예제 #1
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;
}
예제 #2
0
/* ////////////////////////////////////////////////////////////////////////////
   -- Testing chegvdx
*/
int main( int argc, char** argv)
{
    TESTING_INIT();

    /* Constants */
    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;
    
    /* Local variables */
    real_Double_t gpu_time;

    magmaFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;

    #ifdef COMPLEX
    float *rwork;
    magma_int_t lrwork;
    #endif

    float *w1, *w2, result[2]={0,0};
    magma_int_t *iwork;
    magma_int_t N, n2, info, lda, lwork, liwork;
    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");
    float tolulp = opts.tolerance * lapackf77_slamch("P");

    magma_range_t range = MagmaRangeAll;
    if (opts.fraction != 1)
        range = MagmaRangeI;

    // pass ngpu = -1 to test multi-GPU code using 1 gpu
    magma_int_t abs_ngpu = abs( opts.ngpu );
    
    printf("%% itype = %d, jobz = %s, range = %s, uplo = %s, fraction = %6.4f, ngpu = %d\n",
           int(opts.itype), lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
           opts.fraction, int(abs_ngpu) );

    if (opts.itype == 1) {
        printf("%%   N     M   GPU Time (sec)   |AZ-BZD|   |D - D_magma|\n");
    }                                                   
    else if (opts.itype == 2) {                      
        printf("%%   N     M   GPU Time (sec)   |ABZ-ZD|   |D - D_magma|\n");
    }                                                   
    else if (opts.itype == 3) {                      
        printf("%%   N     M   GPU Time (sec)   |BAZ-ZD|   |D - D_magma|\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];
            lda = N;
            n2  = lda*N;
            
            // TODO: test vl-vu range
            magma_int_t m1 = 0;
            float vl = 0;
            float vu = 0;
            magma_int_t il = 0;
            magma_int_t iu = 0;
            if (opts.fraction == 0) {
                il = max( 1, magma_int_t(0.1*N) );
                iu = max( 1, magma_int_t(0.3*N) );
            }
            else {
                il = 1;
                iu = max( 1, magma_int_t(opts.fraction*N) );
            }

            magma_cheevdx_getworksize(N, threads, (opts.jobz == MagmaVec),
                                     &lwork,
                                     #ifdef COMPLEX
                                     &lrwork,
                                     #endif
                                     &liwork);
            /* Allocate host memory for the matrix */
            TESTING_MALLOC_CPU( h_A,    magmaFloatComplex, n2 );
            TESTING_MALLOC_CPU( h_B,    magmaFloatComplex, n2 );
            TESTING_MALLOC_CPU( w1,     float, N );
            TESTING_MALLOC_CPU( w2,     float, N );
            TESTING_MALLOC_CPU( iwork,  magma_int_t, liwork );
            
            TESTING_MALLOC_PIN( h_R,    magmaFloatComplex, n2 );
            TESTING_MALLOC_PIN( h_S,    magmaFloatComplex, n2 );
            TESTING_MALLOC_PIN( h_work, magmaFloatComplex, max( lwork, N*N ));  // check needs N*N
            #ifdef COMPLEX
            TESTING_MALLOC_PIN( rwork,  float, lrwork);
            #endif

            /* Initialize the matrix */
            lapackf77_clarnv( &ione, ISEED, &n2, h_A );
            lapackf77_clarnv( &ione, ISEED, &n2, h_B );
            magma_cmake_hpd( N, h_B, lda );
            magma_cmake_hermitian( N, h_A, lda );

            lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
            lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );

            // ===================================================================
            // Performs operation using MAGMA
            // ===================================================================
            gpu_time = magma_wtime();
            if (opts.ngpu == 1) {
                magma_chegvdx_2stage( opts.itype, opts.jobz, range, opts.uplo,
                                      N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
                                      h_work, lwork,
                                      #ifdef COMPLEX
                                      rwork, lrwork,
                                      #endif
                                      iwork, liwork,
                                      &info );
            }
            else {
                magma_chegvdx_2stage_m( abs_ngpu, opts.itype, opts.jobz, range, opts.uplo,
                                        N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
                                        h_work, lwork,
                                        #ifdef COMPLEX
                                        rwork, lrwork,
                                        #endif
                                        iwork, liwork,
                                        &info );
            }
            gpu_time = magma_wtime() - gpu_time;
            if (info != 0) {
                printf("magma_chegvdx_2stage 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)    | D(with V, magma) - D(w/o V, lapack) | / | D |
                   =================================================================== */
                #ifdef REAL
                float *rwork = h_work + N*N;
                #endif
                
                if ( opts.jobz != MagmaNoVec ) {
                    result[0] = 1.;
                    result[0] /= safe_lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork);
                    result[0] /= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork);
                    
                    if (opts.itype == 1) {
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
                        for (int i=0; i < m1; ++i)
                            blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
                        result[0] *= lapackf77_clange("1", &N, &m1, h_work, &N, rwork)/N;
                    }
                    else if (opts.itype == 2) {
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N);
                        for (int i=0; i < m1; ++i)
                            blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
                        result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
                    }
                    else if (opts.itype == 3) {
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
                        for (int i=0; i < m1; ++i)
                            blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
                        blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
                        result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
                    }
                }
                
                lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
                lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );
                
                lapackf77_chegvd( &opts.itype, "N", lapack_uplo_const(opts.uplo), &N,
                                  h_R, &lda, h_S, &lda, w2,
                                  h_work, &lwork,
                                  #ifdef COMPLEX
                                  rwork, &lrwork,
                                  #endif
                                  iwork, &liwork,
                                  &info );
                if (info != 0) {
                    printf("lapackf77_chegvd returned error %d: %s.\n",
                           (int) info, magma_strerror( info ));
                }
                
                float maxw=0, diff=0;
                for (int j=0; j < m1; j++) {
                    maxw = max(maxw, fabs(w1[j]));
                    maxw = max(maxw, fabs(w2[j]));
                    diff = max(diff, fabs(w1[j] - w2[j]));
                }
                result[1] = diff / (m1*maxw);
            }
            
            /* =====================================================================
               Print execution time
               =================================================================== */
            printf("%5d %5d   %9.4f     ",
                   (int) N, (int) m1, gpu_time);
            if ( opts.check ) {
                bool okay = (result[1] < tolulp);
                if ( opts.jobz != MagmaNoVec ) {
                    okay = okay && (result[0] < tol);
                    printf("   %8.2e", result[0] );
                }
                else {
                    printf("     ---   ");
                }
                printf("        %8.2e  %s\n", result[1], (okay ? "ok" : "failed"));
                status += ! okay;
            }
            else {
                printf("     ---\n");
            }
            
            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 );
            #ifdef COMPLEX
            TESTING_FREE_PIN( rwork );
            #endif
            fflush( stdout );
        }
        if ( opts.niter > 1 ) {
            printf( "\n" );
        }
    }

    opts.cleanup();
    TESTING_FINALIZE();
    return status;
}
예제 #3
0
/* ////////////////////////////////////////////////////////////////////////////
   -- Testing chegst
*/
int main( int argc, char** argv)
{
    TESTING_INIT();
    
    // Constants
    const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
    const magma_int_t ione = 1;

    // Local variables
    real_Double_t gpu_time, cpu_time;
    magmaFloatComplex *h_A, *h_B, *h_R;
    magmaFloatComplex_ptr d_A, d_B;
    float      Anorm, error, work[1];
    magma_int_t N, n2, lda, ldda, info;
    magma_int_t ISEED[4] = {0,0,0,1};
    magma_int_t status = 0;
    
    magma_opts opts;
    opts.parse_opts( argc, argv );
    opts.lapack |= opts.check;  // check (-c) implies lapack (-l)
    
    float tol = opts.tolerance * lapackf77_slamch("E");

    printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
    printf("%% itype   N   CPU time (sec)   GPU time (sec)   |R|     \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;
            ldda   = magma_roundup( lda, opts.align );
            n2     = N*lda;
            
            TESTING_MALLOC_CPU( h_A,     magmaFloatComplex, lda*N );
            TESTING_MALLOC_CPU( h_B,     magmaFloatComplex, lda*N );
            
            TESTING_MALLOC_PIN( h_R,     magmaFloatComplex, lda*N );
            
            TESTING_MALLOC_DEV( d_A,     magmaFloatComplex, ldda*N );
            TESTING_MALLOC_DEV( d_B,     magmaFloatComplex, ldda*N );
            
            /* ====================================================================
               Initialize the matrix
               =================================================================== */
            lapackf77_clarnv( &ione, ISEED, &n2, h_A );
            lapackf77_clarnv( &ione, ISEED, &n2, h_B );
            magma_cmake_hermitian( N, h_A, lda );
            magma_cmake_hpd(       N, h_B, lda );
            magma_cpotrf( opts.uplo, N, h_B, lda, &info );
            if (info != 0) {
                printf("magma_cpotrf returned error %d: %s.\n",
                       (int) info, magma_strerror( info ));
            }
            
            magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
            magma_csetmatrix( N, N, h_B, lda, d_B, ldda );
            
            /* ====================================================================
               Performs operation using MAGMA
               =================================================================== */
            gpu_time = magma_wtime();
            magma_chegst_gpu( opts.itype, opts.uplo, N, d_A, ldda, d_B, ldda, &info );
            gpu_time = magma_wtime() - gpu_time;
            if (info != 0) {
                printf("magma_chegst_gpu returned error %d: %s.\n",
                       (int) info, magma_strerror( info ));
            }
            
            /* =====================================================================
               Performs operation using LAPACK
               =================================================================== */
            if ( opts.lapack ) {
                cpu_time = magma_wtime();
                lapackf77_chegst( &opts.itype, lapack_uplo_const(opts.uplo),
                                  &N, h_A, &lda, h_B, &lda, &info );
                cpu_time = magma_wtime() - cpu_time;
                if (info != 0) {
                    printf("lapackf77_chegst returned error %d: %s.\n",
                           (int) info, magma_strerror( info ));
                }
                
                magma_cgetmatrix( N, N, d_A, ldda, h_R, lda );
                
                blasf77_caxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
                Anorm = safe_lapackf77_clanhe("f", lapack_uplo_const(opts.uplo), &N, h_A, &lda, work );
                error = safe_lapackf77_clanhe("f", lapack_uplo_const(opts.uplo), &N, h_R, &lda, work )
                      / Anorm;
                
                bool okay = (error < tol);
                status += ! okay;
                printf("%3d   %5d   %7.2f          %7.2f          %8.2e   %s\n",
                       (int) opts.itype, (int) N, cpu_time, gpu_time,
                       error, (okay ? "ok" : "failed"));
            }
            else {
                printf("%3d   %5d     ---            %7.2f\n",
                       (int) opts.itype, (int) N, gpu_time );
            }
            
            TESTING_FREE_CPU( h_A );
            TESTING_FREE_CPU( h_B );
            
            TESTING_FREE_PIN( h_R );
            
            TESTING_FREE_DEV( d_A );
            TESTING_FREE_DEV( d_B );
            
            fflush( stdout );
        }
        if ( opts.niter > 1 ) {
            printf( "\n" );
        }
    }

    opts.cleanup();
    TESTING_FINALIZE();
    return status;
}