/* ////////////////////////////////////////////////////////////////////////////
-- Return nb for geqlf based on m
*/
magma_int_t magma_get_sgeqlf_nb( magma_int_t m )
{
magma_int_t arch = magma_getdevice_arch();
if ( arch >= 200 ) { // 2.x Fermi
return magma_get_sgeqrf_nb( m );
}
else { // 1.x
if (m < 1024) return 32;
else if (m < 4032) return 64;
else return 128;
}
}
/**
Purpose
-------
SORGQR generates an M-by-N REAL matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by SGEQRF.
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 REAL array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by SGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau REAL array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by SGEQRF_GPU.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument has an illegal value
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sorgqr2(magma_int_t m, magma_int_t n, magma_int_t k,
float *A, magma_int_t lda,
float *tau,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
magma_int_t nb = magma_get_sgeqrf_nb(min(m, n));
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk; //, iinfo;
magma_int_t lddwork;
float *dA, *dV, *dW, *dT, *T;
float *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_smalloc( &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_smalloc_cpu( &work, lwork );
T = work;
if (work == NULL) {
magma_free( dA );
magma_free_cpu( work );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
float *V = work + (n+nb)*nb;
magma_queue_t stream;
magma_queue_create( &stream );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
/*
lapackf77_sorgqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_slacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_slaset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_slarfb( 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_ssetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_slaset( 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_slaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_ssetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&mi, &ib,
A(i,i), &lda, &tau[i], T, &nb);
magma_ssetmatrix_async( ib, ib,
T, nb,
dT, nb, stream );
// set panel to identity
magmablas_slaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda );
magmablas_slaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda );
magma_queue_sync( stream );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_slarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT, nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_sgetmatrix( 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_sorgqr */
/**
Purpose
-------
SGELS solves the overdetermined, least squares problem
min || A*X - C ||
using the QR factorization A.
The underdetermined problem (m < n) is not currently handled.
Arguments
---------
@param[in]
trans magma_trans_t
- = MagmaNoTrans: the linear system involves A.
Only TRANS=MagmaNoTrans is currently handled.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in,out]
dA REAL array, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, A is overwritten by details of its QR
factorization as returned by SGEQRF3.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in,out]
dB REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) REAL array, dimension MAX(1,LWORK).
On exit, if INFO = 0, HWORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array HWORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_sgeqrf_nb( M, N ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the HWORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgels_driver
********************************************************************/
extern "C" magma_int_t
magma_sgels3_gpu(
magma_trans_t trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, magma_int_t ldda,
magmaFloat_ptr dB, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info)
{
magmaFloat_ptr dT;
float *tau;
magma_int_t min_mn;
magma_int_t nb = magma_get_sgeqrf_nb( m, n );
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
bool lquery = (lwork == -1);
hwork[0] = magma_smake_lwork( lwkopt );
*info = 0;
/* For now, N is the only case working */
if ( trans != MagmaNoTrans )
*info = -1;
else if (m < 0)
*info = -2;
else if (n < 0 || m < n) /* LQ is not handle for now*/
*info = -3;
else if (nrhs < 0)
*info = -4;
else if (ldda < max(1,m))
*info = -6;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
min_mn = min(m,n);
if (min_mn == 0) {
hwork[0] = MAGMA_S_ONE;
return *info;
}
/*
* Allocate temporary buffers
*/
magma_int_t ldtwork = ( 2*min_mn + magma_roundup( n, 32 ) )*nb;
if (nb < nrhs)
ldtwork = ( 2*min_mn + magma_roundup( n, 32 ) )*nrhs;
if (MAGMA_SUCCESS != magma_smalloc( &dT, ldtwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_smalloc_cpu( &tau, min_mn );
if ( tau == NULL ) {
magma_free( dT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_sgeqrf3_gpu( m, n, dA, ldda, tau, dT, info );
if ( *info == 0 ) {
magma_sgeqrs3_gpu( m, n, nrhs,
dA, ldda, tau, dT,
dB, lddb, hwork, lwork, info );
}
magma_free( dT );
magma_free_cpu( tau );
return *info;
}
extern "C" magma_int_t
magma_sgeqrf2_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, size_t dA_offset, magma_int_t ldda,
float *tau,
magma_queue_t* queue,
magma_int_t *info)
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
SGEQRF computes a QR factorization of a real M-by-N matrix A:
A = Q * R.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
TAU (output) REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dA(a_1,a_2) dA, (dA_offset + (a_1) + (a_2)*(ldda))
#define work(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
magmaFloat_ptr dwork;
float *work;
magma_int_t i, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
lwork = (m+n) * nb;
lhwork = lwork - (m)*nb;
if ( MAGMA_SUCCESS != magma_smalloc( &dwork, n*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/*
if ( MAGMA_SUCCESS != magma_smalloc_cpu( &work, lwork ) ) {
*info = MAGMA_ERR_HOST_ALLOC;
magma_free( dwork );
return *info;
}
*/
cl_mem buffer = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR,
sizeof(float)*lwork, NULL, NULL);
work = (float*)clEnqueueMapBuffer(queue[0], buffer, CL_TRUE,
CL_MAP_READ | CL_MAP_WRITE,
0, lwork*sizeof(float),
0, NULL, NULL, NULL);
nbmin = 2;
nx = nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb) {
ib = min(k-i, nb);
rows = m -i;
magma_queue_sync( queue[1] );
magma_sgetmatrix_async(rows, ib, dA(i, i), ldda, work(i), ldwork, queue[0], NULL);
if (i > 0) {
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i ), ldda, dwork,0, lddwork,
dA(old_i, old_i+2*old_ib), ldda, dwork,old_ib, lddwork, queue[1]);
magma_ssetmatrix_async( old_ib, old_ib, work(old_i), ldwork,
dA(old_i, old_i), ldda, queue[1], NULL);
}
magma_queue_sync(queue[0]);
lapackf77_sgeqrf(&rows, &ib, work(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work(i), &ldwork, tau+i, hwork, &ib);
spanel_to_q( MagmaUpper, ib, work(i), ldwork, hwork+ib*ib );
/* download the i-th V matrix */
magma_ssetmatrix_async(rows, ib, work(i), ldwork, dA(i,i), ldda, queue[0], NULL);
/* download the T matrix */
magma_queue_sync( queue[1] );
magma_ssetmatrix_async( ib, ib, hwork, ib, dwork, 0, lddwork, queue[0], NULL);
magma_queue_sync( queue[0] );
if (i + ib < n) {
if (i+nb < k-nx) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dwork,0, lddwork,
dA(i, i+ib), ldda, dwork,ib, lddwork, queue[1]);
sq_to_panel( MagmaUpper, ib, work(i), ldwork, hwork+ib*ib );
}
else {
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dwork,0, lddwork,
dA(i, i+ib), ldda, dwork,ib, lddwork, queue[1]);
sq_to_panel( MagmaUpper, ib, work(i), ldwork, hwork+ib*ib );
magma_ssetmatrix_async(ib, ib, work(i), ldwork, dA(i,i), ldda, queue[1], NULL);
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
magma_free(dwork);
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_sgetmatrix_async(rows, ib, dA(i, i), ldda, work, rows, queue[1], NULL);
magma_queue_sync(queue[1]);
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_ssetmatrix_async(rows, ib, work, rows, dA(i, i), ldda, queue[1], NULL);
}
magma_queue_sync(queue[0]);
magma_queue_sync(queue[1]);
// magma_free_cpu(work);
clEnqueueUnmapMemObject(queue[0], buffer, work, 0, NULL, NULL);
clReleaseMemObject(buffer);
return *info;
} /* magma_sgeqrf2_gpu */
/**
Purpose
-------
SGEQRF computes a QR factorization of a REAL M-by-N matrix A:
A = Q * R. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= max( N*NB, 2*NB*NB ),
where NB can be obtained through magma_get_sgeqrf_nb(M).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf(magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *tau,
float *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)
float *dA, *dwork, *dT;
float c_one = MAGMA_S_ONE;
magma_int_t i, k, lddwork, old_i, old_ib;
magma_int_t ib, ldda;
/* Function Body */
*info = 0;
magma_int_t nb = magma_get_sgeqrf_nb(min(m, n));
// need 2*nb*nb to store T and upper triangle of V simultaneously
magma_int_t lwkopt = max(n*nb, 2*nb*nb);
work[0] = MAGMA_S_MAKE( (float)lwkopt, 0 );
int lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1, lwkopt) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
work[0] = c_one;
return *info;
}
// largest N for larfb is n-nb (trailing matrix lacks 1st panel)
lddwork = ((n+31)/32)*32 - nb;
ldda = ((m+31)/32)*32;
magma_int_t num_gpus = magma_num_gpus();
if ( num_gpus > 1 ) {
/* call multiple-GPU interface */
return magma_sgeqrf4(num_gpus, m, n, A, lda, tau, work, lwork, info);
}
// allocate space for dA, dwork, and dT
if (MAGMA_SUCCESS != magma_smalloc( &dA, n*ldda + nb*lddwork + nb*nb )) {
/* Switch to the "out-of-core" (out of GPU-memory) version */
return magma_sgeqrf_ooc(m, n, A, lda, tau, work, lwork, info);
}
/* Define user stream if current stream is NULL */
magma_queue_t stream[2], current_stream;
magmablasGetKernelStream(¤t_stream);
magma_queue_create( &stream[0] );
if (current_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else {
stream[1] = current_stream;
}
dwork = dA + n*ldda;
dT = dA + n*ldda + nb*lddwork;
if ( (nb > 1) && (nb < k) ) {
/* Use blocked code initially.
Asynchronously send the matrix to the GPU except the first panel. */
magma_ssetmatrix_async( m, n-nb,
A(0,nb), lda,
dA(0,nb), ldda, stream[0] );
old_i = 0;
old_ib = nb;
for (i = 0; i < k-nb; i += nb) {
ib = min(k-i, nb);
if (i > 0) {
/* download i-th panel */
magma_queue_sync( stream[1] );
magma_sgetmatrix_async( m-i, ib,
dA(i,i), ldda,
A(i,i), lda, stream[0] );
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i), ldda, dT, nb,
dA(old_i, old_i+2*old_ib), ldda, dwork, lddwork);
magma_sgetmatrix_async( i, ib,
dA(0,i), ldda,
A(0,i), lda, stream[1] );
magma_queue_sync( stream[0] );
}
magma_int_t rows = m-i;
lapackf77_sgeqrf(&rows, &ib, A(i,i), &lda, tau+i, work, &lwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib, A(i,i), &lda, tau+i, work, &ib);
spanel_to_q(MagmaUpper, ib, A(i,i), lda, work+ib*ib);
/* download the i-th V matrix */
magma_ssetmatrix_async( rows, ib, A(i,i), lda, dA(i,i), ldda, stream[0] );
/* download the T matrix */
magma_queue_sync( stream[1] );
magma_ssetmatrix_async( ib, ib, work, ib, dT, nb, stream[0] );
magma_queue_sync( stream[0] );
if (i + ib < n) {
if (i+ib < k-nb) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left (look-ahead) */
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dT, nb,
dA(i, i+ib), ldda, dwork, lddwork);
sq_to_panel(MagmaUpper, ib, A(i,i), lda, work+ib*ib);
}
else {
/* After last panel, update whole trailing matrix. */
/* Apply H' to A(i:m,i+ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dT, nb,
dA(i, i+ib), ldda, dwork, lddwork);
sq_to_panel(MagmaUpper, ib, A(i,i), lda, work+ib*ib);
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
if (i != 0) {
magma_sgetmatrix_async( m, ib, dA(0,i), ldda, A(0,i), lda, stream[1] );
magma_queue_sync( stream[1] );
}
magma_int_t rows = m-i;
lapackf77_sgeqrf(&rows, &ib, A(i,i), &lda, tau+i, work, &lwork, info);
}
magma_queue_destroy( stream[0] );
if (current_stream == NULL) {
magma_queue_destroy( stream[1] );
magmablasSetKernelStream(NULL);
}
magma_free( dA );
return *info;
} /* magma_sgeqrf */
/* ////////////////////////////////////////////////////////////////////////////
-- Return nb for gelqf based on m
*/
magma_int_t magma_get_sgelqf_nb( magma_int_t m )
{
return magma_get_sgeqrf_nb( m );
}
int magma_get_sgeqlf_nb(int m) {
return magma_get_sgeqrf_nb(m);
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sormqr_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max, dt_size;
float *C, *R, *A, *W, *tau;
float *dC, *dA, *dT;
magma_opts opts;
parse_opts( argc, argv, &opts );
// test all combinations of input parameters
const char* side[] = { MagmaLeftStr, MagmaRightStr };
const char* trans[] = { MagmaTransStr, MagmaNoTransStr };
printf(" M N K side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n");
printf("===============================================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
m = opts.msize[i];
n = opts.nsize[i];
k = opts.ksize[i];
nb = magma_get_sgeqrf_nb( m );
ldc = ((m + 31)/32)*32;
lda = ((max(m,n) + 31)/32)*32;
gflops = FLOPS_SORMQR( m, n, k, *side[iside] ) / 1e9;
if ( *side[iside] == 'L' && m < k ) {
printf( "%5d %5d %5d %-5s %-9s skipping because side=left and m < k\n",
(int) m, (int) n, (int) k, side[iside], trans[itran] );
continue;
}
if ( *side[iside] == 'R' && n < k ) {
printf( "%5d %5d %5d %-5s %-9s skipping because side=right and n < k\n",
(int) m, (int) n, (int) k, side[iside], trans[itran] );
continue;
}
if ( *side[iside] == 'L' ) {
// side = left
lwork_max = (m - k + nb)*(n + nb) + n*nb;
dt_size = ( 2*min(m,k) + ((k + 31)/32)*32 )*nb;
}
else {
// side = right
lwork_max = (n - k + nb)*(m + nb) + m*nb;
dt_size = ( 2*min(n,k) + ((k + 31)/32)*32 )*nb;
}
TESTING_MALLOC_CPU( C, float, ldc*n );
TESTING_MALLOC_CPU( R, float, ldc*n );
TESTING_MALLOC_CPU( A, float, lda*k );
TESTING_MALLOC_CPU( W, float, lwork_max );
TESTING_MALLOC_CPU( tau, float, k );
TESTING_MALLOC_DEV( dC, float, ldc*n );
TESTING_MALLOC_DEV( dA, float, lda*k );
TESTING_MALLOC_DEV( dT, float, dt_size );
// C is full, m x n
size = ldc*n;
lapackf77_slarnv( &ione, ISEED, &size, C );
magma_ssetmatrix( m, n, C, ldc, dC, ldc );
// A is m x k (left) or n x k (right)
lda = (*side[iside] == 'L' ? m : n);
size = lda*k;
lapackf77_slarnv( &ione, ISEED, &size, A );
// compute QR factorization to get Householder vectors in dA, tau, dT
magma_ssetmatrix( lda, k, A, lda, dA, lda );
magma_sgeqrf_gpu( lda, k, dA, lda, tau, dT, &info );
magma_sgetmatrix( lda, k, dA, lda, A, lda );
if (info != 0)
printf("magma_sgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_sormqr( side[iside], trans[itran],
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sormqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_sormqr_gpu( *side[iside], *trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, W, lwork, dT, nb, &info );
if (info != 0)
printf("magma_sormqr_gpu (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_S_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max )
printf("invalid lwork %d, lwork_max %d\n", (int) lwork, (int) lwork_max );
gpu_time = magma_sync_wtime( 0 ); // sync needed for L,N and R,T cases
magma_sormqr_gpu( *side[iside], *trans[itran],
m, n, k,
dA, lda, tau, dC, ldc, W, lwork, dT, nb, &info );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sormqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_sgetmatrix( m, n, dC, ldc, R, ldc );
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_slange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_saxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_slange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %-5s %-9s %7.2f (%7.2f) %7.2f (%7.2f) %8.2e\n",
(int) m, (int) n, (int) k, side[iside], trans[itran],
cpu_perf, cpu_time, gpu_perf, gpu_time, error );
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
}} // end iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return 0;
}
magma_int_t magmaf_get_sgeqrf_nb( magma_int_t *m )
{
return magma_get_sgeqrf_nb( *m );
}
/**
Purpose
-------
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by SGEQRF3_GPU.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in]
dA REAL array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
SGEQRF3_GPU in the first n columns of its array argument A.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in]
tau REAL array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_SGEQRF_GPU.
@param[in,out]
dB REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
dT REAL array that is the output (the 6th argument)
of magma_sgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
matrices for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) REAL array, dimension (LWORK)
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_sgeqrf_nb( M ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgels_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrs3_gpu(
magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, magma_int_t ldda,
float *tau,
magmaFloat_ptr dT,
magmaFloat_ptr dB, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info)
{
#define dA(a_1,a_2) (dA + (a_2)*(ldda) + (a_1))
#define dT(a_1) (dT + (lddwork+(a_1))*nb)
float c_one = MAGMA_S_ONE;
magma_int_t k, lddwork;
magma_int_t nb = magma_get_sgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_S_MAKE( (float)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
lddwork = k;
/* B := Q' * B */
magma_sormqr_gpu( MagmaLeft, MagmaTrans,
m, nrhs, n,
dA(0,0), ldda, tau,
dB, lddb, hwork, lwork, dT, nb, info );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:)
1. Move the (k-1)/nb block diagonal submatrices from dT to R
2. Solve
3. Restore the data format moving data from R back to dT
*/
magmablas_sswapdblk(k-1, nb, dA(0,0), ldda, 1, dT(0), nb, 0);
if ( nrhs == 1 ) {
magma_strsv(MagmaUpper, MagmaNoTrans, MagmaNonUnit,
n, dA(0,0), ldda, dB, 1);
} else {
magma_strsm(MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
n, nrhs, c_one, dA(0,0), ldda, dB, lddb);
}
magmablas_sswapdblk(k-1, nb, dT(0), nb, 0, dA(0,0), ldda, 1);
return *info;
}
/**
Purpose
-------
SGEQRF computes a QR factorization of a real M-by-N matrix A:
A = Q * R.
This version stores the triangular dT matrices used in
the block QR factorization so that they can be applied directly (i.e.,
without being recomputed) later. As a result, the application
of Q is much faster. Also, the upper triangular matrices for V have 0s
in them. The corresponding parts of the upper triangular R are inverted
and stored separately in dT.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
dT (workspace) REAL array on the GPU,
dimension (2*MIN(M, N) + (N+31)/32*32 )*NB,
where NB can be obtained through magma_get_sgeqrf_nb(M).
It starts with MIN(M,N)*NB block that store the triangular T
matrices, followed by the MIN(M,N)*NB block of the diagonal
inverses for the R matrix. The rest of the array is used as workspace.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
float *tau,
magmaFloat_ptr dT,
magma_int_t *info )
{
#define dA(a_1,a_2) (dA + (a_2)*(ldda) + (a_1))
#define dT(a_1) (dT + (a_1)*nb)
#define d_ref(a_1) (dT + ( minmn+(a_1))*nb)
#define dd_ref(a_1) (dT + (2*minmn+(a_1))*nb)
#define work(a_1) (work + (a_1))
#define hwork (work + (nb)*(m))
magma_int_t i, k, minmn, old_i, old_ib, rows, cols;
magma_int_t ib, nb;
magma_int_t ldwork, lddwork, lwork, lhwork;
float *work, *ut;
/* check arguments */
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = minmn = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
lwork = (m + n + nb)*nb;
lhwork = lwork - m*nb;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, lwork )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
ut = hwork+nb*(n);
memset( ut, 0, nb*nb*sizeof(float));
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
ldwork = m;
lddwork= n;
if ( (nb > 1) && (nb < k) ) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nb; i += nb) {
ib = min(k-i, nb);
rows = m -i;
magma_sgetmatrix_async( rows, ib,
dA(i,i), ldda,
work(i), ldwork, stream[1] );
if (i > 0) {
/* Apply H' to A(i:m,i+2*ib:n) from the left */
cols = n-old_i-2*old_ib;
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, cols, old_ib,
dA(old_i, old_i ), ldda, dT(old_i), nb,
dA(old_i, old_i+2*old_ib), ldda, dd_ref(0), lddwork);
/* store the diagonal */
magma_ssetmatrix_async( old_ib, old_ib,
ut, old_ib,
d_ref(old_i), old_ib, stream[0] );
}
magma_queue_sync( stream[1] );
lapackf77_sgeqrf(&rows, &ib, work(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work(i), &ldwork, tau+i, hwork, &ib);
/* Put 0s in the upper triangular part of a panel (and 1s on the
diagonal); copy the upper triangular in ut and invert it. */
magma_queue_sync( stream[0] );
ssplit_diag_block(ib, work(i), ldwork, ut);
magma_ssetmatrix( rows, ib, work(i), ldwork, dA(i,i), ldda );
if (i + ib < n) {
/* Send the triangular factor T to the GPU */
magma_ssetmatrix( ib, ib, hwork, ib, dT(i), nb );
if (i+nb < k-nb) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dT(i), nb,
dA(i, i+ib), ldda, dd_ref(0), lddwork);
}
else {
cols = n-i-ib;
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, cols, ib,
dA(i, i ), ldda, dT(i), nb,
dA(i, i+ib), ldda, dd_ref(0), lddwork);
/* Fix the diagonal block */
magma_ssetmatrix( ib, ib, ut, ib, d_ref(i), ib );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_sgetmatrix( rows, ib, dA(i, i), ldda, work, rows );
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_ssetmatrix( rows, ib, work, rows, dA(i, i), ldda );
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free_pinned( work );
return *info;
} /* magma_sgeqrf_gpu */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrs_gpu
*/
int main( int argc, char** argv)
{
//#if defined(PRECISION_s)
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float matnorm, work[1];
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *hwork, tmp[1];
magmaFloat_ptr d_A, d_B;
/* Matrix size */
magma_int_t M = 0, N = 0, n2;
magma_int_t lda, ldb, ldda, lddb, lworkgpu, lhwork;
magma_int_t size[7] = {1024,2048,3072,4032,5184,6016,7000};
magma_int_t i, info, min_mn, nb, l1, l2;
magma_int_t ione = 1;
magma_int_t nrhs = 3;
magma_int_t ISEED[4] = {0,0,0,1};
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
else if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
else if (strcmp("-nrhs", argv[i])==0)
nrhs = atoi(argv[++i]);
}
if (N>0 && M>0 && M >= N)
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, M, N);
else
{
printf("\nUsage: \n");
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, M, N);
printf(" M has to be >= N, exit.\n");
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, 1024, 1024);
M = N = size[6];
}
ldda = ((M+31)/32)*32;
lddb = ldda;
n2 = M * N;
min_mn = min(M, N);
nb = magma_get_sgeqrf_nb(M);
lda = ldb = M;
lworkgpu = (M-N + nb)*(nrhs+2*nb);
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( tau, float, min_mn );
TESTING_MALLOC_PIN( h_A, float, lda*N );
TESTING_MALLOC_PIN( h_A2, float, lda*N );
TESTING_MALLOC_PIN( h_B, float, ldb*nrhs );
TESTING_MALLOC_PIN( h_X, float, ldb*nrhs );
TESTING_MALLOC_PIN( h_R, float, ldb*nrhs );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, lddb*nrhs );
/*
* Get size for host workspace
*/
lhwork = -1;
lapackf77_sgeqrf(&M, &N, h_A, &M, tau, tmp, &lhwork, &info);
l1 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lhwork = -1;
lapackf77_sormqr( MagmaLeftStr, MagmaTransStr,
&M, &nrhs, &min_mn, h_A, &lda, tau,
h_X, &ldb, tmp, &lhwork, &info);
l2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lhwork = max( max( l1, l2 ), lworkgpu );
TESTING_MALLOC_PIN( hwork, float, lhwork );
printf("\n");
printf(" ||b-Ax|| / (N||A||)\n");
printf(" M N CPU GFlop/s GPU GFlop/s CPU GPU \n");
printf("============================================================\n");
for(i=0; i<7; i++){
if (argc == 1){
M = N = size[i];
}
min_mn= min(M, N);
ldb = lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = (FLOPS_GEQRF( (float)M, (float)N )
+ FLOPS_GEQRS( (float)M, (float)N, (float)nrhs )) / 1e9;
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
n2 = M*nrhs;
lapackf77_slarnv( &ione, ISEED, &n2, h_B );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Warm up to measure the performance */
magma_ssetmatrix( M, N, h_A, 0, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( M, nrhs, h_B, 0, ldb, d_B, 0, lddb, queue );
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, 0, ldda,
d_B, 0, lddb, hwork, lworkgpu, &info, queue);
magma_ssetmatrix( M, N, h_A, 0, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( M, nrhs, h_B, 0, ldb, d_B, 0, lddb, queue );
gpu_time = magma_wtime();
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, 0, ldda,
d_B, 0, lddb, hwork, lworkgpu, &info, queue);
gpu_time = magma_wtime() - gpu_time;
if (info < 0)
printf("Argument %d of magma_sgels had an illegal value.\n", -info);
gpu_perf = gflops / gpu_time;
// Get the solution in h_X
magma_sgetmatrix( N, nrhs, d_B, 0, lddb, h_X, 0, ldb, queue );
// compute the residual
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb);
matnorm = lapackf77_slange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, hwork, &lhwork, &info);
cpu_time = magma_wtime()-cpu_time;
cpu_perf = gflops / cpu_time;
if (info < 0)
printf("Argument %d of lapackf77_sgels had an illegal value.\n", -info);
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb);
printf("%5d %5d %6.1f %6.1f %7.2e %7.2e\n",
M, N, cpu_perf, gpu_perf,
lapackf77_slange("f", &M, &nrhs, h_B, &M, work)/(min_mn*matnorm),
lapackf77_slange("f", &M, &nrhs, h_R, &M, work)/(min_mn*matnorm) );
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_A2 );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_X );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( hwork );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
/* Shutdown */
magma_queue_destroy( queue );
magma_finalize();
}
/**
Purpose
-------
SGEQRF_OOC computes a QR factorization of a REAL M-by-N matrix A:
A = Q * R. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
This is an out-of-core (ooc) version that is similar to magma_sgeqrf but
the difference is that this version can use a GPU even if the matrix
does not fit into the GPU memory at once.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_sgeqrf_nb(M).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf_ooc(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *tau,
float *work, magma_int_t lwork,
magma_int_t *info )
{
#define A(a_1,a_2) ( A + (a_2)*(lda) + (a_1))
#define dA(a_1,a_2) (dA + (a_2)*ldda + (a_1))
float *dA, *dwork;
float c_one = MAGMA_S_ONE;
int k, lddwork, ldda;
*info = 0;
int nb = magma_get_sgeqrf_nb(min(m, n));
int lwkopt = n * nb;
work[0] = MAGMA_S_MAKE( (float)lwkopt, 0 );
int lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,n) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* Check how much memory do we have */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(float);
magma_int_t IB, NB = (magma_int_t)(0.8*freeMem/m);
NB = (NB / nb) * nb;
if (NB >= n)
return magma_sgeqrf(m, n, A, lda, tau, work, lwork, info);
k = min(m,n);
if (k == 0) {
work[0] = c_one;
return *info;
}
lddwork = ((NB+31)/32)*32+nb;
ldda = ((m+31)/32)*32;
if (MAGMA_SUCCESS != magma_smalloc( &dA, (NB + nb)*ldda + nb*lddwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
// magmablasSetKernelStream(stream[1]);
float *ptr = dA + ldda * NB;
dwork = dA + ldda*(NB + nb);
/* start the main loop over the blocks that fit in the GPU memory */
for (int i=0; i < n; i += NB) {
IB = min(n-i, NB);
//printf("Processing %5d columns -- %5d to %5d ... \n", IB, i, i+IB);
/* 1. Copy the next part of the matrix to the GPU */
magma_ssetmatrix_async( (m), IB,
A(0,i), lda,
dA(0,0), ldda, stream[0] );
magma_queue_sync( stream[0] );
/* 2. Update it with the previous transformations */
for (int j=0; j < min(i,k); j += nb) {
magma_int_t ib = min(k-j, nb);
/* Get a panel in ptr. */
// 1. Form the triangular factor of the block reflector
// 2. Send it to the GPU.
// 3. Put 0s in the upper triangular part of V.
// 4. Send V to the GPU in ptr.
// 5. Update the matrix.
// 6. Restore the upper part of V.
magma_int_t rows = m-j;
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib, A(j,j), &lda, tau+j, work, &ib);
magma_ssetmatrix_async( ib, ib,
work, ib,
dwork, lddwork, stream[1] );
spanel_to_q(MagmaUpper, ib, A(j,j), lda, work+ib*ib);
magma_ssetmatrix_async( rows, ib,
A(j,j), lda,
ptr, rows, stream[1] );
magma_queue_sync( stream[1] );
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, IB, ib,
ptr, rows, dwork, lddwork,
dA(j, 0), ldda, dwork+ib, lddwork);
sq_to_panel(MagmaUpper, ib, A(j,j), lda, work+ib*ib);
}
/* 3. Do a QR on the current part */
if (i < k)
magma_sgeqrf2_gpu(m-i, IB, dA(i,0), ldda, tau+i, info);
/* 4. Copy the current part back to the CPU */
magma_sgetmatrix_async( (m), IB,
dA(0,0), ldda,
A(0,i), lda, stream[0] );
}
magma_queue_sync( stream[0] );
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free( dA );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_sgeqrf_ooc */
extern "C" magma_int_t
magma_sgeqrf2_mgpu( magma_int_t num_gpus, magma_int_t m, magma_int_t n,
float **dlA, magma_int_t ldda,
float *tau,
magma_int_t *info )
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
SGEQRF2_MGPU computes a QR factorization of a real M-by-N matrix A:
A = Q * R. This is a GPU interface of the routine.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
TAU (output) REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dlA(dev, i, j) (dlA[dev] + (i) + (j)*(ldda))
#define hpanel(i) (hpanel + (i))
// set to NULL to make cleanup easy: free(NULL) does nothing.
float *dwork[MagmaMaxGPUs]={NULL}, *dpanel[MagmaMaxGPUs]={NULL};
float *hwork=NULL, *hpanel=NULL;
magma_queue_t stream[MagmaMaxGPUs][2]={{NULL}};
magma_event_t panel_event[MagmaMaxGPUs]={NULL};
magma_int_t i, j, min_mn, dev, ldhpanel, lddwork, rows;
magma_int_t ib, nb;
magma_int_t lhwork, lwork;
magma_int_t panel_dev, i_local, i_nb_local, n_local[MagmaMaxGPUs], la_dev, dpanel_offset;
magma_queue_t cqueue;
magmablasGetKernelStream( &cqueue );
magma_device_t cdevice;
magma_getdevice( &cdevice );
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
min_mn = min(m,n);
if (min_mn == 0)
return *info;
nb = magma_get_sgeqrf_nb( m );
/* dwork is (n*nb) --- for T (nb*nb) and slarfb work ((n-nb)*nb) ---
* + dpanel (ldda*nb), on each GPU.
* I think slarfb work could be smaller, max(n_local[:]).
* Oddly, T and slarfb work get stacked on top of each other, both with lddwork=n.
* on GPU that owns panel, set dpanel = dlA(dev,i,i_local).
* on other GPUs, set dpanel = dwork[dev] + dpanel_offset. */
lddwork = n;
dpanel_offset = lddwork*nb;
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
if ( MAGMA_SUCCESS != magma_smalloc( &(dwork[dev]), (lddwork + ldda)*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto CLEANUP;
}
}
/* hwork is MAX( workspace for sgeqrf (n*nb), two copies of T (2*nb*nb) )
* + hpanel (m*nb).
* for last block, need 2*n*nb total. */
ldhpanel = m;
lhwork = max( n*nb, 2*nb*nb );
lwork = max( lhwork + ldhpanel*nb, 2*n*nb );
if ( MAGMA_SUCCESS != magma_smalloc_pinned( &hwork, lwork )) {
*info = MAGMA_ERR_HOST_ALLOC;
goto CLEANUP;
}
hpanel = hwork + lhwork;
/* Set the number of local n for each GPU */
for( dev=0; dev < num_gpus; dev++ ) {
n_local[dev] = ((n/nb)/num_gpus)*nb;
if (dev < (n/nb) % num_gpus)
n_local[dev] += nb;
else if (dev == (n/nb) % num_gpus)
n_local[dev] += n % nb;
}
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
magma_queue_create( &stream[dev][0] );
magma_queue_create( &stream[dev][1] );
magma_event_create( &panel_event[dev] );
}
if ( nb < min_mn ) {
/* Use blocked code initially */
// Note: as written, ib cannot be < nb.
for( i = 0; i < min_mn-nb; i += nb ) {
/* Set the GPU number that holds the current panel */
panel_dev = (i/nb) % num_gpus;
/* Set the local index where the current panel is (j==i) */
i_local = i/(nb*num_gpus)*nb;
ib = min(min_mn-i, nb);
rows = m-i;
/* Send current panel to the CPU, after panel_event indicates it has been updated */
magma_setdevice( panel_dev );
magma_queue_wait_event( stream[panel_dev][1], panel_event[panel_dev] );
magma_sgetmatrix_async( rows, ib,
dlA(panel_dev, i, i_local), ldda,
hpanel(i), ldhpanel, stream[panel_dev][1] );
magma_queue_sync( stream[panel_dev][1] );
// Factor panel
lapackf77_sgeqrf( &rows, &ib, hpanel(i), &ldhpanel, tau+i,
hwork, &lhwork, info );
if ( *info != 0 ) {
fprintf( stderr, "error %d\n", (int) *info );
}
// Form the triangular factor of the block reflector
// H = H(i) H(i+1) . . . H(i+ib-1)
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
hpanel(i), &ldhpanel, tau+i, hwork, &ib );
spanel_to_q( MagmaUpper, ib, hpanel(i), ldhpanel, hwork + ib*ib );
// Send the current panel back to the GPUs
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
if (dev == panel_dev)
dpanel[dev] = dlA(dev, i, i_local);
else
dpanel[dev] = dwork[dev] + dpanel_offset;
magma_ssetmatrix_async( rows, ib,
hpanel(i), ldhpanel,
dpanel[dev], ldda, stream[dev][0] );
}
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
magma_queue_sync( stream[dev][0] );
}
// TODO: if spanel_to_q copied whole block, wouldn't need to restore
// -- just send the copy to the GPUs.
// TODO: also, could zero out the lower triangle and use Azzam's larfb w/ gemm.
/* Restore the panel */
sq_to_panel( MagmaUpper, ib, hpanel(i), ldhpanel, hwork + ib*ib );
if (i + ib < n) {
/* Send the T matrix to the GPU. */
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
magma_ssetmatrix_async( ib, ib,
hwork, ib,
dwork[dev], lddwork, stream[dev][0] );
}
la_dev = (panel_dev+1) % num_gpus;
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
magmablasSetKernelStream( stream[dev][0] );
if (dev == la_dev && i+nb < min_mn-nb) {
// If not last panel,
// for look-ahead panel, apply H' to A(i:m,i+ib:i+2*ib)
i_nb_local = (i+nb)/(nb*num_gpus)*nb;
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dpanel[dev], ldda, // V
dwork[dev], lddwork, // T
dlA(dev, i, i_nb_local), ldda, // C
dwork[dev]+ib, lddwork ); // work
magma_event_record( panel_event[dev], stream[dev][0] );
// for trailing matrix, apply H' to A(i:m,i+2*ib:n)
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n_local[dev]-(i_nb_local+ib), ib,
dpanel[dev], ldda, // V
dwork[dev], lddwork, // T
dlA(dev, i, i_nb_local+ib), ldda, // C
dwork[dev]+ib, lddwork ); // work
}
else {
// for trailing matrix, apply H' to A(i:m,i+ib:n)
i_nb_local = i_local;
if (dev <= panel_dev) {
i_nb_local += ib;
}
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n_local[dev]-i_nb_local, ib,
dpanel[dev], ldda, // V
dwork[dev], lddwork, // T
dlA(dev, i, i_nb_local), ldda, // C
dwork[dev]+ib, lddwork ); // work
}
}
// Restore top of panel (after larfb is done)
magma_setdevice( panel_dev );
magma_ssetmatrix_async( ib, ib,
hpanel(i), ldhpanel,
dlA(panel_dev, i, i_local), ldda, stream[panel_dev][0] );
}
}
}
else {
i = 0;
}
/* Use unblocked code to factor the last or only block row. */
if (i < min_mn) {
rows = m-i;
for( j=i; j < n; j += nb ) {
panel_dev = (j/nb) % num_gpus;
i_local = j/(nb*num_gpus)*nb;
ib = min( n-j, nb );
magma_setdevice( panel_dev );
magma_sgetmatrix( rows, ib,
dlA(panel_dev, i, i_local), ldda,
hwork + (j-i)*rows, rows );
}
// needs lwork >= 2*n*nb:
// needs (m-i)*(n-i) for last block row, bounded by nb*n.
// needs (n-i)*nb for sgeqrf work, bounded by n*nb.
ib = n-i; // total columns in block row
lhwork = lwork - ib*rows;
lapackf77_sgeqrf( &rows, &ib, hwork, &rows, tau+i, hwork + ib*rows, &lhwork, info );
if ( *info != 0 ) {
fprintf( stderr, "error %d\n", (int) *info );
}
for( j=i; j < n; j += nb ) {
panel_dev = (j/nb) % num_gpus;
i_local = j/(nb*num_gpus)*nb;
ib = min( n-j, nb );
magma_setdevice( panel_dev );
magma_ssetmatrix( rows, ib,
hwork + (j-i)*rows, rows,
dlA(panel_dev, i, i_local), ldda );
}
}
CLEANUP:
// free(NULL) does nothing.
// check that queues and events are non-zero before destroying them, though.
for( dev=0; dev < num_gpus; dev++ ) {
magma_setdevice( dev );
if ( stream[dev][0] ) { magma_queue_destroy( stream[dev][0] ); }
if ( stream[dev][1] ) { magma_queue_destroy( stream[dev][1] ); }
if ( panel_event[dev] ) { magma_event_destroy( panel_event[dev] ); }
magma_free( dwork[dev] );
}
magma_free_pinned( hwork );
magma_setdevice( cdevice );
magmablasSetKernelStream( cqueue );
return *info;
} /* magma_sgeqrf2_mgpu */
extern "C" magma_err_t
magma_sgeqrs_gpu(magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, size_t dA_offset, magma_int_t ldda,
float *tau, magmaFloat_ptr dT, size_t dT_offset,
magmaFloat_ptr dB, size_t dB_offset, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info, magma_queue_t queue)
{
/* -- clMagma (version 0.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2014
Purpose
=======
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by SGEQRF_GPU.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. M >= N >= 0.
NRHS (input) INTEGER
The number of columns of the matrix C. NRHS >= 0.
A (input) REAL array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
SGEQRF_GPU in the first n columns of its array argument A.
LDDA (input) INTEGER
The leading dimension of the array A, LDDA >= M.
TAU (input) REAL array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_SGEQRF_GPU.
DB (input/output) REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
DT (input) REAL array that is the output (the 6th argument)
of magma_sgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
inverses for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
LDDB (input) INTEGER
The leading dimension of the array DB. LDDB >= M.
HWORK (workspace/output) REAL array, dimension (LWORK)
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The dimension of the array WORK, LWORK >= max(1,NRHS).
For optimum performance LWORK >= (M-N+NB)*(NRHS + 2*NB), where
NB is the blocksize given by magma_get_sgeqrf_nb( M ).
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
===================================================================== */
#define a_ref(a_1,a_2) dA, (dA_offset + (a_1) + (a_2)*(ldda))
#define d_ref(a_1) dT, (dT_offset + (lddwork+(a_1))*nb)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magmaFloat_ptr dwork;
magma_int_t i, k, lddwork, rows, ib;
magma_int_t ione = 1;
magma_int_t nb = magma_get_sgeqrf_nb(m);
magma_int_t lwkopt = (m-n+nb)*(nrhs+2*nb);
long int lquery = (lwork == -1);
hwork[0] = MAGMA_S_MAKE( (float)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
/* B := Q' * B */
magma_sormqr_gpu( MagmaLeft, MagmaTrans,
m, nrhs, n,
a_ref(0,0), ldda, tau,
dB, dB_offset, lddb, hwork, lwork, dT, dT_offset, nb, info, queue );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:) */
lddwork= k;
int ldtwork;
size_t dwork_offset = 0;
if (nb < k)
{
dwork = dT;
dwork_offset = dT_offset+2*lddwork*nb;
}
else
{
ldtwork = ( 2*k + ((n+31)/32)*32 )*nb;
magma_smalloc( &dwork, ldtwork );
}
// To do: Why did we have this line originally; seems to be a bug (Stan)?
//dwork = dT;
i = (k-1)/nb * nb;
ib = n-i;
rows = m-i;
if ( nrhs == 1 ) {
blasf77_strsv( MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, hwork, &rows,
hwork+rows*ib, &ione);
} else {
blasf77_strsm( MagmaLeftStr, MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, &nrhs,
&c_one, hwork, &rows,
hwork+rows*ib, &rows);
}
// update the solution vector
magma_ssetmatrix( ib, nrhs, hwork+rows*ib, 0, rows, dwork, dwork_offset+i, lddwork, queue );
// update c
if (nrhs == 1)
magma_sgemv( MagmaNoTrans, i, ib,
c_neg_one, a_ref(0, i), ldda,
dwork, dwork_offset+i, 1,
c_one, dB, dB_offset, 1, queue );
else
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, a_ref(0, i), ldda,
dwork, dwork_offset + i, lddwork,
c_one, dB, dB_offset, lddb, queue );
int start = i-nb;
if (nb < k) {
for (i = start; i >=0; i -= nb) {
ib = min(k-i, nb);
rows = m -i;
if (i + ib < n) {
if (nrhs == 1) {
magma_sgemv( MagmaNoTrans, ib, ib,
c_one, d_ref(i), ib,
dB, dB_offset+i, 1,
c_zero, dwork, dwork_offset+i, 1, queue );
magma_sgemv( MagmaNoTrans, i, ib,
c_neg_one, a_ref(0, i), ldda,
dwork, dwork_offset+i, 1,
c_one, dB, dB_offset, 1, queue );
} else {
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
ib, nrhs, ib,
c_one, d_ref(i), ib,
dB, dB_offset+i, lddb,
c_zero, dwork, dwork_offset+i, lddwork, queue );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, a_ref(0, i), ldda,
dwork, dwork_offset+i, lddwork,
c_one, dB, dB_offset, lddb, queue );
}
}
}
}
magma_scopymatrix( (n), nrhs,
dwork, dwork_offset, lddwork,
dB, dB_offset, lddb, queue );
if (nb >= k)
magma_free(dwork);
magma_queue_sync( queue );
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sormqr
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
float *C, *R, *A, *W, *tau;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { MagmaTrans, MagmaNoTrans };
printf(" M N K side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n");
printf("===============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_sgeqrf_nb( m );
ldc = m;
// A is m x k (left) or n x k (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = mm;
gflops = FLOPS_SORMQR( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
// need at least 2*nb*nb for geqrf
lwork_max = max( max( m*nb, n*nb ), 2*nb*nb );
TESTING_MALLOC_CPU( C, float, ldc*n );
TESTING_MALLOC_CPU( R, float, ldc*n );
TESTING_MALLOC_CPU( A, float, lda*k );
TESTING_MALLOC_CPU( W, float, lwork_max );
TESTING_MALLOC_CPU( tau, float, k );
// C is full, m x n
size = ldc*n;
lapackf77_slarnv( &ione, ISEED, &size, C );
lapackf77_slacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*k;
lapackf77_slarnv( &ione, ISEED, &size, A );
// compute QR factorization to get Householder vectors in A, tau
magma_sgeqrf( mm, k, A, lda, tau, W, lwork_max, &info );
if (info != 0)
printf("magma_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_sormqr( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sormqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_sormqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0)
printf("magma_sormqr (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_S_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("optimal lwork %d > lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_sormqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sormqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_slange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_saxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_slange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SGEQRF4 computes a QR factorization of a REAL M-by-N matrix A:
A = Q * R using multiple GPUs. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
Arguments
---------
@param[in]
num_gpus INTEGER
The number of GPUs to be used for the factorization.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_sgeqrf_nb(M).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *tau,
float *work, magma_int_t lwork,
magma_int_t *info )
{
float *da[MagmaMaxGPUs];
float c_one = MAGMA_S_ONE;
int i, k, ldda;
*info = 0;
int nb = magma_get_sgeqrf_nb(min(m, n));
int lwkopt = n * nb;
work[0] = MAGMA_S_MAKE( (float)lwkopt, 0 );
int lquery = (lwork == -1);
if (num_gpus < 0 || num_gpus > MagmaMaxGPUs) {
*info = -1;
} else if (m < 0) {
*info = -2;
} else if (n < 0) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
} else if (lwork < max(1,n) && ! lquery) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
work[0] = c_one;
return *info;
}
ldda = ((m+31)/32)*32;
magma_int_t n_local[MagmaMaxGPUs];
for (i=0; i < num_gpus; i++) {
n_local[i] = ((n/nb)/num_gpus)*nb;
if (i < (n/nb)%num_gpus)
n_local[i] += nb;
else if (i == (n/nb)%num_gpus)
n_local[i] += n%nb;
magma_setdevice(i);
// TODO on failure, free previously allocated memory
if (MAGMA_SUCCESS != magma_smalloc( &da[i], ldda*n_local[i] )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
if (m > nb && n > nb) {
/* Copy the matrix to the GPUs in 1D block cyclic distribution */
magma_ssetmatrix_1D_col_bcyclic(m, n, A, lda, da, ldda, num_gpus, nb);
/* Factor using the GPU interface */
magma_sgeqrf2_mgpu( num_gpus, m, n, da, ldda, tau, info);
/* Copy the matrix back from the GPUs to the CPU */
magma_sgetmatrix_1D_col_bcyclic(m, n, da, ldda, A, lda, num_gpus, nb);
}
else {
lapackf77_sgeqrf(&m, &n, A, &lda, tau, work, &lwork, info);
}
/* Free the allocated GPU memory */
for (i=0; i < num_gpus; i++) {
magma_setdevice(i);
magma_free( da[i] );
}
return *info;
} /* magma_sgeqrf4 */
extern "C" magma_int_t
magma_sgeqrs3_gpu(magma_int_t m, magma_int_t n, magma_int_t nrhs,
float *dA, magma_int_t ldda,
float *tau, float *dT,
float *dB, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by SGEQRF3_GPU.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. M >= N >= 0.
NRHS (input) INTEGER
The number of columns of the matrix C. NRHS >= 0.
A (input) REAL array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
SGEQRF3_GPU in the first n columns of its array argument A.
LDDA (input) INTEGER
The leading dimension of the array A, LDDA >= M.
TAU (input) REAL array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_SGEQRF_GPU.
DB (input/output) REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
DT (input) REAL array that is the output (the 6th argument)
of magma_sgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
matrices for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
LDDB (input) INTEGER
The leading dimension of the array DB. LDDB >= M.
HWORK (workspace/output) REAL array, dimension (LWORK)
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The dimension of the array WORK, LWORK >= max(1,NRHS).
For optimum performance LWORK >= (M-N+NB)*(NRHS + 2*NB), where
NB is the blocksize given by magma_get_sgeqrf_nb( M ).
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
===================================================================== */
#define a_ref(a_1,a_2) (dA+(a_2)*(ldda) + (a_1))
#define d_ref(a_1) (dT+(lddwork+(a_1))*nb)
float c_one = MAGMA_S_ONE;
magma_int_t k, lddwork;
magma_int_t nb = magma_get_sgeqrf_nb(m);
magma_int_t lwkopt = (m-n+nb)*(nrhs+2*nb);
int lquery = (lwork == -1);
hwork[0] = MAGMA_S_MAKE( (float)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
lddwork= k;
/* B := Q' * B */
magma_sormqr_gpu( MagmaLeft, MagmaTrans,
m, nrhs, n,
a_ref(0,0), ldda, tau,
dB, lddb, hwork, lwork, dT, nb, info );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:)
1. Move the block diagonal submatrices from d_ref to R
2. Solve
3. Restore the data format moving data from R back to d_ref
*/
magmablas_sswapdblk(k, nb, a_ref(0,0), ldda, 1, d_ref(0), nb, 0);
if ( nrhs == 1 ) {
magma_strsv(MagmaUpper, MagmaNoTrans, MagmaNonUnit,
n, a_ref(0,0), ldda, dB, 1);
} else {
magma_strsm(MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
n, nrhs, c_one, a_ref(0,0), ldda, dB, lddb);
}
magmablas_sswapdblk(k, nb, d_ref(0), nb, 0, a_ref(0,0), ldda, 1);
return *info;
}
/**
Purpose
-------
SGEQRF computes a QR factorization of a REAL M-by-N matrix A:
A = Q * R. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
This uses 2 queues to overlap communication and computation.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= max( N*NB, 2*NB*NB ),
where NB can be obtained through magma_get_sgeqrf_nb( M, N ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda,
float *tau,
float *work, magma_int_t lwork,
magma_int_t *info )
{
#define A(i_,j_) (A + (i_) + (j_)*lda)
#ifdef HAVE_clBLAS
#define dA(i_,j_) dA, ((i_) + (j_)*ldda + dA_offset)
#define dT(i_,j_) dT, ((i_) + (j_)*nb + dT_offset)
#define dwork(i_) dwork, ((i_) + dwork_offset)
#else
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
#define dT(i_,j_) (dT + (i_) + (j_)*nb)
#define dwork(i_) (dwork + (i_))
#endif
/* Constants */
const float c_one = MAGMA_S_ONE;
/* Local variables */
magmaFloat_ptr dA, dT, dwork;
magma_int_t i, ib, min_mn, ldda, lddwork, old_i, old_ib;
/* Function Body */
*info = 0;
magma_int_t nb = magma_get_sgeqrf_nb( m, n );
// need 2*nb*nb to store T and upper triangle of V simultaneously
magma_int_t lwkopt = max( n*nb, 2*nb*nb );
work[0] = magma_smake_lwork( lwkopt );
bool lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1, lwkopt) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
min_mn = min( m, n );
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
// largest N for larfb is n-nb (trailing matrix lacks 1st panel)
lddwork = magma_roundup( n, 32 ) - nb;
ldda = magma_roundup( m, 32 );
magma_int_t ngpu = magma_num_gpus();
if ( ngpu > 1 ) {
/* call multiple-GPU interface */
return magma_sgeqrf_m( ngpu, m, n, A, lda, tau, work, lwork, info );
}
// allocate space for dA, dwork, and dT
if (MAGMA_SUCCESS != magma_smalloc( &dA, n*ldda + nb*lddwork + nb*nb )) {
/* alloc failed so call non-GPU-resident version */
return magma_sgeqrf_ooc( m, n, A, lda, tau, work, lwork, info );
}
dwork = dA + n*ldda;
dT = dA + n*ldda + nb*lddwork;
magma_queue_t queues[2];
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
if ( (nb > 1) && (nb < min_mn) ) {
/* Use blocked code initially.
Asynchronously send the matrix to the GPU except the first panel. */
magma_ssetmatrix_async( m, n-nb,
A(0,nb), lda,
dA(0,nb), ldda, queues[0] );
old_i = 0;
old_ib = nb;
for (i = 0; i < min_mn-nb; i += nb) {
ib = min( min_mn-i, nb );
if (i > 0) {
/* get i-th panel from device */
magma_queue_sync( queues[1] );
magma_sgetmatrix_async( m-i, ib,
dA(i,i), ldda,
A(i,i), lda, queues[0] );
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i), ldda, dT(0,0), nb,
dA(old_i, old_i+2*old_ib), ldda, dwork(0), lddwork, queues[1] );
magma_sgetmatrix_async( i, ib,
dA(0,i), ldda,
A(0,i), lda, queues[1] );
magma_queue_sync( queues[0] );
}
magma_int_t rows = m-i;
lapackf77_sgeqrf( &rows, &ib, A(i,i), &lda, tau+i, work, &lwork, info );
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib, A(i,i), &lda, tau+i, work, &ib );
magma_spanel_to_q( MagmaUpper, ib, A(i,i), lda, work+ib*ib );
/* put i-th V matrix onto device */
magma_ssetmatrix_async( rows, ib, A(i,i), lda, dA(i,i), ldda, queues[0] );
/* put T matrix onto device */
magma_queue_sync( queues[1] );
magma_ssetmatrix_async( ib, ib, work, ib, dT(0,0), nb, queues[0] );
magma_queue_sync( queues[0] );
if (i + ib < n) {
if (i+ib < min_mn-nb) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left (look-ahead) */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dT(0,0), nb,
dA(i, i+ib), ldda, dwork(0), lddwork, queues[1] );
magma_sq_to_panel( MagmaUpper, ib, A(i,i), lda, work+ib*ib );
}
else {
/* After last panel, update whole trailing matrix. */
/* Apply H' to A(i:m,i+ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dT(0,0), nb,
dA(i, i+ib), ldda, dwork(0), lddwork, queues[1] );
magma_sq_to_panel( MagmaUpper, ib, A(i,i), lda, work+ib*ib );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
/* Use unblocked code to factor the last or only block. */
if (i < min_mn) {
ib = n-i;
if (i != 0) {
magma_sgetmatrix( m, ib, dA(0,i), ldda, A(0,i), lda, queues[1] );
}
magma_int_t rows = m-i;
lapackf77_sgeqrf( &rows, &ib, A(i,i), &lda, tau+i, work, &lwork, info );
}
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
magma_free( dA );
return *info;
} /* magma_sgeqrf */
/**
Purpose
-------
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by SGEQRF_GPU.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in]
dA REAL array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
SGEQRF_GPU in the first n columns of its array argument A.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in]
tau REAL array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_SGEQRF_GPU.
@param[in,out]
dB REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
dT REAL array that is the output (the 6th argument)
of magma_sgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
inverses for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) REAL array, dimension (LWORK)
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_sgeqrf_nb( M ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgels_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrs_gpu(magma_int_t m, magma_int_t n, magma_int_t nrhs,
float *dA, magma_int_t ldda,
float *tau, float *dT,
float *dB, magma_int_t lddb,
float *hwork, magma_int_t lwork,
magma_int_t *info)
{
#define dA(a_1,a_2) (dA + (a_2)*(ldda) + (a_1))
#define dT(a_1) (dT + (lddwork+(a_1))*nb)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *dwork;
magma_int_t i, k, lddwork, rows, ib;
magma_int_t ione = 1;
magma_int_t nb = magma_get_sgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_S_MAKE( (float)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -9;
else if (lwork < lwkopt && ! lquery)
*info = -11;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
/* B := Q' * B */
magma_sormqr_gpu( MagmaLeft, MagmaTrans,
m, nrhs, n,
dA(0,0), ldda, tau,
dB, lddb, hwork, lwork, dT, nb, info );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:) */
lddwork= k;
if (nb < k)
dwork = dT+2*lddwork*nb;
else
dwork = dT;
// To do: Why did we have this line originally; seems to be a bug (Stan)?
// dwork = dT;
i = (k-1)/nb * nb;
ib = n-i;
rows = m-i;
// TODO: this assumes that, on exit from magma_sormqr_gpu, hwork contains
// the last block of A and B (i.e., C in sormqr). This should be fixed.
// Seems this data should already be on the GPU, so could switch to
// magma_strsm and drop the ssetmatrix.
if ( nrhs == 1 ) {
blasf77_strsv( MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, hwork, &rows,
hwork+rows*ib, &ione);
} else {
blasf77_strsm( MagmaLeftStr, MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, &nrhs,
&c_one, hwork, &rows,
hwork+rows*ib, &rows);
}
// update the solution vector
magma_ssetmatrix( ib, nrhs, hwork+rows*ib, rows, dwork+i, lddwork );
// update c
if (nrhs == 1)
magma_sgemv( MagmaNoTrans, i, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
else
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
int start = i-nb;
if (nb < k) {
for (i = start; i >= 0; i -= nb) {
ib = min(k-i, nb);
rows = m -i;
if (i + ib < n) {
if (nrhs == 1) {
magma_sgemv( MagmaNoTrans, ib, ib,
c_one, dT(i), ib,
dB+i, 1,
c_zero, dwork+i, 1);
magma_sgemv( MagmaNoTrans, i, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
}
else {
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
ib, nrhs, ib,
c_one, dT(i), ib,
dB+i, lddb,
c_zero, dwork+i, lddwork);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
}
}
}
}
magma_scopymatrix( (n), nrhs,
dwork, lddwork,
dB, lddb );
return *info;
}
/**
Purpose
-------
SGEQRF computes a QR factorization of a REAL M-by-N matrix A:
A = Q * R using multiple GPUs. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
Arguments
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_sgeqrf_nb( M, N ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf_m(
magma_int_t ngpu,
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *tau,
float *work, magma_int_t lwork,
magma_int_t *info )
{
float *da[MagmaMaxGPUs];
float c_one = MAGMA_S_ONE;
magma_int_t i, min_mn, ldda;
*info = 0;
magma_int_t nb = magma_get_sgeqrf_nb( m, n );
magma_int_t lwkopt = n * nb;
work[0] = magma_smake_lwork( lwkopt );
bool lquery = (lwork == -1);
if (ngpu < 0 || ngpu > MagmaMaxGPUs) {
*info = -1;
} else if (m < 0) {
*info = -2;
} else if (n < 0) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
} else if (lwork < max(1,n) && ! lquery) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
min_mn = min(m,n);
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
ldda = magma_roundup( m, 32 );
magma_int_t n_local[MagmaMaxGPUs];
for (i=0; i < ngpu; i++) {
n_local[i] = ((n/nb)/ngpu)*nb;
if (i < (n/nb)%ngpu)
n_local[i] += nb;
else if (i == (n/nb)%ngpu)
n_local[i] += n%nb;
magma_setdevice(i);
// TODO on failure, free previously allocated memory
if (MAGMA_SUCCESS != magma_smalloc( &da[i], ldda*n_local[i] )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
if (m > nb && n > nb) {
magma_queue_t queues[MagmaMaxGPUs];
for( magma_int_t dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
magma_queue_create( dev, &queues[dev] );
}
/* Copy the matrix to the GPUs in 1D block cyclic distribution */
magma_ssetmatrix_1D_col_bcyclic(m, n, A, lda, da, ldda, ngpu, nb, queues);
for( magma_int_t dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
magma_queue_sync( queues[dev] );
}
/* Factor using the GPU interface */
magma_sgeqrf2_mgpu( ngpu, m, n, da, ldda, tau, info);
/* Copy the matrix back from the GPUs to the CPU */
magma_sgetmatrix_1D_col_bcyclic(m, n, da, ldda, A, lda, ngpu, nb, queues);
for( magma_int_t dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
magma_queue_sync( queues[dev] );
magma_queue_destroy( queues[dev] );
}
}
else {
lapackf77_sgeqrf(&m, &n, A, &lda, tau, work, &lwork, info);
}
/* Free the allocated GPU memory */
for (i=0; i < ngpu; i++) {
magma_setdevice(i);
magma_free( da[i] );
}
magma_setdevice( orig_dev );
return *info;
} /* magma_sgeqrf4 */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgels
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float gpu_error, cpu_error, error, Anorm, work[1];
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1];
float *d_A, *d_B;
magma_int_t M, N, size, nrhs, lda, ldb, ldda, lddb, min_mn, max_mn, nb, info;
magma_int_t lworkgpu, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" ||b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n");
printf(" M N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) CPU GPU \n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if ( M < N ) {
printf( "%5d %5d %5d skipping because M < N is not yet supported.\n", (int) M, (int) N, (int) nrhs );
continue;
}
min_mn = min(M, N);
max_mn = max(M, N);
lda = M;
ldb = max_mn;
ldda = ((M+31)/32)*32;
lddb = ((max_mn+31)/32)*32;
nb = magma_get_sgeqrf_nb(M);
gflops = (FLOPS_SGEQRF( M, N ) + FLOPS_SGEQRS( M, N, nrhs )) / 1e9;
lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb;
// query for workspace size
lhwork = -1;
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
NULL, &lda, NULL, &ldb, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_S_REAL( tmp[0] );
lhwork = max( lhwork, lworkgpu );
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, lda*N );
TESTING_MALLOC_CPU( h_A2, float, lda*N );
TESTING_MALLOC_CPU( h_B, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_R, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_work, float, lhwork );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, lddb*nrhs );
/* Initialize the matrices */
size = lda*N;
lapackf77_slarnv( &ione, ISEED, &size, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
// make random RHS
size = ldb*nrhs;
lapackf77_slarnv( &ione, ISEED, &size, h_B );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
// make consistent RHS
//size = N*nrhs;
//lapackf77_slarnv( &ione, ISEED, &size, h_X );
//blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
// &c_one, h_A, &lda,
// h_X, &ldb,
// &c_zero, h_B, &ldb );
//lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_ssetmatrix( M, N, h_A, lda, d_A, ldda );
magma_ssetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
gpu_time = magma_wtime();
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda,
d_B, lddb, h_work, lworkgpu, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgels_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute the residual
magma_sgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb );
Anorm = lapackf77_slange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sgels returned error %d: %s.\n",
(int) info, magma_strerror( info ));
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb );
cpu_error = lapackf77_slange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm);
gpu_error = lapackf77_slange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
// error relative to LAPACK
size = M*nrhs;
blasf77_saxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_slange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e",
(int) M, (int) N, (int) nrhs,
cpu_perf, cpu_time, gpu_perf, gpu_time, cpu_error, gpu_error, error );
if ( M == N ) {
printf( " %s\n", (gpu_error < tol && error < tol ? "ok" : "failed"));
status += ! (gpu_error < tol && error < tol);
}
else {
printf( " %s\n", (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_A2 );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( h_work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_sgeqrf3_gpu( magma_int_t m, magma_int_t n,
float *dA, magma_int_t ldda,
float *tau, float *dT,
magma_int_t *info )
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
SGEQRF3 computes a QR factorization of a real M-by-N matrix A:
A = Q * R.
This version stores the triangular dT matrices used in
the block QR factorization so that they can be applied directly (i.e.,
without being recomputed) later. As a result, the application
of Q is much faster. Also, the upper triangular matrices for V have 0s
in them and the corresponding parts of the upper triangular R are
stored separately in dT.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
TAU (output) REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
dT (workspace/output) REAL array on the GPU,
dimension (2*MIN(M, N) + (N+31)/32*32 )*NB,
where NB can be obtained through magma_get_sgeqrf_nb(M).
It starts with MIN(M,N)*NB block that store the triangular T
matrices, followed by the MIN(M,N)*NB block of the diagonal
matrices for the R matrix. The rest of the array is used as workspace.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define a_ref(a_1,a_2) (dA+(a_2)*(ldda) + (a_1))
#define t_ref(a_1) (dT+(a_1)*nb)
#define d_ref(a_1) (dT+(minmn+(a_1))*nb)
#define dd_ref(a_1) (dT+(2*minmn+(a_1))*nb)
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
magma_int_t i, k, minmn, old_i, old_ib, rows, cols;
magma_int_t ib, nb;
magma_int_t ldwork, lddwork, lwork, lhwork;
float *work, *ut;
/* check arguments */
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = minmn = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
lwork = (m + n + nb)*nb;
lhwork = lwork - m*nb;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, lwork )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
ut = hwork+nb*(n);
memset( ut, 0, nb*nb*sizeof(float));
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
ldwork = m;
lddwork= n;
if ( (nb > 1) && (nb < k) ) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nb; i += nb) {
ib = min(k-i, nb);
rows = m -i;
magma_sgetmatrix_async( rows, ib,
a_ref(i,i), ldda,
work_ref(i), ldwork, stream[1] );
if (i>0){
/* Apply H' to A(i:m,i+2*ib:n) from the left */
cols = n-old_i-2*old_ib;
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
m-old_i, cols, old_ib,
a_ref(old_i, old_i ), ldda, t_ref(old_i), nb,
a_ref(old_i, old_i+2*old_ib), ldda, dd_ref(0), lddwork);
/* store the diagonal */
magma_ssetmatrix_async( old_ib, old_ib,
ut, old_ib,
d_ref(old_i), old_ib, stream[0] );
}
magma_queue_sync( stream[1] );
lapackf77_sgeqrf(&rows, &ib, work_ref(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work_ref(i), &ldwork, tau+i, hwork, &ib);
/* Put 0s in the upper triangular part of a panel (and 1s on the
diagonal); copy the upper triangular in ut. */
magma_queue_sync( stream[0] );
ssplit_diag_block3(ib, work_ref(i), ldwork, ut);
magma_ssetmatrix( rows, ib, work_ref(i), ldwork, a_ref(i,i), ldda );
if (i + ib < n) {
/* Send the triangular factor T to the GPU */
magma_ssetmatrix( ib, ib, hwork, ib, t_ref(i), nb );
if (i+nb < k-nb){
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
a_ref(i, i ), ldda, t_ref(i), nb,
a_ref(i, i+ib), ldda, dd_ref(0), lddwork);
}
else {
cols = n-i-ib;
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, cols, ib,
a_ref(i, i ), ldda, t_ref(i), nb,
a_ref(i, i+ib), ldda, dd_ref(0), lddwork);
/* Fix the diagonal block */
magma_ssetmatrix( ib, ib, ut, ib, d_ref(i), ib );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_sgetmatrix( rows, ib, a_ref(i, i), ldda, work, rows );
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_ssetmatrix( rows, ib, work, rows, a_ref(i, i), ldda );
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free_pinned( work );
return *info;
/* End of MAGMA_SGEQRF */
} /* magma_sgeqrf */
/**
Purpose
-------
DSGEQRSV solves the least squares problem
min || A*X - B ||,
where A is an M-by-N matrix and X and B are M-by-NRHS matrices.
DSGEQRSV first attempts to factorize the matrix in real SINGLE PRECISION
and use this factorization within an iterative refinement procedure
to produce a solution with real DOUBLE PRECISION norm-wise backward error
quality (see below). If the approach fails the method switches to a
real DOUBLE PRECISION factorization and solve.
The iterative refinement is not going to be a winning strategy if
the ratio real SINGLE PRECISION performance over real DOUBLE PRECISION
performance is too small. A reasonable strategy should take the
number of right-hand sides and the size of the matrix into account.
This might be done with a call to ILAENV in the future. Up to now, we
always try iterative refinement.
The iterative refinement process is stopped if
ITER > ITERMAX
or for all the RHS we have:
RNRM < SQRT(N)*XNRM*ANRM*EPS*BWDMAX
where
o ITER is the number of the current iteration in the iterative
refinement process
o RNRM is the infinity-norm of the residual
o XNRM is the infinity-norm of the solution
o ANRM is the infinity-operator-norm of the matrix A
o EPS is the machine epsilon returned by DLAMCH('Epsilon')
The value ITERMAX and BWDMAX are fixed to 30 and 1.0D+00 respectively.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in,out]
dA DOUBLE PRECISION array on the GPU, dimension (LDDA,N)
On entry, the M-by-N coefficient matrix A.
On exit, if iterative refinement has been successfully used
(info.EQ.0 and ITER.GE.0, see description below), A is
unchanged. If double precision factorization has been used
(info.EQ.0 and ITER.LT.0, see description below), then the
array dA contains the QR factorization of A as returned by
function DGEQRF_GPU.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
@param[in,out]
dB DOUBLE PRECISION array on the GPU, dimension (LDDB,NRHS)
The M-by-NRHS right hand side matrix B.
May be overwritten (e.g., if refinement fails).
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= max(1,M).
@param[out]
dX DOUBLE PRECISION array on the GPU, dimension (LDDX,NRHS)
If info = 0, the N-by-NRHS solution matrix X.
@param[in]
lddx INTEGER
The leading dimension of the array dX. LDDX >= max(1,N).
@param[out]
iter INTEGER
- < 0: iterative refinement has failed, double precision
factorization has been performed
+ -1 : the routine fell back to full precision for
implementation- or machine-specific reasons
+ -2 : narrowing the precision induced an overflow,
the routine fell back to full precision
+ -3 : failure of SGEQRF
+ -31: stop the iterative refinement after the 30th iteration
- > 0: iterative refinement has been successfully used.
Returns the number of iterations
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if info = -i, the i-th argument had an illegal value
@ingroup magma_dgels_driver
********************************************************************/
extern "C" magma_int_t
magma_dsgeqrsv_gpu(
magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaDouble_ptr dA, magma_int_t ldda,
magmaDouble_ptr dB, magma_int_t lddb,
magmaDouble_ptr dX, magma_int_t lddx,
magma_int_t *iter,
magma_int_t *info)
{
#define dB(i,j) (dB + (i) + (j)*lddb)
#define dX(i,j) (dX + (i) + (j)*lddx)
#define dR(i,j) (dR + (i) + (j)*lddr)
#define dSX(i,j) (dSX + (i) + (j)*lddsx)
double c_neg_one = MAGMA_D_NEG_ONE;
double c_one = MAGMA_D_ONE;
magma_int_t ione = 1;
double *hworkd;
float *hworks;
double *tau;
float *stau;
magmaDouble_ptr dworkd;
magmaFloat_ptr dworks;
magmaDouble_ptr dR, dT;
magmaFloat_ptr dSA, dSX, dST;
double Xnrmv, Rnrmv;
double Anrm, Xnrm, Rnrm, cte, eps;
magma_int_t i, j, iiter, lddsa, lddsx, lddr, nb, lhwork, minmn, size, ldworkd;
/* Check arguments */
*iter = 0;
*info = 0;
if ( m < 0 )
*info = -1;
else if ( n < 0 || n > m )
*info = -2;
else if ( nrhs < 0 )
*info = -3;
else if ( ldda < max(1,m))
*info = -5;
else if ( lddb < max(1,m))
*info = -7;
else if ( lddx < max(1,n))
*info = -9;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if ( m == 0 || n == 0 || nrhs == 0 )
return *info;
nb = magma_get_sgeqrf_nb(m);
minmn= min(m, n);
/* dSX contains both B and X, so must be max(m or lddb,n). */
lddsa = ldda;
lddsx = max(lddb,n);
lddr = lddb;
/*
* Allocate temporary buffers
*/
/* dworks(dSA + dSX + dST) */
size = lddsa*n + lddsx*nrhs + ( 2*minmn + ((n+31)/32)*32 )*nb;
if (MAGMA_SUCCESS != magma_smalloc( &dworks, size )) {
fprintf(stderr, "Allocation of dworks failed (%d)\n", (int) size);
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dSA = dworks;
dSX = dSA + lddsa*n;
dST = dSX + lddsx*nrhs;
/* dworkd(dR) = lddr*nrhs */
ldworkd = lddr*nrhs;
if (MAGMA_SUCCESS != magma_dmalloc( &dworkd, ldworkd )) {
magma_free( dworks );
fprintf(stderr, "Allocation of dworkd failed\n");
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dR = dworkd;
/* hworks(workspace for cgeqrs + stau) = min(m,n) + lhworks */
lhwork = (m - n + nb)*(nrhs + nb) + nrhs*nb;
size = lhwork + minmn;
magma_smalloc_cpu( &hworks, size );
if ( hworks == NULL ) {
magma_free( dworks );
magma_free( dworkd );
fprintf(stderr, "Allocation of hworks failed\n");
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
stau = hworks + lhwork;
eps = lapackf77_dlamch("Epsilon");
Anrm = magmablas_dlange(MagmaInfNorm, m, n, dA, ldda, (double*)dworkd );
cte = Anrm * eps * pow((double)n, 0.5) * BWDMAX;
/*
* Convert to single precision
*/
magmablas_dlag2s( m, nrhs, dB, lddb, dSX, lddsx, info );
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
magmablas_dlag2s( m, n, dA, ldda, dSA, lddsa, info );
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
// factor dSA in single precision
magma_sgeqrf_gpu( m, n, dSA, lddsa, stau, dST, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// solve dSA*dSX = dB in single precision
magma_sgeqrs_gpu( m, n, nrhs, dSA, lddsa, stau, dST, dSX, lddsx, hworks, lhwork, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// residual dR = dB - dA*dX in double precision
magmablas_slag2d( n, nrhs, dSX, lddsx, dX, lddx, info );
magmablas_dlacpy( MagmaUpperLower, m, nrhs, dB, lddb, dR, lddr );
if ( nrhs == 1 ) {
magma_dgemv( MagmaNoTrans, m, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( MagmaNoTrans, MagmaNoTrans, m, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
// TODO: use MAGMA_D_ABS( dX(i,j) ) instead of dlange?
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( m, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto REFINEMENT;
}
}
*iter = 0;
/* Free workspaces */
magma_free( dworks );
magma_free( dworkd );
magma_free_cpu( hworks );
return *info;
REFINEMENT:
/* TODO: this iterative refinement algorithm works only for compatibile
* systems (B in colspan of A).
* See Matrix Computations (3rd ed) p. 267 for correct algorithm. */
for( iiter=1; iiter < ITERMAX; ) {
*info = 0;
// convert residual dR to single precision dSX
magmablas_dlag2s( m, nrhs, dR, lddr, dSX, lddsx, info );
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
// solve dSA*dSX = R in single precision
magma_sgeqrs_gpu( m, n, nrhs, dSA, lddsa, stau, dST, dSX, lddsx, hworks, lhwork, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// Add correction and setup residual
// dX += dSX [including conversion] --and--
// dR[1:n] = dB[1:n] (only n rows, not whole m rows! -- useless if m > n)
for( j=0; j < nrhs; j++ ) {
magmablas_dsaxpycp( n, dSX(0,j), dX(0,j), dB(0,j), dR(0,j) );
}
// dR = dB (whole m rows)
magmablas_dlacpy( MagmaUpperLower, m, nrhs, dB, lddb, dR, lddr );
// residual dR = dB - dA*dX in double precision
if ( nrhs == 1 ) {
magma_dgemv( MagmaNoTrans, m, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( MagmaNoTrans, MagmaNoTrans, m, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
/* Check whether the nrhs normwise backward errors satisfy the
* stopping criterion. If yes, set ITER=IITER > 0 and return. */
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( m, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto L20;
}
}
/* If we are here, the nrhs normwise backward errors satisfy
* the stopping criterion, we are good to exit. */
*iter = iiter;
/* Free workspaces */
magma_free( dworks );
magma_free( dworkd );
magma_free_cpu( hworks );
return *info;
L20:
iiter++;
}
/* If we are at this place of the code, this is because we have
* performed ITER=ITERMAX iterations and never satisified the
* stopping criterion. Set up the ITER flag accordingly and follow
* up on double precision routine. */
*iter = -ITERMAX - 1;
FALLBACK:
/* Single-precision iterative refinement failed to converge to a
* satisfactory solution, so we resort to double precision. */
magma_free( dworks );
magma_free_cpu( hworks );
/*
* Allocate temporary buffers
*/
/* dworkd = dT for dgeqrf */
nb = magma_get_dgeqrf_nb( m );
size = (2*min(m, n) + (n+31)/32*32 )*nb;
if ( size > ldworkd ) {
magma_free( dworkd );
if (MAGMA_SUCCESS != magma_dmalloc( &dworkd, size )) {
fprintf(stderr, "Allocation of dworkd2 failed\n");
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
dT = dworkd;
/* hworkd(dtau + workspace for dgeqrs) = min(m,n) + lhwork */
size = lhwork + minmn;
magma_dmalloc_cpu( &hworkd, size );
if ( hworkd == NULL ) {
magma_free( dworkd );
fprintf(stderr, "Allocation of hworkd2 failed\n");
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
tau = hworkd + lhwork;
magma_dgeqrf_gpu( m, n, dA, ldda, tau, dT, info );
if (*info == 0) {
// if m > n, then dB won't fit in dX, so solve with dB and copy n rows to dX
magma_dgeqrs_gpu( m, n, nrhs, dA, ldda, tau, dT, dB, lddb, hworkd, lhwork, info );
magmablas_dlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dX, lddx );
}
magma_free( dworkd );
magma_free_cpu( hworkd );
return *info;
}
extern "C" magma_int_t
magma_sgeqrf_msub(
magma_int_t num_subs, magma_int_t num_gpus,
magma_int_t m, magma_int_t n,
magmaFloat_ptr *dlA, magma_int_t ldda,
float *tau,
magma_queue_t *queues,
magma_int_t *info)
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
SGEQRF2_MGPU computes a QR factorization of a real M-by-N matrix A:
A = Q * R. This is a GPU interface of the routine.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
TAU (output) REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dlA(gpu,a_1,a_2) dlA[gpu], ((a_2)*(ldda) + (a_1))
#define dlA_offset(a_1, a_2) ((a_2)*(ldda) + (a_1))
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
#define hwrk(a_1) ( local_work + (a_1))
#define lhwrk ( local_work + (nb)*(m))
magmaFloat_ptr dwork[MagmaMaxGPUs], panel[MagmaMaxGPUs];
size_t panel_offset[MagmaMaxGPUs];
float *local_work = NULL;
magma_int_t i, j, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
int panel_id = -1, i_local, n_local[MagmaMaxGPUs * MagmaMaxSubs], la_id, displacement,
tot_subs = num_gpus * num_subs;
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
displacement = n * nb;
lwork = (m+n+64) * nb;
lhwork = lwork - (m)*nb;
for (i=0; i<num_gpus; i++) {
if (MAGMA_SUCCESS != magma_smalloc( &(dwork[i]), (n + ldda)*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
/* Set the number of local n for each GPU */
for (i=0; i<tot_subs; i++) {
n_local[i] = ((n/nb)/tot_subs)*nb;
if (i < (n/nb)%tot_subs)
n_local[i] += nb;
else if (i == (n/nb)%tot_subs)
n_local[i] += n%nb;
}
#ifdef USE_PINNED_CLMEMORY
cl_mem buffer = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(float)*lwork, NULL, NULL);
for (j=0; j<num_gpus; j++) {
local_work = (float*)clEnqueueMapBuffer(queues[2*j], buffer, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0,
sizeof(float)*lwork, 0, NULL, NULL, NULL);
}
#else
if (MAGMA_SUCCESS != magma_smalloc_cpu( (&local_work), lwork )) {
*info = -9;
for (i=0; i<num_gpus; i++) {
magma_free( dwork[i] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
#endif
nbmin = 2;
nx = nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb) {
/* Set the GPU number that holds the current panel */
panel_id = (i/nb)%tot_subs;
/* Set the local index where the current panel is */
i_local = i/(nb*tot_subs)*nb;
ib = min(k-i, nb);
rows = m -i;
/* Send current panel to the CPU */
magma_queue_sync(queues[2*(panel_id%num_gpus)]);
magma_sgetmatrix_async( rows, ib,
dlA(panel_id, i, i_local), ldda,
hwrk(i), ldwork,
queues[2*(panel_id%num_gpus)+1], NULL );
if (i > 0) {
/* Apply H' to A(i:m,i+2*ib:n) from the left; this is the look-ahead
application to the trailing matrix */
la_id = panel_id;
/* only the GPU that has next panel is done look-ahead */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n_local[la_id]-i_local-old_ib, old_ib,
panel[la_id%num_gpus], panel_offset[la_id%num_gpus], ldda,
dwork[la_id%num_gpus], 0, lddwork,
dlA(la_id, old_i, i_local+old_ib), ldda,
dwork[la_id%num_gpus], old_ib, lddwork,
queues[2*(la_id%num_gpus)]);
la_id = ((i-nb)/nb)%tot_subs;
magma_ssetmatrix_async( old_ib, old_ib,
hwrk(old_i), ldwork,
panel[la_id%num_gpus], panel_offset[la_id%num_gpus], ldda,
queues[2*(la_id%num_gpus)], NULL );
}
magma_queue_sync( queues[2*(panel_id%num_gpus)+1] );
lapackf77_sgeqrf(&rows, &ib, hwrk(i), &ldwork, tau+i, lhwrk, &lhwork, info);
// Form the triangular factor of the block reflector
// H = H(i) H(i+1) . . . H(i+ib-1)
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
hwrk(i), &ldwork, tau+i, lhwrk, &ib);
spanel_to_q( MagmaUpper, ib, hwrk(i), ldwork, lhwrk+ib*ib );
// Send the current panel back to the GPUs
// Has to be done with asynchronous copies
for (j=0; j<num_gpus; j++) {
if (j == panel_id%num_gpus){
panel[j] = dlA(panel_id, i, i_local);
panel_offset[j] = dlA_offset(i, i_local);
} else {
panel[j] = dwork[j];
panel_offset[j] = displacement;
}
magma_queue_sync( queues[2*j] );
magma_ssetmatrix_async( rows, ib,
hwrk(i), ldwork,
panel[j], panel_offset[j], ldda,
queues[2*j+1], NULL );
/* Send the T matrix to the GPU.
Has to be done with asynchronous copies */
magma_ssetmatrix_async( ib, ib, lhwrk, ib,
dwork[j], 0, lddwork,
queues[2*j+1], NULL );
}
for(j=0; j<num_gpus; j++) {
magma_queue_sync( queues[2*j+1] );
}
if (i + ib < n) {
if (i+nb < k-nx) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left;
This is update for the next panel; part of the look-ahead */
la_id = (panel_id+1)%tot_subs;
int i_loc = (i+nb)/(nb*tot_subs)*nb;
for (j=0; j<tot_subs; j++) {
if (j == la_id)
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
panel[j%num_gpus], panel_offset[j%num_gpus], ldda,
dwork[j%num_gpus], 0, lddwork,
dlA(j, i, i_loc), ldda,
dwork[j%num_gpus], ib, lddwork,
queues[2*(j%num_gpus)]);
else if (j <= panel_id)
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n_local[j]-i_local-ib, ib,
panel[j%num_gpus], panel_offset[j%num_gpus], ldda,
dwork[j%num_gpus], 0, lddwork,
dlA(j, i, i_local+ib), ldda,
dwork[j%num_gpus], ib, lddwork,
queues[2*(j%num_gpus)]);
else
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n_local[j]-i_local, ib,
panel[j%num_gpus], panel_offset[j%num_gpus], ldda,
dwork[j%num_gpus], 0, lddwork,
dlA(j, i, i_local), ldda,
dwork[j%num_gpus], ib, lddwork,
queues[2*(j%num_gpus)]);
}
/* Restore the panel */
sq_to_panel( MagmaUpper, ib, hwrk(i), ldwork, lhwrk+ib*ib );
} else {
/* do the entire update as we exit and there would be no lookahead */
la_id = (panel_id+1)%tot_subs;
int i_loc = (i+nb)/(nb*tot_subs)*nb;
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n_local[la_id]-i_loc, ib,
panel[la_id%num_gpus], panel_offset[la_id%num_gpus], ldda,
dwork[la_id%num_gpus], 0, lddwork,
dlA(la_id, i, i_loc), ldda,
dwork[la_id%num_gpus], ib, lddwork,
queues[2*(la_id%num_gpus)]);
/* Restore the panel */
sq_to_panel( MagmaUpper, ib, hwrk(i), ldwork, lhwrk+ib*ib );
magma_ssetmatrix( ib, ib,
hwrk(i), ldwork,
dlA(panel_id, i, i_local), ldda,
queues[2*(panel_id%num_gpus)]);
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
for (j=0; j<num_gpus; j++) {
magma_free( dwork[j] );
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
lhwork = lwork - rows*ib;
panel_id = (panel_id+1)%tot_subs;
int i_loc = (i)/(nb*tot_subs)*nb;
magma_sgetmatrix( rows, ib,
dlA(panel_id, i, i_loc), ldda,
lhwrk, rows,
queues[2*(panel_id%num_gpus)]);
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, lhwrk, &rows, tau+i, lhwrk+ib*rows, &lhwork, info);
magma_ssetmatrix( rows, ib,
lhwrk, rows,
dlA(panel_id, i, i_loc), ldda,
queues[2*(panel_id%num_gpus)]);
}
#ifdef USE_PINNED_CLMEMORY
#else
magma_free_cpu( local_work );
#endif
return *info;
} /* magma_sgeqrf_msub */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrf
*/
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float error, work[1];
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_R, *tau, *h_work, tmp[1];
magma_int_t M, N, n2, lda, lwork, info, min_mn, nb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4];
/* Initialize */
magma_queue_t queue[2];
magma_device_t devices[MagmaMaxGPUs];
int num = 0;
magma_err_t err;
magma_init();
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t status = 0;
float tol, eps = lapackf77_slamch("E");
tol = opts.tolerance * eps;
opts.lapack |= ( opts.check == 2 ); // check (-c2) implies lapack (-l)
err = magma_get_devices( devices, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
// Create two queues on device opts.device
err = magma_queue_create( devices[opts.device], &queue[0] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( devices[opts.device], &queue[1] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
printf("ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 1 ) {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R-Q'A||_1 / (M*||A||_1) ||I-Q'Q||_1 / M\n");
printf("===============================================================================================\n");
} else {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n");
printf("=======================================================================\n");
}
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
nb = magma_get_sgeqrf_nb(M);
gflops = FLOPS_SGEQRF( M, N ) / 1e9;
lwork = -1;
lapackf77_sgeqrf(&M, &N, h_A, &M, tau, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lwork = max( lwork, max( N*nb, 2*nb*nb ));
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_CPU( h_work, float, lwork );
/* Initialize the matrix */
for ( int j=0; j<4; j++ ) ISEED2[j] = ISEED[j]; // saving seeds
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgeqrf(M, N, h_R, lda, tau, h_work, lwork, &info, queue);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
float *tau;
TESTING_MALLOC_CPU( tau, float, min_mn );
cpu_time = magma_wtime();
lapackf77_sgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
TESTING_FREE_CPU( tau );
}
if ( opts.check == 1 ) {
/* =====================================================================
Check the result
=================================================================== */
magma_int_t lwork = n2+N;
float *h_W1, *h_W2, *h_W3;
float *h_RW, results[2];
TESTING_MALLOC_CPU( h_W1, float, n2 ); // Q
TESTING_MALLOC_CPU( h_W2, float, n2 ); // R
TESTING_MALLOC_CPU( h_W3, float, lwork ); // WORK
TESTING_MALLOC_CPU( h_RW, float, M ); // RWORK
lapackf77_slarnv( &ione, ISEED2, &n2, h_A );
lapackf77_sqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork,
h_RW, results );
results[0] *= eps;
results[1] *= eps;
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
}
status |= ! (results[0] < tol);
TESTING_FREE_CPU( h_W1 );
TESTING_FREE_CPU( h_W2 );
TESTING_FREE_CPU( h_W3 );
TESTING_FREE_CPU( h_RW );
} else if ( opts.check == 2 ) {
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = lapackf77_slange("f", &M, &N, h_A, &lda, work);
blasf77_saxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_slange("f", &M, &N, h_R, &lda, work) / error;
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e",
(int) M, (int) N, gpu_perf, gpu_time, error );
}
printf("%s\n", (error < tol ? "" : " failed"));
status |= ! (error < tol);
}
else {
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, gpu_time);
}
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
magma_queue_destroy( queue[0] );
magma_queue_destroy( queue[1] );
magma_finalize();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const float d_neg_one = MAGMA_D_NEG_ONE;
const float d_one = MAGMA_D_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
const float c_one = MAGMA_S_ONE;
const float c_zero = MAGMA_S_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float Anorm, error=0, error2=0;
float *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloat_ptr d_A, dT;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb, size;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
// version 3 can do either check
if (opts.check == 1 && opts.version == 1) {
opts.check = 2;
printf( "%% version 1 requires check 2 (solve A*x=b)\n" );
}
if (opts.check == 2 && opts.version == 2) {
opts.check = 1;
printf( "%% version 2 requires check 1 (R - Q^H*A)\n" );
}
printf( "%% version %d\n", (int) opts.version );
if ( opts.check == 1 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |R - Q^H*A| |I - Q^H*Q|\n");
printf("%%==============================================================================\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |b - A*x|\n");
printf("%%===============================================================\n");
}
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min( M, N );
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
nb = magma_get_sgeqrf_nb( M, N );
gflops = FLOPS_SGEQRF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgeqrf( &M, &N, NULL, &M, NULL, tmp, &lwork, &info );
lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_work, float, lwork );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
if ( opts.version == 1 || opts.version == 3 ) {
size = (2*min(M, N) + magma_roundup( N, 32 ) )*nb;
TESTING_MALLOC_DEV( dT, float, size );
magmablas_slaset( MagmaFull, size, 1, c_zero, c_zero, dT, size );
}
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
magma_ssetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
nb = magma_get_sgeqrf_nb( M, N );
gpu_time = magma_wtime();
if ( opts.version == 1 ) {
// stores dT, V blocks have zeros, R blocks inverted & stored in dT
magma_sgeqrf_gpu( M, N, d_A, ldda, tau, dT, &info );
}
else if ( opts.version == 2 ) {
// LAPACK complaint arguments
magma_sgeqrf2_gpu( M, N, d_A, ldda, tau, &info );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// stores dT, V blocks have zeros, R blocks stored in dT
magma_sgeqrf3_gpu( M, N, d_A, ldda, tau, dT, &info );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( opts.check == 1 && (opts.version == 2 || opts.version == 3) ) {
if ( opts.version == 3 ) {
// copy diagonal blocks of R back to A
for( int i=0; i < min_mn-nb; i += nb ) {
magma_int_t ib = min( min_mn-i, nb );
magmablas_slacpy( MagmaUpper, ib, ib, &dT[min_mn*nb + i*nb], nb, &d_A[ i + i*ldda ], ldda );
}
}
/* =====================================================================
Check the result, following zqrt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
Only for version 2, which has LAPACK complaint output.
Or for version 3, after restoring diagonal blocks of A above.
=================================================================== */
magma_sgetmatrix( M, N, d_A, ldda, h_R, lda );
magma_int_t ldq = M;
magma_int_t ldr = min_mn;
float *Q, *R;
float *work;
TESTING_MALLOC_CPU( Q, float, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( R, float, ldr*N ); // K by N
TESTING_MALLOC_CPU( work, float, min_mn );
// generate M by K matrix Q, where K = min(M,N)
lapackf77_slacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_sorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_slaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_slacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_sgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_slange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_slange( "1", &min_mn, &N, R, &ldr, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set R = I (K by K identity), then R = I - Q^H*Q
// error = || I - Q^H*Q || / N
lapackf77_slaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_ssyrk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
error2 = safe_lapackf77_slansy( "1", "Upper", &min_mn, R, &ldr, work );
if ( N > 0 )
error2 /= N;
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( R ); R = NULL;
TESTING_FREE_CPU( work ); work = NULL;
}
else if ( opts.check == 2 && M >= N && (opts.version == 1 || opts.version == 3) ) {
/* =====================================================================
Check the result by solving consistent linear system, A*x = b.
Only for versions 1 & 3 with M >= N.
=================================================================== */
magma_int_t lwork2;
float *x, *b, *hwork;
magmaFloat_ptr d_B;
// initialize RHS, b = A*random
TESTING_MALLOC_CPU( x, float, N );
TESTING_MALLOC_CPU( b, float, M );
lapackf77_slarnv( &ione, ISEED, &N, x );
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione );
// copy to GPU
TESTING_MALLOC_DEV( d_B, float, M );
magma_ssetvector( M, b, 1, d_B, 1 );
if ( opts.version == 1 ) {
// allocate hwork
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( hwork );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// allocate hwork
magma_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs3 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( hwork );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
magma_sgetvector( N, d_B, 1, x, 1 );
// compute r = Ax - b, saved in b
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione );
// compute residual |Ax - b| / (max(m,n)*|A|*|x|)
float norm_x, norm_A, norm_r, work[1];
norm_A = lapackf77_slange( "F", &M, &N, h_A, &lda, work );
norm_r = lapackf77_slange( "F", &M, &ione, b, &M, work );
norm_x = lapackf77_slange( "F", &N, &ione, x, &N, work );
TESTING_FREE_CPU( x );
TESTING_FREE_CPU( b );
TESTING_FREE_DEV( d_B );
error = norm_r / (max(M,N) * norm_A * norm_x);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgeqrf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* =====================================================================
Print performance and error.
=================================================================== */
printf("%5d %5d ", (int) M, (int) N );
if ( opts.lapack ) {
printf( "%7.2f (%7.2f)", cpu_perf, cpu_time );
}
else {
printf(" --- ( --- )" );
}
printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time );
if ( opts.check == 1 ) {
bool okay = (error < tol && error2 < tol);
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else if ( opts.check == 2 ) {
if ( M >= N ) {
bool okay = (error < tol);
status += ! okay;
printf( "%10.2e %s\n", error, (okay ? "ok" : "failed") );
}
else {
printf( "(error check only for M >= N)\n" );
}
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
if ( opts.version == 1 || opts.version == 3 ) {
TESTING_FREE_DEV( dT );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SGEQRF computes a QR factorization of a real M-by-N matrix A:
A = Q * R.
This version has LAPACK-complaint arguments.
If the current stream is NULL, this version replaces it with a new
stream to overlap computation with communication.
Other versions (magma_sgeqrf_gpu and magma_sgeqrf3_gpu) store the
intermediate T matrices.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[out]
tau REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgeqrf2_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
float *tau,
magma_int_t *info )
{
#define dA(a_1,a_2) ( dA+(a_2)*(ldda) + (a_1))
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
magmaFloat_ptr dwork;
float *work;
magma_int_t i, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
/* Function Body */
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
lwork = (m+n) * nb;
lhwork = lwork - (m)*nb;
if (MAGMA_SUCCESS != magma_smalloc( &dwork, n*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, lwork )) {
magma_free( dwork );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* Define user stream if current stream is NULL */
magma_queue_t stream[2];
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_queue_create( &stream[0] );
if (orig_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else {
stream[1] = orig_stream;
}
nbmin = 2;
nx = nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb) {
ib = min(k-i, nb);
rows = m -i;
/* download i-th panel */
magma_queue_sync( stream[1] );
magma_sgetmatrix_async( rows, ib,
dA(i,i), ldda,
work_ref(i), ldwork, stream[0] );
if (i > 0) {
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i ), ldda, dwork, lddwork,
dA(old_i, old_i+2*old_ib), ldda, dwork+old_ib, lddwork);
magma_ssetmatrix_async( old_ib, old_ib,
work_ref(old_i), ldwork,
dA(old_i, old_i), ldda, stream[1] );
}
magma_queue_sync( stream[0] );
lapackf77_sgeqrf(&rows, &ib, work_ref(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work_ref(i), &ldwork, tau+i, hwork, &ib);
spanel_to_q( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
/* download the i-th V matrix */
magma_ssetmatrix_async( rows, ib, work_ref(i), ldwork, dA(i,i), ldda, stream[0] );
/* download the T matrix */
magma_queue_sync( stream[1] );
magma_ssetmatrix_async( ib, ib, hwork, ib, dwork, lddwork, stream[0] );
magma_queue_sync( stream[0] );
if (i + ib < n) {
if (i+nb < k-nx) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dwork, lddwork,
dA(i, i+ib), ldda, dwork+ib, lddwork);
sq_to_panel( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
}
else {
magma_slarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dwork, lddwork,
dA(i, i+ib), ldda, dwork+ib, lddwork);
sq_to_panel( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
magma_ssetmatrix_async( ib, ib,
work_ref(i), ldwork,
dA(i,i), ldda, stream[1] );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
magma_free( dwork );
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_sgetmatrix_async( rows, ib, dA(i, i), ldda, work, rows, stream[1] );
magma_queue_sync( stream[1] );
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_ssetmatrix_async( rows, ib, work, rows, dA(i, i), ldda, stream[1] );
}
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
if (orig_stream == NULL) {
magma_queue_destroy( stream[1] );
}
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_sgeqrf2_gpu */
extern "C" magma_int_t
magma_sgeqrf2_mgpu( magma_int_t num_gpus, magma_int_t m, magma_int_t n,
float **dlA, magma_int_t ldda,
float *tau,
magma_int_t *info )
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
SGEQRF2_MGPU computes a QR factorization of a real M-by-N matrix A:
A = Q * R. This is a GPU interface of the routine.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) REAL array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
TAU (output) REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dlA(gpu,a_1,a_2) ( dlA[gpu]+(a_2)*(ldda) + (a_1))
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
#define hwrk_ref(a_1) ( local_work + (a_1))
#define lhwrk ( local_work + (nb)*(m))
float *dwork[4], *panel[4], *local_work;
magma_int_t i, j, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
magma_device_t cdevice;
magma_getdevice(&cdevice);
int panel_gpunum, i_local, n_local[4], la_gpu, displacement;
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
k = min(m,n);
if (k == 0)
return *info;
nb = magma_get_sgeqrf_nb(m);
displacement = n * nb;
lwork = (m+n+64) * nb;
lhwork = lwork - (m)*nb;
for(i=0; i<num_gpus; i++){
#ifdef MultiGPUs
magma_setdevice(i);
#endif
if (MAGMA_SUCCESS != magma_smalloc( &(dwork[i]), (n + ldda)*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
/* Set the number of local n for each GPU */
for(i=0; i<num_gpus; i++){
n_local[i] = ((n/nb)/num_gpus)*nb;
if (i < (n/nb)%num_gpus)
n_local[i] += nb;
else if (i == (n/nb)%num_gpus)
n_local[i] += n%nb;
}
if (MAGMA_SUCCESS != magma_smalloc_pinned( &local_work, lwork )) {
*info = -9;
for(i=0; i<num_gpus; i++){
#ifdef MultiGPUs
magma_setdevice(i);
#endif
magma_free( dwork[i] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
cudaStream_t streaml[4][2];
for(i=0; i<num_gpus; i++){
#ifdef MultiGPUs
magma_setdevice(i);
#endif
magma_queue_create( &streaml[i][0] );
magma_queue_create( &streaml[i][1] );
}
nbmin = 2;
nx = nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb)
{
/* Set the GPU number that holds the current panel */
panel_gpunum = (i/nb)%num_gpus;
/* Set the local index where the current panel is */
i_local = i/(nb*num_gpus)*nb;
ib = min(k-i, nb);
rows = m -i;
/* Send current panel to the CPU */
#ifdef MultiGPUs
magma_setdevice(panel_gpunum);
#endif
magma_sgetmatrix_async( rows, ib,
dlA(panel_gpunum, i, i_local), ldda,
hwrk_ref(i), ldwork, streaml[panel_gpunum][1] );
if (i>0){
/* Apply H' to A(i:m,i+2*ib:n) from the left; this is the look-ahead
application to the trailing matrix */
la_gpu = panel_gpunum;
/* only the GPU that has next panel is done look-ahead */
#ifdef MultiGPUs
magma_setdevice(la_gpu);
#endif
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
m-old_i, n_local[la_gpu]-i_local-old_ib, old_ib,
panel[la_gpu], ldda, dwork[la_gpu], lddwork,
dlA(la_gpu, old_i, i_local+old_ib), ldda,
dwork[la_gpu]+old_ib, lddwork);
la_gpu = ((i-nb)/nb)%num_gpus;
#ifdef MultiGPUs
magma_setdevice(la_gpu);
#endif
magma_ssetmatrix_async( old_ib, old_ib,
hwrk_ref(old_i), ldwork,
panel[la_gpu], ldda, streaml[la_gpu][0] );
}
#ifdef MultiGPUs
magma_setdevice(panel_gpunum);
#endif
magma_queue_sync( streaml[panel_gpunum][1] );
lapackf77_sgeqrf(&rows, &ib, hwrk_ref(i), &ldwork, tau+i, lhwrk, &lhwork, info);
// Form the triangular factor of the block reflector
// H = H(i) H(i+1) . . . H(i+ib-1)
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
hwrk_ref(i), &ldwork, tau+i, lhwrk, &ib);
spanel_to_q( MagmaUpper, ib, hwrk_ref(i), ldwork, lhwrk+ib*ib );
// Send the current panel back to the GPUs
// Has to be done with asynchronous copies
for(j=0; j<num_gpus; j++)
{
#ifdef MultiGPUs
magma_setdevice(j);
#endif
if (j == panel_gpunum)
panel[j] = dlA(j, i, i_local);
else
panel[j] = dwork[j]+displacement;
magma_ssetmatrix_async( rows, ib,
hwrk_ref(i), ldwork,
panel[j], ldda, streaml[j][0] );
}
for(j=0; j<num_gpus; j++)
{
#ifdef MultiGPUs
magma_setdevice(j);
#endif
magma_queue_sync( streaml[j][0] );
}
/* Restore the panel */
sq_to_panel( MagmaUpper, ib, hwrk_ref(i), ldwork, lhwrk+ib*ib );
if (i + ib < n)
{
/* Send the T matrix to the GPU.
Has to be done with asynchronous copies */
for(j=0; j<num_gpus; j++)
{
#ifdef MultiGPUs
magma_setdevice(j);
#endif
magma_ssetmatrix_async( ib, ib,
lhwrk, ib,
dwork[j], lddwork, streaml[j][0] );
}
if (i+nb < k-nx)
{
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left;
This is update for the next panel; part of the look-ahead */
la_gpu = (panel_gpunum+1)%num_gpus;
int i_loc = (i+nb)/(nb*num_gpus)*nb;
for(j=0; j<num_gpus; j++){
#ifdef MultiGPUs
magma_setdevice(j);
#endif
//magma_queue_sync( streaml[j][0] );
if (j==la_gpu)
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
panel[j], ldda, dwork[j], lddwork,
dlA(j, i, i_loc), ldda, dwork[j]+ib, lddwork);
else if (j<=panel_gpunum)
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n_local[j]-i_local-ib, ib,
panel[j], ldda, dwork[j], lddwork,
dlA(j, i, i_local+ib), ldda, dwork[j]+ib, lddwork);
else
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n_local[j]-i_local, ib,
panel[j], ldda, dwork[j], lddwork,
dlA(j, i, i_local), ldda, dwork[j]+ib, lddwork);
}
}
else {
/* do the entire update as we exit and there would be no lookahead */
la_gpu = (panel_gpunum+1)%num_gpus;
int i_loc = (i+nb)/(nb*num_gpus)*nb;
#ifdef MultiGPUs
magma_setdevice(la_gpu);
#endif
magma_slarfb_gpu( MagmaLeft, MagmaTrans, MagmaForward, MagmaColumnwise,
rows, n_local[la_gpu]-i_loc, ib,
panel[la_gpu], ldda, dwork[la_gpu], lddwork,
dlA(la_gpu, i, i_loc), ldda, dwork[la_gpu]+ib, lddwork);
#ifdef MultiGPUs
magma_setdevice(panel_gpunum);
#endif
magma_ssetmatrix( ib, ib,
hwrk_ref(i), ldwork,
dlA(panel_gpunum, i, i_local), ldda );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
for(j=0; j<num_gpus; j++){
#ifdef MultiGPUs
magma_setdevice(j);
#endif
magma_free( dwork[j] );
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
lhwork = lwork - rows*ib;
panel_gpunum = (panel_gpunum+1)%num_gpus;
int i_loc = (i)/(nb*num_gpus)*nb;
#ifdef MultiGPUs
magma_setdevice(panel_gpunum);
#endif
magma_sgetmatrix( rows, ib,
dlA(panel_gpunum, i, i_loc), ldda,
lhwrk, rows );
lhwork = lwork - rows*ib;
lapackf77_sgeqrf(&rows, &ib, lhwrk, &rows, tau+i, lhwrk+ib*rows, &lhwork, info);
magma_ssetmatrix( rows, ib,
lhwrk, rows,
dlA(panel_gpunum, i, i_loc), ldda );
}
for(i=0; i<num_gpus; i++){
#ifdef MultiGPUs
magma_setdevice(i);
#endif
magma_queue_destroy( streaml[i][0] );
magma_queue_destroy( streaml[i][1] );
}
magma_setdevice(cdevice);
magma_free_pinned( local_work );
return *info;
} /* magma_sgeqrf2_mgpu */
extern "C" magma_int_t
magma_sorgqr2(magma_int_t m, magma_int_t n, magma_int_t k,
float *A, magma_int_t lda,
float *tau,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
SORGQR generates an M-by-N REAL matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by SGEQRF.
This version recomputes the T matrices on the CPU and sends them to the GPU.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix Q. M >= 0.
N (input) INTEGER
The number of columns of the matrix Q. M >= N >= 0.
K (input) INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
A (input/output) REAL array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by SGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
LDA (input) INTEGER
The first dimension of the array A. LDA >= max(1,M).
TAU (input) REAL array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by SGEQRF_GPU.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument has an illegal value
===================================================================== */
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(i,j) (dA + (i) + (j)*ldda)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
magma_int_t nb = magma_get_sgeqrf_nb(min(m, n));
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldda;
magma_int_t i, ib, ki, kk; //, iinfo;
magma_int_t lddwork;
float *dA, *dV, *dW, *dT, *T;
float *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_smalloc( &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_smalloc_cpu( &work, lwork );
T = work;
if (work == NULL) {
magma_free( dA );
magma_free_cpu( work );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
float *V = work + (n+nb)*nb;
magma_queue_t stream;
magma_queue_create( &stream );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
/*
lapackf77_sorgqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
*/
lapackf77_slacpy( MagmaUpperLowerStr, &m_kk, &k_kk, A(kk,kk), &lda, V, &m_kk);
lapackf77_slaset( MagmaUpperLowerStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
V, &m_kk, &tau[kk], work, &k_kk);
lapackf77_slarfb( 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_ssetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(kk, kk), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_slaset( 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_slaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
magma_ssetmatrix_async( mi, ib,
A(i, i), lda,
dV, ldda, stream );
lapackf77_slarft( MagmaForwardStr, MagmaColumnwiseStr,
&mi, &ib,
A(i,i), &lda, &tau[i], T, &nb);
magma_ssetmatrix_async( ib, ib,
T, nb,
dT , nb, stream );
// set panel to identity
magmablas_slaset( MagmaUpperLower, i, ib, dA(0, i), ldda );
magmablas_slaset_identity( mi, ib, dA(i, i), ldda );
magma_queue_sync( stream );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_slarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT, nb,
dA(i, i), ldda, dW, lddwork );
}
}
// copy result back to CPU
magma_sgetmatrix( 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_sorgqr */
