void get_QR_error(magma_int_t M, magma_int_t N, magma_int_t min_mn,
magmaDoubleComplex *h_R, magmaDoubleComplex *h_A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex *Q, magma_int_t ldq,
magmaDoubleComplex *R, magma_int_t ldr,
magmaDoubleComplex *h_work, magma_int_t lwork,
double *work, double *error, double *error2)
{
/* h_R:input the factorized matrix by lapack QR,
h_A:input the original matrix copy
tau: input
*/
const double d_neg_one = MAGMA_D_NEG_ONE;
const double d_one = MAGMA_D_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
double Anorm;
magma_int_t info;
// generate M by K matrix Q, where K = min(M,N)
lapackf77_zlacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_zungqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_zlaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_zlacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_zgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_zlange( "1", &M, &N, h_A, &lda, work );
*error = lapackf77_zlange( "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_zlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_zherk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
*error2 = safe_lapackf77_zlanhe( "1", "Upper", &min_mn, R, &ldr, work );
if ( N > 0 )
*error2 /= N;
}
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 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 ZGEQRF_GPU.
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]
dA COMPLEX_16 array A on the GPU, 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 ZGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
ldda INTEGER
The first dimension of the array A. LDDA >= max(1,M).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
dT (workspace) COMPLEX_16 work space array on the GPU,
dimension (2*MIN(M, N) + ceil(N/32)*32 )*NB.
This must be the 6th argument of magma_zgeqrf_gpu
[ note that if N here is bigger than N in magma_zgeqrf_gpu,
the workspace requirement DT in magma_zgeqrf_gpu must be
as specified in this routine ].
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in DT.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr_gpu(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex_ptr dA, magma_int_t ldda,
magmaDoubleComplex *tau,
magmaDoubleComplex_ptr dT, magma_int_t nb,
magma_int_t *info)
{
#define dA(i,j) (dA + (i) + (j)*ldda)
#define dT(j) (dT + (j)*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, lpanel;
magma_int_t i, ib, ki, kk, iinfo;
magma_int_t lddwork;
magmaDoubleComplex_ptr dV, dW;
magmaDoubleComplex *work, *panel;
*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 (ldda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
// 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 zungqr workspace
// (m - kk)*(n - kk) for last block's panel
lwork = n*nb;
lpanel = (m - kk)*(n - kk);
magma_zmalloc_cpu( &work, lwork + lpanel );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
panel = work + lwork;
// Allocate work space on GPU
if (MAGMA_SUCCESS != magma_zmalloc( &dV, ldda*nb )) {
magma_free_cpu( work );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// dT workspace has:
// 2*min(m,n)*nb for T and R^{-1} matrices from geqrf
// roundup(n,32) * nb for dW larfb workspace.
lddwork = min(m,n);
dW = dT + 2*lddwork*nb;
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
magma_zgetmatrix( m_kk, k_kk,
dA(kk, kk), ldda, panel, m_kk, queue );
lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
panel, &m_kk,
&tau[kk], work, &lwork, &iinfo );
magma_zsetmatrix( m_kk, n_kk,
panel, m_kk, dA(kk, kk), ldda, queue );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda, queue );
}
if (kk > 0) {
// Use blocked code
// queue: copy Aii to V --> laset --> laset --> larfb --> [next]
// CPU has no computation
for (i = ki; i >= 0; i -= nb) {
ib = min( nb, k-i );
mi = m - i;
// Copy current panel on the GPU from dA to dV
magma_zcopymatrix_async( mi, ib,
dA(i,i), ldda,
dV, ldda, queue );
// set panel to identity
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda, queue );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda, queue );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT(i), nb,
dA(i, i), ldda, dW, lddwork, queue );
}
}
}
magma_queue_sync( queue );
magma_free( dV );
magma_free_cpu( work );
magma_queue_destroy( queue );
return *info;
} /* magma_zungqr_gpu */
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 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 ZGEQRF.
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 COMPLEX_16 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 ZGEQRF_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 COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
T COMPLEX_16 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_zgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_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 has an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex *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)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldwork;
magma_int_t i, ib, ki, kk, iinfo;
magmaDoubleComplex *work;
*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 );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
// Allocate memory on GPUs for A and workspaces
magma_int_t ldda = ((m + 31) / 32) * 32;
magma_int_t lddwork = ((n + 31) / 32) * 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 };
magmaDoubleComplex *dA[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dT[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dV[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dW[ MagmaMaxGPUs ] = { NULL };
magma_queue_t stream[ MagmaMaxGPUs ] = { NULL };
for( int 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_zmalloc( &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( &stream[d] );
}
trace_init( 1, ngpu, 1, stream );
// 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 zungqr workspace
lwork = n * nb;
magma_zmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto CLEANUP;
}
// 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;
dpanel = (kk / nb) % ngpu;
di = ((kk / nb) / ngpu) * nb;
magma_setdevice( dpanel );
lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
magma_zsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(dpanel, kk, di), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(dpanel, 0, di), ldda );
trace_cpu_end( 0 );
}
if (kk > 0) {
// Use blocked code
// send T to all GPUs
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set T" );
magma_zsetmatrix_async( nb, min(m,n), T, nb, dT[d], nb, stream[d] );
trace_gpu_end( d, 0 );
}
// stream: 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 the GPUs
lapackf77_zlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set V" );
magma_zsetmatrix_async( mi, ib,
A(i, i), lda,
dV[d], ldda, stream[d] );
trace_gpu_end( d, 0 );
}
// set panel to identity
magma_setdevice( dpanel );
magmablasSetKernelStream( stream[dpanel] );
trace_gpu_start( dpanel, 0, "laset", "laset" );
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(dpanel, i, di), ldda );
trace_gpu_end( dpanel, 0 );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( stream[d] );
magma_indices_1D_bcyclic( nb, ngpu, d, i, n, &di, &dn );
trace_gpu_start( d, 0, "larfb", "larfb" );
magma_zlarfb_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 );
trace_gpu_end( d, 0 );
}
}
}
}
// copy result back to CPU
trace_cpu_start( 0, "get", "get A" );
magma_zgetmatrix_1D_col_bcyclic( m, n, dA, ldda, A, lda, ngpu, nb );
trace_cpu_end( 0 );
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "zungqr-n%d-ngpu%d.svg", m, ngpu );
trace_finalize( name, "trace.css" );
#endif
CLEANUP:
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( dA[d] );
magma_queue_destroy( stream[d] );
}
magma_free_cpu( work );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_zungqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zungqr_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double Anorm, error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *hA, *hR, *tau, *h_work;
magmaDoubleComplex_ptr dA, dT;
magma_int_t m, n, k;
magma_int_t n2, lda, ldda, lwork, min_mn, nb, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf(" m n k CPU GFlop/s (sec) GPU GFlop/s (sec) ||R|| / ||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];
k = opts.ksize[itest];
if ( m < n || n < k ) {
printf( "%5d %5d %5d skipping because m < n or n < k\n", (int) m, (int) n, (int) k );
continue;
}
lda = m;
ldda = ((m + 31)/32)*32;
n2 = lda*n;
min_mn = min(m, n);
nb = magma_get_zgeqrf_nb( m );
lwork = (m + 2*n+nb)*nb;
gflops = FLOPS_ZUNGQR( m, n, k ) / 1e9;
TESTING_MALLOC_PIN( hA, magmaDoubleComplex, lda*n );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( hR, magmaDoubleComplex, lda*n );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*n );
TESTING_MALLOC_DEV( dT, magmaDoubleComplex, ( 2*min_mn + ((n + 31)/32)*32 )*nb );
lapackf77_zlarnv( &ione, ISEED, &n2, hA );
lapackf77_zlacpy( MagmaFullStr, &m, &n, hA, &lda, hR, &lda );
Anorm = lapackf77_zlange("f", &m, &n, hA, &lda, work );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// first, get QR factors in both hA and dA
// okay that magma_zgeqrf_gpu has special structure for R; R isn't used here.
magma_zsetmatrix( m, n, hA, lda, dA, ldda );
magma_zgeqrf_gpu( m, n, dA, ldda, tau, dT, &info );
if (info != 0)
printf("magma_zgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_zgetmatrix( m, n, dA, ldda, hA, lda );
gpu_time = magma_wtime();
magma_zungqr_gpu( m, n, k, dA, ldda, tau, dT, nb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zungqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// Get dA back to the CPU to compare with the CPU result.
magma_zgetmatrix( m, n, dA, ldda, hR, lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zungqr( &m, &n, &k, hA, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zungqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute relative error |R|/|A| := |Q_magma - Q_lapack|/|A|
blasf77_zaxpy( &n2, &c_neg_one, hA, &ione, hR, &ione );
error = lapackf77_zlange("f", &m, &n, hR, &lda, work) / Anorm;
bool okay = (error < tol);
status += ! okay;
printf("%5d %5d %5d %7.1f (%7.2f) %7.1f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (okay ? "ok" : "failed"));
}
else {
printf("%5d %5d %5d --- ( --- ) %7.1f (%7.2f) --- \n",
(int) m, (int) n, (int) k,
gpu_perf, gpu_time );
}
TESTING_FREE_PIN( hA );
TESTING_FREE_PIN( h_work );
TESTING_FREE_CPU( hR );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zungqr_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *hA, *hR, *tau, *h_work;
magmaDoubleComplex *dA, *dT;
magma_int_t m, n, k;
magma_int_t n2, lda, ldda, lwork, min_mn, nb, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf(" m n k CPU GFlop/s (sec) GPU GFlop/s (sec) ||R|| / ||A||\n");
printf("=========================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[i];
n = opts.nsize[i];
k = opts.ksize[i];
if ( m < n || n < k ) {
printf( "skipping m %d, n %d, k %d because m < n or n < k\n", (int) m, (int) n, (int) k );
continue;
}
lda = m;
ldda = ((m + 31)/32)*32;
n2 = lda*n;
min_mn = min(m, n);
nb = magma_get_zgeqrf_nb( m );
lwork = (m + 2*n+nb)*nb;
gflops = FLOPS_ZUNGQR( m, n, k ) / 1e9;
TESTING_MALLOC_PIN( hA, magmaDoubleComplex, lda*n );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( hR, magmaDoubleComplex, lda*n );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*n );
TESTING_MALLOC_DEV( dT, magmaDoubleComplex, ( 2*min_mn + ((n + 31)/32)*32 )*nb );
lapackf77_zlarnv( &ione, ISEED, &n2, hA );
lapackf77_zlacpy( MagmaUpperLowerStr, &m, &n, hA, &lda, hR, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix( m, n, hA, lda, dA, ldda );
magma_zgeqrf_gpu( m, n, dA, ldda, tau, dT, &info );
if (info != 0)
printf("magma_zgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
gpu_time = magma_wtime();
magma_zungqr_gpu( m, n, k, dA, ldda, tau, dT, nb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zungqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// Get dA back to the CPU to compare with the CPU result.
magma_zgetmatrix( m, n, dA, ldda, hR, lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
error = lapackf77_zlange("f", &m, &n, hA, &lda, work );
lapackf77_zgeqrf( &m, &n, hA, &lda, tau, h_work, &lwork, &info );
if (info != 0)
printf("lapackf77_zgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
cpu_time = magma_wtime();
lapackf77_zungqr( &m, &n, &k, hA, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zungqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute relative error |R|/|A| := |Q_magma - Q_lapack|/|A|
blasf77_zaxpy( &n2, &c_neg_one, hA, &ione, hR, &ione );
error = lapackf77_zlange("f", &m, &n, hR, &lda, work) / error;
printf("%5d %5d %5d %7.1f (%7.2f) %7.1f (%7.2f) %8.2e\n",
(int) m, (int) n, (int) k,
cpu_perf, cpu_time, gpu_perf, gpu_time, error );
}
else {
printf("%5d %5d %5d --- ( --- ) %7.1f (%7.2f) --- \n",
(int) m, (int) n, (int) k,
gpu_perf, gpu_time );
}
TESTING_FREE_PIN( hA );
TESTING_FREE_PIN( h_work );
TESTING_FREE_CPU( hR );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
