/* Update thread */
extern "C" void *cpu_thread(void *a)
{
magma_int_t i;
t_params *tp = (t_params*)a;
magma_qr_params *mp = (magma_qr_params*)tp->params;
//long magma_int_t t = (long int) a;
long int t = (long int) tp->tid;
magma_int_t M;
magma_int_t N;
magma_int_t K;
cuDoubleComplex *WORK;
loop:
while (mp->sync0 == 0) {
sched_yield();
}
for (i = 0; i < mp->np_gpu; i++)
{
while (mp->p[i] == NULL) {
sched_yield();
}
M=mp->m-i*mp->nb;
N=mp->ob;
K=mp->nb;
if (mp->m >= (mp->n-(mp->nthreads*mp->ob))) {
if (i == (mp->np_gpu - 1)) {
K = mp->n-mp->nthreads*mp->ob-(mp->np_gpu-1)*mp->nb;
}
}
WORK = (cuDoubleComplex*)malloc(sizeof(cuDoubleComplex)*M*N);
lapackf77_zlarfb(MagmaLeftStr, MagmaTransStr, MagmaForwardStr, MagmaColumnwiseStr,
&M,&N,&K,mp->a+i*mp->nb*mp->lda+i*mp->nb,&(mp->lda),mp->t+i*mp->nb*mp->nb,&K,
mp->a+mp->m*mp->n-(mp->nthreads-t)*mp->ob*mp->lda+i*mp->nb,&(mp->lda),WORK,&N);
free(WORK);
}
mp->sync1[t] = 1;
while (mp->sync2 == 0) {
sched_yield();
}
goto loop;
return (void*)NULL;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_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};
double error, work[1];
// test all combinations of input parameters
const char side[] = { MagmaLeft, MagmaRight };
const char trans[] = { MagmaConjTrans, MagmaNoTrans };
const char direct[] = { MagmaForward, MagmaBackward };
const char storev[] = { MagmaColumnwise, MagmaRowwise };
magma_opts opts;
parse_opts( argc, argv, &opts );
printf(" M N K storev side direct trans ||R||_F / ||HC||_F\n");
printf("========================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
M = opts.msize[i];
N = opts.nsize[i];
K = opts.ksize[i];
if ( M < K || N < K || K <= 0 ) {
printf( "skipping M %d, N %d, K %d; 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 ) {
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
magmaDoubleComplex *C, *R, *V, *T, *W;
TESTING_MALLOC( C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC( R, magmaDoubleComplex, ldc*N );
TESTING_MALLOC( V, magmaDoubleComplex, ldv*K );
TESTING_MALLOC( T, magmaDoubleComplex, ldt*K );
TESTING_MALLOC( W, magmaDoubleComplex, ldw*K );
magmaDoubleComplex *dC, *dV, *dT, *dW;
TESTING_DEVALLOC( dC, magmaDoubleComplex, ldc*N );
TESTING_DEVALLOC( dV, magmaDoubleComplex, ldv*K );
TESTING_DEVALLOC( dT, magmaDoubleComplex, ldt*K );
TESTING_DEVALLOC( dW, magmaDoubleComplex, ldw*K );
// C is M x N.
size = ldc*N;
lapackf77_zlarnv( &ione, ISEED, &size, C );
//printf( "C=" ); magma_zprint( 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_zlarnv( &ione, ISEED, &size, V );
if ( storev[istor] == MagmaColumnwise ) {
if ( direct[idir] == MagmaForward ) {
lapackf77_zlaset( MagmaUpperStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_zlaset( 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_zlaset( MagmaLowerStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_zlaset( MagmaUpperStr, &K, &K, &c_zero, &c_one, &V[(nv-K)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_zprint( 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_zlarnv( &ione, ISEED, &size, T );
if ( direct[idir] == MagmaForward ) {
lapackf77_zlaset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_zlaset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_zprint( K, K, T, ldt );
magma_zsetmatrix( M, N, C, ldc, dC, ldc );
magma_zsetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_zsetmatrix( K, K, T, ldt, dT, ldt );
lapackf77_zlarfb( &side[iside], &trans[itran], &direct[idir], &storev[istor],
&M, &N, &K,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_zprint( M, N, C, ldc );
magma_zlarfb_gpu( side[iside], trans[itran], direct[idir], storev[istor],
M, N, K,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_zgetmatrix( M, N, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_zprint( M, N, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_zlange( "Fro", &M, &N, C, &ldc, work );
size = ldc*N;
blasf77_zaxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_zlange( "Fro", &M, &N, R, &ldc, work ) / error;
printf( "%5d %5d %5d %c %c %c %c %8.2e\n",
(int) M, (int) N, (int) K,
storev[istor], side[iside], direct[idir], trans[itran], error );
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 );
}}}}
printf( "\n" );
}
TESTING_FINALIZE();
return 0;
}
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by ZGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A COMPLEX_16 array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by ZGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
dT COMPLEX_16 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_zgeqrf_gpu.
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in DT.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex_ptr dT, magma_int_t nb,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
#define dT(j) (dT + (j)*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk;
magma_int_t lddwork;
magmaDoubleComplex *dA=NULL, *dV=NULL, *dW=NULL;
magmaDoubleComplex *work=NULL;
magma_queue_t queue=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;
}
// 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_zmalloc( &dA, ldda*n + ldda*nb + lddwork*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
dV = dA + ldda*n;
dW = dA + ldda*n + ldda*nb;
// Allocate CPU work space
// n*nb for larfb work
// m*nb for V
// nb*nb for T
lwork = (n + m + nb) * nb;
magma_zmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
magmaDoubleComplex *work_T, *work_V;
work_T = work + n*nb;
work_V = work + n*nb + nb*nb;
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;
// zungqr requires less workspace (n*nb), but is slow if k < zungqr's block size.
// replacing it with the 4 routines below is much faster (e.g., 60x).
//magma_int_t iinfo;
//lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
// A(kk, kk), &lda,
// &tau[kk], work, &lwork, &iinfo );
lapackf77_zlacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, work_V, &m_kk);
lapackf77_zlaset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_zlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
work_V, &m_kk, &tau[kk], work_T, &k_kk);
lapackf77_zlarfb( 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) {
magma_zsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda, queue );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda, queue );
}
}
if (kk > 0) {
// Use blocked code
// queue: 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 dV on the GPU
mi = m - i;
lapackf77_zlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_zsetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, queue );
// set panel to identity
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda, queue );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda, queue );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT(i), nb,
dA(i, i), ldda, dW, lddwork, queue );
}
}
// copy result back to CPU
magma_zgetmatrix( m, n,
dA(0, 0), ldda, A(0, 0), lda, queue );
}
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
magma_free_cpu( work );
return *info;
} /* magma_zungqr */
static void magma_ztile_bulge_applyQ(
magma_int_t core_id, magma_side_t side, magma_int_t n_loc, magma_int_t n, magma_int_t nb, magma_int_t Vblksiz,
magmaDoubleComplex *E, magma_int_t lde,
magmaDoubleComplex *V, magma_int_t ldv,
magmaDoubleComplex *TAU,
magmaDoubleComplex *T, magma_int_t ldt)
//, magma_int_t* info)
{
//%===========================
//% local variables
//%===========================
magma_int_t firstcolj;
magma_int_t bg, rownbm;
magma_int_t st,ed,fst,vlen,vnb,colj;
magma_int_t vpos,tpos;
magma_int_t cur_blksiz,avai_blksiz, ncolinvolvd;
magma_int_t nbgr, colst, coled;
if (n <= 0)
return;
if (n_loc <= 0)
return;
//info = 0;
magma_int_t INFO=0;
magma_int_t nbGblk = magma_ceildiv(n-1, Vblksiz);
/*
* version v1: for each chunck it apply all the V's then move to
* the other chunck. the locality here inside each
* chunck meaning that thread t apply V_k then move
* to V_k+1 which overlap with V_k meaning that the
* E_k+1 overlap with E_k. so here is the
* locality however thread t had to read V_k+1 and
* T_k+1 at each apply. note that all thread if they
* run at same speed they might reading the same V_k
* and T_k at the same time.
* */
magma_int_t nb_loc = 128; //$$$$$$$$
magma_int_t lwork = 2*nb_loc*max(Vblksiz,64);
magmaDoubleComplex *work, *work2;
magma_zmalloc_cpu(&work, lwork);
magma_zmalloc_cpu(&work2, lwork);
magma_int_t nbchunk = magma_ceildiv(n_loc, nb_loc);
/* SIDE LEFT meaning apply E = Q*E = (q_1*q_2*.....*q_n) * E ==> so traverse Vs in reverse order (forward) from q_n to q_1
* each q_i consist of applying V to a block of row E(row_i,:) and applies are overlapped meaning
* that q_i+1 overlap a portion of the E(row_i, :).
* IN parallel E is splitten in vertical block over the threads */
/* SIDE RIGHT meaning apply E = E*Q = E * (q_1*q_2*.....*q_n) ==> so tarverse Vs in normal order (forward) from q_1 to q_n
* each q_i consist of applying V to a block of col E(:, col_i,:) and the applies are overlapped meaning
* that q_i+1 overlap a portion of the E(:, col_i).
* IN parallel E is splitten in horizontal block over the threads */
#ifdef ENABLE_DEBUG
if ((core_id == 0) || (core_id == 1))
printf(" APPLY Q2_cpu zbulge_back_m N %d N_loc %d nbchunk %d NB %d Vblksiz %d SIDE %c \n", n, n_loc, nbchunk, nb, Vblksiz, side);
#endif
for (magma_int_t i = 0; i < nbchunk; i++) {
magma_int_t ib_loc = min(nb_loc, (n_loc - i*nb_loc));
if (side == MagmaLeft) {
for (bg = nbGblk; bg > 0; bg--) {
firstcolj = (bg-1)*Vblksiz + 1;
rownbm = magma_ceildiv((n-(firstcolj+1)),nb);
if (bg == nbGblk) rownbm = magma_ceildiv((n-(firstcolj)),nb); // last blk has size=1 used for complex to handle A(N,N-1)
for (magma_int_t j = rownbm; j > 0; j--) {
vlen = 0;
vnb = 0;
colj = (bg-1)*Vblksiz; // for k=0; I compute the fst and then can remove it from the loop
fst = (rownbm -j)*nb+colj +1;
for (magma_int_t k=0; k < Vblksiz; k++) {
colj = (bg-1)*Vblksiz + k;
st = (rownbm -j)*nb+colj +1;
ed = min(st+nb-1,n-1);
if (st > ed)
break;
if ((st == ed) && (colj != n-2))
break;
vlen=ed-fst+1;
vnb=k+1;
}
colst = (bg-1)*Vblksiz;
magma_bulge_findVTpos(n, nb, Vblksiz, colst, fst, ldv, ldt, &vpos, &tpos);
if ((vlen > 0) && (vnb > 0)) {
lapackf77_zlarfb( "L", "N", "F", "C", &vlen, &ib_loc, &vnb, V(vpos), &ldv, T(tpos), &ldt, E(fst,i*nb_loc), &lde, work, &ib_loc);
}
if (INFO != 0)
printf("ERROR ZUNMQR INFO %d \n", (int) INFO);
}
}
} else if (side == MagmaRight) {
rownbm = magma_ceildiv((n-1),nb);
for (magma_int_t k = 1; k <= rownbm; k++) {
ncolinvolvd = min(n-1, k*nb);
avai_blksiz=min(Vblksiz,ncolinvolvd);
nbgr = magma_ceildiv(ncolinvolvd,avai_blksiz);
for (magma_int_t j = 1; j <= nbgr; j++) {
vlen = 0;
vnb = 0;
cur_blksiz = min(ncolinvolvd-(j-1)*avai_blksiz, avai_blksiz);
colst = (j-1)*avai_blksiz;
coled = colst + cur_blksiz -1;
fst = (rownbm -k)*nb+colst +1;
for (colj=colst; colj <= coled; colj++) {
st = (rownbm -k)*nb+colj +1;
ed = min(st+nb-1,n-1);
if (st > ed)
break;
if ((st == ed) && (colj != n-2))
break;
vlen=ed-fst+1;
vnb=vnb+1;
}
magma_bulge_findVTpos(n, nb, Vblksiz, colst, fst, ldv, ldt, &vpos, &tpos);
if ((vlen > 0) && (vnb > 0)) {
lapackf77_zlarfb( "R", "N", "F", "C", &ib_loc, &vlen, &vnb, V(vpos), &ldv, T(tpos), &ldt, E(i*nb_loc,fst), &lde, work, &ib_loc);
}
}
}
} else {
printf("ERROR SIDE %d \n",side);
}
} // END loop over the chunks
magma_free_cpu(work);
magma_free_cpu(work2);
}
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by ZGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A COMPLEX_16 array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by ZGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
dT COMPLEX_16 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_zgeqrf_gpu.
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in DT.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument has an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr(magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex *dT, magma_int_t nb,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
#define dT(j) (dT + (j)*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk; //, iinfo;
magma_int_t lddwork;
magmaDoubleComplex *dA, *dV, *dW;
magmaDoubleComplex *work;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_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 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_zmalloc( &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_zmalloc_cpu( &work, lwork );
if (work == NULL) {
magma_free( dA );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magmaDoubleComplex *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 zungqr is slow for
// k smaller than the zungqr's blocking size (new version can be up to 60x faster)
lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_zlacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_zlaset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_zlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_zlarfb( 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_zsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda );
}
}
if (kk > 0) {
// Use blocked code
// stream: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
magmablasSetKernelStream( stream );
for (i = ki; i >= 0; i -= nb) {
ib = min(nb, k - i);
// Send current panel to the GPU
mi = m - i;
lapackf77_zlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_zsetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
// set panel to identity
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT(i), nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_zgetmatrix( m, n,
dA(0, 0), ldda, A(0, 0), lda);
}
magma_queue_destroy( stream );
magma_free( dA );
magma_free_cpu( work );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_zungqr */
