void magma_cprint_gpu(
magma_int_t m, magma_int_t n,
const magmaFloatComplex *dA, magma_int_t ldda )
{
magma_int_t info = 0;
if ( m < 0 )
info = -1;
else if ( n < 0 )
info = -2;
else if ( ldda < max(1,m) )
info = -4;
if (info != 0) {
magma_xerbla( __func__, -(info) );
return; //info;
}
magma_int_t lda = m;
magmaFloatComplex* A;
magma_cmalloc_cpu( &A, lda*n );
magma_cgetmatrix( m, n, dA, ldda, A, lda );
magma_cprint( m, n, A, lda );
magma_free_cpu( A );
}
void magma_cprint_gpu( magma_int_t m, magma_int_t n, const magmaFloatComplex *dA, magma_int_t ldda )
{
if ( magma_is_devptr( dA ) == 0 ) {
fprintf( stderr, "ERROR: cprint_gpu called with host pointer.\n" );
exit(1);
}
magma_int_t lda = m;
magmaFloatComplex* A;
magma_cmalloc_cpu( &A, lda*n );
magma_cgetmatrix( m, n, dA, ldda, A, lda );
magma_cprint( m, n, A, lda );
magma_free_cpu( A );
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cprint
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magmaFloatComplex *hA, *dA;
//magma_int_t ione = 1;
//magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t M, N, lda, ldda; //size
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = ((M + 31)/32)*32;
//size = lda*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( hA, magmaFloatComplex, lda *N );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*N );
//lapackf77_clarnv( &ione, ISEED, &size, hA );
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
hA[i + j*lda] = MAGMA_C_MAKE( i + j*0.01, 0. );
}
}
magma_csetmatrix( M, N, hA, lda, dA, ldda );
printf( "A=" );
magma_cprint( M, N, hA, lda );
printf( "dA=" );
magma_cprint_gpu( M, N, dA, ldda );
TESTING_FREE_CPU( hA );
TESTING_FREE_DEV( dA );
}
}
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv)
{
#define hA(i,j) (hA + (i) + (j)*lda)
TESTING_CUDA_INIT();
cuFloatComplex c_zero = MAGMA_C_ZERO;
cuFloatComplex c_one = MAGMA_C_ONE;
cuFloatComplex *hA, *hR, *dA;
//real_Double_t gpu_time, gpu_perf;
//int ione = 1;
//int ISEED[4] = {0, 0, 0, 1};
int nsize[] = { 32, 64, 96, 256, 100, 200, 512 };
int ntest = sizeof(nsize) / sizeof(int);
int n = nsize[ntest-1];
int lda = ((n + 31)/32)*32;
int ntile, nb;
TESTING_MALLOC ( hA, cuFloatComplex, lda*n );
TESTING_MALLOC ( hR, cuFloatComplex, lda*n );
TESTING_DEVALLOC ( dA, cuFloatComplex, lda*n );
for( int t = 0; t < ntest; ++t ) {
n = nsize[t];
lda = ((n + 31)/32)*32;
// initialize matrices; entries are (i.j) for A
float nf = 100.;
for( int j = 0; j < n; ++j ) {
// upper
for( int i = 0; i < j; ++i ) {
*hA(i,j) = MAGMA_C_MAKE( (i + j/nf)/nf, 0. );
}
// lower
for( int i = j; i < n; ++i ) {
*hA(i,j) = MAGMA_C_MAKE( i + j/nf, 0. );
}
}
printf( "A%d = ", n );
magma_cprint( n, n, hA, lda );
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize( MagmaLower, n, dA, lda );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d = ", n );
magma_cprint( n, n, hR, lda );
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize( MagmaUpper, n, dA, lda );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "U%d = ", n );
magma_cprint( n, n, hR, lda );
// -----
//lapackf77_claset( "u", &n, &n, &c_zero, &c_one, hA, &lda );
nb = 64;
ntile = n / nb;
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, nb, nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d = ", n, nb );
magma_cprint( n, n, hR, lda );
nb = 32;
ntile = n / nb;
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, nb, nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d = ", n, nb );
magma_cprint( n, n, hR, lda );
ntile = (n - nb < 0 ? 0 : (n - nb) / (2*nb) + 1);
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, 2*nb, nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d_2m = ", n, nb );
magma_cprint( n, n, hR, lda );
nb = 25;
ntile = n / nb;
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, nb, nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d = ", n, nb );
magma_cprint( n, n, hR, lda );
nb = 25;
ntile = (n - nb < 0 ? 0 : (n - nb) / (3*nb) + 1);
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, nb, 3*nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d_3n = ", n, nb );
magma_cprint( n, n, hR, lda );
nb = 100;
ntile = n / nb;
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaLower, nb, dA, lda, ntile, nb, nb );
magmablas_csymmetrize( MagmaLower, n%nb, &dA[ ntile*nb*(1+lda) ], lda ); // last partial block
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "L%d_%d = ", n, nb );
magma_cprint( n, n, hR, lda );
// -----
nb = 64;
ntile = n / nb;
magma_csetmatrix( n, n, hA, lda, dA, lda );
magmablas_csymmetrize_tiles( MagmaUpper, nb, dA, lda, ntile, nb, nb );
magma_cgetmatrix( n, n, dA, lda, hR, lda );
printf( "U%d_%d = ", n, nb );
magma_cprint( n, n, hR, lda );
}
TESTING_FREE( hA );
TESTING_FREE( hR );
TESTING_DEVFREE( dA );
/* Shutdown */
TESTING_CUDA_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing claswp
*/
int main( int argc, char** argv)
{
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
magmaFloatComplex *h_A1, *h_A2, *h_A3, *h_AT;
magmaFloatComplex_ptr d_A1;
real_Double_t gpu_time, cpu_time1, cpu_time2;
/* Matrix size */
int M=0, N=0, n2, lda, ldat;
int size[7] = {1000,2000,3000,4000,5000,6000,7000};
int i, j;
int ione = 1;
int ISEED[4] = {0,0,0,1};
int *ipiv;
int k1, k2, r, c, incx;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
}
if (M>0 && N>0)
printf(" testing_claswp -M %d -N %d\n\n", M, N);
else
{
printf("\nUsage: \n");
printf(" testing_claswp -M %d -N %d\n\n", 1024, 1024);
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_claswp -M %d -N %d\n\n", 1024, 1024);
M = N = size[6];
}
lda = M;
n2 = M*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A1, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_A2, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_A3, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_AT, magmaFloatComplex, n2 );
TESTING_MALLOC_DEV( d_A1, magmaFloatComplex, n2 );
ipiv = (int*)malloc(M * sizeof(int));
if (ipiv == 0) {
fprintf (stderr, "!!!! host memory allocation error (ipiv)\n");
}
printf("\n\n");
printf(" M N CPU_BLAS (sec) CPU_LAPACK (sec) GPU (sec) \n");
printf("=============================================================================\n");
for(i=0; i<7; i++) {
if(argc == 1){
M = N = size[i];
}
lda = M;
ldat = N;
n2 = M*N;
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A1 );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A1, &lda, h_A2, &lda );
for(r=0;r<M;r++){
for(c=0;c<N;c++){
h_AT[c+r*ldat] = h_A1[r+c*lda];
}
}
magma_csetmatrix( N, M, h_AT, 0, ldat, d_A1, 0, ldat, queue);
for(j=0; j<M; j++) {
ipiv[j] = (int)((rand()*1.*M) / (RAND_MAX * 1.)) + 1;
}
/*
* BLAS swap
*/
/* Column Major */
cpu_time1 = magma_wtime();
for ( j=0; j<M; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1+j, &lda, h_A1+(ipiv[j]-1), &lda);
}
}
cpu_time1 = magma_wtime() - cpu_time1;
/*
* LAPACK laswp
*/
cpu_time2 = magma_wtime();
k1 = 1;
k2 = M;
incx = 1;
lapackf77_claswp(&N, h_A2, &lda, &k1, &k2, ipiv, &incx);
cpu_time2 = magma_wtime() - cpu_time2;
/*
* GPU swap
*/
/* Col swap on transpose matrix*/
gpu_time = magma_wtime();
magma_cpermute_long2(N, d_A1, 0, ldat, ipiv, M, 0, queue);
gpu_time = magma_wtime() - gpu_time;
/* Check Result */
magma_cgetmatrix( N, M, d_A1, 0, ldat, h_AT, 0, ldat, queue);
for(r=0;r<N;r++){
for(c=0;c<M;c++){
h_A3[c+r*lda] = h_AT[r+c*ldat];
}
}
int check_bl, check_bg, check_lg;
check_bl = diffMatrix( h_A1, h_A2, M, N, lda );
check_bg = diffMatrix( h_A1, h_A3, M, N, lda );
check_lg = diffMatrix( h_A2, h_A3, M, N, lda );
printf("%5d %5d %6.2f %6.2f %6.2f %s %s %s\n",
M, N, cpu_time1, cpu_time2, gpu_time,
(check_bl == 0) ? "SUCCESS" : "FAILED",
(check_bg == 0) ? "SUCCESS" : "FAILED",
(check_lg == 0) ? "SUCCESS" : "FAILED");
if(check_lg !=0){
printf("lapack swap results:\n");
magma_cprint(M, N, h_A1, lda);
printf("gpu swap transpose matrix result:\n");
magma_cprint(M, N, h_A3, lda);
}
if (argc != 1)
break;
}
/* clean up */
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A1 );
TESTING_FREE_CPU( h_A2 );
TESTING_FREE_CPU( h_A3 );
TESTING_FREE_CPU( h_AT );
TESTING_FREE_DEV( d_A1 );
magma_queue_destroy( queue );
magma_finalize();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing claset
Code is very similar to testing_clacpy.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, 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;
magmaFloatComplex_ptr d_A;
magmaFloatComplex offdiag, diag;
magma_int_t M, N, size, lda, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull };
printf("%% uplo M N offdiag diag CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("%%===================================================================================\n");
for( int iuplo = 0; iuplo < 3; ++iuplo ) {
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
for( int ival = 0; ival < 4; ++ival ) {
// test combinations of zero & non-zero:
// ival offdiag diag
// 0 0 0
// 1 0 3.14
// 2 1.23 0
// 3 1.23 3.14
offdiag = MAGMA_C_MAKE( 1.2345, 6.7890 ) * (ival / 2);
diag = MAGMA_C_MAKE( 3.1415, 2.7183 ) * (ival % 2);
M = opts.msize[itest];
N = opts.nsize[itest];
//M += 2; // space for insets
//N += 2;
lda = M;
ldda = magma_roundup( M, opts.align );
size = lda*N;
if ( uplo[iuplo] == MagmaLower ) {
// save lower trapezoid (with diagonal)
if ( M > N ) {
gbytes = sizeof(magmaFloatComplex) * (1.*M*N - 0.5*N*(N-1)) / 1e9;
} else {
gbytes = sizeof(magmaFloatComplex) * 0.5*M*(M+1) / 1e9;
}
}
else if ( uplo[iuplo] == MagmaUpper ) {
// save upper trapezoid (with diagonal)
if ( N > M ) {
gbytes = sizeof(magmaFloatComplex) * (1.*M*N - 0.5*M*(M-1)) / 1e9;
} else {
gbytes = sizeof(magmaFloatComplex) * 0.5*N*(N+1) / 1e9;
}
}
else {
// save entire matrix
gbytes = sizeof(magmaFloatComplex) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, size );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, size );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A, ldda );
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
//magmablas_claset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_claset( uplo[iuplo], M, N, offdiag, diag, d_A, ldda );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
//magma_int_t M2 = M-2; // inset by 1 row & col
//magma_int_t N2 = N-2;
//lapackf77_claset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_claset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if ( opts.verbose ) {
printf( "A= " ); magma_cprint( M, N, h_A, lda );
printf( "dA=" ); magma_cprint_gpu( M, N, d_A, ldda );
}
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &M, &N, h_R, &lda, work);
bool okay = (error == 0);
status += ! okay;
printf("%5s %5d %5d %9.4f %6.4f %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
lapack_uplo_const( uplo[iuplo] ), (int) M, (int) N,
real(offdiag), real(diag),
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(okay ? "ok" : "failed") );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
