_magma" and correctness of argument throught "info" magma_int_t *cpu_info = (magma_int_t*) malloc(batchCount*sizeof(magma_int_t)); magma_getvector( batchCount, sizeof(magma_int_t), dinfo_array, 1, cpu_info, 1); for(int i=0; i<batchCount; i++) { if(cpu_info[i] != 0 ){ printf("magma_zposv_batched matrix %d returned internal error %d\n",i, (int)cpu_info[i] ); } } if (info != 0) printf("magma_zposv_batched returned argument error %d: %s.\n", (int) info, magma_strerror( info )); //===================================================================== // Residual //===================================================================== magma_zgetmatrix( N, nrhs*batchCount, d_B, lddb, h_X, ldb ); for(magma_int_t s=0; s<batchCount; s++) { Anorm = lapackf77_zlange("I", &N, &N, h_A + s * lda * N, &lda, work); Xnorm = lapackf77_zlange("I", &N, &nrhs, h_X + s * ldb * nrhs, &ldb, work); blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N, &c_one, h_A + s * lda * N, &lda, h_X + s * ldb * nrhs, &ldb, &c_neg_one, h_B + s * ldb * nrhs, &ldb); Rnorm = lapackf77_zlange("I", &N, &nrhs, h_B + s * ldb * nrhs, &ldb, work); double error = Rnorm/(N*Anorm*Xnorm); if ( isnan(error) || isinf(error) ) { err = error; break; } err = max(err, error); } status += ! (err < tol); /* ==================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); for(magma_int_t s=0; s<batchCount; s++) { lapackf77_zposv( lapack_uplo_const(opts.uplo), &N, &nrhs, h_A + s * lda * N, &lda, h_B + s * ldb * nrhs, &ldb, &info ); } cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_zposv returned err %d: %s.\n", (int) info, magma_strerror( info )); printf( "%10d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int)batchCount, (int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time, err, (err < tol ? "ok" : "failed")); } else { printf( "%10d %5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n", (int)batchCount, (int) N, (int) nrhs, gpu_perf, gpu_time, err, (err < tol ? "ok" : "failed")); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( h_X ); TESTING_FREE_CPU( work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); TESTING_FREE_DEV( dinfo_array ); magma_free(dA_array); magma_free(dB_array); free(cpu_info); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zpotri */ 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; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t N, n2, lda, info; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; double Anorm, error, work[1]; magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) double tol = opts.tolerance * lapackf77_dlamch("E"); printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) ); printf("%% N CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||A||_F\n"); printf("%%================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; gflops = FLOPS_ZPOTRI( N ) / 1e9; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); /* ==================================================================== Initialize the matrix =================================================================== */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); magma_zmake_hpd( N, h_A, lda ); lapackf77_zlacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ if ( opts.warmup ) { magma_zpotrf( opts.uplo, N, h_R, lda, &info ); magma_zpotri( opts.uplo, N, h_R, lda, &info ); lapackf77_zlacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda ); } /* factorize matrix */ magma_zpotrf( opts.uplo, N, h_R, lda, &info ); // check for exact singularity //h_R[ 10 + 10*lda ] = MAGMA_Z_MAKE( 0.0, 0.0 ); gpu_time = magma_wtime(); magma_zpotri( opts.uplo, N, h_R, lda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) { printf("magma_zpotri returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info ); cpu_time = magma_wtime(); lapackf77_zpotri( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) { printf("lapackf77_zpotri returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Check the result compared to LAPACK =================================================================== */ blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); Anorm = lapackf77_zlange("f", &N, &N, h_A, &lda, work); error = lapackf77_zlange("f", &N, &N, h_R, &lda, work) / Anorm; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error, (error < tol ? "ok" : "failed") ); status += ! (error < tol); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) ---\n", (int) N, gpu_perf, gpu_time ); } TESTING_FREE_CPU( h_A ); TESTING_FREE_PIN( h_R ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
int main(int argc, char **argv) { TESTING_INIT(); real_Double_t gflopsF, gflopsS, gpu_perf, gpu_time /*cpu_perf, cpu_time*/; real_Double_t gpu_perfdf, gpu_perfds; real_Double_t gpu_perfsf, gpu_perfss; double error, Rnorm, Anorm; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex *h_A, *h_B, *h_X; magmaDoubleComplex *d_A, *d_B, *d_X, *d_workd; magmaFloatComplex *d_As, *d_Bs, *d_works; double *h_workd; magma_int_t lda, ldb, ldx; magma_int_t N, nrhs, posv_iter, info, size; 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("using: uplo = %s\n", lapack_uplo_const(opts.uplo)); printf(" N NRHS DP-Factor DP-Solve SP-Factor SP-Solve MP-Solve Iter |b-Ax|/|A|\n"); printf("=====================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; ldb = ldx = lda = N; gflopsF = FLOPS_ZPOTRF( N ) / 1e9; gflopsS = gflopsF + FLOPS_ZPOTRS( N, nrhs ) / 1e9; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldx*nrhs ); TESTING_MALLOC_CPU( h_workd, double, N ); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, lda*N ); TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, ldb*nrhs ); TESTING_MALLOC_DEV( d_X, magmaDoubleComplex, ldx*nrhs ); TESTING_MALLOC_DEV( d_works, magmaFloatComplex, lda*(N+nrhs) ); TESTING_MALLOC_DEV( d_workd, magmaDoubleComplex, N*nrhs ); /* Initialize the matrix */ size = lda * N ; lapackf77_zlarnv( &ione, ISEED, &size, h_A ); magma_zmake_hpd( N, h_A, lda ); size = ldb * nrhs ; lapackf77_zlarnv( &ione, ISEED, &size, h_B ); magma_zsetmatrix( N, N, h_A, lda, d_A, lda ); magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb ); //===================================================================== // Mixed Precision Iterative Refinement - GPU //===================================================================== gpu_time = magma_wtime(); magma_zcposv_gpu(opts.uplo, N, nrhs, d_A, lda, d_B, ldb, d_X, ldx, d_workd, d_works, &posv_iter, &info); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflopsS / gpu_time; if (info != 0) printf("magma_zcposv returned error %d: %s.\n", (int) info, magma_strerror( info )); //===================================================================== // Error Computation //===================================================================== magma_zgetmatrix( N, nrhs, d_X, ldx, h_X, ldx ) ; Anorm = lapackf77_zlanhe( "I", lapack_uplo_const(opts.uplo), &N, h_A, &N, h_workd); blasf77_zhemm( "L", lapack_uplo_const(opts.uplo), &N, &nrhs, &c_one, h_A, &lda, h_X, &ldx, &c_neg_one, h_B, &ldb); Rnorm = lapackf77_zlange( "I", &N, &nrhs, h_B, &ldb, h_workd); error = Rnorm / Anorm; //===================================================================== // Double Precision Factor //===================================================================== magma_zsetmatrix( N, N, h_A, lda, d_A, lda ); gpu_time = magma_wtime(); magma_zpotrf_gpu(opts.uplo, N, d_A, lda, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfdf = gflopsF / gpu_time; if (info != 0) printf("magma_zpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); //===================================================================== // Double Precision Solve //===================================================================== magma_zsetmatrix( N, N, h_A, lda, d_A, lda ); magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb ); gpu_time = magma_wtime(); magma_zpotrf_gpu(opts.uplo, N, d_A, lda, &info); magma_zpotrs_gpu(opts.uplo, N, nrhs, d_A, lda, d_B, ldb, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfds = gflopsS / gpu_time; if (info != 0) printf("magma_zpotrs returned error %d: %s.\n", (int) info, magma_strerror( info )); //===================================================================== // Single Precision Factor //===================================================================== d_As = d_works; d_Bs = d_works + lda*N; magma_zsetmatrix( N, N, h_A, lda, d_A, lda ); magma_zsetmatrix( N, nrhs, h_B, ldb, d_B, ldb ); magmablas_zlag2c( N, N, d_A, lda, d_As, N, &info ); magmablas_zlag2c( N, nrhs, d_B, ldb, d_Bs, N, &info ); gpu_time = magma_wtime(); magma_cpotrf_gpu(opts.uplo, N, d_As, N, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfsf = gflopsF / gpu_time; if (info != 0) printf("magma_cpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); //===================================================================== // Single Precision Solve //===================================================================== magmablas_zlag2c(N, N, d_A, lda, d_As, N, &info ); magmablas_zlag2c(N, nrhs, d_B, ldb, d_Bs, N, &info ); gpu_time = magma_wtime(); magma_cpotrf_gpu(opts.uplo, N, d_As, lda, &info); magma_cpotrs_gpu(opts.uplo, N, nrhs, d_As, N, d_Bs, N, &info); gpu_time = magma_wtime() - gpu_time; gpu_perfss = gflopsS / gpu_time; if (info != 0) printf("magma_cpotrs returned error %d: %s.\n", (int) info, magma_strerror( info )); printf("%5d %5d %7.2f %7.2f %7.2f %7.2f %7.2f %4d %8.2e %s\n", (int) N, (int) nrhs, gpu_perfdf, gpu_perfds, gpu_perfsf, gpu_perfss, gpu_perf, (int) posv_iter, error, (error < tol ? "ok" : "failed")); status += ! (error < tol); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( h_X ); TESTING_FREE_CPU( h_workd ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); TESTING_FREE_DEV( d_X ); TESTING_FREE_DEV( d_works ); TESTING_FREE_DEV( d_workd ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhegvd */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gpu_time, cpu_time; magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work; double *rwork, *w1, *w2; double result[4] = {0}; magma_int_t *iwork; magma_int_t N, n2, info, nb, lwork, liwork, lda, lrwork; magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; double d_one = 1.; double d_neg_one = -1.; //double d_ten = 10.; //magma_int_t izero = 0; 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"); 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, uplo = %s\n", (int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo)); printf(" N CPU Time (sec) GPU Time(sec)\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 = N*lda; nb = magma_get_zhetrd_nb(N); lwork = 2*N*nb + N*N; lrwork = 1 + 5*N +2*N*N; liwork = 3 + 5*N; 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( rwork, double, lrwork ); 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 ); /* 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, lapack_uplo_const(opts.uplo), h_B, &lda, h_work, &info); //lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_B, &lda); lapackf77_zlarnv( &ione, ISEED, &n2, h_B ); magma_zmake_hpd( N, h_B, lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda ); /* warmup */ if ( opts.warmup ) { magma_zhegvd( opts.itype, opts.jobz, opts.uplo, N, h_R, lda, h_S, lda, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info ); if (info != 0) printf("magma_zhegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda ); } /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zhegvd( opts.itype, opts.jobz, opts.uplo, N, h_R, lda, h_S, lda, w1, h_work, lwork, rwork, lrwork, iwork, liwork, &info ); gpu_time = magma_wtime() - gpu_time; if (info != 0) printf("magma_zhegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.check ) { /* ===================================================================== Check the results following the LAPACK's [zc]hegvd 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) | I - V V' B | / ( N ) (itype = 1,2) | B - V V' | / ( |B| N ) (itype = 3) (3) | S(with V) - S(w/o V) | / | S | =================================================================== */ double temp1, temp2; //magmaDoubleComplex *tau; if ( opts.itype == 1 || opts.itype == 2 ) { lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_S, &lda); blasf77_zgemm("N", "C", &N, &N, &N, &c_one, h_R, &lda, h_R, &lda, &c_zero, h_work, &N); blasf77_zhemm("R", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_work, &N, &c_one, h_S, &lda); result[1] = lapackf77_zlange("1", &N, &N, h_S, &lda, rwork) / N; } else if ( opts.itype == 3 ) { lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda); blasf77_zherk(lapack_uplo_const(opts.uplo), "N", &N, &N, &d_neg_one, h_R, &lda, &d_one, h_S, &lda); result[1] = lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_S, &lda, rwork) / N / lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_B, &lda, rwork); } result[0] = 1.; result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork); result[0] /= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork); if ( opts.itype == 1 ) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N); result[0] *= lapackf77_zlange("1", &N, &N, h_work, &lda, rwork)/N; } else if ( opts.itype == 2 ) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda); result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N; } else if ( opts.itype == 3 ) { blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N); for(int i=0; i<N; ++i) blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione); blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda); result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N; } /* lapackf77_zhet21( &ione, lapack_uplo_const(opts.uplo), &N, &izero, h_A, &lda, w1, w1, h_R, &lda, h_R, &lda, tau, h_work, rwork, &result[0] ); */ lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda ); lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda ); magma_zhegvd( opts.itype, MagmaNoVec, opts.uplo, N, h_R, lda, h_S, lda, w2, h_work, lwork, rwork, lrwork, iwork, liwork, &info ); if (info != 0) printf("magma_zhegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); temp1 = temp2 = 0; for(int j=0; j<N; j++) { temp1 = max(temp1, absv(w1[j])); temp1 = max(temp1, absv(w2[j])); temp2 = max(temp2, absv(w1[j]-w2[j])); } result[2] = temp2 / (((double)N)*temp1); } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_zhegvd( &opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), &N, h_A, &lda, h_B, &lda, w2, h_work, &lwork, rwork, &lrwork, iwork, &liwork, &info ); cpu_time = magma_wtime() - cpu_time; if (info != 0) printf("lapackf77_zhegvd returned error %d: %s.\n", (int) info, magma_strerror( info )); printf("%5d %7.2f %7.2f\n", (int) N, cpu_time, gpu_time); } else { printf("%5d --- %7.2f\n", (int) N, gpu_time); } /* ===================================================================== Print execution time =================================================================== */ if ( opts.check ) { printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if ( opts.itype==1 ) { printf("(1) | 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("(1) | 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("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") ); } if ( opts.itype==1 || opts.itype==2 ) { printf("(2) | I - Z Z' B | / N = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") ); } else { printf("(2) | B - Z Z' | / (|B| N) = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") ); } printf( "(3) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[2], (result[2] < tolulp ? "ok" : "failed") ); status += ! (result[0] < tol && result[1] < tol && result[2] < tolulp); } TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_B ); TESTING_FREE_CPU( w1 ); TESTING_FREE_CPU( w2 ); TESTING_FREE_CPU( rwork ); TESTING_FREE_CPU( iwork ); TESTING_FREE_PIN( h_R ); TESTING_FREE_PIN( h_S ); TESTING_FREE_PIN( h_work ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zpotf2_gpu */ 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; magmaDoubleComplex_ptr d_A; magma_int_t N, n2, lda, ldda, info; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; double Anorm, error, work[1]; magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); double tol = opts.tolerance * lapackf77_dlamch("E"); opts.lapack |= opts.check; // check (-c) implies lapack (-l) printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) ); printf("%% N CPU Gflop/s (ms) GPU Gflop/s (ms) ||R_magma - R_lapack||_F / ||R_lapack||_F\n"); printf("%%=======================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; n2 = lda*N; ldda = magma_roundup( N, opts.align ); // multiple of 32 by default gflops = FLOPS_ZPOTRF( N ) / 1e9; if ( N > 512 ) { printf( "%5d skipping because zpotf2 does not support N > 512\n", (int) N ); continue; } TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 ); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 ); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); magma_zmake_hpd( N, h_A, lda ); lapackf77_zlacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda ); magma_zsetmatrix( N, N, h_A, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( opts.queue ); magma_zpotf2_gpu( opts.uplo, N, d_A, ldda, opts.queue, &info ); gpu_time = magma_sync_wtime( opts.queue ) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) { printf("magma_zpotf2_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); } if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) { printf("lapackf77_zpotrf returned error %d: %s.\n", (int) info, magma_strerror( info )); } /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_zgetmatrix( N, N, d_A, ldda, h_R, lda ); blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); Anorm = lapackf77_zlange("f", &N, &N, h_A, &lda, work); error = lapackf77_zlange("f", &N, &N, h_R, &lda, work) / Anorm; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., error, (error < tol ? "ok" : "failed")); status += ! (error < tol); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) N, gpu_perf, gpu_time*1000. ); } TESTING_FREE_CPU( h_A ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zpotf2_gpu */ 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; magmaDoubleComplex *d_A; magma_int_t N, n2, lda, ldda, 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_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) printf(" N CPU GFlop/s (ms) GPU GFlop/s (ms) ||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; ldda = ((N+31)/32)*32; gflops = FLOPS_ZPOTRF( N ) / 1e9; TESTING_MALLOC( h_A, magmaDoubleComplex, n2 ); TESTING_HOSTALLOC( h_R, magmaDoubleComplex, n2 ); TESTING_DEVALLOC( d_A, magmaDoubleComplex, ldda*N ); /* 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 ); magma_zsetmatrix( N, N, h_A, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); magma_zpotf2_gpu( opts.uplo, N, d_A, ldda, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_zpotf2_gpu 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 =================================================================== */ magma_zgetmatrix( N, N, d_A, ldda, h_R, lda ); 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\n", (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., error ); } else { printf("%5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) N, gpu_perf, gpu_time*1000. ); } TESTING_FREE( h_A ); TESTING_HOSTFREE( h_R ); TESTING_DEVFREE( d_A ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return 0; }
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; magmaDoubleComplex *d_A; magma_int_t N, n2, lda, ldda, 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; magmaDoubleComplex **d_A_array = NULL; magma_int_t *dinfo_magma; magma_int_t batchCount; magma_queue_t queue = magma_stream; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) batchCount = opts.batchcount; double tol = opts.tolerance * lapackf77_dlamch("E"); printf("BatchCount N CPU GFlop/s (ms) GPU GFlop/s (ms) ||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]; ldda = lda = ((N+31)/32)*32; n2 = lda* N * batchCount; gflops = batchCount * FLOPS_ZPOTRF( N ) / 1e9 ; TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2); TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2); TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda * N * batchCount); TESTING_MALLOC_DEV( dinfo_magma, magma_int_t, batchCount); magma_malloc((void**)&d_A_array, batchCount * sizeof(*d_A_array)); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &n2, h_A ); for(int i=0; i<batchCount; i++) { magma_zmake_hpd( N, h_A + i * lda * N, lda );// need modification } magma_int_t columns = N * batchCount; lapackf77_zlacpy( MagmaUpperLowerStr, &N, &(columns), h_A, &lda, h_R, &lda ); magma_zsetmatrix( N, columns, h_A, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ zset_pointer(d_A_array, d_A, ldda, 0, 0, ldda * N, batchCount, queue); gpu_time = magma_sync_wtime(NULL); info = magma_zpotrf_batched( opts.uplo, N, d_A_array, ldda, dinfo_magma, batchCount, queue); gpu_time = magma_sync_wtime(NULL) - gpu_time; gpu_perf = gflops / gpu_time; magma_int_t *cpu_info = (magma_int_t*) malloc(batchCount*sizeof(magma_int_t)); magma_getvector( batchCount, sizeof(magma_int_t), dinfo_magma, 1, cpu_info, 1); for(int i=0; i<batchCount; i++) { if(cpu_info[i] != 0 ){ printf("magma_zpotrf_batched matrix %d returned internal error %d\n",i, (int)cpu_info[i] ); } } if (info != 0) printf("magma_zpotrf_batched returned argument error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); for(int i=0; i<batchCount; i++) { lapackf77_zpotrf( lapack_uplo_const(opts.uplo), &N, h_A + i * lda * N, &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 =================================================================== */ magma_zgetmatrix( N, columns, d_A, ldda, h_R, lda ); magma_int_t NN = lda*N; char const uplo = 'l'; // lapack_uplo_const(opts.uplo) double err = 0.0; for(int i=0; i<batchCount; i++) { error = lapackf77_zlanhe("f", &uplo, &N, h_A + i * lda*N, &lda, work); blasf77_zaxpy(&NN, &c_neg_one, h_A + i * lda*N, &ione, h_R + i * lda*N, &ione); error = lapackf77_zlanhe("f", &uplo, &N, h_R + i * lda*N, &lda, work) / error; if ( isnan(error) || isinf(error) ) { err = error; break; } err = max(fabs(error),err); } printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n", (int)batchCount, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000., err, (error < tol ? "ok" : "failed")); status += ! (err < tol); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int)batchCount, (int) N, gpu_perf, gpu_time*1000. ); } TESTING_FREE_CPU( h_A ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_A_array ); TESTING_FREE_DEV( dinfo_magma ); free(cpu_info); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing zhegvdx */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gpu_time /*cpu_time*/; magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work; double *w1, *w2, vl=0, vu=0; double result[2] = {0}; magma_int_t *iwork; magma_int_t N, n2, info, il, iu, m1, m2, nb, lwork, liwork; magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_one = MAGMA_Z_ONE; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; #if defined(PRECISION_z) || defined(PRECISION_c) double *rwork; magma_int_t lrwork; #endif //double d_one = 1.; //double d_ten = 10.; 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"); 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, uplo = %s, check = %d, fraction = %6.4f\n", (int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo), (int) opts.check, opts.fraction); printf(" N M GPU Time (sec)\n"); printf("============================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; n2 = N*N; nb = magma_get_zhetrd_nb(N); #if defined(PRECISION_z) || defined(PRECISION_c) lwork = 2*N*nb + N*N; lrwork = 1 + 5*N +2*N*N; #else lwork = 1 + 6*N*nb + 2* N*N; #endif liwork = 3 + 5*N; if ( opts.fraction == 0 ) { il = N / 10; iu = N / 5+il; } else { il = 1; iu = (int) (opts.fraction*N); if (iu < 1) iu = 1; } 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 ); // ================================================================== // 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( opts.itype, opts.jobz, MagmaRangeI, 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 ); if (info != 0) printf("magma_zhegvdx returned error %d: %s.\n", (int) info, magma_strerror( 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( opts.itype, opts.jobz, MagmaRangeI, 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 (info != 0) printf("magma_zhegvdx returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.check ) { /* ===================================================================== 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 double temp1, temp2; 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_zhegvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo, N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2, h_work, lwork, #if defined(PRECISION_z) || defined(PRECISION_c) rwork, lrwork, #endif iwork, liwork, &info ); if (info != 0) printf("magma_zhegvdx returned error %d: %s.\n", (int) info, magma_strerror( 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 / (((double)m2)*temp1); } /* ===================================================================== Print execution time =================================================================== */ printf("%5d %5d %7.2f\n", (int) N, (int) m1, gpu_time); if ( opts.check ) { printf("Testing the eigenvalues and eigenvectors for correctness:\n"); if (opts.itype == 1) { printf("(1) | 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("(1) | 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("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed")); } printf( "(2) | 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" ); } } TESTING_FINALIZE(); return status; }