/* //////////////////////////////////////////////////////////////////////////// -- Testing cgeqrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0; float error, work[1]; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1]; magmaFloatComplex_ptr d_A, dT; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb, size; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4]; magma_opts opts; parse_opts( argc, argv, &opts ); magma_int_t status = 0; float tol; opts.lapack |= (opts.version == 2 && opts.check == 2); // check (-c2) implies lapack (-l) if ( opts.version != 2 && opts.check == 1 ) { printf( "NOTE: version %d requires -c2 check due to the special structure of the\n" "MAGMA cgeqrf results; using -c2.\n\n", (int) opts.version ); opts.check = 2; } printf( "version %d\n", (int) opts.version ); if ( opts.version == 2 ) { if ( opts.check == 1 ) { printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R-Q'A||_1 / (M*||A||_1*eps) ||I-Q'Q||_1 / (M*eps)\n"); printf("=========================================================================================================\n"); } else { printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n"); printf("=======================================================================\n"); } tol = 1.0; } else { printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||Ax-b||_F/(N*||A||_F*||x||_F)\n"); printf("====================================================================================\n"); tol = opts.tolerance * lapackf77_slamch("E"); } 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; gflops = FLOPS_CGEQRF( M, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lwork ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); /* Initialize the matrix */ for ( int j=0; j<4; j++ ) ISEED2[j] = ISEED[j]; // save seeds lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_csetmatrix( M, N, h_R, lda, d_A, 0, ldda, opts.queue ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_wtime(); if ( opts.version == 2 ) { magma_cgeqrf2_gpu( M, N, d_A, 0, ldda, tau, opts.queues2, &info ); } else { nb = magma_get_cgeqrf_nb( M ); size = (2*min(M, N) + (N+31)/32*32 )*nb; TESTING_MALLOC_DEV( dT, magmaFloatComplex, size ); if ( opts.version == 1 ) { magma_cgeqrf_gpu( M, N, d_A, 0, ldda, tau, dT, 0, opts.queue, &info ); } #ifdef HAVE_CUBLAS else if ( opts.version == 3 ) { magma_cgeqrf3_gpu( M, N, d_A, 0, ldda, tau, dT, opts.queue, &info ); } #endif else { printf( "Unknown version %d\n", opts.version ); exit(1); } } gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ magmaFloatComplex *tau2; TESTING_MALLOC_CPU( tau2, magmaFloatComplex, min_mn ); cpu_time = magma_wtime(); lapackf77_cgeqrf(&M, &N, h_A, &lda, tau2, h_work, &lwork, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); TESTING_FREE_CPU( tau2 ); } if ( opts.check == 1 && M >= N ) { /* ===================================================================== Check the result -- only version 1, cqrt02 requires M >= N =================================================================== */ magma_int_t lwork = n2+N; magmaFloatComplex *h_W1, *h_W2, *h_W3; float *h_RW, results[2]; magma_cgetmatrix( M, N, d_A, 0, ldda, h_R, M, opts.queue ); TESTING_MALLOC_CPU( h_W1, magmaFloatComplex, n2 ); // Q TESTING_MALLOC_CPU( h_W2, magmaFloatComplex, n2 ); // R TESTING_MALLOC_CPU( h_W3, magmaFloatComplex, lwork ); // WORK TESTING_MALLOC_CPU( h_RW, float, M ); // RWORK lapackf77_clarnv( &ione, ISEED2, &n2, h_A ); lapackf77_cqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork, h_RW, results ); if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0], results[1] ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e", (int) M, (int) N, gpu_perf, gpu_time, results[0], results[1] ); } // todo also check results[1] < tol? printf(" %s\n", (results[0] < tol ? "ok" : "failed")); status += ! (results[0] < tol); TESTING_FREE_CPU( h_W1 ); TESTING_FREE_CPU( h_W2 ); TESTING_FREE_CPU( h_W3 ); TESTING_FREE_CPU( h_RW ); } else if ( opts.check == 2 && opts.version == 2 ) { /* ===================================================================== Check the result compared to LAPACK -- only version 2 =================================================================== */ magma_cgetmatrix( M, N, d_A, 0, ldda, h_R, M, opts.queue ); error = lapackf77_clange("f", &M, &N, h_A, &lda, work); blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_clange("f", &M, &N, h_R, &lda, work) / error; if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e", (int) M, (int) N, gpu_perf, gpu_time, error ); } printf(" %s\n", (error < tol ? "ok" : "failed")); status += ! (error < tol); } else if ( opts.check == 2 && M >= N ) { /* ===================================================================== Check the result by solving linear system -- only versions 1 & 3, M >= N =================================================================== */ magma_int_t lwork; magmaFloatComplex *x, *b, *hwork; magmaFloatComplex_ptr d_B; const magmaFloatComplex c_zero = MAGMA_C_ZERO; const magmaFloatComplex c_one = MAGMA_C_ONE; const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; const magma_int_t ione = 1; // initialize RHS, b = A*random TESTING_MALLOC_CPU( x, magmaFloatComplex, N ); TESTING_MALLOC_CPU( b, magmaFloatComplex, M ); lapackf77_clarnv( &ione, ISEED, &N, x ); blasf77_cgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione ); // copy to GPU TESTING_MALLOC_DEV( d_B, magmaFloatComplex, M ); magma_csetvector( M, b, 1, d_B, 0, 1, opts.queue ); if ( opts.version == 1 ) { // allocate hwork magma_cgeqrs_gpu( M, N, 1, d_A, 0, ldda, tau, dT, 0, d_B, 0, M, tmp, -1, opts.queue, &info ); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); TESTING_MALLOC_CPU( hwork, magmaFloatComplex, lwork ); // solve linear system magma_cgeqrs_gpu( M, N, 1, d_A, 0, ldda, tau, dT, 0, d_B, 0, M, hwork, lwork, opts.queue, &info ); if (info != 0) printf("magma_cgeqrs returned error %d: %s.\n", (int) info, magma_strerror( info )); TESTING_FREE_CPU( hwork ); } #ifdef HAVE_CUBLAS else if ( opts.version == 3 ) { // allocate hwork magma_cgeqrs3_gpu( M, N, 1, d_A, 0, ldda, tau, dT, 0, d_B, 0, M, tmp, -1, opts.queue, &info ); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); TESTING_MALLOC_CPU( hwork, magmaFloatComplex, lwork ); // solve linear system magma_cgeqrs3_gpu( M, N, 1, d_A, 0, ldda, tau, dT, 0, d_B, 0, M, hwork, lwork, opts.queue, &info ); if (info != 0) printf("magma_cgeqrs3 returned error %d: %s.\n", (int) info, magma_strerror( info )); TESTING_FREE_CPU( hwork ); } #endif else { printf( "Unknown version %d\n", opts.version ); exit(1); } magma_cgetvector( N, d_B, 0, 1, x, 1, opts.queue ); // compute r = Ax - b, saved in b lapackf77_clarnv( &ione, ISEED2, &n2, h_A ); blasf77_cgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione ); // compute residual |Ax - b| / (n*|A|*|x|) float norm_x, norm_A, norm_r, work[1]; norm_A = lapackf77_clange( "F", &M, &N, h_A, &lda, work ); norm_r = lapackf77_clange( "F", &M, &ione, b, &M, work ); norm_x = lapackf77_clange( "F", &N, &ione, x, &N, work ); TESTING_FREE_CPU( x ); TESTING_FREE_CPU( b ); TESTING_FREE_DEV( d_B ); error = norm_r / (N * norm_A * norm_x); if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e", (int) M, (int) N, gpu_perf, gpu_time, error ); } printf(" %s\n", (error < tol ? "ok" : "failed")); status += ! (error < tol); } else { if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) ---", (int) M, (int) N, gpu_perf, gpu_time); } printf("%s\n", (opts.check != 0 ? " (error check only for M >= N)" : "")); } TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); if ( opts.version != 2 ) TESTING_FREE_DEV( dT ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing cgeqrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float error, work[1]; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex *h_A, *h_R, *tau, *dtau, *h_work, tmp[1]; magmaFloatComplex *d_A; float *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_CGEQRF( M, N ) / 1e9; lwork = -1; lapackf77_cgeqrf(&M, &N, h_A, &M, tau, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); TESTING_MALLOC( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC( h_A, magmaFloatComplex, n2 ); TESTING_HOSTALLOC( h_R, magmaFloatComplex, n2 ); TESTING_DEVALLOC( d_A, magmaFloatComplex, ldda*N ); TESTING_DEVALLOC( dtau, magmaFloatComplex, min_mn ); TESTING_DEVALLOC(dwork, float, min_mn ); TESTING_MALLOC( h_work, magmaFloatComplex, lwork ); /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); // warmup magma_cgeqr2_gpu( M, N, d_A, ldda, dtau, dwork, &info ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( 0 ); magma_cgeqr2_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_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_cgeqrf(&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_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_cgetmatrix( M, N, d_A, ldda, h_R, M ); error = lapackf77_clange("f", &M, &N, h_A, &lda, work); blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_clange("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 cgels */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float gpu_error, cpu_error, error, Anorm, work[1]; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1]; magmaFloatComplex *d_A, *d_B; magma_int_t M, N, size, 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; float tol = opts.tolerance * lapackf77_slamch("E"); nrhs = opts.nrhs; printf(" ||b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n"); printf(" M N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) 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; size = lda*N; ldda = ((M+31)/32)*32; lddb = ((max_mn+31)/32)*32; nb = magma_get_cgeqrf_nb(M); gflops = (FLOPS_CGEQRF( M, N ) + FLOPS_CGEQRS( M, N, nrhs )) / 1e9; lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb; // query for workspace size lhwork = -1; lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lhwork, &info); lhwork2 = (magma_int_t) MAGMA_C_REAL( tmp[0] ); lhwork = -1; lapackf77_cunmqr( MagmaLeftStr, MagmaConjTransStr, &M, &nrhs, &min_mn, NULL, &lda, NULL, NULL, &ldb, tmp, &lhwork, &info); lhwork = (magma_int_t) MAGMA_C_REAL( tmp[0] ); lhwork = max( max( lhwork, lhwork2 ), lworkgpu ); TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N ); TESTING_MALLOC_CPU( h_A2, magmaFloatComplex, lda*N ); TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_R, magmaFloatComplex, ldb*nrhs ); TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lhwork ); TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*nrhs ); /* Initialize the matrices */ lapackf77_clarnv( &ione, ISEED, &size, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda ); // make random RHS size = M*nrhs; lapackf77_clarnv( &ione, ISEED, &size, h_B ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb ); // make consistent RHS //size = N*nrhs; //lapackf77_clarnv( &ione, ISEED, &size, h_X ); //blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, // &c_one, h_A, &lda, // h_X, &ldb, // &c_zero, h_B, &ldb ); //lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_csetmatrix( M, N, h_A, lda, d_A, ldda ); magma_csetmatrix( M, nrhs, h_B, ldb, d_B, lddb ); gpu_time = magma_wtime(); magma_cgels3_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_cgels3_gpu returned error %d: %s.\n", (int) info, magma_strerror( info )); // Get the solution in h_X magma_cgetmatrix( N, nrhs, d_B, lddb, h_X, ldb ); // compute the residual blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A, &lda, h_X, &ldb, &c_one, h_R, &ldb); Anorm = lapackf77_clange("f", &M, &N, h_A, &lda, work); /* ===================================================================== Performs operation using LAPACK =================================================================== */ lapackf77_clacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb ); cpu_time = magma_wtime(); lapackf77_cgels( 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_cgels returned error %d: %s.\n", (int) info, magma_strerror( info )); blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N, &c_neg_one, h_A2, &lda, h_X, &ldb, &c_one, h_B, &ldb); cpu_error = lapackf77_clange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm); gpu_error = lapackf77_clange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm); // error relative to LAPACK size = M*nrhs; blasf77_caxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione ); error = lapackf77_clange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm); printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e", (int) M, (int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time, cpu_error, gpu_error, error ); if ( M == N ) { printf( " %s\n", (gpu_error < tol && error < tol ? "ok" : "failed")); status += ! (gpu_error < tol && error < tol); } else { printf( " %s\n", (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_work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_B ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing cgegqr */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float e1, e2, e3, e4, e5, *work; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex c_one = MAGMA_C_ONE; magmaFloatComplex c_zero = MAGMA_C_ZERO; magmaFloatComplex *h_A, *h_R, *tau, *dtau, *h_work, *h_rwork, tmp[1]; magmaFloatComplex *d_A, *dwork; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn; magma_int_t ione = 1, ldwork; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) // versions 1...4 are valid if (opts.version < 1 || opts.version > 4) { printf("Unknown version %d; exiting\n", opts.version ); return -1; } float tol, eps = lapackf77_slamch("E"); tol = 10* opts.tolerance * eps; printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||I-Q'Q||_F / M ||I-Q'Q||_I / M ||A-Q R||_I\n"); printf(" MAGMA / LAPACK MAGMA / LAPACK\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 (N > 128) { printf("%5d %5d skipping because cgegqr requires N <= 128\n", (int) M, (int) N); continue; } if (M < N) { printf("%5d %5d skipping because cgegqr requires M >= N\n", (int) M, (int) N); continue; } min_mn = min(M, N); lda = M; n2 = lda*N; ldda = ((M+31)/32)*32; gflops = FLOPS_CGEQRF( M, N ) / 1e9 + FLOPS_CUNGQR( M, N, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); lwork = max(lwork, 3*N*N); ldwork = N*N; if (opts.version == 2) { ldwork = 3*N*N + min_mn; } TESTING_MALLOC_PIN( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork ); TESTING_MALLOC_PIN(h_rwork, magmaFloatComplex, lwork ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( work, float, M ); TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn ); TESTING_MALLOC_DEV( dwork, magmaFloatComplex, ldwork ); /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); // warmup magma_cgegqr_gpu( 1, M, N, d_A, ldda, dwork, h_work, &info ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( 0 ); magma_cgegqr_gpu( opts.version, M, N, d_A, ldda, dwork, h_rwork, &info ); gpu_time = magma_sync_wtime( 0 ) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_cgegqr returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_cgetmatrix( M, N, d_A, ldda, h_R, M ); // Regenerate R // blasf77_cgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_A, &M, &c_zero, h_rwork, &N); // magma_cprint(N, N, h_work, N); blasf77_ctrmm("r", "u", "n", "n", &M, &N, &c_one, h_rwork, &N, h_R, &M); blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione ); e5 = lapackf77_clange("i", &M, &N, h_R, &M, work) / lapackf77_clange("i", &M, &N, h_A, &lda, work); magma_cgetmatrix( M, N, d_A, ldda, h_R, M ); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); /* Orthogonalize on the CPU */ lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); lapackf77_cungqr(&M, &N, &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_cungqr returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ blasf77_cgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_R, &M, &c_zero, h_work, &N); for(int ii = 0; ii < N*N; ii += N+1 ) { h_work[ii] = MAGMA_C_SUB(h_work[ii], c_one); } e1 = lapackf77_clange("f", &N, &N, h_work, &N, work) / N; e3 = lapackf77_clange("i", &N, &N, h_work, &N, work) / N; blasf77_cgemm("t", "n", &N, &N, &M, &c_one, h_A, &M, h_A, &M, &c_zero, h_work, &N); for(int ii = 0; ii < N*N; ii += N+1 ) { h_work[ii] = MAGMA_C_SUB(h_work[ii], c_one); } e2 = lapackf77_clange("f", &N, &N, h_work, &N, work) / N; e4 = lapackf77_clange("i", &N, &N, h_work, &N, work) / N; printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e / %8.2e %8.2e / %8.2e %8.2e %s\n", (int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, e1, e2, e3, e4, e5, (e1 < tol ? "ok" : "failed")); status += ! (e1 < tol); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, 1000.*gpu_time ); } TESTING_FREE_PIN( tau ); TESTING_FREE_PIN( h_work ); TESTING_FREE_PIN( h_rwork ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_R ); TESTING_FREE_CPU( work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( dtau ); TESTING_FREE_DEV( dwork ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing cgegqr */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float e1, e2, work[1]; magmaFloatComplex *h_A, *h_R, *tau, *dtau, *h_work, tmp[1]; magmaFloatComplex *d_A, *dwork, *ddA, *d_T; 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) ||I - Q'Q||_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_CGEQRF( M, N ) / 1e9 + FLOPS_CUNGQR( M, N, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); lwork = max(lwork, 3*N*N); TESTING_MALLOC_PIN( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn ); TESTING_MALLOC_DEV( dwork, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( ddA, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( d_T, magmaFloatComplex, N*N ); cudaMemset( ddA, 0, N*N*sizeof(magmaFloatComplex) ); cudaMemset( d_T, 0, N*N*sizeof(magmaFloatComplex) ); /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); // warmup magma_cgegqr_gpu( M, N, d_A, ldda, dwork, h_work, &info ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( 0 ); if (opts.version == 2) { int min_mn = min(M, N); int nb = N; cuFloatComplex *dtau = dwork; magma_cgeqr2x3_gpu(&M, &N, d_A, &ldda, dtau, d_T, ddA, (float *)(dwork+min_mn), &info); magma_cgetmatrix( min_mn, 1, dtau, min_mn, tau, min_mn); magma_cungqr_gpu( M, N, N, d_A, ldda, tau, d_T, nb, &info ); } else magma_cgegqr_gpu( M, N, d_A, ldda, dwork, h_work, &info ); gpu_time = magma_sync_wtime( 0 ) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_cgegqr returned error %d: %s.\n", (int) info, magma_strerror( info )); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); /* Orthogonalize on the CPU */ lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); lapackf77_cungqr(&M, &N, &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_cungqr returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_cgetmatrix( M, N, d_A, ldda, h_R, M ); magmaFloatComplex one = MAGMA_C_ONE, zero = MAGMA_C_ZERO; blasf77_cgemm("t", "n", &N, &N, &M, &one, h_R, &M, h_R, &M, &zero, h_work, &N); for(int ii=0; ii<N*N; ii+=(N+1)) h_work[ii] = MAGMA_C_SUB(h_work[ii], one); e1 = lapackf77_clange("f", &N, &N, h_work, &N, work); blasf77_cgemm("t", "n", &N, &N, &M, &one, h_A, &M, h_A, &M, &zero, h_work, &N); for(int ii=0; ii<N*N; ii+=(N+1)) h_work[ii] = MAGMA_C_SUB(h_work[ii], one); e2 = lapackf77_clange("f", &N, &N, h_work, &N, work); printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n", (int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, e1, e2 ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, 1000.*gpu_time ); } TESTING_FREE_PIN( tau ); TESTING_FREE_PIN( h_work ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_R ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( dtau ); TESTING_FREE_DEV( dwork ); TESTING_FREE_DEV( ddA ); TESTING_FREE_DEV( d_T ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return 0; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing cgeqrf_mgpu */ int main( int argc, char** argv ) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0; float error, work[1]; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1]; magmaFloatComplex *d_lA[ MagmaMaxGPUs ]; magma_int_t M, N, n2, lda, ldda, n_local, ngpu; magma_int_t info, min_mn, nb, lhwork; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4]; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= (opts.check == 2); // check (-c2) implies lapack (-l) magma_int_t status = 0; float tol, eps = lapackf77_slamch("E"); tol = opts.tolerance * eps; printf("ngpu %d\n", (int) opts.ngpu ); if ( opts.check == 1 ) { printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R-Q'A||_1 / (M*||A||_1) ||I-Q'Q||_1 / M\n"); printf("================================================================================================\n"); } else { printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F /(M*||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; nb = magma_get_cgeqrf_nb( M ); gflops = FLOPS_CGEQRF( M, N ) / 1e9; // ngpu must be at least the number of blocks ngpu = min( opts.ngpu, int((N+nb-1)/nb) ); if ( ngpu < opts.ngpu ) { printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu ); } // query for workspace size lhwork = -1; lapackf77_cgeqrf( &M, &N, NULL, &M, NULL, tmp, &lhwork, &info ); lhwork = (magma_int_t) MAGMA_C_REAL( tmp[0] ); // Allocate host memory for the matrix TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lhwork ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 ); // Allocate device memory for( int dev = 0; dev < ngpu; dev++ ) { n_local = ((N/nb)/ngpu)*nb; if (dev < (N/nb) % ngpu) n_local += nb; else if (dev == (N/nb) % ngpu) n_local += N % nb; magma_setdevice( dev ); TESTING_MALLOC_DEV( d_lA[dev], magmaFloatComplex, ldda*n_local ); } /* Initialize the matrix */ for ( int j=0; j<4; j++ ) ISEED2[j] = ISEED[j]; // saving seeds lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { magmaFloatComplex *tau2; TESTING_MALLOC_CPU( tau2, magmaFloatComplex, min_mn ); cpu_time = magma_wtime(); lapackf77_cgeqrf( &M, &N, h_A, &M, tau2, h_work, &lhwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapack_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); TESTING_FREE_CPU( tau2 ); } /* ==================================================================== Performs operation using MAGMA =================================================================== */ magma_csetmatrix_1D_col_bcyclic( M, N, h_R, lda, d_lA, ldda, ngpu, nb ); gpu_time = magma_wtime(); magma_cgeqrf2_mgpu( ngpu, M, N, d_lA, ldda, tau, &info ); gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_cgeqrf2 returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_cgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_R, lda, ngpu, nb ); magma_queue_sync( NULL ); if ( opts.check == 1 ) { /* ===================================================================== Check the result =================================================================== */ magma_int_t lwork = n2+N; magmaFloatComplex *h_W1, *h_W2, *h_W3; float *h_RW, results[2]; TESTING_MALLOC_CPU( h_W1, magmaFloatComplex, n2 ); // Q TESTING_MALLOC_CPU( h_W2, magmaFloatComplex, n2 ); // R TESTING_MALLOC_CPU( h_W3, magmaFloatComplex, lwork ); // WORK TESTING_MALLOC_CPU( h_RW, float, M ); // RWORK lapackf77_clarnv( &ione, ISEED2, &n2, h_A ); lapackf77_cqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork, h_RW, results ); results[0] *= eps; results[1] *= eps; if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0],results[1] ); printf("%s\n", (results[0] < tol ? "" : " failed")); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e", (int) M, (int) N, gpu_perf, gpu_time, results[0],results[1] ); printf("%s\n", (results[0] < tol ? "" : " failed")); } status |= ! (results[0] < tol); TESTING_FREE_CPU( h_W1 ); TESTING_FREE_CPU( h_W2 ); TESTING_FREE_CPU( h_W3 ); TESTING_FREE_CPU( h_RW ); } else if ( opts.check == 2 ) { /* ===================================================================== Check the result compared to LAPACK =================================================================== */ error = lapackf77_clange("f", &M, &N, h_A, &lda, work ); blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione ); error = lapackf77_clange("f", &M, &N, h_R, &lda, work ) / (min_mn*error); printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error ); printf("%s\n", (error < tol ? "" : " failed")); status |= ! (error < tol); } else { if ( opts.lapack ) { printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---\n", (int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time ); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, gpu_time); } } TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); for( int dev=0; dev < ngpu; dev++ ){ magma_setdevice( dev ); TESTING_FREE_DEV( d_lA[dev] ); } } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing cgeqrf */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float error, work[1]; magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE; magmaFloatComplex *h_A, *h_T, *h_R, *tau, *h_work, tmp[1]; magmaFloatComplex *d_A, *d_T, *ddA, *dtau; magmaFloatComplex *d_A2, *d_T2, *ddA2, *dtau2; float *dwork, *dwork2; magma_int_t M, N, lda, ldda, lwork, n2, info, min_mn; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; #define BLOCK_SIZE 64 magma_opts opts; parse_opts( argc, argv, &opts ); float tol = 10. * opts.tolerance * lapackf77_slamch("E"); magma_queue_t stream[2]; magma_queue_create( &stream[0] ); magma_queue_create( &stream[1] ); printf("version %d\n", (int) opts.version ); printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R||_F/||A||_F ||R_T||\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 (N > 128) { printf("%5d %5d skipping because cgeqr2x requires N <= 128\n", (int) M, (int) N); continue; } if (M < N) { printf("%5d %5d skipping because cgeqr2x requires M >= N\n", (int) M, (int) N); continue; } min_mn = min(M, N); lda = M; n2 = lda*N; ldda = ((M+31)/32)*32; gflops = (FLOPS_CGEQRF( M, N ) + FLOPS_CGEQRT( M, N )) / 1e9; /* Allocate memory for the matrix */ TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn ); TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 ); TESTING_MALLOC_CPU( h_T, magmaFloatComplex, N*N ); TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 ); TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( d_T, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( ddA, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( dtau, magmaFloatComplex, min_mn ); TESTING_MALLOC_DEV( d_A2, magmaFloatComplex, ldda*N ); TESTING_MALLOC_DEV( d_T2, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( ddA2, magmaFloatComplex, N*N ); TESTING_MALLOC_DEV( dtau2, magmaFloatComplex, min_mn ); TESTING_MALLOC_DEV( dwork, float, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) ); TESTING_MALLOC_DEV( dwork2, float, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) ); // todo replace with magma_claset cudaMemset(ddA, 0, N*N*sizeof(magmaFloatComplex)); cudaMemset(d_T, 0, N*N*sizeof(magmaFloatComplex)); cudaMemset(ddA2, 0, N*N*sizeof(magmaFloatComplex)); cudaMemset(d_T2, 0, N*N*sizeof(magmaFloatComplex)); lwork = -1; lapackf77_cgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] ); lwork = max(lwork, N*N); TESTING_MALLOC_CPU( h_work, magmaFloatComplex, lwork ); /* Initialize the matrix */ lapackf77_clarnv( &ione, ISEED, &n2, h_A ); lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_csetmatrix( M, N, h_R, lda, d_A, ldda ); magma_csetmatrix( M, N, h_R, lda, d_A2, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime(0); if (opts.version == 1) magma_cgeqr2x_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info); else if (opts.version == 2) magma_cgeqr2x2_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info); else if (opts.version == 3) magma_cgeqr2x3_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info); else { printf( "call magma_cgeqr2x4_gpu\n" ); /* Going through NULL stream is faster Going through any stream is slower Doing two streams in parallel is slower than doing them sequentially Queuing happens on the NULL stream - user defined buffers are smaller? */ magma_cgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info, NULL); //magma_cgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info, stream[1]); //magma_cgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, stream[0]); //magma_cgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, NULL); //gflops *= 2; } gpu_time = magma_sync_wtime(0) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) { printf("magma_cgeqr2x_gpu version %d returned error %d: %s.\n", (int) opts.version, (int) info, magma_strerror( info )); } else { if ( opts.check ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); lapackf77_cgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr, &M, &N, h_A, &lda, tau, h_work, &N); //magma_cgeqr2(&M, &N, h_A, &lda, tau, h_work, &info); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_cgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ magma_cgetmatrix( M, N, d_A, ldda, h_R, M ); magma_cgetmatrix( N, N, ddA, N, h_T, N ); // Restore the upper triangular part of A before the check for(int col=0; col < N; col++){ for(int row=0; row <= col; row++) h_R[row + col*M] = h_T[row + col*N]; } error = lapackf77_clange("M", &M, &N, h_A, &lda, work); blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione); error = lapackf77_clange("M", &M, &N, h_R, &lda, work) / (N * error); // Check if T is the same magma_cgetmatrix( N, N, d_T, N, h_T, N ); float terr = 0.; for(int col=0; col < N; col++) for(int row=0; row <= col; row++) terr += ( MAGMA_C_ABS(h_work[row + col*N] - h_T[row + col*N])* MAGMA_C_ABS(h_work[row + col*N] - h_T[row + col*N]) ); terr = magma_ssqrt(terr); printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n", (int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, error, terr, (error < tol ? "ok" : "failed") ); status += ! (error < tol); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, 1000.*gpu_time); } } TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_T ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( d_T ); TESTING_FREE_DEV( ddA ); TESTING_FREE_DEV( dtau ); TESTING_FREE_DEV( dwork ); TESTING_FREE_DEV( d_A2 ); TESTING_FREE_DEV( d_T2 ); TESTING_FREE_DEV( ddA2 ); TESTING_FREE_DEV( dtau2 ); TESTING_FREE_DEV( dwork2 ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } magma_queue_destroy( stream[0] ); magma_queue_destroy( stream[1] ); TESTING_FINALIZE(); return status; }