/*-------------------------------------------------------------------
* Check the orthogonality of Q
*/
static magma_int_t check_orthogonality(magma_int_t M, magma_int_t N, double *Q, magma_int_t LDQ, double eps)
{
double done = 1.0;
double mdone = -1.0;
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
double normQ, result;
magma_int_t info_ortho;
magma_int_t minMN = min(M, N);
double *work;
magma_dmalloc_cpu( &work, minMN );
/* Build the idendity matrix */
double *Id;
magma_dmalloc_cpu( &Id, minMN*minMN );
lapackf77_dlaset("A", &minMN, &minMN, &c_zero, &c_one, Id, &minMN);
/* Perform Id - Q'Q */
if (M >= N)
blasf77_dsyrk("U", "C", &N, &M, &done, Q, &LDQ, &mdone, Id, &N);
else
blasf77_dsyrk("U", "N", &M, &N, &done, Q, &LDQ, &mdone, Id, &M);
normQ = lapackf77_dlansy("I", "U", &minMN, Id, &minMN, work);
result = normQ / (minMN * eps);
printf( " %12.2e", result*eps );
//printf(" ======================================================\n");
//printf(" ||Id-Q'*Q||_oo / (minMN*eps) : %15.3E \n", result );
//printf(" ======================================================\n");
if ( isnan(result) || isinf(result) || (result > 60.0) ) {
//printf("-- Orthogonality is suspicious ! \n");
info_ortho=1;
}
else {
//printf("-- Orthogonality is CORRECT ! \n");
info_ortho=0;
}
magma_free_cpu(work);
magma_free_cpu(Id);
return info_ortho;
}
/**
Purpose
-------
DORGQR generates an M-by-N DOUBLE_PRECISION 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 DGEQRF.
This version recomputes the T matrices on the CPU and sends them to the 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]
A DOUBLE_PRECISION 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 DGEQRF_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 DOUBLE_PRECISION array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by DGEQRF_GPU.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument has an illegal value
@ingroup magma_dgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_dorgqr2(magma_int_t m, magma_int_t n, magma_int_t k,
double *A, magma_int_t lda,
double *tau,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
magma_int_t nb = magma_get_dgeqrf_nb(min(m, n));
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk; //, iinfo;
magma_int_t lddwork;
double *dA, *dV, *dW, *dT, *T;
double *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;
}
// 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 GPU work space
// ldda*n for matrix dA
// ldda*nb for dV
// lddwork*nb for dW larfb workspace
ldda = ((m + 31) / 32) * 32;
lddwork = ((n + 31) / 32) * 32;
if (MAGMA_SUCCESS != magma_dmalloc( &dA, ldda*n + ldda*nb + lddwork*nb + nb*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dV = dA + ldda*n;
dW = dA + ldda*n + ldda*nb;
dT = dA + ldda*n + ldda*nb + lddwork*nb;
// Allocate CPU work space
lwork = (n+m+nb) * nb;
magma_dmalloc_cpu( &work, lwork );
T = work;
if (work == NULL) {
magma_free( dA );
magma_free_cpu( work );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
double *V = work + (n+nb)*nb;
magma_queue_t stream;
magma_queue_create( &stream );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
/*
lapackf77_dorgqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_dlacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_dlaset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_dlarfb( MagmaLeftStr, MagmaNoTransStr, MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &n_kk, &k_kk,
V, &m_kk, work, &k_kk, A(kk, kk), &lda, work+k_kk*k_kk, &n_kk );
if (kk > 0) {
magma_dsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_dlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda );
}
}
if (kk > 0) {
// Use blocked code
// stream: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
magmablasSetKernelStream( stream );
for (i = ki; i >= 0; i -= nb) {
ib = min(nb, k - i);
// Send current panel to the GPU
mi = m - i;
lapackf77_dlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_dsetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&mi, &ib,
A(i,i), &lda, &tau[i], T, &nb);
magma_dsetmatrix_async( ib, ib,
T, nb,
dT, nb, stream );
// set panel to identity
magmablas_dlaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda );
magmablas_dlaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda );
magma_queue_sync( stream );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_dlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT, nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_dgetmatrix( m, n,
dA(0, 0), ldda, A(0, 0), lda);
}
magmablasSetKernelStream( NULL );
magma_queue_destroy( stream );
magma_free( dA );
magma_free_cpu( work );
return *info;
} /* magma_dorgqr */
/***************************************************************************//**
Purpose
-------
DORGQR generates an M-by-N DOUBLE PRECISION 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 DGEQRF.
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 DOUBLE PRECISION 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 DGEQRF_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 DOUBLE PRECISION array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by DGEQRF_GPU.
@param[in]
T DOUBLE PRECISION 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_dgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in DGEQRF_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_ungqr
*******************************************************************************/
extern "C" magma_int_t
magma_dorgqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
double *A, magma_int_t lda,
double *tau,
double *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)
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_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;
double *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 };
double *dA[ MagmaMaxGPUs ] = { NULL };
double *dT[ MagmaMaxGPUs ] = { NULL };
double *dV[ MagmaMaxGPUs ] = { NULL };
double *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_dmalloc( &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_dmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
double *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;
// dorgqr requires less workspace (n*nb), but is slow if k < dorgqr's block size.
// replacing it with the 4 routines below is much faster (e.g., 60x).
//magma_int_t iinfo;
//lapackf77_dorgqr( &m_kk, &n_kk, &k_kk,
// A(kk, kk), &lda,
// &tau[kk], work, &lwork, &iinfo );
lapackf77_dlacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, work_V, &m_kk);
lapackf77_dlaset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
work_V, &m_kk, &tau[kk], work_T, &k_kk);
lapackf77_dlarfb( 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_dsetmatrix( m_kk, jb,
A(kk, j), lda,
dA(d, kk, di), ldda, queues[d] );
// Set A(1:kk,kk+1:n) to zero.
magmablas_dlaset( 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_dsetmatrix_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_dlaset( "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_dsetmatrix_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_dlaset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda, queues[dpanel] );
magmablas_dlaset( 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_dlarfb_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_dgetmatrix_1D_col_bcyclic( ngpu, m, n, nb, dA, ldda, A, lda, queues );
trace_cpu_end( 0 );
}
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "dorgqr-n%lld-ngpu%lld.svg", (long long) m, (long long) 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_dorgqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dlarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_CUDA_INIT();
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
printf( "\nUsage: %s -M m -N n -K k\n\n", argv[0] );
magma_int_t m = 500;
magma_int_t n = 300;
magma_int_t k = 32;
for( int i = 1; i < argc; i++ ) {
if (strcmp("-M", argv[i]) == 0 && i+1 < argc) {
m = atoi( argv[++i] );
}
else if (strcmp("-N", argv[i]) == 0 && i+1 < argc) {
n = atoi( argv[++i] );
}
else if (strcmp("-K", argv[i]) == 0 && i+1 < argc) {
k = atoi( argv[++i] );
}
else {
printf( "invalid argument: %s\n", argv[i] );
exit(1);
}
}
if ( k <= 0 || k > m || k > n ) {
printf( "requires 0 < k <= min(m,n)\n" );
exit(1);
}
magma_int_t ldc = m;
magma_int_t ldv = max(m,n);
magma_int_t ldt = k;
magma_int_t ldw = max(m,n);
magma_int_t nv;
ldc = ((ldc+31)/32)*32;
ldv = ((ldv+31)/32)*32;
ldt = ((ldt+31)/32)*32;
ldw = ((ldw+31)/32)*32;
// Allocate memory for matrices
double *C, *R, *V, *T, *W;
TESTING_MALLOC( C, double, ldc*n );
TESTING_MALLOC( R, double, ldc*n );
TESTING_MALLOC( V, double, ldv*k );
TESTING_MALLOC( T, double, ldt*k );
TESTING_MALLOC( W, double, ldw*k );
double *dC, *dV, *dT, *dW;
TESTING_DEVALLOC( dC, double, ldc*n );
TESTING_DEVALLOC( dV, double, ldv*k );
TESTING_DEVALLOC( dT, double, ldt*k );
TESTING_DEVALLOC( dW, double, ldw*k );
magma_int_t size;
magma_int_t iseed[4] = { 1, 2, 3, 4 };
double error, work[1];
// test all combinations of input parameters
const char* side[] = { MagmaLeftStr, MagmaRightStr };
const char* trans[] = { MagmaTransStr, MagmaNoTransStr };
const char* direct[] = { MagmaForwardStr, MagmaBackwardStr };
const char* storev[] = { MagmaColumnwiseStr, MagmaRowwiseStr };
printf(" M N K storev side direct trans ||R||_F / ||HC||_F\n");
printf("==================================================================================\n");
for( int istor = 0; istor < 2; ++istor ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int idir = 0; idir < 2; ++idir ) {
for( int itran = 0; itran < 2; ++itran ) {
//printf( "# ----------\n" );
//printf( "# %-10s %-10s %-10s %-10s\n", storev[istor], side[iside], direct[idir], trans[itran] );
// C is full
size = ldc*n;
lapackf77_dlarnv( &ione, iseed, &size, C );
//printf( "C=" ); magma_dprint( m, n, C, ldc );
// V is ldv x nv. See larfb docs for description.
ldv = (*side[iside] == 'L' ? m : n);
nv = k;
size = ldv*nv;
lapackf77_dlarnv( &ione, iseed, &size, V );
if ( *storev[istor] == MagmaColumnwise ) {
if ( *direct[idir] == MagmaForward ) {
lapackf77_dlaset( MagmaUpperStr, &k, &k, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_dlaset( MagmaLowerStr, &k, &k, &c_zero, &c_one, &V[(ldv-k)], &ldv );
}
}
else {
// rowwise, swap V's dimensions
std::swap( ldv, nv );
if ( *direct[idir] == MagmaForward ) {
lapackf77_dlaset( MagmaLowerStr, &k, &k, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_dlaset( MagmaUpperStr, &k, &k, &c_zero, &c_one, &V[(nv-k)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_dprint( ldv, nv, V, ldv );
// T is upper triangular for forward, and lower triangular for backward
magma_int_t k1 = k-1;
size = ldt*k;
lapackf77_dlarnv( &ione, iseed, &size, T );
if ( *direct[idir] == MagmaForward ) {
lapackf77_dlaset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_dlaset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_dprint( k, k, T, ldt );
magma_dsetmatrix( m, n, C, ldc, dC, ldc );
magma_dsetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_dsetmatrix( k, k, T, ldt, dT, ldt );
lapackf77_dlarfb( side[iside], trans[itran], direct[idir], storev[istor],
&m, &n, &k,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_dprint( m, n, C, ldc );
magma_dlarfb_gpu( *side[iside], *trans[itran], *direct[idir], *storev[istor],
m, n, k,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_dgetmatrix( m, n, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_dprint( m, n, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_dlange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_daxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_dlange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %-10s %-10s %-10s %-10s %8.2e\n",
(int) m, (int) n, (int) k,
storev[istor], side[iside], direct[idir], trans[itran], error );
}}}}
// Memory clean up
TESTING_FREE( C );
TESTING_FREE( R );
TESTING_FREE( V );
TESTING_FREE( T );
TESTING_FREE( W );
TESTING_DEVFREE( dC );
TESTING_DEVFREE( dV );
TESTING_DEVFREE( dT );
TESTING_DEVFREE( dW );
// Shutdown
TESTING_CUDA_FINALIZE();
return 0;
}
/**
Purpose
-------
DLAHR2 reduces the first NB columns of a real general n-BY-(n-k+1)
matrix A so that elements below the k-th subdiagonal are zero. The
reduction is performed by an orthogonal similarity transformation
Q' * A * Q. The routine returns the matrices V and T which determine
Q as a block reflector I - V*T*V', and also the matrix Y = A * V.
(Note this is different than LAPACK, which computes Y = A * V * T.)
This is an auxiliary routine called by DGEHRD.
Arguments
---------
@param[in]
n INTEGER
The order of the matrix A.
@param[in]
k INTEGER
The offset for the reduction. Elements below the k-th
subdiagonal in the first NB columns are reduced to zero.
K < N.
@param[in]
nb INTEGER
The number of columns to be reduced.
@param[in,out]
A DOUBLE_PRECISION array, dimension (LDA,N-K+1)
On entry, the n-by-(n-k+1) general matrix A.
On exit, the elements on and above the k-th subdiagonal in
the first NB columns are overwritten with the corresponding
elements of the reduced matrix; the elements below the k-th
subdiagonal, with the array TAU, represent the matrix Q as a
product of elementary reflectors. The other columns of A are
unchanged. 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 (NB)
The scalar factors of the elementary reflectors. See Further
Details.
@param[out]
T DOUBLE_PRECISION array, dimension (LDT,NB)
The upper triangular matrix T.
@param[in]
ldt INTEGER
The leading dimension of the array T. LDT >= NB.
@param[out]
Y DOUBLE_PRECISION array, dimension (LDY,NB)
The n-by-nb matrix Y.
@param[in]
ldy INTEGER
The leading dimension of the array Y. LDY >= N.
@param[in,out]
data Structure with pointers to dA, dT, dV, dW, dY
which are distributed across multiple GPUs.
Further Details
---------------
The matrix Q is represented as a product of nb elementary reflectors
Q = H(1) H(2) . . . H(nb).
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+k-1) = 0, v(i+k) = 1; v(i+k+1:n) is stored on exit in
A(i+k+1:n,i), and tau in TAU(i).
The elements of the vectors v together form the (n-k+1)-by-nb matrix
V which is needed, with T and Y, to apply the transformation to the
unreduced part of the matrix, using an update of the form:
A := (I - V*T*V') * (A - Y*T*V').
The contents of A on exit are illustrated by the following example
with n = 7, k = 3 and nb = 2:
@verbatim
( a a a a a )
( a a a a a )
( a a a a a )
( h h a a a )
( v1 h a a a )
( v1 v2 a a a )
( v1 v2 a 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.
@ingroup magma_dgeev_aux
********************************************************************/
extern "C" magma_int_t
magma_dlahr2_m(
magma_int_t n, magma_int_t k, magma_int_t nb,
double *A, magma_int_t lda,
double *tau,
double *T, magma_int_t ldt,
double *Y, magma_int_t ldy,
struct dgehrd_data* data )
{
#define A( i, j ) ( A + (i) + (j)*lda)
#define Y( i, j ) ( Y + (i) + (j)*ldy)
#define T( i, j ) ( T + (i) + (j)*ldt)
#define dA( d, i, j ) (data->A [d] + (i) + (j)*ldda)
#define dTi( d ) (data->Ti[d])
#define dV( d, i, j ) (data->V [d] + (i) + (j)*ldv )
#define dVd( d, i, j ) (data->Vd[d] + (i) + (j)*ldvd)
#define dY( d, i, j ) (data->Y [d] + (i) + (j)*ldda)
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double tmp;
magma_int_t ngpu = data->ngpu;
magma_int_t ldda = data->ldda;
magma_int_t ldv = data->ldv;
magma_int_t ldvd = data->ldvd;
magma_int_t ione = 1;
magma_int_t d, dki1, dn, nblocks, gblock, lblock, lgid;
magma_int_t n_k_i_1, n_k;
double scale;
magma_int_t i;
double ei = MAGMA_D_ZERO;
magma_int_t info_data = 0;
magma_int_t *info = &info_data;
if (n < 0) {
*info = -1;
} else if (k < 0 || k >= n) {
*info = -2;
} else if (nb < 1 || nb > n) {
*info = -3;
} else if (lda < max(1,n)) {
*info = -5;
} else if (ldt < nb) {
*info = -8;
} else if (ldy < max(1,n)) {
*info = -10;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
// adjust from 1-based indexing
k -= 1;
// Function Body
if (n <= 1)
return 0;
// zero out current top block of V on all GPUs
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( data->streams[d] );
magmablas_dlaset( MagmaFull, nb, nb, c_zero, c_zero, dV(d,k,0), ldv );
}
// set all Y=0
lapackf77_dlaset( "Full", &n, &nb, &c_zero, &c_zero, Y, &ldy );
for (i = 0; i < nb; ++i) {
n_k_i_1 = n - k - i - 1;
n_k = n - k;
if (i > 0) {
// Finish applying I - V * T * V' on right
tmp = MAGMA_D_NEGATE( tau[i-1] );
blasf77_daxpy( &n_k, &tmp, Y(k,i-1), &ione, A(k,i), &ione );
// Apply I - V * T' * V' to this column (call it b) from the
// left, using the last column of T as workspace, w.
//
// Let V = ( V1 ) and b = ( b1 ) (first i-1 rows)
// ( V2 ) ( b2 )
// where V1 is unit lower triangular
// w := b1 = A(k+1:k+i, i)
blasf77_dcopy( &i,
A(k+1,i), &ione,
T(0,nb-1), &ione );
// w := V1' * b1 = VA(k+1:k+i, 0:i-1)' * w
blasf77_dtrmv( "Lower", "Conj", "Unit", &i,
A(k+1,0), &lda,
T(0,nb-1), &ione );
// w := w + V2'*b2 = w + VA(k+i+1:n-1, 0:i-1)' * A(k+i+1:n-1, i)
blasf77_dgemv( "Conj", &n_k_i_1, &i,
&c_one, A(k+i+1,0), &lda,
A(k+i+1,i), &ione,
&c_one, T(0,nb-1), &ione );
// w := T'*w = T(0:i-1, 0:i-1)' * w
blasf77_dtrmv( "Upper", "Conj", "Non-unit", &i,
T(0,0), &ldt,
T(0,nb-1), &ione );
// b2 := b2 - V2*w = A(k+i+1:n-1, i) - VA(k+i+1:n-1, 0:i-1) * w
blasf77_dgemv( "No trans", &n_k_i_1, &i,
&c_neg_one, A(k+i+1,0), &lda,
T(0,nb-1), &ione,
&c_one, A(k+i+1,i), &ione );
// w := V1*w = VA(k+1:k+i, 0:i-1) * w
blasf77_dtrmv( "Lower", "No trans", "Unit", &i,
A(k+1,0), &lda,
T(0,nb-1), &ione );
// b1 := b1 - w = A(k+1:k+i-1, i) - w
blasf77_daxpy( &i,
&c_neg_one, T(0,nb-1), &ione,
A(k+1,i), &ione );
// Restore diagonal element, saved below during previous iteration
*A(k+i,i-1) = ei;
}
// Generate the elementary reflector H(i) to annihilate A(k+i+1:n-1,i)
lapackf77_dlarfg( &n_k_i_1,
A(k+i+1,i),
A(k+i+2,i), &ione, &tau[i] );
// Save diagonal element and set to one, to simplify multiplying by V
ei = *A(k+i+1,i);
*A(k+i+1,i) = c_one;
// compute yi = A vi = sum_g A{d} vi{d}
nblocks = (n-1) / nb / ngpu + 1;
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( data->streams[d] );
// dV(k+i+1:n-1, i) = VA(k+i:n, i)
magma_dsetvector_async( n_k_i_1,
A(k+i+1,i), 1,
dV(d, k+i+1, i), 1, data->streams[d] );
// copy column of dV -> dVd, using block cyclic distribution.
// This assumes V and Vd have been padded so that
// a 2D matrix copy doesn't access them out-of-bounds
gblock = k / nb;
lblock = gblock / ngpu;
lgid = gblock % ngpu;
if ( d < lgid ) {
lblock += 1;
}
// treat V as (nb*ngpu) x nblock matrix, and Vd as nb x nblock matrix
magmablas_dlacpy( MagmaFull, nb, nblocks-lblock,
dV (d, d*nb + lblock*nb*ngpu, i), nb*ngpu,
dVd(d, 0 + lblock*nb, i), nb );
// convert global indices (k) to local indices (dk)
magma_indices_1D_bcyclic( nb, ngpu, d, k+i+1, n, &dki1, &dn );
// dY(k:n, i) = dA(k:n, k+i+1:n) * dV(k+i+1:n, i)
// skip if matrix is empty
// each GPU copies to different temporary vector in Y,
// which are summed in separate loop below
if ( dn-dki1 > 0 ) {
magma_dgemv( MagmaNoTrans, n-k, dn-dki1,
c_one, dA (d, k, dki1), ldda,
dVd(d, dki1, i), 1,
c_zero, dY (d, k, i), 1 );
// copy vector to host, storing in column nb+d of Y
// as temporary space (Y has >= nb+ngpu columns)
magma_dgetvector_async( n-k,
dY(d, k, i), 1,
Y(k, nb+d), 1, data->streams[d] );
}
}
// while GPU is doing above Ag*v...
// Compute T(0:i,i) = [ -tau T V' vi ]
// [ tau ]
// T(0:i-1, i) = -tau VA(k+i+1:n-1, 0:i-1)' VA(k+i+1:n-1, i)
scale = MAGMA_D_NEGATE( tau[i] );
blasf77_dgemv( "Conj", &n_k_i_1, &i,
&scale, A(k+i+1,0), &lda,
A(k+i+1,i), &ione,
&c_zero, T(0,i), &ione );
// T(0:i-1, i) = T(0:i-1, 0:i-1) * T(0:i-1, i)
blasf77_dtrmv( "Upper", "No trans", "Non-unit", &i,
T(0,0), &ldt,
T(0,i), &ione );
*T(i,i) = tau[i];
// apply reflectors to next column, A(i+1), on right only.
// one axpy will be required to finish this, in the next iteration above
if ( i > 0 && i+1 < nb ) {
// Update next column, A(k:n,i+1), applying Q on right.
// One axpy will be required to finish this, in the next iteration
// above, after yi is computed.
// This updates one more row than LAPACK does (row k),
// making block above panel an even multiple of nb.
// Use last column of T as workspace, w.
magma_int_t i1 = i+1;
// If real, conjugate row of V, and undo afterwards
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_dlacgv( &i1, A(k+i1,0), &lda );
#endif
// w = T(0:i, 0:i+1) * VA(k+i+1, 0:i+1)'
// T is now rectangular, so we use gemv instead of trmv as in lapack.
blasf77_dgemv( "No trans", &i, &i1,
&c_one, T(0,0), &ldt,
A(k+i1,0), &lda,
&c_zero, T(0,nb-1), &ione );
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_dlacgv( &i1, A(k+i1,0), &lda );
#endif
// A(k:n, i+1) -= Y(k:n, 0:i) * w
blasf77_dgemv( "No trans", &n_k, &i,
&c_neg_one, Y(k,0), &ldy,
T(0,nb-1), &ione,
&c_one, A(k,i1), &ione );
}
// yi = sum_g yi{d}
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_queue_sync( data->streams[d] );
magma_indices_1D_bcyclic( nb, ngpu, d, k+i+1, n, &dki1, &dn );
if ( dn-dki1 > 0 ) {
// yi = yi + yi{d}
blasf77_daxpy( &n_k, &c_one, Y(k,nb+d), &ione, Y(k,i), &ione );
}
}
}
// Restore diagonal element
*A(k+nb,nb-1) = ei;
// compute Y = Am V = sum_g Am{d} V{d} --- top part, Y(0:k-1,:)
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( data->streams[d] );
// convert global indices (k) to local indices (dk)
magma_indices_1D_bcyclic( nb, ngpu, d, k+1, n, &dki1, &dn );
// dY(0:k, :) = dA(0:k, k+i+1:n-1) * dV(k+i+1:n-1, :)
// skip if matrix is empty
// each GPU copies to different temporary block in Y,
// which are summed in separate loop below
if ( dn-dki1 > 0 ) {
magma_dgemm( MagmaNoTrans, MagmaNoTrans, k, nb, dn-dki1,
c_one, dA (d, 0, dki1), ldda,
dVd(d, dki1, 0), ldvd,
c_zero, dY (d, 0, 0), ldda );
// copy result to host, storing in columns [nb + nb*d : nb + nb*(d+1)] of Y
// as temporary space (Y has nb + nb*ngpu columns)
magma_dgetmatrix_async( k, nb,
dY(d, 0, 0), ldda,
Y(0,nb+nb*d), ldy, data->streams[d] );
}
}
// Y = sum_g Y{d}
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_queue_sync( 0 );
magma_indices_1D_bcyclic( nb, ngpu, d, k+1, n, &dki1, &dn );
if ( dn-dki1 > 0 ) {
// Y = Y + Am V
for( i = 0; i < nb; ++i ) {
blasf77_daxpy( &k, &c_one, Y(0,nb+nb*d+i), &ione, Y(0,i), &ione );
}
}
}
// copy Y and T matrices to GPUs
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_dsetmatrix_async( n, nb, Y, ldy, dY(d, 0, 0), ldda, data->streams[d] );
magma_dsetmatrix_async( nb, nb, T, nb, dTi(d), nb, data->streams[d] );
}
return 0;
} /* magma_dlahr2 */
/**
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 >= max(1,N).
For optimum performance 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]
T DOUBLE_PRECISION array, 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, for later use in magma_dorghr.
@ingroup magma_dgeev_comp
********************************************************************/
extern "C" magma_int_t
magma_dgehrd_m(
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,
double *T,
magma_int_t *info)
{
#define A( i, j ) (A + (i) + (j)*lda)
#define dA( d, i, j ) (data.A[d] + (i) + (j)*ldda)
double c_one = MAGMA_D_ONE;
double c_zero = MAGMA_D_ZERO;
magma_int_t nb = magma_get_dgehrd_nb(n);
magma_int_t nh, iws, ldda, min_lblocks, max_lblocks, last_dev, d;
magma_int_t dpanel, di, nlocal, i, i2, ib, ldwork;
magma_int_t iinfo;
magma_int_t lquery;
struct dgehrd_data data;
int ngpu = magma_num_gpus();
*info = 0;
iws = n*(nb + nb*ngpu);
work[0] = MAGMA_D_MAKE( iws, 0 );
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 < max(1,n) && ! 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;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
// 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;
// set T to zero
lapackf77_dlaset( "Full", &nb, &n, &c_zero, &c_zero, T, &nb );
// set to null, to simplify cleanup code
for( d = 0; d < ngpu; ++d ) {
data.A[d] = NULL;
data.streams[d] = NULL;
}
// 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
// allocate memory on GPUs for A and workspaces
ldda = ((n+31)/32)*32;
min_lblocks = (n / nb) / ngpu;
max_lblocks = ((n-1) / nb) / ngpu + 1;
last_dev = (n / nb) % ngpu;
// V and Vd need to be padded for copying in mdlahr2
data.ngpu = ngpu;
data.ldda = ldda;
data.ldv = nb*max_lblocks*ngpu;
data.ldvd = nb*max_lblocks;
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
nlocal = min_lblocks*nb;
if ( d < last_dev ) {
nlocal += nb;
}
else if ( d == last_dev ) {
nlocal += (n % nb);
}
ldwork = nlocal*ldda // A
+ nb*data.ldv // V
+ nb*data.ldvd // Vd
+ nb*ldda // Y
+ nb*ldda // W
+ nb*nb; // Ti
if ( MAGMA_SUCCESS != magma_dmalloc( &data.A[d], ldwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto CLEANUP;
}
data.V [d] = data.A [d] + nlocal*ldda;
data.Vd[d] = data.V [d] + nb*data.ldv;
data.Y [d] = data.Vd[d] + nb*data.ldvd;
data.W [d] = data.Y [d] + nb*ldda;
data.Ti[d] = data.W [d] + nb*ldda;
magma_queue_create( &data.streams[d] );
}
// Copy the matrix to GPUs
magma_dsetmatrix_1D_col_bcyclic( n, n, A, lda, data.A, ldda, ngpu, nb );
// round ilo down to block boundary
ilo = (ilo/nb)*nb;
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)
dpanel = (i / nb) % ngpu;
di = ((i / nb) / ngpu) * nb;
if ( i > ilo ) {
magma_setdevice( dpanel );
magma_dgetmatrix( ihi-i, nb,
dA(dpanel, i, di), ldda,
A(i,i), lda );
}
// add 1 to i for 1-based index
magma_dlahr2_m( ihi, i+1, nb, A(0,i), lda,
&tau[i], &T[i*nb], nb, work, n, &data );
magma_dlahru_m( n, ihi, i, nb, A, lda, &data );
// copy first i rows above panel to host
magma_setdevice( dpanel );
magma_dgetmatrix_async( i, nb,
dA(dpanel, 0, di), ldda,
A(0,i), lda, data.streams[dpanel] );
}
// Copy remainder to host, block-by-block
for( i2 = i; i2 < n; i2 += nb ) {
ib = min( nb, n-i2 );
d = (i2 / nb) % ngpu;
di = (i2 / nb) / ngpu * nb;
magma_setdevice( d );
magma_dgetmatrix( n, ib,
dA(d, 0, di), ldda,
A(0,i2), lda );
}
}
// 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_D_MAKE( iws, 0 );
CLEANUP:
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( data.A[d] );
magma_queue_destroy( data.streams[d] );
}
magma_setdevice( orig_dev );
return *info;
} /* magma_dgehrd */
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
/**
Purpose
-------
DLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to DLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code 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]
k INTEGER
The number of terms in the rational function to be solved by
DLAED4. K >= 0.
@param[in]
n INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N > K).
@param[in]
n1 INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
@param[out]
d DOUBLE PRECISION array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
@param[out]
Q DOUBLE PRECISION array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
@param[in]
ldq INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
@param[in]
rho DOUBLE PRECISION
The value of the parameter in the rank one update equation.
RHO >= 0 required.
@param[in,out]
dlamda DOUBLE PRECISION array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
@param[in]
Q2 DOUBLE PRECISION array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
TODO what is LDQ2?
@param[in]
indx INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see DLAED2).
The rows of the eigenvectors found by DLAED4 must be likewise
permuted before the matrix multiply can take place.
@param[in]
ctot INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
@param[in,out]
w DOUBLE PRECISION array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
@param
s (workspace) DOUBLE PRECISION array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
@param[out]
indxq INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
@param
dwork (workspace) DOUBLE PRECISION array, dimension (3*N*N/2+3*N)
@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.
TODO verify range, vl, vu, il, iu -- copied from dlaex1.
@param[in]
vl DOUBLE PRECISION
@param[in]
vu DOUBLE PRECISION
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]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = 1, an eigenvalue did not converge
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_dsyev_aux
********************************************************************/
extern "C" magma_int_t
magma_dlaex3(magma_int_t k, magma_int_t n, magma_int_t n1, double* d,
double* Q, magma_int_t ldq, double rho,
double* dlamda, double* Q2, magma_int_t* indx,
magma_int_t* ctot, double* w, double* s, magma_int_t* indxq,
double* dwork,
magma_range_t range, double vl, double vu, magma_int_t il, magma_int_t iu,
magma_int_t* info )
{
#define Q(i_,j_) (Q + (i_) + (j_)*ldq)
double d_one = 1.;
double d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
magma_int_t iil, iiu, rk;
double* dq2= dwork;
double* ds = dq2 + n*(n/2+1);
double* dq = ds + n*(n/2+1);
magma_int_t lddq = n/2 + 1;
magma_int_t i, iq2, j, n12, n2, n23, tmp, lq2;
double temp;
magma_int_t alleig, valeig, indeig;
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
*info = 0;
if (k < 0)
*info=-1;
else if (n < k)
*info=-2;
else if (ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if (*info != 0) {
magma_xerbla(__func__, -(*info));
return *info;
}
// Quick return if possible
if (k == 0)
return *info;
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
lq2 = iq2 + n2 * n23;
magma_dsetvector_async( lq2, Q2, 1, dq2, 1, NULL );
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for (i = ib; i < ie; ++i)
dlamda[i]=lapackf77_dlamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = ib; j < ie; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_dlaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0) {
#pragma omp critical (info)
*info=iinfo;
break;
}
}
#pragma omp barrier
if (*info == 0) {
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_dlamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_dvrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_dirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2) {
#pragma omp single
{
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if (k != 1) {
// Compute updated W.
blasf77_dcopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_dcopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for (j = 0; j < k; ++j) {
magma_int_t i_tmp = min(j, ie);
for (i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for (i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if (id < tot) {
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else {
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for (j = ib; j < ie; ++j) {
for (i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = magma_cblas_dnrm2( k, s+id*k, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return *info;
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
for (i = 0; i < k; ++i)
dlamda[i]=lapackf77_dlamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = 0; j < k; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_dlaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0)
*info=iinfo;
}
if (*info != 0)
return *info;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_dlamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_dvrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_dirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2) {
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if (k != 1) {
// Compute updated W.
blasf77_dcopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_dcopy( &k, Q, &tmp, w, &ione);
for (j = 0; j < k; ++j) {
for (i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for (i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for (j = iil-1; j < iiu; ++j) {
for (i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = magma_cblas_dnrm2( k, s, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#endif //_OPENMP
// Compute the updated eigenvectors.
timer_start( time );
magma_queue_sync( NULL );
if (rk != 0) {
if ( n23 != 0 ) {
if (rk < magma_get_dlaed3_k()) {
lapackf77_dlacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_dgemm("N", "N", &n2, &rk, &n23, &d_one, &Q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
} else {
magma_dsetmatrix( n23, rk, Q(ctot[0],iil-1), ldq, ds, n23 );
magma_dgemm( MagmaNoTrans, MagmaNoTrans, n2, rk, n23, d_one, &dq2[iq2], n2, ds, n23, d_zero, dq, lddq);
magma_dgetmatrix( n2, rk, dq, lddq, Q(n1,iil-1), ldq );
}
} else
lapackf77_dlaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 != 0 ) {
if (rk < magma_get_dlaed3_k()) {
lapackf77_dlacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_dgemm("N", "N", &n1, &rk, &n12, &d_one, Q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
} else {
magma_dsetmatrix( n12, rk, Q(0,iil-1), ldq, ds, n12 );
magma_dgemm( MagmaNoTrans, MagmaNoTrans, n1, rk, n12, d_one, dq2, n1, ds, n12, d_zero, dq, lddq);
magma_dgetmatrix( n1, rk, dq, lddq, Q(0,iil-1), ldq );
}
} else
lapackf77_dlaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
timer_stop( time );
timer_printf( "gemms = %6.2f\n", time );
return *info;
} /* magma_dlaex3 */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgeqrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const double d_neg_one = MAGMA_D_NEG_ONE;
const double d_one = MAGMA_D_ONE;
const double c_neg_one = MAGMA_D_NEG_ONE;
const double c_one = MAGMA_D_ONE;
const double c_zero = MAGMA_D_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double Anorm, error=0, error2=0;
double *h_A, *h_R, *tau, *h_work, tmp[1];
magmaDouble_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;
parse_opts( argc, argv, &opts );
magma_int_t status = 0;
double tol = opts.tolerance * lapackf77_dlamch("E");
printf( "version %d\n", (int) opts.version );
if ( opts.version == 2 ) {
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 = ((M+31)/32)*32;
gflops = FLOPS_DGEQRF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_dgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_D_REAL( tmp[0] );
TESTING_MALLOC_CPU( tau, double, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_work, double, lwork );
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( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_dsetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
if ( opts.version == 2 ) {
// LAPACK complaint arguments
magma_dgeqrf2_gpu( M, N, d_A, ldda, tau, &info );
}
else {
nb = magma_get_dgeqrf_nb( M );
size = (2*min(M, N) + (N+31)/32*32 )*nb;
TESTING_MALLOC_DEV( dT, double, size );
if ( opts.version == 1 ) {
// stores dT, V blocks have zeros, R blocks inverted & stored in dT
magma_dgeqrf_gpu( M, N, d_A, ldda, tau, dT, &info );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// stores dT, V blocks have zeros, R blocks stored in dT
magma_dgeqrf3_gpu( M, N, d_A, ldda, tau, dT, &info );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
exit(1);
}
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check && opts.version == 2 ) {
/* =====================================================================
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.
=================================================================== */
magma_dgetmatrix( M, N, d_A, ldda, h_R, lda );
magma_int_t ldq = M;
magma_int_t ldr = min_mn;
double *Q, *R;
double *work;
TESTING_MALLOC_CPU( Q, double, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( R, double, ldr*N ); // K by N
TESTING_MALLOC_CPU( work, double, min_mn );
// generate M by K matrix Q, where K = min(M,N)
lapackf77_dlacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_dorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_dlaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_dlacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_dgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_dlange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_dlange( "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_dlaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_dsyrk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
error2 = lapackf77_dlansy( "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 && M >= N ) {
/* =====================================================================
Check the result by solving consistent linear system, A*x = b.
Only for versions 1 & 3 with M >= N.
=================================================================== */
magma_int_t lwork;
double *x, *b, *hwork;
magmaDouble_ptr d_B;
const double c_zero = MAGMA_D_ZERO;
const double c_one = MAGMA_D_ONE;
const double c_neg_one = MAGMA_D_NEG_ONE;
const magma_int_t ione = 1;
// initialize RHS, b = A*random
TESTING_MALLOC_CPU( x, double, N );
TESTING_MALLOC_CPU( b, double, M );
lapackf77_dlarnv( &ione, ISEED, &N, x );
blasf77_dgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione );
// copy to GPU
TESTING_MALLOC_DEV( d_B, double, M );
magma_dsetvector( M, b, 1, d_B, 1 );
if ( opts.version == 1 ) {
// allocate hwork
magma_dgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork = (magma_int_t)MAGMA_D_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, double, lwork );
// solve linear system
magma_dgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork, &info );
if (info != 0)
printf("magma_dgeqrs 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_dgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork = (magma_int_t)MAGMA_D_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, double, lwork );
// solve linear system
magma_dgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork, &info );
if (info != 0)
printf("magma_dgeqrs3 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
TESTING_FREE_CPU( hwork );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
exit(1);
}
magma_dgetvector( N, d_B, 1, x, 1 );
// compute r = Ax - b, saved in b
blasf77_dgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione );
// compute residual |Ax - b| / (n*|A|*|x|)
double norm_x, norm_A, norm_r, work[1];
norm_A = lapackf77_dlange( "F", &M, &N, h_A, &lda, work );
norm_r = lapackf77_dlange( "F", &M, &ione, b, &M, work );
norm_x = lapackf77_dlange( "F", &N, &ione, x, &N, work );
TESTING_FREE_CPU( x );
TESTING_FREE_CPU( b );
TESTING_FREE_DEV( d_B );
error = norm_r / (N * norm_A * norm_x);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dgeqrf(&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_dgeqrf 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 ) {
if ( opts.version == 2 ) {
bool okay = (error < tol && error2 < tol);
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else 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 != 2 )
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dlaset
Code is very similar to testing_dlacpy.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];
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_R;
double *d_A;
double offdiag = MAGMA_D_MAKE( 1.2000, 6.7000 );
double diag = MAGMA_D_MAKE( 3.1415, 2.7183 );
magma_int_t M, N, size, lda, ldb, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull };
printf("uplo M N 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 ) {
M = opts.msize[itest];
N = opts.nsize[itest];
//M += 2; // space for insets
//N += 2;
lda = M;
ldb = lda;
ldda = ((M+31)/32)*32;
size = lda*N;
if ( uplo[iuplo] == MagmaLower || uplo[iuplo] == MagmaUpper ) {
// save triangle (with diagonal)
// TODO wrong for trapezoid
gbytes = sizeof(double) * 0.5*N*(N+1) / 1e9;
}
else {
// save entire matrix
gbytes = sizeof(double) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, double, size );
TESTING_MALLOC_CPU( h_R, double, size );
TESTING_MALLOC_DEV( d_A, double, 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_D_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
//magmablas_dlaset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_dlaset( uplo[iuplo], M, N, offdiag, diag, d_A, ldda );
gpu_time = magma_sync_wtime( 0 ) - 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_dlaset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_dlaset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_dgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_daxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_dlange("f", &M, &N, h_R, &lda, work);
printf("%4c %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
lapacke_uplo_const( uplo[iuplo] ), (int) M, (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" );
}
}
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgeqrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, error2;
double c_zero = MAGMA_D_ZERO;
double c_neg_one = MAGMA_D_NEG_ONE;
double c_one = MAGMA_D_ONE;
double *h_A, *h_T, *h_R, *tau, *h_work, tmp[1];
magmaDouble_ptr d_A, d_T, ddA, dtau;
magmaDouble_ptr d_A2, d_T2, ddA2, dtau2;
magmaDouble_ptr dwork, dwork2;
magma_int_t M, N, lda, ldda, lwork, n2, info, min_mn;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
#define BLOCK_SIZE 64
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = 10. * opts.tolerance * lapackf77_dlamch("E");
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
printf("version %d\n", (int) opts.version );
printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||R - Q^H*A|| ||R_T||\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 (N > 128) {
printf("%5d %5d skipping because dgeqr2x requires N <= 128\n",
(int) M, (int) N);
continue;
}
if (M < N) {
printf("%5d %5d skipping because dgeqr2x requires M >= N\n",
(int) M, (int) N);
continue;
}
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = (FLOPS_DGEQRF( M, N ) + FLOPS_DGEQRT( M, N )) / 1e9;
/* Allocate memory for the matrix */
TESTING_MALLOC_CPU( tau, double, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_T, double, N*N );
TESTING_MALLOC_PIN( h_R, double, n2 );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
TESTING_MALLOC_DEV( d_T, double, N*N );
TESTING_MALLOC_DEV( ddA, double, N*N );
TESTING_MALLOC_DEV( dtau, double, min_mn );
TESTING_MALLOC_DEV( d_A2, double, ldda*N );
TESTING_MALLOC_DEV( d_T2, double, N*N );
TESTING_MALLOC_DEV( ddA2, double, N*N );
TESTING_MALLOC_DEV( dtau2, double, min_mn );
TESTING_MALLOC_DEV( dwork, double, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) );
TESTING_MALLOC_DEV( dwork2, double, max(5*min_mn, (BLOCK_SIZE*2+2)*min_mn) );
// todo replace with magma_dlaset
magmablas_dlaset( MagmaFull, N, N, c_zero, c_zero, ddA, N );
magmablas_dlaset( MagmaFull, N, N, c_zero, c_zero, d_T, N );
magmablas_dlaset( MagmaFull, N, N, c_zero, c_zero, ddA2, N );
magmablas_dlaset( MagmaFull, N, N, c_zero, c_zero, d_T2, N );
lwork = -1;
lapackf77_dgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_D_REAL( tmp[0] );
lwork = max(lwork, N*N);
TESTING_MALLOC_CPU( h_work, double, lwork );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_dsetmatrix( M, N, h_R, lda, d_A, ldda );
magma_dsetmatrix( M, N, h_R, lda, d_A2, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime(0);
if (opts.version == 1)
magma_dgeqr2x_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else if (opts.version == 2)
magma_dgeqr2x2_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else if (opts.version == 3)
magma_dgeqr2x3_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info);
else {
printf( "call magma_dgeqr2x4_gpu\n" );
/*
Going through NULL stream is faster
Going through any stream is slower
Doing two streams in parallel is slower than doing them sequentially
Queuing happens on the NULL stream - user defined buffers are smaller?
*/
magma_dgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, NULL, &info);
//magma_dgeqr2x4_gpu(M, N, d_A, ldda, dtau, d_T, ddA, dwork, &info, stream[1]);
//magma_dgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, stream[0]);
//magma_dgeqr2x4_gpu(M, N, d_A2, ldda, dtau2, d_T2, ddA2, dwork2, &info, NULL);
//gflops *= 2;
}
gpu_time = magma_sync_wtime(0) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgeqr2x_gpu version %d returned error %d: %s.\n",
(int) opts.version, (int) info, magma_strerror( info ));
}
else {
if ( opts.check ) {
/* =====================================================================
Check the result, following zqrt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
=================================================================== */
magma_dgetmatrix( M, N, d_A, ldda, h_R, M );
magma_dgetmatrix( N, N, ddA, N, h_T, N );
magma_dgetmatrix( min_mn, 1, dtau, min_mn, tau, min_mn );
// Restore the upper triangular part of A before the check
for(int col=0; col < N; col++){
for(int row=0; row <= col; row++)
h_R[row + col*M] = h_T[row + col*N];
}
magma_int_t ldq = M;
magma_int_t ldr = min_mn;
double *Q, *R;
double *work;
TESTING_MALLOC_CPU( Q, double, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( R, double, ldr*N ); // K by N
TESTING_MALLOC_CPU( work, double, min_mn );
// generate M by K matrix Q, where K = min(M,N)
lapackf77_dlacpy( "Lower", &M, &min_mn, h_R, &M, Q, &ldq );
lapackf77_dorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_dlaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_dlacpy( "Upper", &min_mn, &N, h_R, &M, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_dgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
double Anorm = lapackf77_dlange( "1", &M, &N, h_A, &lda, work );
error2 = lapackf77_dlange( "1", &min_mn, &N, R, &ldr, work );
if ( N > 0 && Anorm > 0 )
error2 /= (N*Anorm);
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( R ); R = NULL;
TESTING_FREE_CPU( work ); work = NULL;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
//lapackf77_dgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
lapackf77_dlacpy( MagmaUpperLowerStr, &M, &N, h_R, &M, h_A, &lda );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&M, &N, h_A, &lda, tau, h_work, &N);
//magma_dgeqr2(&M, &N, h_A, &lda, tau, h_work, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
// Restore the upper triangular part of A before the check
for(int col=0; col < N; col++){
for(int row=0; row <= col; row++)
h_R[row + col*M] = h_T[row + col*N];
}
error = lapackf77_dlange("M", &M, &N, h_A, &lda, work);
blasf77_daxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_dlange("M", &M, &N, h_R, &lda, work) / (N * error);
// Check if T is the same
magma_dgetmatrix( N, N, d_T, N, h_T, N );
double terr = 0.;
for(int col=0; col < N; col++)
for(int row=0; row <= col; row++)
terr += ( MAGMA_D_ABS(h_work[row + col*N] - h_T[row + col*N])*
MAGMA_D_ABS(h_work[row + col*N] - h_T[row + col*N]) );
terr = sqrt( terr );
// If comparison to LAPACK fail, check || R - Q^H*A || / (N * ||A||)
// and print fail if both fails, otherwise print ok (*)
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
error2, terr, (error2 < tol ? "ok" : "failed" ));
status += ! (error2 < tol);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, 1000.*gpu_time);
}
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_T );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_T );
TESTING_FREE_DEV( ddA );
TESTING_FREE_DEV( dtau );
TESTING_FREE_DEV( dwork );
TESTING_FREE_DEV( d_A2 );
TESTING_FREE_DEV( d_T2 );
TESTING_FREE_DEV( ddA2 );
TESTING_FREE_DEV( dtau2 );
TESTING_FREE_DEV( dwork2 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsygvd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
double *h_A, *h_R, *h_B, *h_S, *h_work;
double *w1, *w2;
magma_int_t *iwork;
magma_int_t N, n2, info, nb, lwork, liwork, lda;
double result[4] = {0};
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double d_zero = 0.;
double d_one = 1.;
double d_neg_one = -1.;
//magma_int_t izero = 0;
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");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, uplo = %s\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo));
printf(" N CPU Time (sec) GPU Time(sec)\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 = N*lda;
nb = magma_get_dsytrd_nb(N);
lwork = 1 + 6*N*nb + 2* N*N;
liwork = 3 + 5*N;
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_B, double, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, double, n2 );
TESTING_MALLOC_PIN( h_S, double, n2 );
TESTING_MALLOC_PIN( h_work, double, lwork );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlarnv( &ione, ISEED, &n2, h_B );
magma_dmake_hpd( N, h_B, lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
/* warmup */
if ( opts.warmup ) {
magma_dsygvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dsygvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_dsygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvd routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
| A B Z - Z D | / ( |A||Z| N ) (itype = 2)
| B A Z - Z D | / ( |A||Z| N ) (itype = 3)
(2) | I - V V' B | / ( N ) (itype = 1,2)
| B - V V' | / ( |B| N ) (itype = 3)
(3) | S(with V) - S(w/o V) | / | S |
=================================================================== */
double temp1, temp2;
//double *tau;
if ( opts.itype == 1 || opts.itype == 2 ) {
lapackf77_dlaset( "A", &N, &N, &d_zero, &c_one, h_S, &lda);
blasf77_dgemm("N", "C", &N, &N, &N, &c_one, h_R, &lda, h_R, &lda, &d_zero, h_work, &N);
blasf77_dsymm("R", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_work, &N, &c_one, h_S, &lda);
result[1] = lapackf77_dlange("1", &N, &N, h_S, &lda, h_work) / N;
}
else if ( opts.itype == 3 ) {
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda);
blasf77_dsyrk(lapack_uplo_const(opts.uplo), "N", &N, &N, &d_neg_one, h_R, &lda, &d_one, h_S, &lda);
result[1] = lapackf77_dlansy("1", lapack_uplo_const(opts.uplo), &N, h_S, &lda, h_work) / N
/ lapackf77_dlansy("1", lapack_uplo_const(opts.uplo), &N, h_B, &lda, h_work);
}
result[0] = 1.;
result[0] /= lapackf77_dlansy("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, h_work);
result[0] /= lapackf77_dlange("1", &N, &N, h_R, &lda, h_work);
if ( opts.itype == 1 ) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &d_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
result[0] *= lapackf77_dlange("1", &N, &N, h_work, &N, &temp1)/N;
}
else if ( opts.itype == 2 ) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_R, &lda, &d_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_dlange("1", &N, &N, h_R, &lda, &temp1)/N;
}
else if ( opts.itype == 3 ) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &d_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_dlange("1", &N, &N, h_R, &lda, &temp1)/N;
}
/*
lapackf77_dsyt21(&ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, w1,
h_R, &lda,
h_R, &lda,
tau, h_work, rwork, &result[0]);
*/
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
magma_dsygvd( opts.itype, MagmaNoVec, opts.uplo,
N, h_R, lda, h_S, lda, w2,
h_work, lwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j<N; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[2] = temp2 / temp1;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dsygvd( &opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, h_B, &lda, w2,
h_work, &lwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_dsygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution time
=================================================================== */
if ( opts.check ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if ( opts.itype==1 ) {
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==2 ) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==3 ) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
if ( opts.itype==1 || opts.itype==2 ) {
printf("(2) | I - Z Z' B | / N = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
else {
printf("(2) | B - Z Z' | / (|B| N) = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
printf( "(3) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[2], (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0] < tol && result[1] < tol && result[2] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_dorgqr2(magma_int_t m, magma_int_t n, magma_int_t k,
double *A, magma_int_t lda,
double *tau,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
DORGQR generates an M-by-N DOUBLE_PRECISION 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 DGEQRF.
This version recomputes the T matrices on the CPU and sends them to the GPU.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix Q. M >= 0.
N (input) INTEGER
The number of columns of the matrix Q. M >= N >= 0.
K (input) INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
A (input/output) DOUBLE_PRECISION 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 DGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
LDA (input) INTEGER
The first dimension of the array A. LDA >= max(1,M).
TAU (input) DOUBLE_PRECISION array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by DGEQRF_GPU.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument has an illegal value
===================================================================== */
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
magma_int_t nb = magma_get_dgeqrf_nb(min(m, n));
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk; //, iinfo;
magma_int_t lddwork;
double *dA, *dV, *dW, *dT, *T;
double *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;
}
// 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 GPU work space
// ldda*n for matrix dA
// ldda*nb for dV
// lddwork*nb for dW larfb workspace
ldda = ((m + 31) / 32) * 32;
lddwork = ((n + 31) / 32) * 32;
if (MAGMA_SUCCESS != magma_dmalloc( &dA, ldda*n + ldda*nb + lddwork*nb + nb*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dV = dA + ldda*n;
dW = dA + ldda*n + ldda*nb;
dT = dA + ldda*n + ldda*nb + lddwork*nb;
// Allocate CPU work space
lwork = (n+m+nb) * nb;
magma_dmalloc_cpu( &work, lwork );
T = work;
if (work == NULL) {
magma_free( dA );
magma_free_cpu( work );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
double *V = work + (n+nb)*nb;
magma_queue_t stream;
magma_queue_create( &stream );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
/*
lapackf77_dorgqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_dlacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_dlaset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_dlarfb( MagmaLeftStr, MagmaNoTransStr, MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &n_kk, &k_kk,
V, &m_kk, work, &k_kk, A(kk, kk), &lda, work+k_kk*k_kk, &n_kk );
if (kk > 0) {
magma_dsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_dlaset( MagmaUpperLower, kk, n - kk, dA(0, kk), ldda );
}
}
if (kk > 0) {
// Use blocked code
// stream: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
magmablasSetKernelStream( stream );
for (i = ki; i >= 0; i -= nb) {
ib = min(nb, k - i);
// Send current panel to the GPU
mi = m - i;
lapackf77_dlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_dsetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&mi, &ib,
A(i,i), &lda, &tau[i], T, &nb);
magma_dsetmatrix_async( ib, ib,
T, nb,
dT , nb, stream );
// set panel to identity
magmablas_dlaset( MagmaUpperLower, i, ib, dA(0, i), ldda );
magmablas_dlaset_identity( mi, ib, dA(i, i), ldda );
magma_queue_sync( stream );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_dlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT, nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_dgetmatrix( m, n,
dA(0, 0), ldda, A(0, 0), lda);
}
magmablasSetKernelStream( NULL );
magma_queue_destroy( stream );
magma_free( dA );
magma_free_cpu( work );
return *info;
} /* magma_dorgqr */
extern "C" magma_int_t
magma_dlaex3_m(magma_int_t nrgpu,
magma_int_t k, magma_int_t n, magma_int_t n1, double* d,
double* q, magma_int_t ldq, double rho,
double* dlamda, double* q2, magma_int_t* indx,
magma_int_t* ctot, double* w, double* s, magma_int_t* indxq,
double** dwork, magma_queue_t stream[MagmaMaxGPUs][2],
char range, double vl, double vu, magma_int_t il, magma_int_t iu,
magma_int_t* info )
{
/*
Purpose
=======
DLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to DLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code 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
=========
K (input) INTEGER
The number of terms in the rational function to be solved by
DLAED4. K >= 0.
N (input) INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N>K).
N1 (input) INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
D (output) DOUBLE PRECISION array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
Q (output) DOUBLE PRECISION array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
LDQ (input) INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
RHO (input) DOUBLE PRECISION
The value of the parameter in the rank one update equation.
RHO >= 0 required.
DLAMDA (input/output) DOUBLE PRECISION array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
Q2 (input) DOUBLE PRECISION array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
INDX (input) INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see DLAED2).
The rows of the eigenvectors found by DLAED4 must be likewise
permuted before the matrix multiply can take place.
CTOT (input) INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
W (input/output) DOUBLE PRECISION array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
S (workspace) DOUBLE PRECISION array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
INDXQ (output) INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
DWORK (devices workspaces) DOUBLE PRECISION array of arrays,
dimension NRGPU.
if NRGPU = 1 the dimension of the first workspace
should be (3*N*N/2+3*N)
otherwise the NRGPU workspaces should have the size
ceil((N-N1) * (N-N1) / floor(nrgpu/2)) +
NB * ((N-N1) + (N-N1) / floor(nrgpu/2))
STREAM (device stream) magma_queue_t array,
dimension (MagmaMaxGPUs,2)
INFO (output) INTEGER
= 0: successful exit.
< 0: if INFO = -i, the i-th argument had an illegal value.
> 0: if INFO = 1, an eigenvalue did not converge
Further Details
===============
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
===================================================================== */
if (nrgpu==1){
magma_setdevice(0);
magma_dlaex3(k, n, n1, d, q, ldq, rho,
dlamda, q2, indx, ctot, w, s, indxq,
*dwork, range, vl, vu, il, iu, info );
return MAGMA_SUCCESS;
}
double d_one = 1.;
double d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
char range_[] = {range, 0};
magma_int_t iil, iiu, rk;
magma_int_t n1_loc, n2_loc, ib, nb, ib2, igpu;
magma_int_t ni_loc[MagmaMaxGPUs];
magma_int_t i,ind,iq2,j,n12,n2,n23,tmp,lq2;
double temp;
magma_int_t alleig, valeig, indeig;
alleig = lapackf77_lsame(range_, "A");
valeig = lapackf77_lsame(range_, "V");
indeig = lapackf77_lsame(range_, "I");
*info = 0;
if(k < 0)
*info=-1;
else if(n < k)
*info=-2;
else if(ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if(*info != 0){
magma_xerbla(__func__, -(*info));
return MAGMA_ERR_ILLEGAL_VALUE;
}
// Quick return if possible
if(k == 0)
return MAGMA_SUCCESS;
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
//#define CHECK_CPU
#ifdef CHECK_CPU
double *hwS[2][MagmaMaxGPUs], *hwQ[2][MagmaMaxGPUs], *hwQ2[MagmaMaxGPUs];
#define hQ2(id) (hwQ2[id])
#define hS(id, ii) (hwS[ii][id])
#define hQ(id, ii) (hwQ[ii][id])
#endif
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
lq2 = iq2 + n2 * n23;
n1_loc = (n1-1) / (nrgpu/2) + 1;
n2_loc = (n2-1) / (nrgpu/2) + 1;
nb = magma_get_dlaex3_m_nb();
if (n1 >= magma_get_dlaex3_m_k()){
#ifdef CHECK_CPU
for (igpu = 0; igpu < nrgpu; ++igpu){
magma_dmalloc_pinned( &(hwS[0][igpu]), n2*nb );
magma_dmalloc_pinned( &(hwS[1][igpu]), n2*nb );
magma_dmalloc_pinned( &(hwQ2[igpu]), n2*n2_loc );
magma_dmalloc_pinned( &(hwQ[0][igpu]), n2_loc*nb );
magma_dmalloc_pinned( &(hwQ[1][igpu]), n2_loc*nb );
}
#endif
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
ni_loc[igpu] = min(n1_loc, n1 - igpu/2 * n1_loc);
#ifdef CHECK_CPU
lapackf77_dlacpy("A", &ni_loc[igpu], &n12, q2+n1_loc*(igpu/2), &n1, hQ2(igpu), &n1_loc);
#endif
magma_setdevice(igpu);
magma_dsetmatrix_async( ni_loc[igpu], n12,
q2+n1_loc*(igpu/2), n1,
dQ2(igpu), n1_loc, stream[igpu][0] );
ni_loc[igpu+1] = min(n2_loc, n2 - igpu/2 * n2_loc);
#ifdef CHECK_CPU
lapackf77_dlacpy("A", &ni_loc[igpu+1], &n23, q2+iq2+n2_loc*(igpu/2), &n2, hQ2(igpu+1), &n2_loc);
#endif
magma_setdevice(igpu+1);
magma_dsetmatrix_async( ni_loc[igpu+1], n23,
q2+iq2+n2_loc*(igpu/2), n2,
dQ2(igpu+1), n2_loc, stream[igpu+1][0] );
}
}
//
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
magma_timestr_t start, end;
start = get_current_time();
#endif
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for(i = ib; i < ie; ++i)
dlamda[i]=lapackf77_dlamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for(j = ib; j < ie; ++j){
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_dlaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if(iinfo != 0){
#pragma omp critical (info)
*info=iinfo;
break;
}
}
#pragma omp barrier
if(*info == 0){
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_dlamrg( &k, &nk, d, &ione , &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_dvrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_dirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2){
#pragma omp single
{
for(j = 0; j < k; ++j){
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if(k != 1){
// Compute updated W.
blasf77_dcopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_dcopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for(j = 0; j < k; ++j){
magma_int_t i_tmp = min(j, ie);
for(i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for(i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for(i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if(id < tot){
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else{
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for(j = ib; j < ie; ++j){
for(i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = cblas_dnrm2( k, s+id*k, 1);
for(i = 0; i < k; ++i){
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return MAGMA_SUCCESS; //??????
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("eigenvalues/vector D+zzT = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
magma_timestr_t start, end;
start = get_current_time();
#endif
for(i = 0; i < k; ++i)
dlamda[i]=lapackf77_dlamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for(j = 0; j < k; ++j){
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_dlaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if(iinfo != 0)
*info=iinfo;
}
if(*info != 0)
return MAGMA_SUCCESS;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_dlamrg( &k, &nk, d, &ione , &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_dvrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_dirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2){
for(j = 0; j < k; ++j){
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if(k != 1){
// Compute updated W.
blasf77_dcopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_dcopy( &k, q, &tmp, w, &ione);
for(j = 0; j < k; ++j){
for(i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for(i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for(i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for(j = iil-1; j < iiu; ++j){
for(i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = cblas_dnrm2( k, s, 1);
for(i = 0; i < k; ++i){
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("eigenvalues/vector D+zzT = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
#endif //_OPENMP
// Compute the updated eigenvectors.
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
start = get_current_time();
#endif
if(rk > 0){
if (n1 < magma_get_dlaex3_m_k()){
// stay on the CPU
if( n23 != 0 ){
lapackf77_dlacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_dgemm("N", "N", &n2, &rk, &n23, &d_one, &q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
}
else
lapackf77_dlaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if( n12 != 0 ) {
lapackf77_dlacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_dgemm("N", "N", &n1, &rk, &n12, &d_one, q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
}
else
lapackf77_dlaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
else {
//use the gpus
ib = min(nb, rk);
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_dsetmatrix_async( n23, ib,
Q(ctot[0],iil-1), ldq,
dS(igpu+1,0), n23, stream[igpu+1][0] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_dsetmatrix_async( n12, ib,
Q(0,iil-1), ldq,
dS(igpu,0), n12, stream[igpu][0] );
}
}
for (i = 0; i<rk; i+=nb){
ib = min(nb, rk - i);
ind = (i/nb)%2;
if (i+nb<rk){
ib2 = min(nb, rk - i - nb);
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_dsetmatrix_async( n23, ib2,
Q(ctot[0],iil-1+i+nb), ldq,
dS(igpu+1,(ind+1)%2), n23, stream[igpu+1][(ind+1)%2] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_dsetmatrix_async( n12, ib2,
Q(0,iil-1+i+nb), ldq,
dS(igpu,(ind+1)%2), n12, stream[igpu][(ind+1)%2] );
}
}
}
// Ensure that the data is copied on gpu since we will overwrite it.
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
if (n23 != 0) {
#ifdef CHECK_CPU
lapackf77_dlacpy("A", &n23, &ib, Q(ctot[0],iil-1+i), &ldq, hS(igpu+1,ind), &n23);
#endif
magma_setdevice(igpu+1);
magma_queue_sync( stream[igpu+1][ind] );
}
if (n12 != 0) {
#ifdef CHECK_CPU
lapackf77_dlacpy("A", &n12, &ib, Q(0,iil-1+i), &ldq, hS(igpu,ind), &n12);
#endif
magma_setdevice(igpu);
magma_queue_sync( stream[igpu][ind] );
}
}
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
if (n23 != 0) {
#ifdef CHECK_CPU
blasf77_dgemm("N", "N", &ni_loc[igpu+1], &ib, &n23, &d_one, hQ2(igpu+1), &n2_loc,
hS(igpu+1,ind), &n23, &d_zero, hQ(igpu+1, ind), &n2_loc);
#endif
magma_setdevice(igpu+1);
magmablasSetKernelStream(stream[igpu+1][ind]);
magma_dgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu+1], ib, n23, d_one, dQ2(igpu+1), n2_loc,
dS(igpu+1, ind), n23, d_zero, dQ(igpu+1, ind), n2_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu+1, cpu_gpu_ddiff(ni_loc[igpu+1], ib, hQ(igpu+1, ind), n2_loc, dQ(igpu+1, ind), n2_loc));
#endif
}
if (n12 != 0) {
#ifdef CHECK_CPU
blasf77_dgemm("N", "N", &ni_loc[igpu], &ib, &n12, &d_one, hQ2(igpu), &n1_loc,
hS(igpu,ind%2), &n12, &d_zero, hQ(igpu, ind%2), &n1_loc);
#endif
magma_setdevice(igpu);
magmablasSetKernelStream(stream[igpu][ind]);
magma_dgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu], ib, n12, d_one, dQ2(igpu), n1_loc,
dS(igpu, ind), n12, d_zero, dQ(igpu, ind), n1_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu, cpu_gpu_ddiff(ni_loc[igpu], ib, hQ(igpu, ind), n1_loc, dQ(igpu, ind), n1_loc));
#endif
}
}
for (igpu = 0; igpu < nrgpu-1; igpu += 2){
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_dgetmatrix( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
Q(n1+n2_loc*(igpu/2),iil-1+i), ldq );
// magma_dgetmatrix_async( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
// Q(n1+n2_loc*(igpu/2),iil-1+i), ldq, stream[igpu+1][ind] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_dgetmatrix( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
Q(n1_loc*(igpu/2),iil-1+i), ldq );
// magma_dgetmatrix_async( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
// Q(n1_loc*(igpu/2),iil-1+i), ldq, stream[igpu][ind] );
}
}
}
for (igpu = 0; igpu < nrgpu; ++igpu){
#ifdef CHECK_CPU
magma_free_pinned( hwS[1][igpu] );
magma_free_pinned( hwS[0][igpu] );
magma_free_pinned( hwQ2[igpu] );
magma_free_pinned( hwQ[1][igpu] );
magma_free_pinned( hwQ[0][igpu] );
#endif
magma_setdevice(igpu);
magmablasSetKernelStream(NULL);
magma_queue_sync( stream[igpu][0] );
magma_queue_sync( stream[igpu][1] );
}
if( n23 == 0 )
lapackf77_dlaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if( n12 == 0 )
lapackf77_dlaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
}
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("gemms = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
return MAGMA_SUCCESS;
} /*magma_dlaed3_m*/
/**
Purpose
-------
DSTEDX computes some eigenvalues and, optionally, eigenvectors of a
symmetric tridiagonal matrix using the divide and conquer method.
This code 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. See DLAEX3 for details.
Arguments
---------
@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]
n INTEGER
The dimension of the symmetric tridiagonal matrix. N >= 0.
@param[in]
vl DOUBLE PRECISION
@param[in]
vu DOUBLE PRECISION
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[in,out]
d DOUBLE PRECISION array, dimension (N)
On entry, the diagonal elements of the tridiagonal matrix.
On exit, if INFO = 0, the eigenvalues in ascending order.
@param[in,out]
e DOUBLE PRECISION array, dimension (N-1)
On entry, the subdiagonal elements of the tridiagonal matrix.
On exit, E has been destroyed.
@param[in,out]
Z DOUBLE PRECISION array, dimension (LDZ,N)
On exit, if INFO = 0, Z contains the orthonormal eigenvectors
of the symmetric tridiagonal matrix.
@param[in]
ldz INTEGER
The leading dimension of the array Z. LDZ >= max(1,N).
@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 dimension of the array WORK.
If N > 1 then LWORK >= ( 1 + 4*N + N**2 ).
Note that if N is less than or
equal to the minimum divide size, usually 25, then LWORK need
only be max(1,2*(N-1)).
\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]
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.
LIWORK >= ( 3 + 5*N ).
Note that if N is less than or
equal to the minimum divide size, usually 25, then LIWORK
need only be 1.
\n
If LIWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal size of the IWORK array,
returns this value as the first entry of the IWORK array, and
no error message related to LIWORK is issued by XERBLA.
@param
dwork (workspace) DOUBLE PRECISION array, dimension (3*N*N/2+3*N)
@param[out]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: 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).
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_dsyev_comp
********************************************************************/
extern "C" magma_int_t
magma_dstedx(
magma_range_t range, magma_int_t n, double vl, double vu,
magma_int_t il, magma_int_t iu, double *d, double *e,
double *Z, magma_int_t ldz,
double *work, magma_int_t lwork,
magma_int_t *iwork, magma_int_t liwork,
magmaDouble_ptr dwork,
magma_int_t *info)
{
#define Z(i_,j_) (Z + (i_) + (j_)*ldz)
double d_zero = 0.;
double d_one = 1.;
magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t alleig, indeig, valeig, lquery;
magma_int_t i, j, k, m;
magma_int_t liwmin, lwmin;
magma_int_t start, end, smlsiz;
double eps, orgnrm, p, tiny;
// Test the input parameters.
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
lquery = (lwork == -1 || liwork == -1);
*info = 0;
if (! (alleig || valeig || indeig)) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldz < max(1,n)) {
*info = -10;
} else {
if (valeig) {
if (n > 0 && vu <= vl) {
*info = -4;
}
} else if (indeig) {
if (il < 1 || il > max(1,n)) {
*info = -5;
} else if (iu < min(n,il) || iu > n) {
*info = -6;
}
}
}
if (*info == 0) {
// Compute the workspace requirements
smlsiz = magma_get_smlsize_divideconquer();
if ( n <= 1 ) {
lwmin = 1;
liwmin = 1;
} else {
lwmin = 1 + 4*n + n*n;
liwmin = 3 + 5*n;
}
work[0] = magma_dmake_lwork( lwmin );
iwork[0] = liwmin;
if (lwork < lwmin && ! lquery) {
*info = -12;
} else if (liwork < liwmin && ! lquery) {
*info = -14;
}
}
if (*info != 0) {
magma_xerbla( __func__, -(*info));
return *info;
} else if (lquery) {
return *info;
}
// Quick return if possible
if (n == 0)
return *info;
if (n == 1) {
*Z = 1.;
return *info;
}
/* determine the number of threads *///not needed here to be checked Azzam
//magma_int_t threads = magma_get_parallel_numthreads();
//magma_int_t mklth = magma_get_lapack_numthreads();
//magma_set_lapack_numthreads(mklth);
#ifdef ENABLE_DEBUG
//printf(" D&C is using %d threads\n", threads);
#endif
// If N is smaller than the minimum divide size (SMLSIZ+1), then
// solve the problem with another solver.
if (n < smlsiz) {
lapackf77_dsteqr("I", &n, d, e, Z, &ldz, work, info);
} else {
lapackf77_dlaset("F", &n, &n, &d_zero, &d_one, Z, &ldz);
//Scale.
orgnrm = lapackf77_dlanst("M", &n, d, e);
if (orgnrm == 0) {
work[0] = magma_dmake_lwork( lwmin );
iwork[0] = liwmin;
return *info;
}
eps = lapackf77_dlamch( "Epsilon" );
if (alleig) {
start = 0;
while ( start < n ) {
// Let FINISH be the position of the next subdiagonal entry
// such that E( END ) <= TINY or FINISH = N if no such
// subdiagonal exists. The matrix identified by the elements
// between START and END constitutes an independent
// sub-problem.
for (end = start+1; end < n; ++end) {
tiny = eps * sqrt( MAGMA_D_ABS(d[end-1]*d[end]));
if (MAGMA_D_ABS(e[end-1]) <= tiny)
break;
}
// (Sub) Problem determined. Compute its size and solve it.
m = end - start;
if (m == 1) {
start = end;
continue;
}
if (m > smlsiz) {
// Scale
orgnrm = lapackf77_dlanst("M", &m, &d[start], &e[start]);
lapackf77_dlascl("G", &izero, &izero, &orgnrm, &d_one, &m, &ione, &d[start], &m, info);
magma_int_t mm = m-1;
lapackf77_dlascl("G", &izero, &izero, &orgnrm, &d_one, &mm, &ione, &e[start], &mm, info);
magma_dlaex0( m, &d[start], &e[start], Z(start, start), ldz, work, iwork, dwork, MagmaRangeAll, vl, vu, il, iu, info);
if ( *info != 0) {
return *info;
}
// Scale Back
lapackf77_dlascl("G", &izero, &izero, &d_one, &orgnrm, &m, &ione, &d[start], &m, info);
} else {
lapackf77_dsteqr( "I", &m, &d[start], &e[start], Z(start, start), &ldz, work, info);
if (*info != 0) {
*info = (start+1) *(n+1) + end;
}
}
start = end;
}
// If the problem split any number of times, then the eigenvalues
// will not be properly ordered. Here we permute the eigenvalues
// (and the associated eigenvectors) into ascending order.
if (m < n) {
// Use Selection Sort to minimize swaps of eigenvectors
for (i = 1; i < n; ++i) {
k = i-1;
p = d[i-1];
for (j = i; j < n; ++j) {
if (d[j] < p) {
k = j;
p = d[j];
}
}
if (k != i-1) {
d[k] = d[i-1];
d[i-1] = p;
blasf77_dswap(&n, Z(0,i-1), &ione, Z(0,k), &ione);
}
}
}
} else {
// Scale
lapackf77_dlascl("G", &izero, &izero, &orgnrm, &d_one, &n, &ione, d, &n, info);
magma_int_t nm = n-1;
lapackf77_dlascl("G", &izero, &izero, &orgnrm, &d_one, &nm, &ione, e, &nm, info);
magma_dlaex0( n, d, e, Z, ldz, work, iwork, dwork, range, vl, vu, il, iu, info);
if ( *info != 0) {
return *info;
}
// Scale Back
lapackf77_dlascl("G", &izero, &izero, &d_one, &orgnrm, &n, &ione, d, &n, info);
}
}
work[0] = magma_dmake_lwork( lwmin );
iwork[0] = liwmin;
return *info;
} /* magma_dstedx */
