/*-------------------------------------------------------------------
* Check the orthogonality of Q
*/
static magma_int_t check_orthogonality(magma_int_t M, magma_int_t N, magmaFloatComplex *Q, magma_int_t LDQ, float eps)
{
float done = 1.0;
float mdone = -1.0;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
float normQ, result;
magma_int_t info_ortho;
magma_int_t minMN = min(M, N);
float *work = (float *)malloc(minMN*sizeof(float));
/* Build the idendity matrix */
magmaFloatComplex *Id = (magmaFloatComplex *) malloc(minMN*minMN*sizeof(magmaFloatComplex));
lapackf77_claset("A", &minMN, &minMN, &c_zero, &c_one, Id, &minMN);
/* Perform Id - Q'Q */
if (M >= N)
blasf77_cherk("U", "C", &N, &M, &done, Q, &LDQ, &mdone, Id, &N);
else
blasf77_cherk("U", "N", &M, &N, &done, Q, &LDQ, &mdone, Id, &M);
normQ = lapackf77_clanhe("I", "U", &minMN, Id, &minMN, work);
result = normQ / (minMN * 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;
}
free(work); free(Id);
return info_ortho;
}
extern "C" magma_int_t
magma_cungqr(magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *dT, magma_int_t nb,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
CUNGQR generates an M-by-N COMPLEX 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 CGEQRF.
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) COMPLEX 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 CGEQRF_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) COMPLEX array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by CGEQRF_GPU.
DT (input) COMPLEX array on the GPU device.
DT contains the T matrices used in blocking the elementary
reflectors H(i), e.g., this can be the 6th argument of
magma_cgeqrf_gpu.
NB (input) INTEGER
This is the block size used in CGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in DT.
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)
#define dT(j) (dT + (j)*nb)
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
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;
magmaFloatComplex *dA, *dV, *dW;
magmaFloatComplex *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_cmalloc( &dA, ldda*n + ldda*nb + lddwork*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dV = dA + ldda*n;
dW = dA + ldda*n + ldda*nb;
// Allocate CPU work space
lwork = (n+m+nb) * nb;
magma_cmalloc_cpu( &work, lwork );
if (work == NULL) {
magma_free( dA );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magmaFloatComplex *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;
/*
// Replacing this with the following 4 routines works but cungqr is slow for
// k smaller than the cungqr's blocking size (new version can be up to 60x faster)
lapackf77_cungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_clacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_claset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_clarfb( 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_csetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_claset( 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_claset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_csetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
// set panel to identity
magmablas_claset( MagmaUpperLower, i, ib, dA(0, i), ldda );
magmablas_claset_identity( mi, ib, dA(i, i), ldda );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT(i), nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_cgetmatrix( 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_cungqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing clarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t M, N, K, size, ldc, ldv, ldt, ldw, nv;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float error, work[1];
magma_int_t status = 0;
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans [] = { MagmaConjTrans, MagmaNoTrans };
magma_direct_t direct[] = { MagmaForward, MagmaBackward };
magma_storev_t storev[] = { MagmaColumnwise, MagmaRowwise };
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf(" M N K storev side direct trans ||R||_F / ||HC||_F\n");
printf("========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
K = opts.ksize[itest];
if ( M < K || N < K || K <= 0 ) {
printf( "%5d %5d %5d skipping because clarfb requires M >= K, N >= K, K >= 0\n",
(int) M, (int) N, (int) K );
continue;
}
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 ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
ldc = ((M+31)/32)*32;
ldt = ((K+31)/32)*32;
ldw = (side[iside] == MagmaLeft ? N : M);
// (ldv, nv) get swapped later if rowwise
ldv = (side[iside] == MagmaLeft ? M : N);
nv = K;
// Allocate memory for matrices
magmaFloatComplex *C, *R, *V, *T, *W;
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( V, magmaFloatComplex, ldv*K );
TESTING_MALLOC_CPU( T, magmaFloatComplex, ldt*K );
TESTING_MALLOC_CPU( W, magmaFloatComplex, ldw*K );
magmaFloatComplex_ptr dC, dV, dT, dW;
TESTING_MALLOC_DEV( dC, magmaFloatComplex, ldc*N );
TESTING_MALLOC_DEV( dV, magmaFloatComplex, ldv*K );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, ldt*K );
TESTING_MALLOC_DEV( dW, magmaFloatComplex, ldw*K );
// C is M x N.
size = ldc*N;
lapackf77_clarnv( &ione, ISEED, &size, C );
//printf( "C=" ); magma_cprint( M, N, C, ldc );
// V is ldv x nv. See larfb docs for description.
// if column-wise and left, M x K
// if column-wise and right, N x K
// if row-wise and left, K x M
// if row-wise and right, K x N
size = ldv*nv;
lapackf77_clarnv( &ione, ISEED, &size, V );
if ( storev[istor] == MagmaColumnwise ) {
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( 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_claset( MagmaLowerStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, &V[(nv-K)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_cprint( ldv, nv, V, ldv );
// T is K x K, upper triangular for forward, and lower triangular for backward
magma_int_t k1 = K-1;
size = ldt*K;
lapackf77_clarnv( &ione, ISEED, &size, T );
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_claset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_cprint( K, K, T, ldt );
magma_csetmatrix( M, N, C, ldc, dC, ldc );
magma_csetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_csetmatrix( K, K, T, ldt, dT, ldt );
lapackf77_clarfb( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
lapack_direct_const( direct[idir] ), lapack_storev_const( storev[istor] ),
&M, &N, &K,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_cprint( M, N, C, ldc );
magma_clarfb_gpu( side[iside], trans[itran], direct[idir], storev[istor],
M, N, K,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_cgetmatrix( M, N, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_cprint( M, N, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_clange( "Fro", &M, &N, C, &ldc, work );
size = ldc*N;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "Fro", &M, &N, R, &ldc, work ) / error;
printf( "%5d %5d %5d %c %c %c %c %8.2e %s\n",
(int) M, (int) N, (int) K,
lapacke_storev_const(storev[istor]), lapacke_side_const(side[iside]),
lapacke_direct_const(direct[idir]), lapacke_trans_const(trans[itran]),
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( V );
TESTING_FREE_CPU( T );
TESTING_FREE_CPU( W );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dV );
TESTING_FREE_DEV( dT );
TESTING_FREE_DEV( dW );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}}}}
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgelqf_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const float d_neg_one = MAGMA_D_NEG_ONE;
const float d_one = MAGMA_D_ONE;
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float Anorm, error=0, error2=0;
magmaFloatComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloatComplex_ptr d_A;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |L - A*Q^H| |I - Q*Q^H|\n");
printf("%%==============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
n2 = lda*N;
nb = magma_get_cgeqrf_nb( M, N );
gflops = FLOPS_CGELQF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_cgelqf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_C_REAL( tmp[0] );
lwork = max( lwork, M*nb );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_R, lda, d_A, ldda, opts.queue );
gpu_time = magma_wtime();
magma_cgelqf_gpu( M, N, d_A, ldda, tau, h_work, lwork, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cgelqf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the result, following zlqt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
=================================================================== */
if ( opts.check ) {
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda, opts.queue );
magma_int_t ldq = min_mn;
magma_int_t ldl = M;
magmaFloatComplex *Q, *L;
float *work;
TESTING_MALLOC_CPU( Q, magmaFloatComplex, ldq*N ); // K by N
TESTING_MALLOC_CPU( L, magmaFloatComplex, ldl*min_mn ); // M by K
TESTING_MALLOC_CPU( work, float, min_mn );
// generate K by N matrix Q, where K = min(M,N)
lapackf77_clacpy( "Upper", &min_mn, &N, h_R, &lda, Q, &ldq );
lapackf77_cunglq( &min_mn, &N, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy N by K matrix L
lapackf77_claset( "Upper", &M, &min_mn, &c_zero, &c_zero, L, &ldl );
lapackf77_clacpy( "Lower", &M, &min_mn, h_R, &lda, L, &ldl );
// error = || L - A*Q^H || / (N * ||A||)
blasf77_cgemm( "NoTrans", "Conj", &M, &min_mn, &N,
&c_neg_one, h_A, &lda, Q, &ldq, &c_one, L, &ldl );
Anorm = lapackf77_clange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_clange( "1", &M, &min_mn, L, &ldl, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set L = I (K by K), then L = I - Q*Q^H
// error = || I - Q*Q^H || / N
lapackf77_claset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, L, &ldl );
blasf77_cherk( "Upper", "NoTrans", &min_mn, &N, &d_neg_one, Q, &ldq, &d_one, L, &ldl );
error2 = safe_lapackf77_clanhe( "1", "Upper", &min_mn, L, &ldl, work );
if ( N > 0 )
error2 /= N;
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( L ); L = NULL;
TESTING_FREE_CPU( work ); work = NULL;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cgelqf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapack_cgelqf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* =====================================================================
Print performance and error.
=================================================================== */
printf("%5d %5d ", (int) M, (int) N );
if ( opts.lapack ) {
printf( "%7.2f (%7.2f)", cpu_perf, cpu_time );
}
else {
printf(" --- ( --- )" );
}
printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time );
if ( opts.check ) {
bool okay = (error < tol && error2 < tol);
printf( "error %.4g, error2 %.4g, tol %.4g, okay %d\n", error, error2, tol, okay );
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
CUNGQR generates an M-by-N COMPLEX 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 CGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A COMPLEX 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 CGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by CGEQRF_GPU.
@param[in]
T COMPLEX 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_cgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in CGEQRF_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_cgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_cungqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *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)
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_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;
magmaFloatComplex *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 };
magmaFloatComplex *dA[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *dT[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *dV[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *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_cmalloc( &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_cmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
magmaFloatComplex *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;
// cungqr requires less workspace (n*nb), but is slow if k < cungqr's block size.
// replacing it with the 4 routines below is much faster (e.g., 60x).
//magma_int_t iinfo;
//lapackf77_cungqr( &m_kk, &n_kk, &k_kk,
// A(kk, kk), &lda,
// &tau[kk], work, &lwork, &iinfo );
lapackf77_clacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, work_V, &m_kk);
lapackf77_claset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
work_V, &m_kk, &tau[kk], work_T, &k_kk);
lapackf77_clarfb( 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_csetmatrix( m_kk, jb,
A(kk, j), lda,
dA(d, kk, di), ldda, queues[d] );
// Set A(1:kk,kk+1:n) to zero.
magmablas_claset( 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_csetmatrix_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_claset( "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_csetmatrix_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_claset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda, queues[dpanel] );
magmablas_claset( 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_clarfb_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_cgetmatrix_1D_col_bcyclic( m, n, dA, ldda, A, lda, ngpu, nb, queues );
trace_cpu_end( 0 );
}
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "cungqr-n%d-ngpu%d.svg", m, ngpu );
trace_finalize( name, "trace.css" );
#endif
cleanup:
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( dA[d] );
magma_queue_destroy( queues[d] );
}
magma_free_cpu( work );
magma_setdevice( orig_dev );
return *info;
} /* magma_cungqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing claset
Code is very similar to testing_clacpy.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex *d_A;
magmaFloatComplex offdiag = MAGMA_C_MAKE( 1.2000, 6.7000 );
magmaFloatComplex diag = MAGMA_C_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(magmaFloatComplex) * 0.5*N*(N+1) / 1e9;
}
else {
// save entire matrix
gbytes = sizeof(magmaFloatComplex) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, size );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, size );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
//magmablas_claset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_claset( uplo[iuplo], M, N, offdiag, diag, d_A, ldda );
gpu_time = magma_sync_wtime( 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_claset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_claset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &M, &N, h_R, &lda, work);
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;
}
int main(int argc, char **argv)
{
TESTING_INIT();
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
real_Double_t gflops, magma_perf, magma_time, cpu_perf, cpu_time;
float magma_error, work[1];
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaFloatComplex alpha = MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y, *Ymagma;
magmaFloatComplex_ptr dA, dX, dY, dwork;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf(" N MAGMA Gflop/s (ms) CPU Gflop/s (ms) MAGMA error\n");
printf("=========================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_CSYMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaFloatComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaFloatComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaFloatComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaFloatComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaFloatComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaFloatComplex, ldwork );
magmablas_claset( MagmaFull, ldwork, 1, MAGMA_C_NAN, MAGMA_C_NAN, dwork, ldwork );
magmablas_claset( MagmaFull, ldda, N, MAGMA_C_NAN, MAGMA_C_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
magma_cmake_hermitian( N, A, lda );
// should not use data from the opposite triangle -- fill with NAN to check
magma_int_t N1 = N-1;
if ( opts.uplo == MagmaUpper ) {
lapackf77_claset( "Lower", &N1, &N1, &MAGMA_C_NAN, &MAGMA_C_NAN, &A[1], &lda );
}
else {
lapackf77_claset( "Upper", &N1, &N1, &MAGMA_C_NAN, &MAGMA_C_NAN, &A[lda], &lda );
}
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* Note: CUBLAS does not implement csymv */
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_csetmatrix( N, N, A, lda, dA, ldda );
magma_csetvector( N, X, incx, dX, incx );
magma_csetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
if ( opts.version == 1 ) {
magmablas_csymv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue );
}
else {
// non-work interface (has added overhead)
magmablas_csymv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
}
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
lapackf77_csymv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_caxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "M", &N, &ione, Ymagma, &N, work ) / N;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N,
magma_perf, 1000.*magma_time,
cpu_perf, 1000.*cpu_time,
magma_error, (magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_cungqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *T, magma_int_t nb,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
CUNGQR generates an M-by-N COMPLEX 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 CGEQRF.
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) COMPLEX 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 CGEQRF_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) COMPLEX array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by CGEQRF_GPU.
T (input) COMPLEX 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_cgeqrf_gpu (except stored on the CPU, not the GPU).
NB (input) INTEGER
This is the block size used in CGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in T.
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(d,i,j) (dA[d] + (i) + (j)*ldda)
#define dT(d,i,j) (dT[d] + (i) + (j)*nb)
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldwork;
magma_int_t i, ib, ki, kk, iinfo;
magmaFloatComplex *work;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
magma_int_t di, dn;
int dpanel;
int ngpu = magma_num_gpus();
int doriginal;
magma_getdevice( &doriginal );
// Allocate memory on GPUs for A and workspaces
magma_int_t ldda = ((m + 31) / 32) * 32;
magma_int_t lddwork = ((n + 31) / 32) * 32;
magma_int_t min_lblocks = (n / nb) / ngpu; // min. blocks per gpu
magma_int_t last_dev = (n / nb) % ngpu; // device with last block
magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 };
magmaFloatComplex *dA[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *dT[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *dV[ MagmaMaxGPUs ] = { NULL };
magmaFloatComplex *dW[ MagmaMaxGPUs ] = { NULL };
magma_queue_t stream[ MagmaMaxGPUs ] = { NULL };
for( int d = 0; d < ngpu; ++d ) {
// example with n = 75, nb = 10, ngpu = 3
// min_lblocks = 2
// last_dev = 1
// gpu 0: 2 blocks, cols: 0- 9, 30-39, 60-69
// gpu 1: 1+ blocks, cols: 10-19, 40-49, 70-74 (partial)
// gpu 2: 1 block , cols: 20-29, 50-59
magma_setdevice( d );
nlocal[d] = min_lblocks*nb;
if ( d < last_dev ) {
nlocal[d] += nb;
}
else if ( d == last_dev ) {
nlocal[d] += (n % nb);
}
ldwork = nlocal[d]*ldda // dA
+ nb*m // dT
+ nb*ldda // dV
+ nb*lddwork; // dW
if ( MAGMA_SUCCESS != magma_cmalloc( &dA[d], ldwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto CLEANUP;
}
dT[d] = dA[d] + nlocal[d]*ldda;
dV[d] = dT[d] + nb*m;
dW[d] = dV[d] + nb*ldda;
magma_queue_create( &stream[d] );
}
trace_init( 1, ngpu, 1, stream );
// first kk columns are handled by blocked method.
// ki is start of 2nd-to-last block
if ((nb > 1) && (nb < k)) {
ki = (k - nb - 1) / nb * nb;
kk = min(k, ki + nb);
} else {
ki = 0;
kk = 0;
}
// Allocate CPU work space
// n*nb for cungqr workspace
lwork = n * nb;
magma_cmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto CLEANUP;
}
// Use unblocked code for the last or only block.
if (kk < n) {
trace_cpu_start( 0, "ungqr", "ungqr last block" );
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
dpanel = (kk / nb) % ngpu;
di = ((kk / nb) / ngpu) * nb;
magma_setdevice( dpanel );
lapackf77_cungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
magma_csetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(dpanel, kk, di), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_claset( MagmaUpperLower, kk, n - kk, dA(dpanel, 0, di), ldda );
trace_cpu_end( 0 );
}
if (kk > 0) {
// Use blocked code
// send T to all GPUs
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set T" );
magma_csetmatrix_async( nb, min(m,n), T, nb, dT[d], nb, stream[d] );
trace_gpu_end( d, 0 );
}
// stream: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
for( i = ki; i >= 0; i -= nb ) {
ib = min(nb, k - i);
mi = m - i;
dpanel = (i / nb) % ngpu;
di = ((i / nb) / ngpu) * nb;
// Send current panel to the GPUs
lapackf77_claset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set V" );
magma_csetmatrix_async( mi, ib,
A(i, i), lda,
dV[d], ldda, stream[d] );
trace_gpu_end( d, 0 );
}
// set panel to identity
magma_setdevice( dpanel );
magmablasSetKernelStream( stream[dpanel] );
trace_gpu_start( dpanel, 0, "laset", "laset" );
magmablas_claset( MagmaUpperLower, i, ib, dA(dpanel, 0, di), ldda );
magmablas_claset_identity( mi, ib, dA(dpanel, i, di), ldda );
trace_gpu_end( dpanel, 0 );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( stream[d] );
magma_indices_1D_bcyclic( nb, ngpu, d, i, n, &di, &dn );
trace_gpu_start( d, 0, "larfb", "larfb" );
magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, dn-di, ib,
dV[d], ldda, dT(d,0,i), nb,
dA(d, i, di), ldda, dW[d], lddwork );
trace_gpu_end( d, 0 );
}
}
}
}
// copy result back to CPU
trace_cpu_start( 0, "get", "get A" );
magma_cgetmatrix_1D_col_bcyclic( m, n, dA, ldda, A, lda, ngpu, nb );
trace_cpu_end( 0 );
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "cungqr-n%d-ngpu%d.svg", m, ngpu );
trace_finalize( name, "trace.css" );
#endif
CLEANUP:
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( NULL );
magma_free( dA[d] );
dA[d] = NULL;
if ( stream[d] != NULL ) {
magma_queue_destroy( stream[d] );
}
}
magma_free_cpu( work );
magma_setdevice( doriginal );
return *info;
} /* magma_cungqr */
magma_int_t magma_ctrevc3_mt(
magma_side_t side, magma_vec_t howmany,
magma_int_t *select, // logical in Fortran
magma_int_t n,
magmaFloatComplex *T, magma_int_t ldt,
magmaFloatComplex *VL, magma_int_t ldvl,
magmaFloatComplex *VR, magma_int_t ldvr,
magma_int_t mm, magma_int_t *mout,
magmaFloatComplex *work, magma_int_t lwork,
float *rwork, magma_int_t *info )
{
#define T(i,j) ( T + (i) + (j)*ldt )
#define VL(i,j) (VL + (i) + (j)*ldvl)
#define VR(i,j) (VR + (i) + (j)*ldvr)
#define work(i,j) (work + (i) + (j)*n)
// .. Parameters ..
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magma_int_t nbmin = 16, nbmax = 128;
const magma_int_t ione = 1;
// .. Local Scalars ..
magma_int_t allv, bothv, leftv, over, rightv, somev;
magma_int_t i, ii, is, j, k, ki, iv, n2, nb, nb2, version;
float ovfl, remax, smin, smlnum, ulp, unfl;
// Decode and test the input parameters
bothv = (side == MagmaBothSides);
rightv = (side == MagmaRight) || bothv;
leftv = (side == MagmaLeft ) || bothv;
allv = (howmany == MagmaAllVec);
over = (howmany == MagmaBacktransVec);
somev = (howmany == MagmaSomeVec);
// Set mout to the number of columns required to store the selected
// eigenvectors.
if ( somev ) {
*mout = 0;
for( j=0; j < n; ++j ) {
if ( select[j] ) {
*mout += 1;
}
}
}
else {
*mout = n;
}
*info = 0;
if ( ! rightv && ! leftv )
*info = -1;
else if ( ! allv && ! over && ! somev )
*info = -2;
else if ( n < 0 )
*info = -4;
else if ( ldt < max( 1, n ) )
*info = -6;
else if ( ldvl < 1 || ( leftv && ldvl < n ) )
*info = -8;
else if ( ldvr < 1 || ( rightv && ldvr < n ) )
*info = -10;
else if ( mm < *mout )
*info = -11;
else if ( lwork < max( 1, 2*n ) )
*info = -14;
if ( *info != 0 ) {
magma_xerbla( __func__, -(*info) );
return *info;
}
// Quick return if possible.
if ( n == 0 ) {
return *info;
}
// Use blocked version (2) if sufficient workspace.
// Requires 1 vector to save diagonal elements, and 2*nb vectors for x and Q*x.
// (Compared to dtrevc3, rwork stores 1-norms.)
// Zero-out the workspace to avoid potential NaN propagation.
nb = 2;
if ( lwork >= n + 2*n*nbmin ) {
version = 2;
nb = (lwork - n) / (2*n);
nb = min( nb, nbmax );
nb2 = 1 + 2*nb;
lapackf77_claset( "F", &n, &nb2, &c_zero, &c_zero, work, &n );
}
else {
version = 1;
}
// Set the constants to control overflow.
unfl = lapackf77_slamch( "Safe minimum" );
ovfl = 1. / unfl;
lapackf77_slabad( &unfl, &ovfl );
ulp = lapackf77_slamch( "Precision" );
smlnum = unfl*( n / ulp );
// Store the diagonal elements of T in working array work.
for( i=0; i < n; ++i ) {
*work(i,0) = *T(i,i);
}
// Compute 1-norm of each column of strictly upper triangular
// part of T to control overflow in triangular solver.
rwork[0] = 0.;
for( j=1; j < n; ++j ) {
rwork[j] = cblas_scasum( j, T(0,j), ione );
}
// launch threads -- each single-threaded MKL
magma_int_t nthread = magma_get_parallel_numthreads();
magma_int_t lapack_nthread = magma_get_lapack_numthreads();
magma_set_lapack_numthreads( 1 );
magma_queue queue;
queue.launch( nthread );
//printf( "nthread %d, %d\n", nthread, lapack_nthread );
// NB = N/thread, rounded up to multiple of 16,
// but avoid multiples of page size, e.g., 512*8 bytes = 4096.
magma_int_t NB = magma_int_t( ceil( ceil( ((float)n) / nthread ) / 16. ) * 16. );
if ( NB % 512 == 0 ) {
NB += 32;
}
magma_timer_t time_total=0, time_trsv=0, time_gemm=0, time_gemv=0, time_trsv_sum=0, time_gemm_sum=0, time_gemv_sum=0;
timer_start( time_total );
if ( rightv ) {
// ============================================================
// Compute right eigenvectors.
// iv is index of column in current block.
// Non-blocked version always uses iv=1;
// blocked version starts with iv=nb, goes down to 1.
// (Note the "0-th" column is used to store the original diagonal.)
iv = 1;
if ( version == 2 ) {
iv = nb;
}
timer_start( time_trsv );
is = *mout - 1;
for( ki=n-1; ki >= 0; --ki ) {
if ( somev ) {
if ( ! select[ki] ) {
continue;
}
}
smin = max( ulp*( MAGMA_C_ABS1( *T(ki,ki) ) ), smlnum );
// --------------------------------------------------------
// Complex right eigenvector
*work(ki,iv) = c_one;
// Form right-hand side.
for( k=0; k < ki; ++k ) {
*work(k,iv) = -(*T(k,ki));
}
// Solve upper triangular system:
// [ T(1:ki-1,1:ki-1) - T(ki,ki) ]*X = scale*work.
if ( ki > 0 ) {
queue.push_task( new magma_clatrsd_task(
MagmaUpper, MagmaNoTrans, MagmaNonUnit, MagmaTrue,
ki, T, ldt, *T(ki,ki),
work(0,iv), work(ki,iv), rwork ));
}
// Copy the vector x or Q*x to VR and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VR and normalize
queue.sync();
n2 = ki+1;
blasf77_ccopy( &n2, work(0,iv), &ione, VR(0,is), &ione );
ii = blasf77_icamax( &n2, VR(0,is), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VR(ii,is) );
blasf77_csscal( &n2, &remax, VR(0,is), &ione );
for( k=ki+1; k < n; ++k ) {
*VR(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
queue.sync();
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemv );
if ( ki > 0 ) {
blasf77_cgemv( "n", &n, &ki, &c_one,
VR, &ldvr,
work(0, iv), &ione,
work(ki,iv), VR(0,ki), &ione );
}
time_gemv_sum += timer_stop( time_gemv );
ii = blasf77_icamax( &n, VR(0,ki), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VR(ii,ki) );
blasf77_csscal( &n, &remax, VR(0,ki), &ione );
timer_start( time_trsv );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out below vector
for( k=ki+1; k < n; ++k ) {
*work(k,iv) = c_zero;
}
// Columns iv:nb of work are valid vectors.
// When the number of vectors stored reaches nb,
// or if this was last vector, do the GEMM
if ( (iv == 1) || (ki == 0) ) {
queue.sync();
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemm );
nb2 = nb-iv+1;
n2 = ki+nb-iv+1;
// split gemm into multiple tasks, each doing one block row
for( i=0; i < n; i += NB ) {
magma_int_t ib = min( NB, n-i );
queue.push_task( new cgemm_task(
MagmaNoTrans, MagmaNoTrans, ib, nb2, n2, c_one,
VR(i,0), ldvr,
work(0,iv ), n, c_zero,
work(i,nb+iv), n ));
}
queue.sync();
time_gemm_sum += timer_stop( time_gemm );
// normalize vectors
// TODO if somev, should copy vectors individually to correct location.
for( k = iv; k <= nb; ++k ) {
ii = blasf77_icamax( &n, work(0,nb+k), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *work(ii,nb+k) );
blasf77_csscal( &n, &remax, work(0,nb+k), &ione );
}
lapackf77_clacpy( "F", &n, &nb2, work(0,nb+iv), &n, VR(0,ki), &ldvr );
iv = nb;
timer_start( time_trsv );
}
else {
iv -= 1;
}
} // blocked back-transform
is -= 1;
}
}
timer_stop( time_trsv );
timer_stop( time_total );
timer_printf( "trevc trsv %.4f, gemm %.4f, gemv %.4f, total %.4f\n",
time_trsv_sum, time_gemm_sum, time_gemv_sum, time_total );
if ( leftv ) {
// ============================================================
// Compute left eigenvectors.
// iv is index of column in current block.
// Non-blocked version always uses iv=1;
// blocked version starts with iv=1, goes up to nb.
// (Note the "0-th" column is used to store the original diagonal.)
iv = 1;
is = 0;
for( ki=0; ki < n; ++ki ) {
if ( somev ) {
if ( ! select[ki] ) {
continue;
}
}
smin = max( ulp*MAGMA_C_ABS1( *T(ki,ki) ), smlnum );
// --------------------------------------------------------
// Complex left eigenvector
*work(ki,iv) = c_one;
// Form right-hand side.
for( k = ki + 1; k < n; ++k ) {
*work(k,iv) = -MAGMA_C_CNJG( *T(ki,k) );
}
// Solve conjugate-transposed triangular system:
// [ T(ki+1:n,ki+1:n) - T(ki,ki) ]**H * X = scale*work.
// TODO what happens with T(k,k) - lambda is small? Used to have < smin test.
if ( ki < n-1 ) {
n2 = n-ki-1;
queue.push_task( new magma_clatrsd_task(
MagmaUpper, MagmaConjTrans, MagmaNonUnit, MagmaTrue,
n2, T(ki+1,ki+1), ldt, *T(ki,ki),
work(ki+1,iv), work(ki,iv), rwork ));
}
// Copy the vector x or Q*x to VL and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VL and normalize
queue.sync();
n2 = n-ki;
blasf77_ccopy( &n2, work(ki,iv), &ione, VL(ki,is), &ione );
ii = blasf77_icamax( &n2, VL(ki,is), &ione ) + ki - 1;
remax = 1. / MAGMA_C_ABS1( *VL(ii,is) );
blasf77_csscal( &n2, &remax, VL(ki,is), &ione );
for( k=0; k < ki; ++k ) {
*VL(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
queue.sync();
if ( ki < n-1 ) {
n2 = n-ki-1;
blasf77_cgemv( "n", &n, &n2, &c_one,
VL(0,ki+1), &ldvl,
work(ki+1,iv), &ione,
work(ki, iv), VL(0,ki), &ione );
}
ii = blasf77_icamax( &n, VL(0,ki), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VL(ii,ki) );
blasf77_csscal( &n, &remax, VL(0,ki), &ione );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out above vector
// could go from (ki+1)-NV+1 to ki
for( k=0; k < ki; ++k ) {
*work(k,iv) = c_zero;
}
// Columns 1:iv of work are valid vectors.
// When the number of vectors stored reaches nb,
// or if this was last vector, do the GEMM
if ( (iv == nb) || (ki == n-1) ) {
queue.sync();
n2 = n-(ki+1)+iv;
// split gemm into multiple tasks, each doing one block row
for( i=0; i < n; i += NB ) {
magma_int_t ib = min( NB, n-i );
queue.push_task( new cgemm_task(
MagmaNoTrans, MagmaNoTrans, ib, iv, n2, c_one,
VL(i,ki-iv+1), ldvl,
work(ki-iv+1,1), n, c_zero,
work(i,nb+1), n ));
}
queue.sync();
// normalize vectors
for( k=1; k <= iv; ++k ) {
ii = blasf77_icamax( &n, work(0,nb+k), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *work(ii,nb+k) );
blasf77_csscal( &n, &remax, work(0,nb+k), &ione );
}
lapackf77_clacpy( "F", &n, &iv, work(0,nb+1), &n, VL(0,ki-iv+1), &ldvl );
iv = 1;
}
else {
iv += 1;
}
} // blocked back-transform
is += 1;
}
}
// close down threads
queue.quit();
magma_set_lapack_numthreads( lapack_nthread );
return *info;
} // End of CTREVC
/***************************************************************************//**
Purpose
-------
CUNGQR generates an M-by-N COMPLEX 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 CGEQRF.
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 COMPLEX 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 CGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by CGEQRF_GPU.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument has an illegal value
@ingroup magma_ungqr
*******************************************************************************/
extern "C" magma_int_t
magma_cungqr2(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magma_int_t nb = magma_get_cgeqrf_nb( 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;
magmaFloatComplex *dA, *dV, *dW, *dT, *T;
magmaFloatComplex *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 = magma_roundup( m, 32 );
lddwork = magma_roundup( n, 32 );
if (MAGMA_SUCCESS != magma_cmalloc( &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_cmalloc_cpu( &work, lwork );
T = work;
if (work == NULL) {
magma_free( dA );
magma_free_cpu( work );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magmaFloatComplex *V = work + (n+nb)*nb;
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
/*
lapackf77_cungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_clacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_claset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_clarfb( 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_csetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda, queue );
// Set A(1:kk,kk+1:n) to zero.
magmablas_claset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda, queue );
}
}
if (kk > 0) {
// Use blocked code
// queue: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
for (i = ki; i >= 0; i -= nb) {
ib = min(nb, k - i);
// Send current panel to the GPU
mi = m - i;
lapackf77_claset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_csetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, queue );
lapackf77_clarft( MagmaForwardStr, MagmaColumnwiseStr,
&mi, &ib,
A(i,i), &lda, &tau[i], T, &nb);
magma_csetmatrix_async( ib, ib,
T, nb,
dT, nb, queue );
// set panel to identity
magmablas_claset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda, queue );
magmablas_claset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda, queue );
magma_queue_sync( queue );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_clarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT, nb,
dA(i, i), ldda, dW, lddwork, queue );
}
}
// copy result back to CPU
magma_cgetmatrix( m, n,
dA(0, 0), ldda, A(0, 0), lda, queue );
}
magma_queue_destroy( queue );
magma_free( dA );
magma_free_cpu( work );
return *info;
} /* magma_cungqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing claset
Code is very similar to testing_clacpy.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex_ptr d_A;
magmaFloatComplex offdiag, diag;
magma_int_t M, N, size, lda, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper, MagmaFull };
printf("%% uplo M N offdiag diag CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("%%===================================================================================\n");
for( int iuplo = 0; iuplo < 3; ++iuplo ) {
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
for( int ival = 0; ival < 4; ++ival ) {
// test combinations of zero & non-zero:
// ival offdiag diag
// 0 0 0
// 1 0 3.14
// 2 1.23 0
// 3 1.23 3.14
offdiag = MAGMA_C_MAKE( 1.2345, 6.7890 ) * (ival / 2);
diag = MAGMA_C_MAKE( 3.1415, 2.7183 ) * (ival % 2);
M = opts.msize[itest];
N = opts.nsize[itest];
//M += 2; // space for insets
//N += 2;
lda = M;
ldda = magma_roundup( M, opts.align );
size = lda*N;
if ( uplo[iuplo] == MagmaLower ) {
// save lower trapezoid (with diagonal)
if ( M > N ) {
gbytes = sizeof(magmaFloatComplex) * (1.*M*N - 0.5*N*(N-1)) / 1e9;
} else {
gbytes = sizeof(magmaFloatComplex) * 0.5*M*(M+1) / 1e9;
}
}
else if ( uplo[iuplo] == MagmaUpper ) {
// save upper trapezoid (with diagonal)
if ( N > M ) {
gbytes = sizeof(magmaFloatComplex) * (1.*M*N - 0.5*M*(M-1)) / 1e9;
} else {
gbytes = sizeof(magmaFloatComplex) * 0.5*N*(N+1) / 1e9;
}
}
else {
// save entire matrix
gbytes = sizeof(magmaFloatComplex) * 1.*M*N / 1e9;
}
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, size );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, size );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( M, N, h_A, lda, d_A, ldda );
magmablasSetKernelStream( opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
//magmablas_claset( uplo[iuplo], M-2, N-2, offdiag, diag, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_claset( uplo[iuplo], M, N, offdiag, diag, d_A, ldda );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
//magma_int_t M2 = M-2; // inset by 1 row & col
//magma_int_t N2 = N-2;
//lapackf77_claset( lapack_uplo_const( uplo[iuplo] ), &M2, &N2, &offdiag, &diag, h_A+1+lda, &lda );
lapackf77_claset( lapack_uplo_const( uplo[iuplo] ), &M, &N, &offdiag, &diag, h_A, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if ( opts.verbose ) {
printf( "A= " ); magma_cprint( M, N, h_A, lda );
printf( "dA=" ); magma_cprint_gpu( M, N, d_A, ldda );
}
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( M, N, d_A, ldda, h_R, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &M, &N, h_R, &lda, work);
bool okay = (error == 0);
status += ! okay;
printf("%5s %5d %5d %9.4f %6.4f %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
lapack_uplo_const( uplo[iuplo] ), (int) M, (int) N,
real(offdiag), real(diag),
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(okay ? "ok" : "failed") );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing clarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_CUDA_INIT();
cuFloatComplex c_zero = MAGMA_C_ZERO;
cuFloatComplex c_one = MAGMA_C_ONE;
cuFloatComplex c_neg_one = MAGMA_C_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
cuFloatComplex *C, *R, *V, *T, *W;
TESTING_MALLOC( C, cuFloatComplex, ldc*n );
TESTING_MALLOC( R, cuFloatComplex, ldc*n );
TESTING_MALLOC( V, cuFloatComplex, ldv*k );
TESTING_MALLOC( T, cuFloatComplex, ldt*k );
TESTING_MALLOC( W, cuFloatComplex, ldw*k );
cuFloatComplex *dC, *dV, *dT, *dW;
TESTING_DEVALLOC( dC, cuFloatComplex, ldc*n );
TESTING_DEVALLOC( dV, cuFloatComplex, ldv*k );
TESTING_DEVALLOC( dT, cuFloatComplex, ldt*k );
TESTING_DEVALLOC( dW, cuFloatComplex, ldw*k );
magma_int_t size;
magma_int_t iseed[4] = { 1, 2, 3, 4 };
float error, work[1];
// test all combinations of input parameters
const char* side[] = { MagmaLeftStr, MagmaRightStr };
const char* trans[] = { MagmaConjTransStr, 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_clarnv( &ione, iseed, &size, C );
//printf( "C=" ); magma_cprint( 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_clarnv( &ione, iseed, &size, V );
if ( *storev[istor] == MagmaColumnwise ) {
if ( *direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaUpperStr, &k, &k, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( 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_claset( MagmaLowerStr, &k, &k, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( MagmaUpperStr, &k, &k, &c_zero, &c_one, &V[(nv-k)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_cprint( 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_clarnv( &ione, iseed, &size, T );
if ( *direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_claset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_cprint( k, k, T, ldt );
magma_csetmatrix( m, n, C, ldc, dC, ldc );
magma_csetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_csetmatrix( k, k, T, ldt, dT, ldt );
lapackf77_clarfb( side[iside], trans[itran], direct[idir], storev[istor],
&m, &n, &k,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_cprint( m, n, C, ldc );
magma_clarfb_gpu( *side[iside], *trans[itran], *direct[idir], *storev[istor],
m, n, k,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_cgetmatrix( m, n, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_cprint( m, n, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_clange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "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
-------
CLAHR2 reduces the first NB columns of a complex 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 CGEHRD.
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 COMPLEX 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 COMPLEX array, dimension (NB)
The scalar factors of the elementary reflectors. See Further
Details.
@param[out]
T COMPLEX 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 COMPLEX 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 complex scalar, and v is a complex 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_cgeev_aux
********************************************************************/
extern "C" magma_int_t
magma_clahr2_m(
magma_int_t n, magma_int_t k, magma_int_t nb,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *T, magma_int_t ldt,
magmaFloatComplex *Y, magma_int_t ldy,
struct cgehrd_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)
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex 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;
magmaFloatComplex scale;
magma_int_t i;
magmaFloatComplex ei = MAGMA_C_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 *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
// zero out current top block of V on all GPUs
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( data->streams[d] );
magmablas_claset( MagmaFull, nb, nb, c_zero, c_zero, dV(d,k,0), ldv );
}
// set all Y=0
lapackf77_claset( "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_C_NEGATE( tau[i-1] );
blasf77_caxpy( &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_ccopy( &i,
A(k+1,i), &ione,
T(0,nb-1), &ione );
// w := V1' * b1 = VA(k+1:k+i, 0:i-1)' * w
blasf77_ctrmv( "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_cgemv( "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_ctrmv( "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_cgemv( "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_ctrmv( "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_caxpy( &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_clarfg( &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_csetvector_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_clacpy( 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_cgemv( 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_cgetvector_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_C_NEGATE( tau[i] );
blasf77_cgemv( "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_ctrmv( "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 complex, conjugate row of V, and undo afterwards
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_clacgv( &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_cgemv( "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_clacgv( &i1, A(k+i1,0), &lda );
#endif
// A(k:n, i+1) -= Y(k:n, 0:i) * w
blasf77_cgemv( "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_caxpy( &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_cgemm( 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_cgetmatrix_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_caxpy( &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_csetmatrix_async( n, nb, Y, ldy, dY(d, 0, 0), ldda, data->streams[d] );
magma_csetmatrix_async( nb, nb, T, nb, dTi(d), nb, data->streams[d] );
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_clahr2 */
/**
Purpose:
---------
CUNGLQ generates an M-by-N complex matrix Q with orthonormal rows,
which is defined as the first M rows of a product of K elementary
reflectors of order N
Q = H(k)**H . . . H(2)**H H(1)**H
as returned by CGELQF.
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. N >= M.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. M >= K >= 0.
@param[in,out]
A COMPLEX array, dimension (LDA,N)
On entry, the i-th row must contain the vector which defines
the elementary reflector H(i), for i = 1,2,...,k, as returned
by CGELQF in the first k rows of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by CGELQF.
@param[out]
work COMPLEX array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= NB*NB, where NB is
the optimal blocksize.
If LWORK = -1, 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]
info INTEGER
- = 0: successful exit;
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_cgelqf_comp
********************************************************************/
extern "C" magma_int_t
magma_cunglq(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *work, magma_int_t lwork,
magma_int_t *info)
{
#define A(i_,j_) ( A + (i_) + (j_)*lda)
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
#define tau(i_) (tau + (i_))
// Constants
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
// Local variables
bool lquery;
magma_int_t i, ib, ki, ldda, lddwork, lwkopt, mib, nb, n_i;
magma_queue_t queue = NULL;
magmaFloatComplex_ptr dA = NULL;
magmaFloatComplex* work2 = NULL;
// Test the input arguments
*info = 0;
nb = magma_get_cgelqf_nb( m, n );
lwkopt = nb*nb;
work[0] = magma_cmake_lwork( lwkopt );
lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0 || n < m) {
*info = -2;
} else if (k < 0 || k > m) {
*info = -3;
} else if (lda < max( 1, m )) {
*info = -5;
} else if (lwork < max( 1, lwkopt ) && ! lquery) {
*info = -8;
//printf( "m %d, n %d, nb %d: lwork %d, required %d\n", m, n, nb, lwork, lwkopt );
//*info = 0;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
// Quick return if possible
if (m <= 0) {
work[0] = c_one;
return *info;
}
//if (lwork < lwkopt) {
// magma_cmalloc_cpu( &work2, lwkopt );
//}
//else {
// work2 = work;
//}
work2 = work;
// Allocate GPU work space
// ldda*n for matrix dA
// nb*n for dV
// lddwork*nb for dW larfb workspace
ldda = magma_roundup( m, 32 );
lddwork = magma_roundup( m, 32 );
if (MAGMA_SUCCESS != magma_cmalloc( &dA, ldda*n + n*nb + lddwork*nb + nb*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
magmaFloatComplex_ptr dV; dV = dA + ldda*n;
magmaFloatComplex_ptr dW; dW = dA + ldda*n + n*nb;
magmaFloatComplex_ptr dT; dT = dA + ldda*n + n*nb + lddwork*nb;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
magmablas_claset( MagmaFull, m, n, MAGMA_C_NAN, MAGMA_C_NAN, dA, ldda, queue );
// all columns are handled by blocked method.
// ki is start of last (partial) block
ki = ((k - 1) / nb) * nb;
// Use blocked code
for( i=ki; i >= 0; i -= nb ) {
ib = min( nb, k-i );
// first block has extra rows to update
mib = ib;
if ( i == ki ) {
mib = m - i;
}
// Send current panel of V (block row) to the GPU
lapackf77_claset( "Lower", &ib, &ib, &c_zero, &c_one, A(i,i), &lda );
// TODO: having this _async was causing numerical errors. Why?
magma_csetmatrix( ib, n-i,
A(i,i), lda,
dV, nb, queue );
// Form the triangular factor of the block reflector
// H = H(i) H(i+1) . . . H(i+ib-1)
n_i = n - i;
lapackf77_clarft( MagmaForwardStr, MagmaRowwiseStr, &n_i, &ib,
A(i,i), &lda, &tau[i], work2, &nb );
magma_csetmatrix_async( ib, ib,
work2, nb,
dT, nb, queue );
// set panel of A (block row) to identity
magmablas_claset( MagmaFull, mib, i, c_zero, c_zero, dA(i,0), ldda, queue );
magmablas_claset( MagmaFull, mib, n-i, c_zero, c_one, dA(i,i), ldda, queue );
if (i < m) {
// Apply H**H to A(i:m,i:n) from the right
magma_clarfb_gpu( MagmaRight, MagmaConjTrans, MagmaForward, MagmaRowwise,
m-i, n-i, ib,
dV, nb, dT, nb,
dA(i,i), ldda, dW, lddwork, queue );
}
}
// copy result back to CPU
magma_cgetmatrix( m, n,
dA(0,0), ldda, A(0,0), lda, queue );
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
//if (work2 != work) {
// magma_free_cpu( work2 );
//}
work[0] = magma_cmake_lwork( lwkopt );
return *info;
}
/**
Purpose
-------
CGEHRD reduces a COMPLEX 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 CGEBAL; 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 COMPLEX 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 COMPLEX 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) COMPLEX array, dimension (LWORK)
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The length of the array WORK. LWORK >= N*NB.
where NB is the optimal blocksize.
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued by XERBLA.
@param[out]
T COMPLEX 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 complex scalar, and v is a complex 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_cunghr.
@ingroup magma_cgeev_comp
********************************************************************/
extern "C" magma_int_t
magma_cgehrd_m(
magma_int_t n, magma_int_t ilo, magma_int_t ihi,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *tau,
magmaFloatComplex *work, magma_int_t lwork,
magmaFloatComplex *T,
magma_int_t *info)
{
#define A( i, j ) (A + (i) + (j)*lda)
#define dA( d, i, j ) (data.A[d] + (i) + (j)*ldda)
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magma_int_t nb = magma_get_cgehrd_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 cgehrd_data data;
magma_int_t ngpu = magma_num_gpus();
*info = 0;
iws = n*(nb + nb*ngpu);
work[0] = magma_cmake_lwork( iws );
lquery = (lwork == -1);
if (n < 0) {
*info = -1;
} else if (ilo < 1 || ilo > max(1,n)) {
*info = -2;
} else if (ihi < min(ilo,n) || ihi > n) {
*info = -3;
} else if (lda < max(1,n)) {
*info = -5;
} else if (lwork < iws && ! lquery) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
// Adjust from 1-based indexing
ilo -= 1;
// Quick return if possible
nh = ihi - ilo;
if (nh <= 1) {
work[0] = c_one;
return *info;
}
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_claset( "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.queues[d] = NULL;
}
// Now requires lwork >= iws; else dT won't be computed in unblocked code.
// If not enough workspace, use unblocked code
//if ( lwork < iws ) {
// nb = 1;
//}
if (nb == 1 || nb >= nh) {
// Use unblocked code below
i = ilo;
}
else {
// Use blocked code
// allocate memory on GPUs for A and workspaces
ldda = magma_roundup( n, 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 mclahr2
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_cmalloc( &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( d, &data.queues[d] );
}
// Copy the matrix to GPUs
magma_csetmatrix_1D_col_bcyclic( n, n, A, lda, data.A, ldda, ngpu, nb, data.queues );
// 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_cgetmatrix( ihi-i, nb,
dA(dpanel, i, di), ldda,
A(i,i), lda, data.queues[dpanel] );
}
// add 1 to i for 1-based index
magma_clahr2_m( ihi, i+1, nb, A(0,i), lda,
&tau[i], &T[i*nb], nb, work, n, &data );
magma_clahru_m( n, ihi, i, nb, A, lda, &data );
// copy first i rows above panel to host
magma_setdevice( dpanel );
magma_cgetmatrix_async( i, nb,
dA(dpanel, 0, di), ldda,
A(0,i), lda, data.queues[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_cgetmatrix( n, ib,
dA(d, 0, di), ldda,
A(0,i2), lda, data.queues[d] );
}
}
// Use unblocked code to reduce the rest of the matrix
// add 1 to i for 1-based index
i += 1;
lapackf77_cgehd2(&n, &i, &ihi, A, &lda, tau, work, &iinfo);
work[0] = magma_cmake_lwork( iws );
CLEANUP:
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( data.A[d] );
magma_queue_destroy( data.queues[d] );
}
magma_setdevice( orig_dev );
return *info;
} /* magma_cgehrd */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvd
*/
int main( int argc, char** argv)
{
TESTING_CUDA_INIT();
cuFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;
float *rwork, *w1, *w2;
magma_int_t *iwork;
float gpu_time, cpu_time;
magma_timestr_t start, end;
/* Matrix size */
magma_int_t N=0, n2;
magma_int_t size[4] = {1024,2048,4100,6001};
magma_int_t i, itype, info;
magma_int_t ione = 1, izero = 0;
magma_int_t five = 5;
cuFloatComplex c_zero = MAGMA_C_ZERO;
cuFloatComplex c_one = MAGMA_C_ONE;
cuFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float d_one = 1.;
float d_neg_one = -1.;
float d_ten = 10.;
magma_int_t ISEED[4] = {0,0,0,1};
const char *uplo = MagmaLowerStr;
const char *jobz = MagmaVectorsStr;
itype = 1;
magma_int_t checkres;
float result[4];
int flagN = 0;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0){
N = atoi(argv[++i]);
if (N>0){
printf(" testing_chegvd -N %d\n\n", (int) N);
flagN=1;
}
else {
printf("\nUsage: \n");
printf(" testing_chegvd -N %d\n\n", (int) N);
exit(1);
}
}
if (strcmp("-itype", argv[i])==0){
itype = atoi(argv[++i]);
if (itype>0 && itype <= 3){
printf(" testing_chegvd -itype %d\n\n", (int) itype);
}
else {
printf("\nUsage: \n");
printf(" testing_chegvd -itype %d\n\n", (int) itype);
exit(1);
}
}
if (strcmp("-L", argv[i])==0){
uplo = MagmaLowerStr;
printf(" testing_chegvd -L");
}
if (strcmp("-U", argv[i])==0){
uplo = MagmaUpperStr;
printf(" testing_chegvd -U");
}
}
} else {
printf("\nUsage: \n");
printf(" testing_chegvd -L/U -N %d -itype %d\n\n", 1024, 1);
}
if(!flagN)
N = size[3];
checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL;
n2 = N * N;
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, cuFloatComplex, n2);
TESTING_MALLOC( h_B, cuFloatComplex, n2);
TESTING_MALLOC( w1, float , N);
TESTING_MALLOC( w2, float , N);
TESTING_HOSTALLOC(h_R, cuFloatComplex, n2);
TESTING_HOSTALLOC(h_S, cuFloatComplex, n2);
magma_int_t nb = magma_get_chetrd_nb(N);
magma_int_t lwork = 2*N*nb + N*N;
magma_int_t lrwork = 1 + 5*N +2*N*N;
magma_int_t liwork = 3 + 5*N;
TESTING_HOSTALLOC(h_work, cuFloatComplex, lwork);
TESTING_MALLOC( rwork, float, lrwork);
TESTING_MALLOC( iwork, magma_int_t, liwork);
printf(" N CPU Time(s) GPU Time(s) \n");
printf("===================================\n");
for(i=0; i<4; i++){
if (!flagN){
N = size[i];
n2 = N*N;
}
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
//lapackf77_clatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten,
// &d_one, &N, &N, uplo, h_B, &N, h_work, &info);
//lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_B, &N);
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
/* increase the diagonal */
{
magma_int_t i, j;
for(i=0; i<N; i++) {
MAGMA_C_SET2REAL( h_B[i*N+i], MAGMA_C_REAL(h_B[i*N+i]) + 1.*N );
MAGMA_C_SET2REAL( h_A[i*N+i], MAGMA_C_REAL(h_A[i*N+i]) );
}
}
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_chegvd(itype, jobz[0], uplo[0],
N, h_R, N, h_S, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
start = get_current_time();
magma_chegvd(itype, jobz[0], uplo[0],
N, h_R, N, h_S, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
/* =====================================================================
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 |
=================================================================== */
float temp1, temp2;
cuFloatComplex *tau;
if (itype == 1 || itype == 2){
lapackf77_claset( "A", &N, &N, &c_zero, &c_one, h_S, &N);
blasf77_cgemm("N", "C", &N, &N, &N, &c_one, h_R, &N, h_R, &N, &c_zero, h_work, &N);
blasf77_chemm("R", uplo, &N, &N, &c_neg_one, h_B, &N, h_work, &N, &c_one, h_S, &N);
result[1]= lapackf77_clange("1", &N, &N, h_S, &N, rwork) / N;
}
else if (itype == 3){
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N);
blasf77_cherk(uplo, "N", &N, &N, &d_neg_one, h_R, &N, &d_one, h_S, &N);
result[1]= lapackf77_clanhe("1",uplo, &N, h_S, &N, rwork) / N / lapackf77_clanhe("1",uplo, &N, h_B, &N, rwork);
}
result[0] = 1.;
result[0] /= lapackf77_clanhe("1",uplo, &N, h_A, &N, rwork);
result[0] /= lapackf77_clange("1",&N , &N, h_R, &N, rwork);
if (itype == 1){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_work, &N, rwork)/N;
}
else if (itype == 2){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N;
}
else if (itype == 3){
blasf77_chemm("L", uplo, &N, &N, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", uplo, &N, &N, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_clange("1", &N, &N, h_R, &N, rwork)/N;
}
/* lapackf77_chet21(&ione, uplo, &N, &izero,
h_A, &N,
w1, w1,
h_R, &N,
h_R, &N,
tau, h_work, rwork, &result[0]);
*/
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_chegvd(itype, 'N', uplo[0],
N, h_R, N, h_S, N, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&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
=================================================================== */
start = get_current_time();
lapackf77_chegvd(&itype, jobz, uplo,
&N, h_A, &N, h_B, &N, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info);
end = get_current_time();
if (info < 0)
printf("Argument %d of chegvd had an illegal value.\n", (int) -info);
cpu_time = GetTimerValue(start,end)/1000.;
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %6.2f %6.2f\n",
(int) N, cpu_time, gpu_time);
if ( checkres ){
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if(itype==1)
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==2)
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==3)
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
if(itype==1 || itype ==2)
printf("(2) | I - Z Z' B | / N = %e\n", result[1]);
else
printf("(2) | B - Z Z' | / (|B| N) = %e\n", result[1]);
printf("(3) | D(w/ Z)-D(w/o Z)|/ |D| = %e\n\n", result[2]);
}
if (flagN)
break;
}
/* Memory clean up */
TESTING_FREE( h_A);
TESTING_FREE( h_B);
TESTING_FREE( w1);
TESTING_FREE( w2);
TESTING_FREE( rwork);
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
TESTING_HOSTFREE( h_S);
/* Shutdown */
TESTING_CUDA_FINALIZE();
}
