Пример #1
0
/* ////////////////////////////////////////////////////////////////////////////
   -- Testing cgelqf_gpu
*/
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, *h_work, tmp[1];
    magmaFloatComplex *d_A;
    magma_int_t M, N, n2, lda, lwork, info, min_mn, nb;
    magma_int_t ione     = 1;
    magma_int_t ISEED[4] = {0,0,0,1};

    magma_opts opts;
    parse_opts( argc, argv, &opts );
    
    printf("    M     N   CPU GFlop/s (sec)   GPU GFlop/s (sec)   ||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;
            nb     = magma_get_cgeqrf_nb(M);
            gflops = FLOPS_CGELQF( M, N ) / 1e9;
            
            // query for workspace size
            lwork = -1;
            lapackf77_cgelqf(&M, &N, h_A, &M, tau, tmp, &lwork, &info);
            lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
            lwork = max( lwork, M*nb );
            
            TESTING_MALLOC(    tau,    magmaFloatComplex, min_mn );
            TESTING_MALLOC(    h_A,    magmaFloatComplex, n2     );
            TESTING_HOSTALLOC( h_R,    magmaFloatComplex, n2     );
            TESTING_DEVALLOC(  d_A,    magmaFloatComplex, lda*N  );
            TESTING_HOSTALLOC( 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 );
            
            /* ====================================================================
               Performs operation using MAGMA
               =================================================================== */
            magma_csetmatrix( M, N, h_R, lda, d_A, lda );
            gpu_time = magma_wtime();
            magma_cgelqf_gpu( M, N, d_A, lda, tau, h_work, lwork, &info);
            gpu_time = magma_wtime() - gpu_time;
            gpu_perf = gflops / gpu_time;
            if (info != 0)
                printf("magma_cgelqf_gpu returned error %d: %s.\n",
                       (int) info, magma_strerror( info ));
            
            /* =====================================================================
               Performs operation using LAPACK
               =================================================================== */
            cpu_time = magma_wtime();
            lapackf77_cgelqf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
            cpu_time = magma_wtime() - cpu_time;
            cpu_perf = gflops / cpu_time;
            if (info != 0)
                printf("lapack_cgelqf returned error %d: %s.\n",
                       (int) info, magma_strerror( info ));
            
            /* =====================================================================
               Check the result compared to LAPACK
               =================================================================== */
            magma_cgetmatrix( M, N, d_A, lda, h_R, lda );
            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, cpu_time, gpu_perf, gpu_time, error );
            
            TESTING_FREE( tau );
            TESTING_FREE( h_A );
            TESTING_HOSTFREE( h_R );
            TESTING_HOSTFREE( h_work );
        }
        if ( opts.niter > 1 ) {
            printf( "\n" );
        }
    }

    TESTING_FINALIZE();
    return 0;
}
Пример #2
0
/* ////////////////////////////////////////////////////////////////////////////
   -- Testing cgelqf_gpu
*/
int main( int argc, char** argv)
{
    TESTING_INIT();

    const float             d_neg_one = MAGMA_D_NEG_ONE;
    const float             d_one     = MAGMA_D_ONE;
    const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
    const magmaFloatComplex c_one     = MAGMA_C_ONE;
    const magmaFloatComplex c_zero    = MAGMA_C_ZERO;
    const magma_int_t        ione      = 1;
    
    real_Double_t    gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
    float           Anorm, error=0, error2=0;
    magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];
    magmaFloatComplex_ptr d_A;
    magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb;
    magma_int_t ISEED[4] = {0,0,0,1};
    magma_int_t status = 0;

    magma_opts opts;
    opts.parse_opts( argc, argv );

    float tol = opts.tolerance * lapackf77_slamch("E");
    
    printf("%%   M     N   CPU Gflop/s (sec)   GPU Gflop/s (sec)   |L - A*Q^H|   |I - Q*Q^H|\n");
    printf("%%==============================================================================\n");
    for( int itest = 0; itest < opts.ntest; ++itest ) {
        for( int iter = 0; iter < opts.niter; ++iter ) {
            M = opts.msize[itest];
            N = opts.nsize[itest];
            min_mn = min(M, N);
            lda    = M;
            ldda   = magma_roundup( M, opts.align );  // multiple of 32 by default
            n2     = lda*N;
            nb     = magma_get_cgeqrf_nb( M, N );
            gflops = FLOPS_CGELQF( M, N ) / 1e9;
            
            // query for workspace size
            lwork = -1;
            lapackf77_cgelqf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
            lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
            lwork = max( lwork, M*nb );
            
            TESTING_MALLOC_CPU( tau,    magmaFloatComplex, min_mn );
            TESTING_MALLOC_CPU( h_A,    magmaFloatComplex, n2     );
            
            TESTING_MALLOC_PIN( h_R,    magmaFloatComplex, n2     );
            TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork  );
            
            TESTING_MALLOC_DEV( d_A,    magmaFloatComplex, ldda*N );
            
            /* Initialize the matrix */
            lapackf77_clarnv( &ione, ISEED, &n2, h_A );
            lapackf77_clacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
            
            /* ====================================================================
               Performs operation using MAGMA
               =================================================================== */
            magma_csetmatrix( M, N, h_R, lda, d_A, ldda, opts.queue );
            gpu_time = magma_wtime();
            magma_cgelqf_gpu( M, N, d_A, ldda, tau, h_work, lwork, &info);
            gpu_time = magma_wtime() - gpu_time;
            gpu_perf = gflops / gpu_time;
            if (info != 0) {
                printf("magma_cgelqf_gpu returned error %d: %s.\n",
                       (int) info, magma_strerror( info ));
            }
            
            /* =====================================================================
               Check the result, following zlqt01 except using the reduced Q.
               This works for any M,N (square, tall, wide).
               =================================================================== */
            if ( opts.check ) {
                magma_cgetmatrix( M, N, d_A, ldda, h_R, lda, opts.queue );
                
                magma_int_t ldq = min_mn;
                magma_int_t ldl = M;
                magmaFloatComplex *Q, *L;
                float *work;
                TESTING_MALLOC_CPU( Q,    magmaFloatComplex, ldq*N );       // K by N
                TESTING_MALLOC_CPU( L,    magmaFloatComplex, ldl*min_mn );  // M by K
                TESTING_MALLOC_CPU( work, float,             min_mn );
                
                // generate K by N matrix Q, where K = min(M,N)
                lapackf77_clacpy( "Upper", &min_mn, &N, h_R, &lda, Q, &ldq );
                lapackf77_cunglq( &min_mn, &N, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
                assert( info == 0 );
                
                // copy N by K matrix L
                lapackf77_claset( "Upper", &M, &min_mn, &c_zero, &c_zero, L, &ldl );
                lapackf77_clacpy( "Lower", &M, &min_mn, h_R, &lda,        L, &ldl );
                
                // error = || L - A*Q^H || / (N * ||A||)
                blasf77_cgemm( "NoTrans", "Conj", &M, &min_mn, &N,
                               &c_neg_one, h_A, &lda, Q, &ldq, &c_one, L, &ldl );
                Anorm = lapackf77_clange( "1", &M, &N,      h_A, &lda, work );
                error = lapackf77_clange( "1", &M, &min_mn, L,   &ldl, work );
                if ( N > 0 && Anorm > 0 )
                    error /= (N*Anorm);
                
                // set L = I (K by K), then L = I - Q*Q^H
                // error = || I - Q*Q^H || / N
                lapackf77_claset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, L, &ldl );
                blasf77_cherk( "Upper", "NoTrans", &min_mn, &N, &d_neg_one, Q, &ldq, &d_one, L, &ldl );
                error2 = safe_lapackf77_clanhe( "1", "Upper", &min_mn, L, &ldl, work );
                if ( N > 0 )
                    error2 /= N;
                
                TESTING_FREE_CPU( Q    );  Q    = NULL;
                TESTING_FREE_CPU( L    );  L    = NULL;
                TESTING_FREE_CPU( work );  work = NULL;
            }
            
            /* =====================================================================
               Performs operation using LAPACK
               =================================================================== */
            if ( opts.lapack ) {
                cpu_time = magma_wtime();
                lapackf77_cgelqf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info );
                cpu_time = magma_wtime() - cpu_time;
                cpu_perf = gflops / cpu_time;
                if (info != 0) {
                    printf("lapack_cgelqf returned error %d: %s.\n",
                           (int) info, magma_strerror( info ));
                }
            }
            
            /* =====================================================================
               Print performance and error.
               =================================================================== */
            printf("%5d %5d   ", (int) M, (int) N );
            if ( opts.lapack ) {
                printf( "%7.2f (%7.2f)", cpu_perf, cpu_time );
            }
            else {
                printf("  ---   (  ---  )" );
            }
            printf( "   %7.2f (%7.2f)   ", gpu_perf, gpu_time );
            if ( opts.check ) {
                bool okay = (error < tol && error2 < tol);
                printf( "error %.4g, error2 %.4g, tol %.4g, okay %d\n", error, error2, tol, okay );
                status += ! okay;
                printf( "%11.2e   %11.2e   %s\n", error, error2, (okay ? "ok" : "failed") );
            }
            else {
                printf( "    ---\n" );
            }
            
            TESTING_FREE_CPU( tau );
            TESTING_FREE_CPU( h_A );
            
            TESTING_FREE_PIN( h_R    );
            TESTING_FREE_PIN( h_work );
            
            TESTING_FREE_DEV( d_A );
            fflush( stdout );
        }
        if ( opts.niter > 1 ) {
            printf( "\n" );
        }
    }

    opts.cleanup();
    TESTING_FINALIZE();
    return status;
}
Пример #3
0
/* ////////////////////////////////////////////////////////////////////////////
   -- Testing cgelqf
*/
int main( int argc, char** argv)
{
    TESTING_CUDA_INIT();

    magma_timestr_t       start, end;
    float           flops, gpu_perf, cpu_perf;
    float           matnorm, work[1];
    cuFloatComplex  c_neg_one = MAGMA_C_NEG_ONE;
    cuFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];

    /* Matrix size */
    magma_int_t M = 0, N = 0, n2, lda, lwork;
    magma_int_t size[10] = {1024,2048,3072,4032,5184,6016,7040,8064,9088,10112};

    magma_int_t i, info, min_mn, nb;
    magma_int_t ione     = 1;
    magma_int_t ISEED[4] = {0,0,0,1};

    if (argc != 1){
        for(i = 1; i<argc; i++){
            if (strcmp("-N", argv[i])==0)
                N = atoi(argv[++i]);
            else if (strcmp("-M", argv[i])==0)
                M = atoi(argv[++i]);
        }
        if ( M == 0 ) {
            M = N;
        }
        if ( N == 0 ) {
            N = M;
        }
        if (N>0 && M>0)
            printf("  testing_cgelqf -M %d -N %d\n\n", (int) M, (int) N);
        else
            {
                printf("\nUsage: \n");
                printf("  testing_cgelqf -M %d -N %d\n\n", (int) M, (int) N);
                exit(1);
            }
    }
    else {
        printf("\nUsage: \n");
        printf("  testing_cgelqf -M %d -N %d\n\n", 1024, 1024);
        M = N = size[9];
    }

    n2  = M * N;
    min_mn = min(M, N);
    nb = magma_get_cgeqrf_nb(M);

    TESTING_MALLOC(    tau, cuFloatComplex, min_mn );
    TESTING_MALLOC(    h_A, cuFloatComplex, n2     );
    TESTING_HOSTALLOC( h_R, cuFloatComplex, n2     );

    lwork = -1;
    lapackf77_cgelqf(&M, &N, h_A, &M, tau, tmp, &lwork, &info);
    lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
    lwork = max( lwork, M*nb );

    TESTING_HOSTALLOC( h_work, cuFloatComplex, lwork );

    printf("  M     N   CPU GFlop/s   GPU GFlop/s    ||R||_F / ||A||_F\n");
    printf("==========================================================\n");
    for(i=0; i<10; i++){
        if (argc == 1){
            M = N = size[i];
        }
        min_mn= min(M, N);
        lda   = M;
        n2    = lda*N;
        flops = FLOPS( (float)M, (float)N ) / 1000000;

        /* Initialize the matrix */
        lapackf77_clarnv( &ione, ISEED, &n2, h_A );
        lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );

        /* ====================================================================
           Performs operation using MAGMA
           =================================================================== */
        start = get_current_time();
        magma_cgelqf( M, N, h_R, lda, tau, h_work, lwork, &info);
        end = get_current_time();
        if (info < 0)
            printf("Argument %d of magma_cgelqf had an illegal value.\n", (int) -info);
        
        gpu_perf = flops / GetTimerValue(start, end);

        /* =====================================================================
           Performs operation using LAPACK
           =================================================================== */
        start = get_current_time();
        lapackf77_cgelqf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
        end = get_current_time();
        if (info < 0)
            printf("Argument %d of lapack_cgelqf had an illegal value.\n", (int) -info);
        
        cpu_perf = flops / GetTimerValue(start, end);

        /* =====================================================================
           Check the result compared to LAPACK
           =================================================================== */
        matnorm = lapackf77_clange("f", &M, &N, h_A, &lda, work);
        blasf77_caxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);

        printf("%5d %5d  %6.2f         %6.2f        %e\n",
               (int) M, (int) N, cpu_perf, gpu_perf,
               lapackf77_clange("f", &M, &N, h_R, &lda, work) / matnorm);

        if (argc != 1)
            break;
    }

    /* Memory clean up */
    TESTING_FREE( tau );
    TESTING_FREE( h_A );
    TESTING_HOSTFREE( h_R );
    TESTING_HOSTFREE( h_work );

    /* Shutdown */
    TESTING_CUDA_FINALIZE();
}