/* ////////////////////////////////////////////////////////////////////////////
-- Testing dprint
*/
int main( int argc, char** argv)
{
TESTING_INIT();
double *hA;
magmaDouble_ptr 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;
opts.parse_opts( argc, argv );
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 = magma_roundup( M, opts.align ); // multiple of 32 by default
//size = lda*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( hA, double, lda *N );
TESTING_MALLOC_DEV( dA, double, ldda*N );
//lapackf77_dlarnv( &ione, ISEED, &size, hA );
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
hA[i + j*lda] = MAGMA_D_MAKE( i + j*0.01, 0. );
}
}
magma_dsetmatrix( M, N, hA, lda, dA, ldda );
printf( "A=" );
magma_dprint( M, N, hA, lda );
printf( "dA=" );
magma_dprint_gpu( M, N, dA, ldda );
TESTING_FREE_CPU( hA );
TESTING_FREE_DEV( dA );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
// ------------------------------------------------------------
// Solve dA * dX = dB, where dA and dX are stored in GPU device memory.
// Internally, MAGMA uses a hybrid CPU + GPU algorithm.
void gpu_interface( magma_int_t n, magma_int_t nrhs )
{
magmaDoubleComplex *dA=NULL, *dX=NULL;
magma_int_t *ipiv=NULL;
magma_int_t ldda = magma_roundup( n, 32 ); // round up to multiple of 32 for best GPU performance
magma_int_t lddx = ldda;
magma_int_t info = 0;
magma_queue_t queue=NULL;
// magma_*malloc routines for GPU memory are type-safe,
// but you can use cudaMalloc if you prefer.
magma_zmalloc( &dA, ldda*n );
magma_zmalloc( &dX, lddx*nrhs );
magma_imalloc_cpu( &ipiv, n ); // ipiv always on CPU
if ( dA == NULL || dX == NULL || ipiv == NULL ) {
fprintf( stderr, "malloc failed\n" );
goto cleanup;
}
magma_int_t dev = 0;
magma_queue_create( dev, &queue );
// Replace these with your code to initialize A and X
zfill_matrix_gpu( n, n, dA, ldda, queue );
zfill_rhs_gpu( n, nrhs, dX, lddx, queue );
magma_zgesv_gpu( n, 1, dA, ldda, ipiv, dX, ldda, &info );
if ( info != 0 ) {
fprintf( stderr, "magma_zgesv_gpu failed with info=%d\n", info );
}
// TODO: use result in dX
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
magma_free( dX );
magma_free_cpu( ipiv );
}
/**
Purpose
-------
SGELS solves the overdetermined, least squares problem
min || A*X - C ||
using the QR factorization A.
The underdetermined problem (m < n) is not currently handled.
Arguments
---------
@param[in]
trans magma_trans_t
- = MagmaNoTrans: the linear system involves A.
Only TRANS=MagmaNoTrans is currently handled.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in,out]
dA REAL array, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, A is overwritten by details of its QR
factorization as returned by SGEQRF3.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in,out]
dB REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) REAL array, dimension MAX(1,LWORK).
On exit, if INFO = 0, HWORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array HWORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_sgeqrf_nb( M, N ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the HWORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgels_driver
********************************************************************/
extern "C" magma_int_t
magma_sgels3_gpu(
magma_trans_t trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, magma_int_t ldda,
magmaFloat_ptr dB, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info)
{
magmaFloat_ptr dT;
float *tau;
magma_int_t min_mn;
magma_int_t nb = magma_get_sgeqrf_nb( m, n );
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
bool lquery = (lwork == -1);
hwork[0] = magma_smake_lwork( lwkopt );
*info = 0;
/* For now, N is the only case working */
if ( trans != MagmaNoTrans )
*info = -1;
else if (m < 0)
*info = -2;
else if (n < 0 || m < n) /* LQ is not handle for now*/
*info = -3;
else if (nrhs < 0)
*info = -4;
else if (ldda < max(1,m))
*info = -6;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
min_mn = min(m,n);
if (min_mn == 0) {
hwork[0] = MAGMA_S_ONE;
return *info;
}
/*
* Allocate temporary buffers
*/
magma_int_t ldtwork = ( 2*min_mn + magma_roundup( n, 32 ) )*nb;
if (nb < nrhs)
ldtwork = ( 2*min_mn + magma_roundup( n, 32 ) )*nrhs;
if (MAGMA_SUCCESS != magma_smalloc( &dT, ldtwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_smalloc_cpu( &tau, min_mn );
if ( tau == NULL ) {
magma_free( dT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_sgeqrf3_gpu( m, n, dA, ldda, tau, dT, info );
if ( *info == 0 ) {
magma_sgeqrs3_gpu( m, n, nrhs,
dA, ldda, tau, dT,
dB, lddb, hwork, lwork, info );
}
magma_free( dT );
magma_free_cpu( tau );
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrf
*/
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 float c_neg_one = MAGMA_S_NEG_ONE;
const float c_one = MAGMA_S_ONE;
const float c_zero = MAGMA_S_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;
float *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloat_ptr d_A, dT;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb, size;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
// version 3 can do either check
if (opts.check == 1 && opts.version == 1) {
opts.check = 2;
printf( "%% version 1 requires check 2 (solve A*x=b)\n" );
}
if (opts.check == 2 && opts.version == 2) {
opts.check = 1;
printf( "%% version 2 requires check 1 (R - Q^H*A)\n" );
}
printf( "%% version %d\n", (int) opts.version );
if ( opts.check == 1 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |R - Q^H*A| |I - Q^H*Q|\n");
printf("%%==============================================================================\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |b - A*x|\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;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
nb = magma_get_sgeqrf_nb( M, N );
gflops = FLOPS_SGEQRF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgeqrf( &M, &N, NULL, &M, NULL, tmp, &lwork, &info );
lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_work, float, lwork );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
if ( opts.version == 1 || opts.version == 3 ) {
size = (2*min(M, N) + magma_roundup( N, 32 ) )*nb;
TESTING_MALLOC_DEV( dT, float, size );
magmablas_slaset( MagmaFull, size, 1, c_zero, c_zero, dT, size );
}
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
magma_ssetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
nb = magma_get_sgeqrf_nb( M, N );
gpu_time = magma_wtime();
if ( opts.version == 1 ) {
// stores dT, V blocks have zeros, R blocks inverted & stored in dT
magma_sgeqrf_gpu( M, N, d_A, ldda, tau, dT, &info );
}
else if ( opts.version == 2 ) {
// LAPACK complaint arguments
magma_sgeqrf2_gpu( M, N, d_A, ldda, tau, &info );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// stores dT, V blocks have zeros, R blocks stored in dT
magma_sgeqrf3_gpu( M, N, d_A, ldda, tau, dT, &info );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( opts.check == 1 && (opts.version == 2 || opts.version == 3) ) {
if ( opts.version == 3 ) {
// copy diagonal blocks of R back to A
for( int i=0; i < min_mn-nb; i += nb ) {
magma_int_t ib = min( min_mn-i, nb );
magmablas_slacpy( MagmaUpper, ib, ib, &dT[min_mn*nb + i*nb], nb, &d_A[ i + i*ldda ], ldda );
}
}
/* =====================================================================
Check the result, following zqrt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
Only for version 2, which has LAPACK complaint output.
Or for version 3, after restoring diagonal blocks of A above.
=================================================================== */
magma_sgetmatrix( M, N, d_A, ldda, h_R, lda );
magma_int_t ldq = M;
magma_int_t ldr = min_mn;
float *Q, *R;
float *work;
TESTING_MALLOC_CPU( Q, float, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( R, float, ldr*N ); // K by N
TESTING_MALLOC_CPU( work, float, min_mn );
// generate M by K matrix Q, where K = min(M,N)
lapackf77_slacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_sorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_slaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_slacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_sgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_slange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_slange( "1", &min_mn, &N, R, &ldr, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set R = I (K by K identity), then R = I - Q^H*Q
// error = || I - Q^H*Q || / N
lapackf77_slaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_ssyrk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
error2 = safe_lapackf77_slansy( "1", "Upper", &min_mn, R, &ldr, work );
if ( N > 0 )
error2 /= N;
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( R ); R = NULL;
TESTING_FREE_CPU( work ); work = NULL;
}
else if ( opts.check == 2 && M >= N && (opts.version == 1 || opts.version == 3) ) {
/* =====================================================================
Check the result by solving consistent linear system, A*x = b.
Only for versions 1 & 3 with M >= N.
=================================================================== */
magma_int_t lwork2;
float *x, *b, *hwork;
magmaFloat_ptr d_B;
// initialize RHS, b = A*random
TESTING_MALLOC_CPU( x, float, N );
TESTING_MALLOC_CPU( b, float, M );
lapackf77_slarnv( &ione, ISEED, &N, x );
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione );
// copy to GPU
TESTING_MALLOC_DEV( d_B, float, M );
magma_ssetvector( M, b, 1, d_B, 1 );
if ( opts.version == 1 ) {
// allocate hwork
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs 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_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs3 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( hwork );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
magma_sgetvector( N, d_B, 1, x, 1 );
// compute r = Ax - b, saved in b
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione );
// compute residual |Ax - b| / (max(m,n)*|A|*|x|)
float norm_x, norm_A, norm_r, work[1];
norm_A = lapackf77_slange( "F", &M, &N, h_A, &lda, work );
norm_r = lapackf77_slange( "F", &M, &ione, b, &M, work );
norm_x = lapackf77_slange( "F", &N, &ione, x, &N, work );
TESTING_FREE_CPU( x );
TESTING_FREE_CPU( b );
TESTING_FREE_DEV( d_B );
error = norm_r / (max(M,N) * norm_A * norm_x);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgeqrf( &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_sgeqrf 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 == 1 ) {
bool okay = (error < tol && error2 < tol);
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else if ( opts.check == 2 ) {
if ( M >= N ) {
bool okay = (error < tol);
status += ! okay;
printf( "%10.2e %s\n", error, (okay ? "ok" : "failed") );
}
else {
printf( "(error check only for M >= N)\n" );
}
}
else {
printf( " ---\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 == 1 || opts.version == 3 ) {
TESTING_FREE_DEV( dT );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgels
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double gpu_error, cpu_error, error, Anorm, work[1];
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1];
magmaDouble_ptr 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;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t status = 0;
double tol = opts.tolerance * lapackf77_dlamch("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;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
lddb = magma_roundup( max_mn, opts.align ); // multiple of 32 by default
nb = magma_get_dgeqrf_nb( M, N );
gflops = (FLOPS_DGEQRF( M, N ) + FLOPS_DGEQRS( M, N, nrhs )) / 1e9;
lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb;
// query for workspace size
lhwork = -1;
lapackf77_dgels( MagmaNoTransStr, &M, &N, &nrhs,
NULL, &lda, NULL, &ldb, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_D_REAL( tmp[0] );
lhwork = max( lhwork, lworkgpu );
TESTING_MALLOC_CPU( tau, double, min_mn );
TESTING_MALLOC_CPU( h_A, double, lda*N );
TESTING_MALLOC_CPU( h_A2, double, lda*N );
TESTING_MALLOC_CPU( h_B, double, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, double, ldb*nrhs );
TESTING_MALLOC_CPU( h_R, double, ldb*nrhs );
TESTING_MALLOC_CPU( h_work, double, lhwork );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
TESTING_MALLOC_DEV( d_B, double, lddb*nrhs );
/* Initialize the matrices */
size = lda*N;
lapackf77_dlarnv( &ione, ISEED, &size, h_A );
lapackf77_dlacpy( MagmaFullStr, &M, &N, h_A, &lda, h_A2, &lda );
// make random RHS
size = ldb*nrhs;
lapackf77_dlarnv( &ione, ISEED, &size, h_B );
lapackf77_dlacpy( MagmaFullStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
// make consistent RHS
//size = N*nrhs;
//lapackf77_dlarnv( &ione, ISEED, &size, h_X );
//blasf77_dgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
// &c_one, h_A, &lda,
// h_X, &ldb,
// &c_zero, h_B, &ldb );
//lapackf77_dlacpy( MagmaFullStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix( M, N, h_A, lda, d_A, ldda );
magma_dsetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
gpu_time = magma_wtime();
magma_dgels_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_dgels_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// compute the residual
magma_dgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
blasf77_dgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb );
Anorm = lapackf77_dlange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_dlacpy( MagmaFullStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_dgels( 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_dgels returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
blasf77_dgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb );
cpu_error = lapackf77_dlange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm);
gpu_error = lapackf77_dlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
// error relative to LAPACK
size = M*nrhs;
blasf77_daxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_dlange("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 );
bool okay;
if ( M == N ) {
okay = (gpu_error < tol && error < tol);
}
else {
okay = (error < tol);
}
status += ! okay;
printf( " %s\n", (okay ? "ok" : "failed"));
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" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlanhe
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A;
double *h_work;
magmaDoubleComplex_ptr d_A;
magmaDouble_ptr d_work;
magma_int_t i, j, N, n2, lda, ldda;
magma_int_t idist = 3; // normal distribution (otherwise max norm is always ~ 1)
magma_int_t ISEED[4] = {0,0,0,1};
double error, norm_magma, norm_lapack;
magma_int_t status = 0;
magma_int_t lapack_nan_fail = 0;
magma_int_t lapack_inf_fail = 0;
bool mkl_warning = false;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tol2;
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper };
magma_norm_t norm[] = { MagmaInfNorm, MagmaOneNorm, MagmaMaxNorm, MagmaFrobeniusNorm };
// Double-Complex inf-norm not supported on Tesla (CUDA arch 1.x)
#if defined(PRECISION_z)
magma_int_t arch = magma_getdevice_arch();
if ( arch < 200 ) {
printf("!!!! NOTE: Double-Complex %s and %s norm are not supported\n"
"!!!! on CUDA architecture %d; requires arch >= 200.\n"
"!!!! It should report \"parameter number 1 had an illegal value\" below.\n\n",
MagmaInfNormStr, MagmaOneNormStr, (int) arch );
for( int inorm = 0; inorm < 2; ++inorm ) {
for( int iuplo = 0; iuplo < 2; ++iuplo ) {
printf( "Testing that magmablas_zlanhe( %s, %s, ... ) returns -1 error...\n",
lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ));
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], 1, NULL, 1, NULL, 1 );
if ( norm_magma != -1 ) {
printf( "expected magmablas_zlanhe to return -1 error, but got %f\n", norm_magma );
status = 1;
}
}}
printf( "...return values %s\n\n", (status == 0 ? "ok" : "failed") );
}
#endif
#ifdef MAGMA_WITH_MKL
// MKL 11.1 has bug in multi-threaded zlanhe; use single thread to work around.
// MKL 11.2 corrects it for inf, one, max norm.
// MKL 11.2 still segfaults for Frobenius norm, which is not tested here
// because MAGMA doesn't implement Frobenius norm yet.
MKLVersion mkl_version;
mkl_get_version( &mkl_version );
magma_int_t la_threads = magma_get_lapack_numthreads();
bool mkl_single_thread = (mkl_version.MajorVersion <= 11 && mkl_version.MinorVersion < 2);
if ( mkl_single_thread ) {
printf( "\nNote: using single thread to work around MKL zlanhe bug.\n\n" );
}
#endif
printf("%% N norm uplo CPU GByte/s (ms) GPU GByte/s (ms) error nan inf\n");
printf("%%=================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int inorm = 0; inorm < 3; ++inorm ) { /* < 4 for Frobenius */
for( int iuplo = 0; iuplo < 2; ++iuplo ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align );
// read upper or lower triangle
gbytes = 0.5*(N+1)*N*sizeof(magmaDoubleComplex) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_work, double, N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( d_work, double, N );
/* Initialize the matrix */
lapackf77_zlarnv( &idist, ISEED, &n2, h_A );
magma_zsetmatrix( N, N, h_A, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gbytes / gpu_time;
if (norm_magma == -1) {
printf( "%5d %4c skipped because %s norm isn't supported\n",
(int) N, lapacke_norm_const( norm[inorm] ), lapack_norm_const( norm[inorm] ));
goto cleanup;
}
else if (norm_magma < 0) {
printf("magmablas_zlanhe returned error %f: %s.\n",
norm_magma, magma_strerror( (int) norm_magma ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
#ifdef MAGMA_WITH_MKL
if ( mkl_single_thread ) {
// work around MKL bug in multi-threaded zlanhe
magma_set_lapack_numthreads( 1 );
}
#endif
cpu_time = magma_wtime();
norm_lapack = lapackf77_zlanhe(
lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if (norm_lapack < 0) {
printf("lapackf77_zlanhe returned error %f: %s.\n",
norm_lapack, magma_strerror( (int) norm_lapack ));
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = fabs( norm_magma - norm_lapack ) / norm_lapack;
tol2 = tol;
if ( norm[inorm] == MagmaMaxNorm ) {
// max-norm depends on only one element, so for Real precisions,
// MAGMA and LAPACK should exactly agree (tol2 = 0),
// while Complex precisions incur roundoff in cuCabs.
#ifdef REAL
tol2 = 0;
#endif
}
bool okay; okay = (error <= tol2);
status += ! okay;
mkl_warning |= ! okay;
/* ====================================================================
Check for NAN and INF propagation
=================================================================== */
#define h_A(i_, j_) (h_A + (i_) + (j_)*lda)
#define d_A(i_, j_) (d_A + (i_) + (j_)*ldda)
i = rand() % N;
j = rand() % N;
magma_int_t tmp;
if ( uplo[iuplo] == MagmaLower && i < j ) {
tmp = i;
i = j;
j = tmp;
}
else if ( uplo[iuplo] == MagmaUpper && i > j ) {
tmp = i;
i = j;
j = tmp;
}
*h_A(i,j) = MAGMA_Z_NAN;
magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 );
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
bool nan_okay; nan_okay = isnan(norm_magma);
bool la_nan_okay; la_nan_okay = isnan(norm_lapack);
lapack_nan_fail += ! la_nan_okay;
status += ! nan_okay;
*h_A(i,j) = MAGMA_Z_INF;
magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 );
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
bool inf_okay; inf_okay = isinf(norm_magma);
bool la_inf_okay; la_inf_okay = isinf(norm_lapack);
lapack_inf_fail += ! la_inf_okay;
status += ! inf_okay;
#ifdef MAGMA_WITH_MKL
if ( mkl_single_thread ) {
// end single thread to work around MKL bug
magma_set_lapack_numthreads( la_threads );
}
#endif
printf("%5d %4c %4c %7.2f (%7.2f) %7.2f (%7.2f) %#9.3g %-6s %6s%1s %6s%1s\n",
(int) N,
lapacke_norm_const( norm[inorm] ),
lapacke_uplo_const( uplo[iuplo] ),
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
error,
(okay ? "ok" : "failed"),
(nan_okay ? "ok" : "failed"), (la_nan_okay ? " " : "*"),
(inf_okay ? "ok" : "failed"), (la_inf_okay ? " " : "*"));
cleanup:
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_work );
fflush( stdout );
} // end iter
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iuplo, inorm
printf( "\n" );
}
// don't print "failed" here because then run_tests.py thinks MAGMA failed
if ( lapack_nan_fail ) {
printf( "* Warning: LAPACK did not pass NAN propagation test; upgrade to LAPACK version >= 3.4.2 (Sep. 2012)\n" );
}
if ( lapack_inf_fail ) {
printf( "* Warning: LAPACK did not pass INF propagation test\n" );
}
if ( mkl_warning ) {
printf("* MKL (e.g., 11.1) has a bug in zlanhe with multiple threads;\n"
" corrected in 11.2 for one, inf, max norms, but still in Frobenius norm.\n"
" Try again with MKL_NUM_THREADS=1.\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgesv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double error, Rnorm, Anorm, Xnorm, *work;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_B, *h_X;
magmaDouble_ptr d_A, d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf("%% N NRHS CPU Gflop/s (sec) GPU Gflop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
lddb = ldda;
gflops = ( FLOPS_DGETRF( N, N ) + FLOPS_DGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, double, lda*N );
TESTING_MALLOC_CPU( h_B, double, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, double, ldb*nrhs );
TESTING_MALLOC_CPU( work, double, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
TESTING_MALLOC_DEV( d_B, double, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_dlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B );
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_dsetmatrix( N, nrhs, h_B, ldb, d_B, lddb, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dgesv_gpu( N, nrhs, d_A, ldda, ipiv, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Residual
//=====================================================================
magma_dgetmatrix( N, nrhs, d_B, lddb, h_X, ldb, opts.queue );
Anorm = lapackf77_dlange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_dlange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_dgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_dlange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dgesv( &N, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_dgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
=======
SSYTRF_nopiv computes the LDLt factorization of a real symmetric
matrix A. This version does not require work space on the GPU passed
as input. GPU memory is allocated in the routine.
The factorization has the form
A = U^H * D * U, if UPLO = MagmaUpper, or
A = L * D * L^H, if UPLO = MagmaLower,
where U is an upper triangular matrix, L is lower triangular, and
D is a diagonal matrix.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of A is stored;
- = MagmaLower: Lower triangle of A is stored.
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the symmetric matrix A. If UPLO = MagmaUpper, the leading
N-by-N upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization A = U^H D U or A = L D L^H.
\n
Higher performance is achieved if A is in pinned memory.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -6, the GPU memory allocation failed
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_ssysv_comp
******************************************************************* */
extern "C" magma_int_t
magma_ssytrf_nopiv(
magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *info)
{
#define A(i, j) ( A +(j)*lda + (i))
#define dA(i, j) (dA +(j)*ldda + (i))
#define dW(i, j) (dW +(j)*ldda + (i))
#define dWt(i, j) (dW +(j)*nb + (i))
/* Constants */
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
/* Local variables */
bool upper = (uplo == MagmaUpper);
magma_int_t j, k, jb, ldda, nb, ib, iinfo;
magmaFloat_ptr dA;
magmaFloat_ptr dW;
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return MAGMA_ERR_ILLEGAL_VALUE;
}
/* Quick return */
if ( n == 0 )
return MAGMA_SUCCESS;
ldda = magma_roundup( n, 32 );
nb = magma_get_ssytrf_nopiv_nb(n);
ib = min(32, nb); // inner-block for diagonal factorization
if ((MAGMA_SUCCESS != magma_smalloc(&dA, n *ldda)) ||
(MAGMA_SUCCESS != magma_smalloc(&dW, nb*ldda))) {
/* alloc failed so call the non-GPU-resident version */
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_device_t cdev;
magma_queue_t queues[2];
magma_event_t event;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
magma_event_create( &event );
trace_init( 1, 1, 2, queues );
/* Use hybrid blocked code. */
if (upper) {
//=========================================================
// Compute the LDLt factorization A = U'*D*U without pivoting.
// copy matrix to GPU
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(j+jb, jb, A(0, j), lda, dA(0, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
if ( j != 0) {
//magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, queues[1]);
}
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(queues[1]);
trace_cpu_start( 0, "potrf", "potrf" );
magma_ssytrf_nopiv_cpu( MagmaUpper, jb, ib, A(j, j), lda, info );
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
// copy j-th column of U back to CPU
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(j, jb, dA(0, j), ldda, A(0, j), lda, queues[1]);
trace_gpu_end( 0, 1 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaUnit,
jb, (n-j-jb),
c_one, dA(j, j), ldda,
dA(j, j+jb), ldda,
queues[0] );
magma_scopymatrix( jb, n-j-jb, dA( j, j+jb ), ldda, dWt( 0, j+jb ), nb, queues[0] );
// update the trailing submatrix with D
magmablas_slascl_diag( MagmaUpper, jb, n-j-jb,
dA(j, j), ldda,
dA(j, j+jb), ldda,
queues[0], &iinfo);
trace_gpu_end( 0, 0 );
// update the trailing submatrix with U and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k < n; k += nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm( MagmaConjTrans, MagmaNoTrans, kb, n-k, jb,
c_neg_one, dWt(0, k), nb,
dA(j, k), ldda,
c_one, dA(k, k), ldda,
queues[0]);
if (k == j+jb) {
// magma_event_record( event, queues[0] );
magma_queue_sync( queues[0] );
}
}
trace_gpu_end( 0, 0 );
}
}
} else {
//=========================================================
// Compute the LDLt factorization A = L*D*L' without pivoting.
// copy the matrix to GPU
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async((n-j), jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
if (j != 0) {
//magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, queues[1]);
}
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(queues[1]);
trace_cpu_start( 0, "potrf", "potrf" );
magma_ssytrf_nopiv_cpu( MagmaLower, jb, ib, A(j, j), lda, info );
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
// copy j-th row of L back to CPU
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(jb, j, dA(j, 0), ldda, A(j, 0), lda, queues[1]);
trace_gpu_end( 0, 1 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaUnit,
(n-j-jb), jb,
c_one, dA(j, j), ldda,
dA(j+jb, j), ldda,
queues[0] );
magma_scopymatrix( n-j-jb,jb, dA( j+jb, j ), ldda, dW( j+jb, 0 ), ldda, queues[0] );
// update the trailing submatrix with D
magmablas_slascl_diag( MagmaLower, n-j-jb, jb,
dA(j, j), ldda,
dA(j+jb, j), ldda,
queues[0], &iinfo );
trace_gpu_end( 0, 0 );
// update the trailing submatrix with L and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k < n; k += nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm( MagmaNoTrans, MagmaConjTrans, n-k, kb, jb,
c_neg_one, dA(k, j), ldda,
dW(k, 0), ldda,
c_one, dA(k, k), ldda,
queues[0] );
if (k == j+jb) {
//magma_event_record( event, queues[0] );
magma_queue_sync(queues[0]);
}
}
trace_gpu_end( 0, 0 );
}
}
}
trace_finalize( "ssytrf.svg","trace.css" );
magma_queue_destroy(queues[0]);
magma_queue_destroy(queues[1]);
magma_event_destroy( event );
magma_free(dW);
magma_free(dA);
return MAGMA_SUCCESS;
} /* magma_ssytrf_nopiv */
/**
Purpose
-------
SSYGVD computes all the eigenvalues, and optionally, the eigenvectors
of a real generalized symmetric-definite eigenproblem, of the form
A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and
B are assumed to be symmetric and B is also positive definite.
If eigenvectors are desired, it uses a divide and conquer algorithm.
The divide and conquer algorithm makes very mild assumptions about
floating point arithmetic. It will work on machines with a guard
digit in add/subtract, or on those binary machines without guard
digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or
Cray-2. It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@param[in]
itype INTEGER
Specifies the problem type to be solved:
= 1: A*x = (lambda)*B*x
= 2: A*B*x = (lambda)*x
= 3: B*A*x = (lambda)*x
@param[in]
range magma_range_t
- = MagmaRangeAll: all eigenvalues will be found.
- = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU]
will be found.
- = MagmaRangeI: the IL-th through IU-th eigenvalues will be found.
@param[in]
jobz magma_vec_t
- = MagmaNoVec: Compute eigenvalues only;
- = MagmaVec: Compute eigenvalues and eigenvectors.
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangles of A and B are stored;
- = MagmaLower: Lower triangles of A and B are stored.
@param[in]
n INTEGER
The order of the matrices A and B. N >= 0.
@param[in,out]
A REAL array, dimension (LDA, N)
On entry, the symmetric matrix A. If UPLO = MagmaUpper, the
leading N-by-N upper triangular part of A contains the
upper triangular part of the matrix A. If UPLO = MagmaLower,
the leading N-by-N lower triangular part of A contains
the lower triangular part of the matrix A.
\n
On exit, if JOBZ = MagmaVec, then if INFO = 0, A contains the
matrix Z of eigenvectors. The eigenvectors are normalized
as follows:
if ITYPE = 1 or 2, Z**T*B*Z = I;
if ITYPE = 3, Z**T*inv(B)*Z = I.
If JOBZ = MagmaNoVec, then on exit the upper triangle (if UPLO=MagmaUpper)
or the lower triangle (if UPLO=MagmaLower) of A, including the
diagonal, is destroyed.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[in,out]
B REAL array, dimension (LDB, N)
On entry, the symmetric matrix B. If UPLO = MagmaUpper, the
leading N-by-N upper triangular part of B contains the
upper triangular part of the matrix B. If UPLO = MagmaLower,
the leading N-by-N lower triangular part of B contains
the lower triangular part of the matrix B.
\n
On exit, if INFO <= N, the part of B containing the matrix is
overwritten by the triangular factor U or L from the Cholesky
factorization B = U**T*U or B = L*L**T.
@param[in]
ldb INTEGER
The leading dimension of the array B. LDB >= max(1,N).
@param[in]
vl REAL
@param[in]
vu REAL
If RANGE=MagmaRangeV, the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeI.
@param[in]
il INTEGER
@param[in]
iu INTEGER
If RANGE=MagmaRangeI, the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeV.
@param[out]
m INTEGER
The total number of eigenvalues found. 0 <= M <= N.
If RANGE = MagmaRangeAll, M = N, and if RANGE = MagmaRangeI, M = IU-IL+1.
@param[out]
w REAL array, dimension (N)
If INFO = 0, the eigenvalues in ascending order.
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The length of the array WORK.
If N <= 1, LWORK >= 1.
If JOBZ = MagmaNoVec and N > 1, LWORK >= 2*N + N*NB.
If JOBZ = MagmaVec and N > 1, LWORK >= max( 2*N + N*NB, 1 + 6*N + 2*N**2 ).
NB can be obtained through magma_get_ssytrd_nb(N).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal sizes of the WORK and
IWORK arrays, returns these values as the first entries of
the WORK and IWORK arrays, and no error message
related to LWORK or LIWORK is issued by XERBLA.
@param[out]
iwork (workspace) INTEGER array, dimension (MAX(1,LIWORK))
On exit, if INFO = 0, IWORK[0] returns the optimal LIWORK.
@param[in]
liwork INTEGER
The dimension of the array IWORK.
If N <= 1, LIWORK >= 1.
If JOBZ = MagmaNoVec and N > 1, LIWORK >= 1.
If JOBZ = MagmaVec and N > 1, LIWORK >= 3 + 5*N.
\n
If LIWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal sizes of the WORK
and IWORK arrays, returns these values as the first entries
of the WORK and IWORK arrays, and no error message
related to LWORK or LIWORK is issued by XERBLA.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
- > 0: SPOTRF or SSYEVD returned an error code:
<= N: if INFO = i and JOBZ = MagmaNoVec, then the algorithm
failed to converge; i off-diagonal elements of an
intermediate tridiagonal form did not converge to
zero;
if INFO = i and JOBZ = MagmaVec, then the algorithm
failed to compute an eigenvalue while working on
the submatrix lying in rows and columns INFO/(N+1)
through mod(INFO,N+1);
> N: if INFO = N + i, for 1 <= i <= N, then the leading
minor of order i of B is not positive definite.
The factorization of B could not be completed and
no eigenvalues or eigenvectors were computed.
Further Details
---------------
Based on contributions by
Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA
Modified so that no backsubstitution is performed if SSYEVD fails to
converge (NEIG in old code could be greater than N causing out of
bounds reference to A - reported by Ralf Meyer). Also corrected the
description of INFO and the test on ITYPE. Sven, 16 Feb 05.
@ingroup magma_ssygv_driver
********************************************************************/
extern "C" magma_int_t
magma_ssygvdx_m(
magma_int_t ngpu,
magma_int_t itype, magma_vec_t jobz, magma_range_t range, magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda,
float *B, magma_int_t ldb,
float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t *m, float *w,
float *work, magma_int_t lwork,
#ifdef COMPLEX
float *rwork, magma_int_t lrwork,
#endif
magma_int_t *iwork, magma_int_t liwork,
magma_int_t *info)
{
/* Constants */
float c_one = MAGMA_S_ONE;
/* Local variables */
const char* uplo_ = lapack_uplo_const( uplo );
const char* jobz_ = lapack_vec_const( jobz );
magma_int_t lower;
magma_trans_t trans;
magma_int_t wantz;
magma_int_t lquery;
magma_int_t alleig, valeig, indeig;
magma_int_t lwmin;
magma_int_t liwmin;
wantz = (jobz == MagmaVec);
lower = (uplo == MagmaLower);
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
lquery = (lwork == -1 || liwork == -1);
*info = 0;
if (itype < 1 || itype > 3) {
*info = -1;
} else if (! (alleig || valeig || indeig)) {
*info = -2;
} else if (! (wantz || (jobz == MagmaNoVec))) {
*info = -3;
} else if (! (lower || (uplo == MagmaUpper))) {
*info = -4;
} else if (n < 0) {
*info = -5;
} else if (lda < max(1,n)) {
*info = -7;
} else if (ldb < max(1,n)) {
*info = -9;
} else {
if (valeig) {
if (n > 0 && vu <= vl) {
*info = -11;
}
} else if (indeig) {
if (il < 1 || il > max(1,n)) {
*info = -12;
} else if (iu < min(n,il) || iu > n) {
*info = -13;
}
}
}
magma_int_t nb = magma_get_ssytrd_nb( n );
if ( n <= 1 ) {
lwmin = 1;
liwmin = 1;
}
else if ( wantz ) {
lwmin = max( 2*n + n*nb, 1 + 6*n + 2*n*n );
liwmin = 3 + 5*n;
}
else {
lwmin = 2*n + n*nb;
liwmin = 1;
}
work[0] = magma_smake_lwork( lwmin );
iwork[0] = liwmin;
if (lwork < lwmin && ! lquery) {
*info = -17;
} else if (liwork < liwmin && ! lquery) {
*info = -19;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info));
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (n == 0) {
return *info;
}
/* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */
if (n <= 128) {
#ifdef ENABLE_DEBUG
printf("--------------------------------------------------------------\n");
printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb);
printf("--------------------------------------------------------------\n");
#endif
lapackf77_ssygvd(&itype, jobz_, uplo_,
&n, A, &lda, B, &ldb,
w, work, &lwork,
iwork, &liwork, info);
*m = n;
return *info;
}
magma_timer_t time=0;
timer_start( time );
magma_spotrf_m(ngpu, uplo, n, B, ldb, info);
if (*info != 0) {
*info = n + *info;
return *info;
}
timer_stop( time );
timer_printf( "time spotrf = %6.2f\n", time );
timer_start( time );
/* Transform problem to standard eigenvalue problem and solve. */
magma_ssygst_m(ngpu, itype, uplo, n, A, lda, B, ldb, info);
timer_stop( time );
timer_printf( "time ssygst = %6.2f\n", time );
timer_start( time );
magma_ssyevdx_m(ngpu, jobz, range, uplo, n, A, lda, vl, vu, il, iu, m, w, work, lwork, iwork, liwork, info);
timer_stop( time );
timer_printf( "time ssyevd = %6.2f\n", time );
if (wantz && *info == 0) {
timer_start( time );
/* Backtransform eigenvectors to the original problem. */
if (itype == 1 || itype == 2) {
/* For A*x=(lambda)*B*x and A*B*x=(lambda)*x;
backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */
if (lower) {
trans = MagmaTrans;
} else {
trans = MagmaNoTrans;
}
magma_strsm_m( ngpu, MagmaLeft, uplo, trans, MagmaNonUnit,
n, *m, c_one, B, ldb, A, lda );
}
else if (itype == 3) {
/* For B*A*x=(lambda)*x;
backtransform eigenvectors: x = L*y or U'*y */
if (lower) {
trans = MagmaNoTrans;
} else {
trans = MagmaTrans;
}
#ifdef ENABLE_DEBUG
printf("--- the multi GPU version is falling back to 1 GPU to perform the last TRMM since there is no TRMM_mgpu --- \n");
#endif
float *dA=NULL, *dB=NULL;
magma_int_t ldda = magma_roundup( n, 32 );
magma_int_t lddb = ldda;
if (MAGMA_SUCCESS != magma_smalloc( &dA, ldda*(*m) ) ||
MAGMA_SUCCESS != magma_smalloc( &dB, lddb*n ) ) {
magma_free( dA );
magma_free( dB );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
magma_ssetmatrix( n, n, B, ldb, dB, lddb, queue );
magma_ssetmatrix( n, (*m), A, lda, dA, ldda, queue );
magma_strmm( MagmaLeft, uplo, trans, MagmaNonUnit,
n, (*m), c_one, dB, lddb, dA, ldda, queue );
magma_sgetmatrix( n, (*m), dA, ldda, A, lda, queue );
magma_queue_destroy( queue );
magma_free( dA );
magma_free( dB );
}
timer_stop( time );
timer_printf( "time setmatrices trsm/mm + getmatrices = %6.2f\n", time );
}
work[0] = magma_smake_lwork( lwmin );
iwork[0] = liwmin;
return *info;
} /* magma_ssygvd_m */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgetri
*/
int main( int argc, char** argv )
{
TESTING_INIT();
// constants
const double c_zero = MAGMA_D_ZERO;
const double c_one = MAGMA_D_ONE;
const double c_neg_one = MAGMA_D_NEG_ONE;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double *h_A, *h_Ainv, *h_R, *work;
magmaDouble_ptr d_A, dwork;
magma_int_t N, n2, lda, ldda, info, lwork, ldwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double tmp;
double error, rwork[1];
magma_int_t *ipiv;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% N CPU Gflop/s (sec) GPU Gflop/s (sec) ||I - A*A^{-1}||_1 / (N*cond(A))\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
ldwork = N * magma_get_dgetri_nb( N );
gflops = FLOPS_DGETRI( N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_dgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0) {
printf("lapackf77_dgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = magma_int_t( MAGMA_D_REAL( tmp ));
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, double, lwork );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_Ainv, double, n2 );
TESTING_MALLOC_CPU( h_R, double, n2 );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
TESTING_MALLOC_DEV( dwork, double, ldwork );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_dgetrf_gpu( N, N, d_A, ldda, ipiv, &info );
magma_dgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
if (info != 0) {
printf("magma_dgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// check for exact singularity
//h_Ainv[ 10 + 10*lda ] = MAGMA_D_MAKE( 0.0, 0.0 );
//magma_dsetmatrix( N, N, h_Ainv, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dgetri_gpu( N, d_A, ldda, ipiv, dwork, ldwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgetri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dgetri( &N, h_Ainv, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_dgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf( "%5d --- ( --- ) %7.2f (%7.2f)",
(int) N, gpu_perf, gpu_time );
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
magma_dgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
// compute 1-norm condition number estimate, following LAPACK's zget03
double normA, normAinv, rcond;
normA = lapackf77_dlange( "1", &N, &N, h_A, &lda, rwork );
normAinv = lapackf77_dlange( "1", &N, &N, h_Ainv, &lda, rwork );
if ( normA <= 0 || normAinv <= 0 ) {
rcond = 0;
error = 1 / (tol/opts.tolerance); // == 1/eps
}
else {
rcond = (1 / normA) / normAinv;
// R = I
// R -= A*A^{-1}
// err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm
lapackf77_dlaset( "full", &N, &N, &c_zero, &c_one, h_R, &lda );
blasf77_dgemm( "no", "no", &N, &N, &N,
&c_neg_one, h_A, &lda,
h_Ainv, &lda,
&c_one, h_R, &lda );
error = lapackf77_dlange( "1", &N, &N, h_R, &lda, rwork );
error = error * rcond / N;
}
bool okay = (error < tol);
status += ! okay;
printf( " %8.2e %s\n",
error, (okay ? "ok" : "failed"));
}
else {
printf( "\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_Ainv );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgetrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float error;
float *h_A;
magmaFloat_ptr d_A;
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, info, min_mn;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% version %d\n", (int) opts.version );
if ( opts.check == 2 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |PA-LU|/(N*|A|)\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;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_SGETRF( M, N ) / 1e9;
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( opts, M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_sgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( opts, M, N, h_A, lda );
if ( opts.version == 2 ) {
// no pivoting versions, so set ipiv to identity
for (magma_int_t i=0; i < min_mn; ++i ) {
ipiv[i] = i+1;
}
}
magma_ssetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_wtime();
if ( opts.version == 1 ) {
magma_sgetrf_gpu( M, N, d_A, ldda, ipiv, &info);
}
else if ( opts.version == 2 ) {
magma_sgetrf_nopiv_gpu( M, N, d_A, ldda, &info);
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the factorization
=================================================================== */
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 );
}
if ( opts.check == 2 ) {
magma_sgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_residual( opts, M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else if ( opts.check ) {
magma_sgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_LU_error( opts, M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf(" --- \n");
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_time, gpu_perf;
magmaDoubleComplex *h_A, *h_R, *h_work;
magmaDoubleComplex *tau;
double *D, *E;
magma_int_t N, n2, lda, ldda, lwork, ldt, info, nstream;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
// TODO add these options to parse_opts
magma_int_t NE = 0;
magma_int_t distblk = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t WANTZ = (opts.jobz == MagmaVec);
double tol = opts.tolerance * lapackf77_dlamch("E");
if (opts.nb == 0)
opts.nb = 64; //magma_get_zhetrd_he2hb_nb(N);
if (NE < 1)
NE = N; //64; //magma_get_zhetrd_he2hb_nb(N);
nstream = max(3, opts.ngpu+2);
magma_queue_t streams[MagmaMaxGPUs][20];
magmaDoubleComplex_ptr da[MagmaMaxGPUs], dT1[MagmaMaxGPUs];
if ((distblk == 0) || (distblk < opts.nb))
distblk = max(256, opts.nb);
printf("%% ngpu %d, distblk %d, NB %d, nstream %d\n",
(int) opts.ngpu, (int) distblk, (int) opts.nb, (int) nstream);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( int i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
}
magma_setdevice( 0 );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldt = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
n2 = lda*N;
/* We suppose the magma NB is bigger than lapack NB */
lwork = N*opts.nb;
//gflops = ....?
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_int_t mlocal = ((N / distblk) / opts.ngpu + 1) * distblk;
magma_setdevice( dev );
TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*opts.nb );
}
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Copy the matrix to the GPU */
magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, opts.ngpu, distblk);
//magmaDoubleComplex_ptr dabis;
//TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N );
//magma_zsetmatrix(N, N, h_R, lda, dabis, ldda);
for (int count=0; count < 1; ++count) {
magma_setdevice(0);
gpu_time = magma_wtime();
if (opts.version == 30) {
// see src/obsolete and magmablas/obsolete
printf( "magma_zhetrd_he2hb_mgpu_spec not compiled\n" );
//magma_zhetrd_he2hb_mgpu_spec(
// opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
// da, ldda, dT1, opts.nb, opts.ngpu, distblk,
// streams, nstream, opts.nthread, &info);
} else {
nstream = 3;
magma_zhetrd_he2hb_mgpu(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
}
// magma_zhetrd_he2hb(opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, dT1[0], &info);
gpu_time = magma_wtime() - gpu_time;
printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n",
N, opts.nb, distblk, opts.ngpu, opts.version, gpu_time);
}
magma_setdevice(0);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice(dev);
magma_device_sync();
}
magma_setdevice(0);
magmablasSetKernelStream( NULL );
// todo neither of these is declared in headers
// magma_zhetrd_bhe2trc_v5(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// magma_zhetrd_bhe2trc(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// todo where is this timer started?
// gpu_time = magma_wtime() - gpu_time;
// todo what are the gflops?
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zhetrd_he2hb returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
int nt = min(12, opts.nthread);
magma_set_lapack_numthreads(nt);
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#ifdef COMPLEX
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, nt);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( lapack_vec_const(opts.jobz), &MATYPE, &N, &opts.nb,
AINIT, &lda, &NOTHING, &NOTHING, D2, D,
h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
magma_set_lapack_numthreads(1);
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) opts.nthread, (int) N, (int) opts.nb, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif // PRECISION_z || PRECISION_d
#endif // CHECKEIG
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time );
printf("%%===========================================================================\n\n\n");
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
TESTING_FREE_DEV( da[dev] );
TESTING_FREE_DEV( dT1[dev] );
}
magma_setdevice( 0 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
for( int i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
TESTING_FINALIZE();
return status;
}
/***************************************************************************//**
Purpose
-------
SGELQF computes an LQ factorization of a REAL M-by-N matrix dA:
dA = L * Q.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and below the diagonal of the array
contain the m-by-min(m,n) lower trapezoidal matrix L (L is
lower triangular if m <= n); the elements above the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors (see Further Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= max(1,M).
For optimum performance LWORK >= M*NB, where NB is the
optimal blocksize.
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(k) . . . H(2) H(1), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:n) is stored on exit in A(i,i+1:n),
and tau in TAU(i).
@ingroup magma_gelqf
*******************************************************************************/
extern "C" magma_int_t
magma_sgelqf_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
float *tau,
float *work, magma_int_t lwork,
magma_int_t *info)
{
/* Constants */
const float c_one = MAGMA_S_ONE;
const magma_int_t ione = 1;
MAGMA_UNUSED( ione ); // used only for real
/* Local variables */
magmaFloat_ptr dAT=NULL;
magma_int_t min_mn, maxm, maxn, nb;
magma_int_t iinfo;
*info = 0;
nb = magma_get_sgelqf_nb( m, n );
min_mn = min( m, n );
work[0] = magma_smake_lwork( m*nb );
bool lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,m) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
magma_int_t lddat = maxn;
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
if ( m == n ) {
dAT = dA;
lddat = ldda;
magmablas_stranspose_inplace( m, dAT, ldda, queue );
}
else {
if (MAGMA_SUCCESS != magma_smalloc( &dAT, maxm*maxn ) ) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
magmablas_stranspose( m, n, dA, ldda, dAT, lddat, queue );
}
magma_sgeqrf2_gpu( n, m, dAT, lddat, tau, &iinfo );
assert( iinfo >= 0 );
if ( iinfo > 0 ) {
*info = iinfo;
}
// conjugate tau
#ifdef COMPLEX
lapackf77_slacgv( &min_mn, tau, &ione );
#endif
if ( m == n ) {
magmablas_stranspose_inplace( m, dAT, lddat, queue );
}
else {
magmablas_stranspose( n, m, dAT, lddat, dA, ldda, queue );
magma_free( dAT );
}
cleanup:
magma_queue_destroy( queue );
return *info;
} /* magma_sgelqf_gpu */
/**
Purpose
-------
SGETRF_NOPIV_GPU computes an LU factorization of a general M-by-N
matrix A without any pivoting.
The factorization has the form
A = L * U
where L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf_nopiv_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *info )
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, (dA_offset + (i_)*nb + (j_)*nb*ldda)
#else
#define dA(i_, j_) (dA + (i_)*nb + (j_)*nb*ldda)
#endif
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, mindim;
magma_int_t j, rows, s, ldwork;
float *work;
/* Check arguments */
*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;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
mindim = min( m, n );
nb = magma_get_sgetrf_nb( m, n );
s = mindim / nb;
magma_queue_t queues[2];
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
if ( MAGMA_SUCCESS != magma_smalloc_cpu( &work, m*n )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_sgetmatrix( m, n, dA(0,0), ldda, work, m, queues[0] );
magma_sgetrf_nopiv( m, n, work, m, info );
magma_ssetmatrix( m, n, work, m, dA(0,0), ldda, queues[0] );
magma_free_cpu( work );
}
else {
/* Use hybrid blocked code. */
maxm = magma_roundup( m, 32 );
ldwork = maxm;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, ldwork*nb )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
for( j=0; j < s; j++ ) {
// get j-th panel from device
magma_queue_sync( queues[1] );
magma_sgetmatrix_async( m-j*nb, nb, dA(j,j), ldda, work, ldwork, queues[0] );
if ( j > 0 ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n - (j+1)*nb,
c_one, dA(j-1,j-1), ldda,
dA(j-1,j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-j*nb, n-(j+1)*nb, nb,
c_neg_one, dA(j, j-1), ldda,
dA(j-1,j+1), ldda,
c_one, dA(j, j+1), ldda, queues[1] );
}
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queues[0] );
magma_sgetrf_nopiv( rows, nb, work, ldwork, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
// send j-th panel to device
magma_ssetmatrix_async( m-j*nb, nb, work, ldwork, dA(j, j), ldda, queues[0] );
magma_queue_sync( queues[0] );
// do the small non-parallel computations (next panel update)
if ( s > j+1 ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dA(j, j ), ldda,
dA(j, j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(j+1)*nb, nb, nb,
c_neg_one, dA(j+1, j ), ldda,
dA(j, j+1), ldda,
c_one, dA(j+1, j+1), ldda, queues[1] );
}
else {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n-s*nb,
c_one, dA(j, j ), ldda,
dA(j, j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(j+1)*nb, n-(j+1)*nb, nb,
c_neg_one, dA(j+1, j ), ldda,
dA(j, j+1), ldda,
c_one, dA(j+1, j+1), ldda, queues[1] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
magma_sgetmatrix( rows, nb0, dA(s,s), ldda, work, ldwork, queues[1] );
// do the cpu part
magma_sgetrf_nopiv( rows, nb0, work, ldwork, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + s*nb;
// send j-th panel to device
magma_ssetmatrix( rows, nb0, work, ldwork, dA(s,s), ldda, queues[1] );
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb0, n-s*nb-nb0,
c_one, dA(s,s), ldda,
dA(s,s)+nb0, ldda, queues[1] );
}
magma_free_pinned( work );
}
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
return *info;
} /* magma_sgetrf_nopiv_gpu */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing slacpy_batched
Code is very similar to testing_sgeadd_batched.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];
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_B;
magmaFloat_ptr d_A, d_B;
float **hAarray, **hBarray, **dAarray, **dBarray;
magma_int_t M, N, mb, nb, size, lda, ldda, mstride, nstride, ntile;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts( MagmaOptsBatched );
opts.parse_opts( argc, argv );
mb = (opts.nb == 0 ? 32 : opts.nb);
nb = (opts.nb == 0 ? 64 : opts.nb);
mstride = 2*mb;
nstride = 3*nb;
printf("%% mb=%d, nb=%d, mstride=%d, nstride=%d\n", (int) mb, (int) nb, (int) mstride, (int) nstride );
printf("%% M N ntile CPU Gflop/s (ms) GPU Gflop/s (ms) check\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];
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
size = lda*N;
if ( N < nb || M < nb ) {
ntile = 0;
} else {
ntile = min( (M - nb)/mstride + 1,
(N - nb)/nstride + 1 );
}
gbytes = 2.*mb*nb*ntile / 1e9;
TESTING_MALLOC_CPU( h_A, float, lda *N );
TESTING_MALLOC_CPU( h_B, float, lda *N );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, ldda*N );
TESTING_MALLOC_CPU( hAarray, float*, ntile );
TESTING_MALLOC_CPU( hBarray, float*, ntile );
TESTING_MALLOC_DEV( dAarray, float*, ntile );
TESTING_MALLOC_DEV( dBarray, float*, ntile );
lapackf77_slarnv( &ione, ISEED, &size, h_A );
lapackf77_slarnv( &ione, ISEED, &size, h_B );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_ssetmatrix( M, N, h_A, lda, d_A, ldda );
magma_ssetmatrix( M, N, h_B, lda, d_B, ldda );
// setup pointers
for( magma_int_t tile = 0; tile < ntile; ++tile ) {
magma_int_t offset = tile*mstride + tile*nstride*ldda;
hAarray[tile] = &d_A[offset];
hBarray[tile] = &d_B[offset];
}
magma_setvector( ntile, sizeof(float*), hAarray, 1, dAarray, 1 );
magma_setvector( ntile, sizeof(float*), hBarray, 1, dBarray, 1 );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_slacpy_batched( MagmaFull, mb, nb, dAarray, ldda, dBarray, ldda, ntile, opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
for( magma_int_t tile = 0; tile < ntile; ++tile ) {
magma_int_t offset = tile*mstride + tile*nstride*lda;
lapackf77_slacpy( MagmaFullStr, &mb, &nb,
&h_A[offset], &lda,
&h_B[offset], &lda );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_sgetmatrix( M, N, d_B, ldda, h_A, lda );
blasf77_saxpy(&size, &c_neg_one, h_A, &ione, h_B, &ione);
error = lapackf77_slange("f", &M, &N, h_B, &lda, work);
bool okay = (error == 0);
status += ! okay;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) M, (int) N, (int) ntile,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(okay ? "ok" : "failed") );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_CPU( hAarray );
TESTING_FREE_CPU( hBarray );
TESTING_FREE_DEV( dAarray );
TESTING_FREE_DEV( dBarray );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
/**
Purpose
-------
SORGQR generates an M-by-N REAL matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by SGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A REAL array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by SGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau REAL array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by SGEQRF_GPU.
@param[in]
T REAL array, dimension (NB, min(M,N)).
T contains the T matrices used in blocking the elementary
reflectors H(i), e.g., this can be the 6th argument of
magma_sgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in SGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in T.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sorgqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
float *A, magma_int_t lda,
float *tau,
float *T, magma_int_t nb,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(d,i,j) (dA[d] + (i) + (j)*ldda)
#define dT(d,i,j) (dT[d] + (i) + (j)*nb)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldwork;
magma_int_t d, i, ib, j, jb, ki, kk;
float *work=NULL;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
magma_int_t di, dn;
magma_int_t dpanel;
magma_int_t ngpu = magma_num_gpus();
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
// Allocate memory on GPUs for A and workspaces
magma_int_t ldda = magma_roundup( m, 32 );
magma_int_t lddwork = magma_roundup( n, 32 );
magma_int_t min_lblocks = (n / nb) / ngpu; // min. blocks per gpu
magma_int_t last_dev = (n / nb) % ngpu; // device with last block
magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 };
float *dA[ MagmaMaxGPUs ] = { NULL };
float *dT[ MagmaMaxGPUs ] = { NULL };
float *dV[ MagmaMaxGPUs ] = { NULL };
float *dW[ MagmaMaxGPUs ] = { NULL };
magma_queue_t queues[ MagmaMaxGPUs ] = { NULL };
for( d = 0; d < ngpu; ++d ) {
// example with n = 75, nb = 10, ngpu = 3
// min_lblocks = 2
// last_dev = 1
// gpu 0: 2 blocks, cols: 0- 9, 30-39, 60-69
// gpu 1: 1+ blocks, cols: 10-19, 40-49, 70-74 (partial)
// gpu 2: 1 block, cols: 20-29, 50-59
magma_setdevice( d );
nlocal[d] = min_lblocks*nb;
if ( d < last_dev ) {
nlocal[d] += nb;
}
else if ( d == last_dev ) {
nlocal[d] += (n % nb);
}
ldwork = nlocal[d]*ldda // dA
+ nb*m // dT
+ nb*ldda // dV
+ nb*lddwork; // dW
if ( MAGMA_SUCCESS != magma_smalloc( &dA[d], ldwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
dT[d] = dA[d] + nlocal[d]*ldda;
dV[d] = dT[d] + nb*m;
dW[d] = dV[d] + nb*ldda;
magma_queue_create( d, &queues[d] );
}
trace_init( 1, ngpu, 1, queues );
// first kk columns are handled by blocked method.
// ki is start of 2nd-to-last block
if ((nb > 1) && (nb < k)) {
ki = (k - nb - 1) / nb * nb;
kk = min(k, ki + nb);
} else {
ki = 0;
kk = 0;
}
// Allocate CPU work space
// n*nb for larfb work
// m*nb for V
// nb*nb for T
lwork = (n + m + nb) * nb;
magma_smalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
float *work_T, *work_V;
work_T = work + n*nb;
work_V = work + n*nb + nb*nb;
// Use unblocked code for the last or only block.
if (kk < n) {
trace_cpu_start( 0, "ungqr", "ungqr last block" );
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
// sorgqr requires less workspace (n*nb), but is slow if k < sorgqr's block size.
// replacing it with the 4 routines below is much faster (e.g., 60x).
//magma_int_t iinfo;
//lapackf77_sorgqr( &m_kk, &n_kk, &k_kk,
// A(kk, kk), &lda,
// &tau[kk], work, &lwork, &iinfo );
lapackf77_slacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, work_V, &m_kk);
lapackf77_slaset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
work_V, &m_kk, &tau[kk], work_T, &k_kk);
lapackf77_slarfb( MagmaLeftStr, MagmaNoTransStr, MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &n_kk, &k_kk,
work_V, &m_kk, work_T, &k_kk, A(kk, kk), &lda, work, &n_kk );
if (kk > 0) {
for( j=kk; j < n; j += nb ) {
jb = min( n-j, nb );
d = (j / nb) % ngpu;
di = ((j / nb) / ngpu) * nb;
magma_setdevice( d );
magma_ssetmatrix( m_kk, jb,
A(kk, j), lda,
dA(d, kk, di), ldda, queues[d] );
// Set A(1:kk,kk+1:n) to zero.
magmablas_slaset( MagmaFull, kk, jb, c_zero, c_zero, dA(d, 0, di), ldda, queues[d] );
}
}
trace_cpu_end( 0 );
}
if (kk > 0) {
// Use blocked code
// send T to all GPUs
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set T" );
magma_ssetmatrix_async( nb, min(m,n), T, nb, dT[d], nb, queues[d] );
trace_gpu_end( d, 0 );
}
// queue: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
for( i = ki; i >= 0; i -= nb ) {
ib = min(nb, k - i);
mi = m - i;
dpanel = (i / nb) % ngpu;
di = ((i / nb) / ngpu) * nb;
// Send current panel to dV on the GPUs
lapackf77_slaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set V" );
magma_ssetmatrix_async( mi, ib,
A(i, i), lda,
dV[d], ldda, queues[d] );
trace_gpu_end( d, 0 );
}
// set panel to identity
magma_setdevice( dpanel );
trace_gpu_start( dpanel, 0, "laset", "laset" );
magmablas_slaset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda, queues[dpanel] );
magmablas_slaset( MagmaFull, mi, ib, c_zero, c_one, dA(dpanel, i, di), ldda, queues[dpanel] );
trace_gpu_end( dpanel, 0 );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_indices_1D_bcyclic( nb, ngpu, d, i, n, &di, &dn );
trace_gpu_start( d, 0, "larfb", "larfb" );
magma_slarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, dn-di, ib,
dV[d], ldda, dT(d,0,i), nb,
dA(d, i, di), ldda, dW[d], lddwork, queues[d] );
trace_gpu_end( d, 0 );
}
}
}
// copy result back to CPU
trace_cpu_start( 0, "get", "get A" );
magma_sgetmatrix_1D_col_bcyclic( m, n, dA, ldda, A, lda, ngpu, nb, queues );
trace_cpu_end( 0 );
}
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "sorgqr-n%d-ngpu%d.svg", m, ngpu );
trace_finalize( name, "trace.css" );
#endif
cleanup:
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( dA[d] );
magma_queue_destroy( queues[d] );
}
magma_free_cpu( work );
magma_setdevice( orig_dev );
return *info;
} /* magma_sorgqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dlag2s and slag2d
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
float serror, swork[1];
double c_neg_one = MAGMA_D_NEG_ONE;
float s_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t m, n, lda, ldda, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
float *SA, *SR;
double *A, *R;
magmaFloat_ptr dSA;
magmaDouble_ptr dA;
magma_opts opts;
opts.parse_opts( argc, argv );
printf("%% func M N CPU GB/s (ms) GPU GB/s (ms) ||R||_F\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];
lda = m;
ldda = magma_roundup( m, opts.align ); // multiple of 32 by default
// m*n double-real loads and m*n single-real stores (and vice-versa for slag2d)
gbytes = (real_Double_t) m*n * (sizeof(double) + sizeof(float)) / 1e9;
size = ldda*n; // ldda >= lda
TESTING_MALLOC_CPU( SA, float, size );
TESTING_MALLOC_CPU( A, double, size );
TESTING_MALLOC_CPU( SR, float, size );
TESTING_MALLOC_CPU( R, double, size );
TESTING_MALLOC_DEV( dSA, float, size );
TESTING_MALLOC_DEV( dA, double, size );
lapackf77_dlarnv( &ione, ISEED, &size, A );
lapackf77_slarnv( &ione, ISEED, &size, SA );
magma_dsetmatrix( m, n, A, lda, dA, ldda, opts.queue );
magma_ssetmatrix( m, n, SA, lda, dSA, ldda, opts.queue );
/* =====================================================================
Performs operation using LAPACK dlag2s
=================================================================== */
cpu_time = magma_wtime();
lapackf77_dlag2s( &m, &n, A, &lda, SA, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if (info != 0) {
printf("lapackf77_dlag2s returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA dlag2s
=================================================================== */
gpu_time = magma_sync_wtime( opts.queue );
magmablas_dlag2s( m, n, dA, ldda, dSA, ldda, opts.queue, &info );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
if (info != 0) {
printf("magmablas_dlag2s returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_sgetmatrix( m, n, dSA, ldda, SR, lda, opts.queue );
/* =====================================================================
compute error |SA_magma - SA_lapack|
should be zero if both are IEEE compliant
=================================================================== */
blasf77_saxpy( &size, &s_neg_one, SA, &ione, SR, &ione );
serror = lapackf77_slange( "Fro", &m, &n, SR, &lda, swork );
printf( "dlag2s %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
serror, (serror == 0 ? "ok" : "failed") );
status += ! (serror == 0);
/* =====================================================================
Reset matrices
=================================================================== */
lapackf77_dlarnv( &ione, ISEED, &size, A );
lapackf77_slarnv( &ione, ISEED, &size, SA );
magma_dsetmatrix( m, n, A, lda, dA, ldda, opts.queue );
magma_ssetmatrix( m, n, SA, lda, dSA, ldda, opts.queue );
/* =====================================================================
Performs operation using LAPACK slag2d
=================================================================== */
cpu_time = magma_wtime();
lapackf77_slag2d( &m, &n, SA, &lda, A, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if (info != 0) {
printf("lapackf77_slag2d returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA slag2d
=================================================================== */
magma_ssetmatrix( m, n, SA, lda, dSA, ldda, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_slag2d( m, n, dSA, ldda, dA, ldda, opts.queue, &info );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
if (info != 0) {
printf("magmablas_slag2d returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_dgetmatrix( m, n, dA, ldda, R, lda, opts.queue );
/* =====================================================================
compute error |A_magma - A_lapack|
should be zero if both are IEEE compliant
=================================================================== */
blasf77_daxpy( &size, &c_neg_one, A, &ione, R, &ione );
error = lapackf77_dlange( "Fro", &m, &n, R, &lda, work );
printf( "slag2d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
error, (error == 0 ? "ok" : "failed") );
status += ! (error == 0);
TESTING_FREE_CPU( SA );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( SR );
TESTING_FREE_CPU( R );
TESTING_FREE_DEV( dSA );
TESTING_FREE_DEV( dA );
printf( "\n" );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsyr2k
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double cublas_error, Cnorm, work[1];
magma_int_t N, K;
magma_int_t Ak, An, Bk, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double *h_A, *h_B, *h_C, *h_Ccublas;
magmaDouble_ptr d_A, d_B, d_C;
double c_neg_one = MAGMA_D_NEG_ONE;
double alpha = MAGMA_D_MAKE( 0.29, -0.86 );
double beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
#ifdef COMPLEX
if (opts.transA == MagmaTrans) {
opts.transA = MagmaConjTrans;
printf("%% WARNING: transA = MagmaTrans changed to MagmaConjTrans\n");
}
#endif
printf("%% If running lapack (option --lapack), CUBLAS error is computed\n"
"%% relative to CPU BLAS result.\n\n");
printf("%% uplo = %s, transA = %s\n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA) );
printf("%% N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("%%=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.msize[itest];
K = opts.ksize[itest];
gflops = FLOPS_DSYR2K(K, N) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = An = N;
Ak = K;
ldb = Bn = N;
Bk = K;
} else {
lda = An = K;
Ak = N;
ldb = Bn = K;
Bk = N;
}
ldc = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
lddb = magma_roundup( ldb, opts.align ); // multiple of 32 by default
lddc = magma_roundup( ldc, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
sizeB = ldb*Ak;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, double, lda*Ak );
TESTING_MALLOC_CPU( h_B, double, ldb*Bk );
TESTING_MALLOC_CPU( h_C, double, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, double, ldc*N );
TESTING_MALLOC_DEV( d_A, double, ldda*Ak );
TESTING_MALLOC_DEV( d_B, double, lddb*Bk );
TESTING_MALLOC_DEV( d_C, double, lddc*N );
/* Initialize the matrices */
lapackf77_dlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_dlarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_dsetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_dsetmatrix( Bn, Bk, h_B, ldb, d_B, lddb );
magma_dsetmatrix( N, N, h_C, ldc, d_C, lddc );
magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasDsyr2k( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA), N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
#else
magma_dsyr2k( opts.uplo, opts.transA, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_dgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_dsyr2k( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_dlange( "M", &N, &N, h_C, &ldc, work );
blasf77_daxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_dlange( "M", &N, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Ccublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgetrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error;
double *h_A, *h_R;
magmaDouble_ptr d_A;
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, info, min_mn;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% M N CPU Gflop/s (ms) GPU Gflop/s (ms) Copy time (ms) ||PA-LU||/(||A||*N)\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;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_DGETRF( M, N ) / 1e9;
if ( N > 512 ) {
printf( "%5d %5d skipping because dgetf2 does not support N > 512\n", (int) M, (int) N );
continue;
}
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_PIN( h_R, double, n2 );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
real_Double_t set_time = magma_wtime();
magma_dsetmatrix( M, N, h_R, lda, d_A, ldda, opts.queue );
set_time = magma_wtime() - set_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_dgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime( opts.queue );
magma_dgetf2_gpu( M, N, d_A, ldda, ipiv, opts.queue, &info );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgetf2_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
real_Double_t get_time = magma_wtime();
magma_dgetmatrix( M, N, d_A, ldda, h_A, lda, opts.queue );
get_time = magma_wtime() - get_time;
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f",
(int) M, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
set_time*1000.+get_time*1000.);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %7.2f",
(int) M, (int) N, gpu_perf, gpu_time*1000., set_time*1000.+get_time*1000. );
}
if ( opts.check ) {
magma_dgetmatrix( M, N, d_A, ldda, h_A, lda, opts.queue );
error = get_LU_error( M, N, h_R, lda, h_A, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
}
else {
printf(" --- \n");
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
It uses 2 queues to overlap communication and computation.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
ipiv INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *ipiv,
magma_int_t *info)
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, ((i_)*nb + (j_)*nb*ldda + dA_offset)
#define dAT(i_, j_) dAT, ((i_)*nb*lddat + (j_)*nb + dAT_offset)
#define dwork(i_) dwork, (i_)
#else
#define dA(i_, j_) ( dA + (i_)*nb + (j_)*nb*ldda)
#define dAT(i_, j_) ( dAT + (i_)*nb*lddat + (j_)*nb)
#define dwork(i_) (dwork + (i_))
#endif
// Constants
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
// Local variables
float *work;
magmaFloat_ptr dA, dAT, dwork;
magma_int_t iinfo, nb;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
nb = magma_get_sgetrf_nb( m, n );
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_sgetrf( &m, &n, A, &lda, ipiv, info );
}
else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, lddat, maxdim;
magma_int_t i, j, rows, cols, s = min(m, n)/nb;
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
maxdim = max( maxm, maxn );
lddat = maxn;
ldda = maxm;
/* set number of GPUs */
magma_int_t ngpu = magma_num_gpus();
if ( ngpu > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magma_queue_t queues[2] = { NULL, NULL };
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
/* check the memory requirement */
size_t mem_size = magma_queue_mem_size( queues[0] );
mem_size /= sizeof(float);
magma_int_t h = 1+(2+ngpu);
magma_int_t ngpu2 = ngpu;
magma_int_t NB = (magma_int_t)(0.8*mem_size/maxm - h*nb);
const char* ngr_nb_char = getenv("MAGMA_NGR_NB");
if ( ngr_nb_char != NULL )
NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
if ( ngpu > ceil((float)NB/nb) ) {
ngpu2 = (magma_int_t)ceil((float)NB/nb);
h = 1+(2+ngpu2);
NB = (magma_int_t)(0.8*mem_size/maxm - h*nb);
}
if ( ngpu2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
// dwork is nb*maxm for panel, and maxdim*maxdim for A
if (MAGMA_SUCCESS != magma_smalloc( &dwork, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
ldda = lddat = maxdim;
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
dAT = dA;
magmablas_stranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
// dwork is nb*maxm for panel, and maxm*maxn for A
if (MAGMA_SUCCESS != magma_smalloc( &dwork, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
if (MAGMA_SUCCESS != magma_smalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dwork );
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magmablas_stranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queues[0] );
}
lapackf77_sgetrf( &m, &nb, work, &lda, ipiv, &iinfo );
for( j = 0; j < s; j++ ) {
// get j-th panel from device
cols = maxm - j*nb;
if (j > 0) {
magmablas_stranspose( nb, cols, dAT(j,j), lddat, dwork(0), cols, queues[0] );
magma_queue_sync( queues[0] );
magma_sgetmatrix_async( m-j*nb, nb, dwork(0), cols, work, lda, queues[1] );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queues[0] );
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queues[1] );
lapackf77_sgetrf( &rows, &nb, work, &lda, ipiv+j*nb, &iinfo );
}
if (*info == 0 && iinfo > 0)
*info = iinfo + j*nb;
// put j-th panel onto device
magma_ssetmatrix_async( m-j*nb, nb, work, lda, dwork(0), cols, queues[1] );
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_slaswp( n, dAT(0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queues[0] );
magma_queue_sync( queues[1] );
magmablas_stranspose( cols, nb, dwork(0), cols, dAT(j,j), lddat, queues[0] );
// do the small non-parallel computations (next panel update)
if (s > (j+1)) {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
else {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_stranspose( nb0, rows, dAT(s,s), lddat, dwork(0), cols, queues[0] );
magma_sgetmatrix_async( rows, nb0, dwork(0), cols, work, lda, queues[0] );
magma_queue_sync( queues[0] );
// do the cpu part
lapackf77_sgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo );
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_slaswp( n, dAT(0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queues[0] );
// put j-th panel onto device
magma_ssetmatrix_async( rows, nb0, work, lda, dwork(0), cols, queues[0] );
magmablas_stranspose( rows, nb0, dwork(0), cols, dAT(s,s), lddat, queues[0] );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), lddat,
dAT(s, s)+nb0, lddat, queues[0] );
}
// undo transpose
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_stranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
magma_sgetmatrix( m, n, dAT(0,0), lddat, A, lda, queues[0] );
}
else {
magmablas_stranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queues[0] );
magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queues[0] );
magma_free( dAT );
}
magma_free( dwork );
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
}
return *info;
} /* magma_sgetrf */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlaset
Code is very similar to testing_zlacpy.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex_ptr d_A;
magmaDoubleComplex 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_Z_MAKE( 1.2345, 6.7890 ) * (ival / 2);
diag = MAGMA_Z_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(magmaDoubleComplex) * (1.*M*N - 0.5*N*(N-1)) / 1e9;
} else {
gbytes = sizeof(magmaDoubleComplex) * 0.5*M*(M+1) / 1e9;
}
}
else if ( uplo[iuplo] == MagmaUpper ) {
// save upper trapezoid (with diagonal)
if ( N > M ) {
gbytes = sizeof(magmaDoubleComplex) * (1.*M*N - 0.5*M*(M-1)) / 1e9;
} else {
gbytes = sizeof(magmaDoubleComplex) * 0.5*N*(N+1) / 1e9;
}
}
else {
// save entire matrix
gbytes = sizeof(magmaDoubleComplex) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, size );
TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, size );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, 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_Z_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix( M, N, h_A, lda, d_A, ldda );
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
//magmablas_zlaset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_zlaset( 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_zlaset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_zlaset( 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_zprint( M, N, h_A, lda );
printf( "dA=" ); magma_zprint_gpu( M, N, d_A, ldda );
}
/* =====================================================================
Check the result
=================================================================== */
magma_zgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_zaxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_zlange("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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssymmetrize
Code is very similar to testing_stranspose.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];
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_R;
magmaFloat_ptr d_A;
magma_int_t N, size, lda, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N CPU GByte/s (ms) GPU GByte/s (ms) check\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( N, opts.align ); // multiple of 32 by default
size = lda*N;
// load strictly lower triangle, save strictly upper triangle
gbytes = sizeof(float) * 1.*N*(N-1) / 1e9;
TESTING_MALLOC_CPU( h_A, float, size );
TESTING_MALLOC_CPU( h_R, float, size );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < N; ++i ) {
h_A[i + j*lda] = MAGMA_S_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_ssetmatrix( N, N, h_A, lda, d_A, ldda );
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
//magmablas_ssymmetrize( opts.uplo, N-2, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_ssymmetrize( opts.uplo, N, d_A, ldda );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using naive in-place algorithm
(LAPACK doesn't implement symmetrize)
=================================================================== */
cpu_time = magma_wtime();
//for( int j = 1; j < N-1; ++j ) { // inset by 1 row & col
// for( int i = 1; i < j; ++i ) {
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < j; ++i ) {
if ( opts.uplo == MagmaLower ) {
h_A[i + j*lda] = MAGMA_S_CONJ( h_A[j + i*lda] );
}
else {
h_A[j + i*lda] = MAGMA_S_CONJ( h_A[i + j*lda] );
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_sgetmatrix( N, N, d_A, ldda, h_R, lda );
blasf77_saxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_slange("f", &N, &N, h_R, &lda, work);
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing strmv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float cublas_error, Cnorm, work[1];
magma_int_t N;
magma_int_t Ak;
magma_int_t sizeA;
magma_int_t lda, ldda;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_x, *h_xcublas;
magmaFloat_ptr d_A, d_x;
float c_neg_one = MAGMA_S_NEG_ONE;
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("%% If running lapack (option --lapack), CUBLAS error is computed\n"
"%% relative to CPU BLAS result.\n\n");
printf("%% uplo = %s, transA = %s, diag = %s \n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA),
lapack_diag_const(opts.diag) );
printf("%% N CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("%%=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
gflops = FLOPS_STRMM(opts.side, N, 1) / 1e9;
lda = N;
Ak = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
TESTING_MALLOC_CPU( h_A, float, lda*Ak );
TESTING_MALLOC_CPU( h_x, float, N );
TESTING_MALLOC_CPU( h_xcublas, float, N );
TESTING_MALLOC_DEV( d_A, float, ldda*Ak );
TESTING_MALLOC_DEV( d_x, float, N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &N, h_x );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_ssetmatrix( Ak, Ak, h_A, lda, d_A, ldda, opts.queue );
magma_ssetvector( N, h_x, 1, d_x, 1, opts.queue );
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasStrmv( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA),
cublas_diag_const(opts.diag),
N,
d_A, ldda,
d_x, 1 );
#else
magma_strmv( opts.uplo, opts.transA, opts.diag,
N,
d_A, 0, ldda,
d_x, 0, 1, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_sgetvector( N, d_x, 1, h_xcublas, 1, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_strmv( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&N,
h_A, &lda,
h_x, &ione );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_slange( "M", &N, &ione, h_x, &N, work );
blasf77_saxpy( &N, &c_neg_one, h_x, &ione, h_xcublas, &ione );
cublas_error = lapackf77_slange( "M", &N, &ione, h_xcublas, &N, work ) / Cnorm;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_x );
TESTING_FREE_CPU( h_xcublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_x );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgetri_batched
*/
int main( int argc, char** argv)
{
TESTING_INIT();
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_Ainv, *h_R, *work;
magmaDoubleComplex_ptr d_A, d_invA;
magmaDoubleComplex_ptr *dA_array;
magmaDoubleComplex_ptr *dinvA_array;
magma_int_t **dipiv_array;
magma_int_t *dinfo_array;
magma_int_t *ipiv, *cpu_info;
magma_int_t *d_ipiv, *d_info;
magma_int_t N, n2, lda, ldda, info, info1, info2, lwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex tmp;
double error, rwork[1];
magma_int_t columns;
magma_int_t status = 0;
magma_opts opts( MagmaOptsBatched );
opts.parse_opts( argc, argv );
magma_int_t batchCount = opts.batchcount;
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% batchCount N CPU Gflop/s (ms) GPU Gflop/s (ms) ||I - A*A^{-1}||_1 / (N*cond(A))\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N * batchCount;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
// This is the correct flops but since this getri_batched is based on
// 2 trsm = getrs and to know the real flops I am using the getrs one
//gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRI( N ))/ 1e9 * batchCount;
gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, N ))/ 1e9 * batchCount;
// query for workspace size
lwork = -1;
lapackf77_zgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0) {
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = magma_int_t( MAGMA_Z_REAL( tmp ));
TESTING_MALLOC_CPU( cpu_info, magma_int_t, batchCount );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N * batchCount );
TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork*batchCount );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_Ainv, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N * batchCount );
TESTING_MALLOC_DEV( d_invA, magmaDoubleComplex, ldda*N * batchCount );
TESTING_MALLOC_DEV( d_ipiv, magma_int_t, N * batchCount );
TESTING_MALLOC_DEV( d_info, magma_int_t, batchCount );
TESTING_MALLOC_DEV( dA_array, magmaDoubleComplex*, batchCount );
TESTING_MALLOC_DEV( dinvA_array, magmaDoubleComplex*, batchCount );
TESTING_MALLOC_DEV( dinfo_array, magma_int_t, batchCount );
TESTING_MALLOC_DEV( dipiv_array, magma_int_t*, batchCount );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
columns = N * batchCount;
lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_Ainv, &lda );
magma_zsetmatrix( N, columns, h_R, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zset_pointer( dA_array, d_A, ldda, 0, 0, ldda * N, batchCount, opts.queue );
magma_zset_pointer( dinvA_array, d_invA, ldda, 0, 0, ldda * N, batchCount, opts.queue );
magma_iset_pointer( dipiv_array, d_ipiv, 1, 0, 0, N, batchCount, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
info1 = magma_zgetrf_batched( N, N, dA_array, ldda, dipiv_array, dinfo_array, batchCount, opts.queue);
info2 = magma_zgetri_outofplace_batched( N, dA_array, ldda, dipiv_array, dinvA_array, ldda, dinfo_array, batchCount, opts.queue);
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
// check correctness of results throught "dinfo_magma" and correctness of argument throught "info"
magma_getvector( batchCount, sizeof(magma_int_t), dinfo_array, 1, cpu_info, 1, opts.queue );
for (magma_int_t i=0; i < batchCount; i++)
{
if (cpu_info[i] != 0 ) {
printf("magma_zgetrf_batched matrix %d returned error %d\n", (int) i, (int)cpu_info[i] );
}
}
if (info1 != 0) printf("magma_zgetrf_batched returned argument error %d: %s.\n", (int) info1, magma_strerror( info1 ));
if (info2 != 0) printf("magma_zgetri_batched returned argument error %d: %s.\n", (int) info2, magma_strerror( info2 ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
#if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP)
magma_int_t nthreads = magma_get_lapack_numthreads();
magma_set_lapack_numthreads(1);
magma_set_omp_numthreads(nthreads);
#pragma omp parallel for schedule(dynamic)
#endif
for (int i=0; i < batchCount; i++)
{
magma_int_t locinfo;
lapackf77_zgetrf(&N, &N, h_Ainv + i*lda*N, &lda, ipiv + i*N, &locinfo);
if (locinfo != 0) {
printf("lapackf77_zgetrf returned error %d: %s.\n",
(int) locinfo, magma_strerror( locinfo ));
}
lapackf77_zgetri(&N, h_Ainv + i*lda*N, &lda, ipiv + i*N, work + i*lwork, &lwork, &locinfo );
if (locinfo != 0) {
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) locinfo, magma_strerror( locinfo ));
}
}
#if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP)
magma_set_lapack_numthreads(nthreads);
#endif
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
printf("%10d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) batchCount, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000. );
}
else {
printf("%10d %5d --- ( --- ) %7.2f (%7.2f)",
(int) batchCount, (int) N, gpu_perf, gpu_time*1000. );
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
magma_igetvector( N*batchCount, d_ipiv, 1, ipiv, 1, opts.queue );
magma_zgetmatrix( N, N*batchCount, d_invA, ldda, h_Ainv, lda, opts.queue );
error = 0;
for (magma_int_t i=0; i < batchCount; i++)
{
for (magma_int_t k=0; k < N; k++) {
if (ipiv[i*N+k] < 1 || ipiv[i*N+k] > N )
{
printf("error for matrix %d ipiv @ %d = %d\n", (int) i, (int) k, (int) ipiv[i*N+k]);
error = -1;
}
}
if (error == -1) {
break;
}
// compute 1-norm condition number estimate, following LAPACK's zget03
double normA, normAinv, rcond, err;
normA = lapackf77_zlange( "1", &N, &N, h_A + i*lda*N, &lda, rwork );
normAinv = lapackf77_zlange( "1", &N, &N, h_Ainv + i*lda*N, &lda, rwork );
if ( normA <= 0 || normAinv <= 0 ) {
rcond = 0;
err = 1 / (tol/opts.tolerance); // == 1/eps
}
else {
rcond = (1 / normA) / normAinv;
// R = I
// R -= A*A^{-1}
// err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm
lapackf77_zlaset( "full", &N, &N, &c_zero, &c_one, h_R + i*lda*N, &lda );
blasf77_zgemm( "no", "no", &N, &N, &N, &c_neg_one,
h_A + i*lda*N, &lda,
h_Ainv + i*lda*N, &lda, &c_one,
h_R + i*lda*N, &lda );
err = lapackf77_zlange( "1", &N, &N, h_R + i*lda*N, &lda, rwork );
err = err * rcond / N;
}
if ( isnan(err) || isinf(err) ) {
error = err;
break;
}
error = max( err, error );
}
bool okay = (error < tol);
status += ! okay;
printf(" %8.2e %s\n", error, (okay ? "ok" : "failed") );
}
else {
printf("\n");
}
TESTING_FREE_CPU( cpu_info );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_Ainv );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_invA );
TESTING_FREE_DEV( d_ipiv );
TESTING_FREE_DEV( d_info );
TESTING_FREE_DEV( dA_array );
TESTING_FREE_DEV( dinvA_array );
TESTING_FREE_DEV( dinfo_array );
TESTING_FREE_DEV( dipiv_array );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
double magma_error, dev_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t M, N, Xm, Ym, lda, ldda, sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ydev, *Ymagma;
magmaDoubleComplex_ptr dA, dX, dY;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% trans = %s\n", lapack_trans_const(opts.transA) );
#ifdef HAVE_CUBLAS
printf("%% M N MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
printf("%% M N %s Gflop/s (ms) CPU Gflop/s (ms) %s error\n",
g_platform_str, g_platform_str );
#endif
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];
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_ZGEMV( M, N ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N;
sizeX = incx*Xm;
sizeY = incy*Ym;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ydev, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( M, N, A, lda, dA, ldda, opts.queue );
magma_zsetvector( Xm, X, incx, dX, incx, opts.queue );
magma_zsetvector( Ym, Y, incy, dY, incy, opts.queue );
dev_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasZgemv( opts.handle, cublas_trans_const(opts.transA),
M, N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
#else
magma_zgemv( opts.transA, M, N,
alpha, dA, ldda,
dX, incx,
beta, dY, incy );
#endif
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
dev_perf = gflops / dev_time;
magma_zgetvector( Ym, dY, incy, Ydev, incy, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetvector( Ym, Y, incy, dY, incy, opts.queue );
magma_time = magma_sync_wtime( opts.queue );
magmablas_zgemv( opts.transA, M, N, alpha, dA, ldda, dX, incx, beta, dY, incy, opts.queue );
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( Ym, dY, incy, Ymagma, incy, opts.queue );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zgemv( lapack_trans_const(opts.transA), &M, &N,
&alpha, A, &lda,
X, &incx,
&beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
double Anorm = lapackf77_zlange( "F", &M, &N, A, &lda, work );
double Xnorm = lapackf77_zlange( "F", &Xm, &ione, X, &Xm, work );
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ydev, &incy );
dev_error = lapackf77_zlange( "F", &Ym, &ione, Ydev, &Ym, work ) / (Anorm * Xnorm);
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "F", &Ym, &ione, Ymagma, &Ym, work ) / (Anorm * Xnorm);
bool okay = (magma_error < tol) && (dev_error < tol);
status += ! okay;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(okay ? "ok" : "failed"));
#else
bool okay = (dev_error < tol);
status += ! okay;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(okay ? "ok" : "failed"));
#endif
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ydev );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
DGEHRD reduces a DOUBLE PRECISION general matrix A to upper Hessenberg form H by
an orthogonal similarity transformation: Q' * A * Q = H . This version
stores the triangular matrices used in the factorization so that they can
be applied directly (i.e., without being recomputed) later. As a result,
the application of Q is much faster.
Arguments
---------
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in]
ilo INTEGER
@param[in]
ihi INTEGER
It is assumed that A is already upper triangular in rows
and columns 1:ILO-1 and IHI+1:N. ILO and IHI are normally
set by a previous call to DGEBAL; otherwise they should be
set to 1 and N respectively. See Further Details.
1 <= ILO <= IHI <= N, if N > 0; ILO=1 and IHI=0, if N=0.
@param[in,out]
A DOUBLE PRECISION array, dimension (LDA,N)
On entry, the N-by-N general matrix to be reduced.
On exit, the upper triangle and the first subdiagonal of A
are overwritten with the upper Hessenberg matrix H, and the
elements below the first subdiagonal, with the array TAU,
represent the orthogonal matrix Q as a product of elementary
reflectors. See Further Details.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
tau DOUBLE PRECISION array, dimension (N-1)
The scalar factors of the elementary reflectors (see Further
Details). Elements 1:ILO-1 and IHI:N-1 of TAU are set to
zero.
@param[out]
work (workspace) DOUBLE PRECISION array, dimension (LWORK)
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The length of the array WORK. LWORK >= N*NB,
where NB is the optimal blocksize.
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued by XERBLA.
@param[out]
dT DOUBLE PRECISION array on the GPU, dimension NB*N,
where NB is the optimal blocksize. It stores the NB*NB blocks
of the triangular T matrices used in the reduction.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value.
Further Details
---------------
The matrix Q is represented as a product of (ihi-ilo) elementary
reflectors
Q = H(ilo) H(ilo+1) . . . H(ihi-1).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i) = 0, v(i+1) = 1 and v(ihi+1:n) = 0; v(i+2:ihi) is stored on
exit in A(i+2:ihi,i), and tau in TAU(i).
The contents of A are illustrated by the following example, with
n = 7, ilo = 2 and ihi = 6:
@verbatim
on entry, on exit,
( a a a a a a a ) ( a a h h h h a )
( a a a a a a ) ( a h h h h a )
( a a a a a a ) ( h h h h h h )
( a a a a a a ) ( v2 h h h h h )
( a a a a a a ) ( v2 v3 h h h h )
( a a a a a a ) ( v2 v3 v4 h h h )
( a ) ( a )
@endverbatim
where a denotes an element of the original matrix A, h denotes a
modified element of the upper Hessenberg matrix H, and vi denotes an
element of the vector defining H(i).
This implementation follows the hybrid algorithm and notations described in
S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg
form through hybrid GPU-based computing," University of Tennessee Computer
Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219),
May 24, 2009.
This version stores the T matrices in dT, for later use in magma_dorghr.
@ingroup magma_dgeev_comp
********************************************************************/
extern "C" magma_int_t
magma_dgehrd(
magma_int_t n, magma_int_t ilo, magma_int_t ihi,
double *A, magma_int_t lda,
double *tau,
double *work, magma_int_t lwork,
magmaDouble_ptr dT,
magma_int_t *info)
{
#define A(i_,j_) ( A + (i_) + (j_)*lda)
#ifdef HAVE_clBLAS
#define dA(i_,j_) dwork, ((i_) + (j_)*ldda + nb*ldda*2)
#define dT(i_,j_) dT, ((i_) + (j_)*nb + dT_offset)
#define dV(i_,j_) dwork, ((i_) + (j_)*ldda + nb*ldda)
#define dwork(i_) dwork, ((i_))
#else
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
#define dT(i_,j_) (dT + (i_) + (j_)*nb)
#define dV(i_,j_) (dV + (i_) + (j_)*ldda)
#define dwork(i_) (dwork + (i_))
#endif
// Constants
const double c_one = MAGMA_D_ONE;
const double c_zero = MAGMA_D_ZERO;
// Local variables
magma_int_t nb = magma_get_dgehrd_nb( n );
magma_int_t ldda = magma_roundup( n, 32 );
magma_int_t i, nh, iws;
magma_int_t iinfo;
magma_int_t lquery;
*info = 0;
iws = n*nb;
work[0] = magma_dmake_lwork( iws );
lquery = (lwork == -1);
if (n < 0) {
*info = -1;
} else if (ilo < 1 || ilo > max(1,n)) {
*info = -2;
} else if (ihi < min(ilo,n) || ihi > n) {
*info = -3;
} else if (lda < max(1,n)) {
*info = -5;
} else if (lwork < iws && ! lquery) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
// Adjust from 1-based indexing
ilo -= 1;
// Quick return if possible
nh = ihi - ilo;
if (nh <= 1) {
work[0] = c_one;
return *info;
}
// Now requires lwork >= iws; else dT won't be computed in unblocked code.
// If not enough workspace, use unblocked code
//if ( lwork < iws ) {
// nb = 1;
//}
if (nb == 1 || nb > nh) {
// Use unblocked code below
i = ilo;
}
else {
// Use blocked code
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// GPU workspace is:
// nb*ldda for dwork for dlahru
// nb*ldda for dV
// n*ldda for dA
magmaDouble_ptr dwork;
if (MAGMA_SUCCESS != magma_dmalloc( &dwork, 2*nb*ldda + n*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
double *dV = dwork + nb*ldda;
double *dA = dwork + nb*ldda*2;
double *T;
magma_dmalloc_cpu( &T, nb*nb );
if ( T == NULL ) {
magma_free( dwork );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
// zero first block of V, which is lower triangular
magmablas_dlaset( MagmaFull, nb, nb, c_zero, c_zero, dV(0,0), ldda, queue );
// Set elements 0:ILO-1 and IHI-1:N-2 of TAU to zero
for (i = 0; i < ilo; ++i)
tau[i] = c_zero;
for (i = max(0,ihi-1); i < n-1; ++i)
tau[i] = c_zero;
assert( nb % 4 == 0 );
for (i=0; i < nb*nb; i += 4)
T[i] = T[i+1] = T[i+2] = T[i+3] = c_zero;
magmablas_dlaset( MagmaFull, nb, n, c_zero, c_zero, dT(0,0), nb, queue );
// Copy the matrix to the GPU
magma_dsetmatrix( n, n-ilo, A(0,ilo), lda, dA(0,0), ldda, queue );
for (i = ilo; i < ihi-1 - nb; i += nb) {
// Reduce columns i:i+nb-1 to Hessenberg form, returning the
// matrices V and T of the block reflector H = I - V*T*V'
// which performs the reduction, and also the matrix Y = A*V*T
// Get the current panel (no need for the 1st iteration)
magma_dgetmatrix( ihi-i, nb,
dA(i,i-ilo), ldda,
A(i,i), lda, queue );
// add 1 to i for 1-based index
magma_dlahr2( ihi, i+1, nb,
dA(0,i-ilo), ldda,
dV(0,0), ldda,
A(0,i), lda,
&tau[i], T, nb, work, n, queue );
// Copy T from the CPU to dT on the GPU
magma_dsetmatrix( nb, nb, T, nb, dT(0,i-ilo), nb, queue );
magma_dlahru( n, ihi, i, nb,
A(0,i), lda,
dA(0,i-ilo), ldda, // dA
dA(i,i-ilo), ldda, // dY, stored over current panel
dV(0,0), ldda,
dT(0,i-ilo), dwork(0), queue );
}
// Copy remainder to host
magma_dgetmatrix( n, n-i,
dA(0,i-ilo), ldda,
A(0,i), lda, queue );
magma_free( dwork );
magma_free_cpu( T );
magma_queue_destroy( queue );
}
// Use unblocked code to reduce the rest of the matrix
// add 1 to i for 1-based index
i += 1;
lapackf77_dgehd2(&n, &i, &ihi, A, &lda, tau, work, &iinfo);
work[0] = magma_dmake_lwork( iws );
return *info;
} /* magma_dgehrd */
/**
Purpose
-------
CGEQRF_OOC computes a QR factorization of a COMPLEX M-by-N matrix A:
A = Q * R. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
This is an out-of-core (ooc) version that is similar to magma_cgeqrf but
the difference is that this version can use a GPU even if the matrix
does not fit into the GPU memory at once.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A COMPLEX array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau COMPLEX array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) COMPLEX array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_cgeqrf_nb( M, N ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_cgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_cgeqrf_ooc(
magma_int_t m, magma_int_t n,
magmaFloatComplex *A, magma_int_t lda, magmaFloatComplex *tau,
magmaFloatComplex *work, magma_int_t lwork,
magma_int_t *info )
{
#define A(i_,j_) ( A + (i_) + (j_)*lda )
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
/* Constants */
const magmaFloatComplex c_one = MAGMA_C_ONE;
/* Local variables */
magmaFloatComplex_ptr dA, dwork;
magma_int_t i, ib, IB, j, min_mn, lddwork, ldda, rows;
magma_int_t nb = magma_get_cgeqrf_nb( m, n );
magma_int_t lwkopt = n * nb;
work[0] = magma_cmake_lwork( lwkopt );
bool lquery = (lwork == -1);
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,n) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Check how much memory do we have */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(magmaFloatComplex);
magma_int_t NB = magma_int_t(0.8*freeMem/m);
NB = (NB / nb) * nb;
if (NB >= n)
return magma_cgeqrf(m, n, A, lda, tau, work, lwork, info);
min_mn = min(m,n);
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
lddwork = magma_roundup( NB, 32 ) + nb;
ldda = magma_roundup( m, 32 );
if (MAGMA_SUCCESS != magma_cmalloc( &dA, (NB + nb)*ldda + nb*lddwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_queue_t queues[2];
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
magmaFloatComplex_ptr ptr = dA + ldda*NB;
dwork = dA + ldda*(NB + nb);
/* start the main loop over the blocks that fit in the GPU memory */
for (i=0; i < n; i += NB) {
IB = min( n-i, NB );
//printf("Processing %5d columns -- %5d to %5d ... \n", IB, i, i+IB);
/* 1. Copy the next part of the matrix to the GPU */
magma_csetmatrix_async( m, IB,
A(0,i), lda,
dA(0,0), ldda, queues[0] );
magma_queue_sync( queues[0] );
/* 2. Update it with the previous transformations */
for (j=0; j < min(i,min_mn); j += nb) {
ib = min( min_mn-j, nb );
/* Get a panel in ptr. */
// 1. Form the triangular factor of the block reflector
// 2. Send it to the GPU.
// 3. Put 0s in the upper triangular part of V.
// 4. Send V to the GPU in ptr.
// 5. Update the matrix.
// 6. Restore the upper part of V.
rows = m-j;
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib, A(j,j), &lda, tau+j, work, &ib);
magma_csetmatrix_async( ib, ib,
work, ib,
dwork, lddwork, queues[1] );
magma_cpanel_to_q( MagmaUpper, ib, A(j,j), lda, work+ib*ib );
magma_csetmatrix_async( rows, ib,
A(j,j), lda,
ptr, rows, queues[1] );
magma_queue_sync( queues[1] );
magma_clarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, IB, ib,
ptr, rows, dwork, lddwork,
dA(j, 0), ldda, dwork+ib, lddwork, queues[1] );
magma_cq_to_panel( MagmaUpper, ib, A(j,j), lda, work+ib*ib );
}
/* 3. Do a QR on the current part */
if (i < min_mn)
magma_cgeqrf2_gpu( m-i, IB, dA(i,0), ldda, tau+i, info );
/* 4. Copy the current part back to the CPU */
magma_cgetmatrix_async( m, IB,
dA(0,0), ldda,
A(0,i), lda, queues[0] );
}
magma_queue_sync( queues[0] );
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
magma_free( dA );
return *info;
} /* magma_cgeqrf_ooc */
// --------------------
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_time=0, cpu_perf=0, gpu_time, gpu_perf, mgpu_time, mgpu_perf, cuda_time, cuda_perf;
double Ynorm, error=0, error2=0, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t n_local[MagmaMaxGPUs];
magma_int_t N, Noffset, lda, ldda, blocks, lhwork, ldwork, matsize, vecsize;
magma_int_t incx = 1;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ylapack, *Ycublas, *Ymagma, *Ymagma1, *hwork;
magmaDoubleComplex_ptr dA, dX, dY;
magmaDoubleComplex_ptr d_lA[MagmaMaxGPUs], dwork[MagmaMaxGPUs];
magma_device_t dev;
magma_queue_t queues[MagmaMaxGPUs];
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
double tol = opts.tolerance * lapackf77_dlamch("E");
magma_int_t nb = 64; // required by magmablas_zhemv_mgpu implementation
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_create( dev, &queues[dev] );
}
// currently, tests all offsets in the offsets array;
// comment out loop below to test a specific offset.
magma_int_t offset = opts.offset;
magma_int_t offsets[] = { 0, 1, 31, 32, 33, 63, 64, 65, 100, 200 };
magma_int_t noffsets = sizeof(offsets) / sizeof(*offsets);
printf("%% uplo = %s, ngpu %d, block size = %d, offset %d\n",
lapack_uplo_const(opts.uplo), (int) opts.ngpu, (int) nb, (int) offset );
printf( "%% BLAS CUBLAS MAGMA 1 GPU MAGMA MGPU Error rel Error rel\n"
"%% N offset Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) to CUBLAS to LAPACK\n"
"%%==================================================================================================================\n" );
for( int itest = 0; itest < opts.ntest; ++itest ) {
// comment out these two lines & end of loop to test a specific offset
for( int ioffset=0; ioffset < noffsets; ioffset += 1 ) {
offset = offsets[ioffset];
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
Noffset = N + offset;
lda = Noffset;
ldda = magma_roundup( Noffset, opts.align ); // multiple of 32 by default
matsize = Noffset*ldda;
vecsize = (Noffset-1)*incx + 1;
gflops = FLOPS_ZHEMV( N ) / 1e9;
blocks = magma_ceildiv( N + (offset % nb), nb );
lhwork = N*opts.ngpu;
ldwork = ldda*(blocks + 1);
TESTING_MALLOC_CPU( A, magmaDoubleComplex, matsize );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma1, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ylapack, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( X, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( hwork, magmaDoubleComplex, lhwork );
magma_setdevice( opts.device );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, matsize );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, vecsize );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, vecsize );
// TODO make magma_zmalloc_bcyclic helper function?
for( dev=0; dev < opts.ngpu; dev++ ) {
n_local[dev] = ((Noffset/nb)/opts.ngpu)*nb;
if (dev < (Noffset/nb) % opts.ngpu)
n_local[dev] += nb;
else if (dev == (Noffset/nb) % opts.ngpu)
n_local[dev] += Noffset % nb;
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], magmaDoubleComplex, ldda*n_local[dev] );
TESTING_MALLOC_DEV( dwork[dev], magmaDoubleComplex, ldwork );
}
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &matsize, A );
magma_zmake_hermitian( Noffset, A, lda );
lapackf77_zlarnv( &ione, ISEED, &vecsize, X );
lapackf77_zlarnv( &ione, ISEED, &vecsize, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_setdevice( opts.device );
magma_zsetmatrix( Noffset, Noffset, A, lda, dA, ldda, opts.queue );
magma_zsetvector( Noffset, X, incx, dX, incx, opts.queue );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
cuda_time = magma_sync_wtime(0);
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N,
&alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
&beta, dY + offset, incx );
cuda_time = magma_sync_wtime(0) - cuda_time;
cuda_perf = gflops / cuda_time;
magma_zgetvector( Noffset, dY, incx, Ycublas, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (1 GPU)
=================================================================== */
magma_setdevice( opts.device );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_zhemv_work( opts.uplo, N,
alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
beta, dY + offset, incx, dwork[ opts.device ], ldwork,
opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_zgetvector( Noffset, dY, incx, Ymagma1, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (multi-GPU)
=================================================================== */
magma_zsetmatrix_1D_col_bcyclic( Noffset, Noffset, A, lda, d_lA, ldda, opts.ngpu, nb, queues );
blasf77_zcopy( &Noffset, Y, &incx, Ymagma, &incx );
// workspaces do NOT need to be zero -- set to NAN to prove
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork[dev], ldwork, opts.queue );
}
lapackf77_zlaset( "Full", &lhwork, &ione, &MAGMA_Z_NAN, &MAGMA_Z_NAN, hwork, &lhwork );
mgpu_time = magma_sync_wtime(0);
magma_int_t info;
info = magmablas_zhemv_mgpu(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
info = magmablas_zhemv_mgpu_sync(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_sync returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
mgpu_time = magma_sync_wtime(0) - mgpu_time;
mgpu_perf = gflops / mgpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
blasf77_zcopy( &Noffset, Y, &incx, Ylapack, &incx );
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N,
&alpha, A + offset + offset*lda, &lda,
X + offset, &incx,
&beta, Ylapack + offset, &incx );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Compute the Difference LAPACK vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ylapack, &incx );
error2 = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ) / Ynorm;
}
/* =====================================================================
Compute the Difference Cublas vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ycublas, &incx );
error = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work ) / Ynorm;
bool okay = (error < tol && error2 < tol);
status += ! okay;
if ( opts.lapack ) {
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) N, (int) offset,
cpu_perf, cpu_time*1000.,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, error2, (okay ? "ok" : "failed") );
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e --- %s\n",
(int) N, (int) offset,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, (okay ? "ok" : "failed") );
}
/* Free Memory */
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_CPU( Ymagma1 );
TESTING_FREE_CPU( Ylapack );
TESTING_FREE_PIN( X );
TESTING_FREE_PIN( hwork );
magma_setdevice( opts.device );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
for( dev=0; dev < opts.ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
TESTING_FREE_DEV( dwork[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
// comment out these two lines line & top of loop test a specific offset
} // end for ioffset
printf( "\n" );
}
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_destroy( queues[dev] );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
