/**
Purpose
-------
DGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
It 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 DOUBLE PRECISION array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
\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]
ipiv INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
@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.
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_dgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_dgetrf(
magma_int_t m, magma_int_t n,
double *A, magma_int_t lda,
magma_int_t *ipiv,
magma_int_t *info)
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, ((i_)*nb + (j_)*nb*ldda + dA_offset)
#define dAT(i_, j_) dAT, ((i_)*nb*lddat + (j_)*nb + dAT_offset)
#define dwork(i_) dwork, (i_)
#else
#define dA(i_, j_) ( dA + (i_)*nb + (j_)*nb*ldda)
#define dAT(i_, j_) ( dAT + (i_)*nb*lddat + (j_)*nb)
#define dwork(i_) (dwork + (i_))
#endif
// Constants
const double c_one = MAGMA_D_ONE;
const double c_neg_one = MAGMA_D_NEG_ONE;
// Local variables
double *work;
magmaDouble_ptr dA, dAT, dwork;
magma_int_t iinfo, nb;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
nb = magma_get_dgetrf_nb( m, n );
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_dgetrf( &m, &n, A, &lda, ipiv, info );
}
else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, lddat, maxdim;
magma_int_t i, j, rows, cols, s = min(m, n)/nb;
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
maxdim = max( maxm, maxn );
lddat = maxn;
ldda = maxm;
/* set number of GPUs */
magma_int_t ngpu = magma_num_gpus();
if ( ngpu > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_dgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magma_queue_t queues[2] = { NULL, NULL };
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
/* check the memory requirement */
size_t mem_size = magma_queue_mem_size( queues[0] );
mem_size /= sizeof(double);
magma_int_t h = 1+(2+ngpu);
magma_int_t ngpu2 = ngpu;
magma_int_t NB = (magma_int_t)(0.8*mem_size/maxm - h*nb);
const char* ngr_nb_char = getenv("MAGMA_NGR_NB");
if ( ngr_nb_char != NULL )
NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
if ( ngpu > ceil((double)NB/nb) ) {
ngpu2 = (magma_int_t)ceil((double)NB/nb);
h = 1+(2+ngpu2);
NB = (magma_int_t)(0.8*mem_size/maxm - h*nb);
}
if ( ngpu2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_dgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
// dwork is nb*maxm for panel, and maxdim*maxdim for A
if (MAGMA_SUCCESS != magma_dmalloc( &dwork, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
ldda = lddat = maxdim;
magma_dsetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
dAT = dA;
magmablas_dtranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
// dwork is nb*maxm for panel, and maxm*maxn for A
if (MAGMA_SUCCESS != magma_dmalloc( &dwork, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
magma_dsetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
if (MAGMA_SUCCESS != magma_dmalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dwork );
magma_dgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magmablas_dtranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queues[0] );
}
lapackf77_dgetrf( &m, &nb, work, &lda, ipiv, &iinfo );
for( j = 0; j < s; j++ ) {
// get j-th panel from device
cols = maxm - j*nb;
if (j > 0) {
magmablas_dtranspose( nb, cols, dAT(j,j), lddat, dwork(0), cols, queues[0] );
magma_queue_sync( queues[0] );
magma_dgetmatrix_async( m-j*nb, nb, dwork(0), cols, work, lda, queues[1] );
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queues[0] );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queues[0] );
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queues[1] );
lapackf77_dgetrf( &rows, &nb, work, &lda, ipiv+j*nb, &iinfo );
}
if (*info == 0 && iinfo > 0)
*info = iinfo + j*nb;
// put j-th panel onto device
magma_dsetmatrix_async( m-j*nb, nb, work, lda, dwork(0), cols, queues[1] );
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_dlaswp( n, dAT(0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queues[0] );
magma_queue_sync( queues[1] );
magmablas_dtranspose( cols, nb, dwork(0), cols, dAT(j,j), lddat, queues[0] );
// do the small non-parallel computations (next panel update)
if (s > (j+1)) {
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
else {
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_dtranspose( nb0, rows, dAT(s,s), lddat, dwork(0), cols, queues[0] );
magma_dgetmatrix_async( rows, nb0, dwork(0), cols, work, lda, queues[0] );
magma_queue_sync( queues[0] );
// do the cpu part
lapackf77_dgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo );
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_dlaswp( n, dAT(0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queues[0] );
// put j-th panel onto device
magma_dsetmatrix_async( rows, nb0, work, lda, dwork(0), cols, queues[0] );
magmablas_dtranspose( rows, nb0, dwork(0), cols, dAT(s,s), lddat, queues[0] );
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), lddat,
dAT(s, s)+nb0, lddat, queues[0] );
}
// undo transpose
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_dtranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
magma_dgetmatrix( m, n, dAT(0,0), lddat, A, lda, queues[0] );
}
else {
magmablas_dtranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queues[0] );
magma_dgetmatrix( m, n, dA(0,0), ldda, A, lda, queues[0] );
magma_free( dAT );
}
magma_free( dwork );
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
}
return *info;
} /* magma_dgetrf */
extern "C" magma_int_t
magma_dgelqf( magma_int_t m, magma_int_t n,
double *a, magma_int_t lda, double *tau,
double *work, magma_int_t lwork, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
DGELQF computes an LQ factorization of a DOUBLE_PRECISION M-by-N matrix A:
A = L * Q.
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.
A (input/output) DOUBLE_PRECISION array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and below the diagonal of the array
contain the m-by-min(m,n) lower trapezoidal matrix L (L is
lower triangular if m <= n); the elements above the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors (see Further Details).
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,M).
TAU (output) DOUBLE_PRECISION array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
WORK (workspace/output) DOUBLE_PRECISION array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LWORK (input) INTEGER
The dimension of the array WORK. LWORK >= max(1,M).
For optimum performance LWORK >= M*NB, where NB is the
optimal blocksize.
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.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -10 internal GPU memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(k) . . . H(2) H(1), 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:n) is stored on exit in A(i,i+1:n),
and tau in TAU(i).
===================================================================== */
#define a_ref(a_1,a_2) ( a+(a_2)*(lda) + (a_1))
double *dA, *dAT;
double c_one = MAGMA_D_ONE;
magma_int_t maxm, maxn, maxdim, nb;
magma_int_t iinfo, ldda;
int lquery;
/* Function Body */
*info = 0;
nb = magma_get_dgelqf_nb(m);
work[0] = MAGMA_D_MAKE( (double)(m*nb), 0 );
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,m) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (min(m, n) == 0) {
work[0] = c_one;
return *info;
}
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
if (maxdim*maxdim < 2*maxm*maxn)
{
ldda = maxdim;
if (MAGMA_SUCCESS != magma_dmalloc( &dA, maxdim*maxdim )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_dsetmatrix( m, n, a, lda, dA, ldda );
dAT = dA;
magmablas_dtranspose_inplace( ldda, dAT, ldda );
}
else
{
ldda = maxn;
if (MAGMA_SUCCESS != magma_dmalloc( &dA, 2*maxn*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_dsetmatrix( m, n, a, lda, dA, maxm );
dAT = dA + maxn * maxm;
magmablas_dtranspose2( dAT, ldda, dA, maxm, m, n );
}
magma_dgeqrf2_gpu(n, m, dAT, ldda, tau, &iinfo);
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_dtranspose_inplace( ldda, dAT, ldda );
magma_dgetmatrix( m, n, dA, ldda, a, lda );
} else {
magmablas_dtranspose2( dA, maxm, dAT, ldda, n, m );
magma_dgetmatrix( m, n, dA, maxm, a, lda );
}
magma_free( dA );
return *info;
} /* magma_dgelqf */
/**
Purpose
-------
DGELQF computes an LQ factorization of a DOUBLE_PRECISION M-by-N matrix dA:
dA = L * Q.
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 DOUBLE_PRECISION array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and below the diagonal of the array
contain the m-by-min(m,n) lower trapezoidal matrix L (L is
lower triangular if m <= n); the elements above the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors (see Further Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
@param[out]
tau DOUBLE_PRECISION array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) DOUBLE_PRECISION 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(1,M).
For optimum performance LWORK >= M*NB, where NB is the
optimal blocksize.
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@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(k) . . . H(2) H(1), 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:n) is stored on exit in A(i,i+1:n),
and tau in TAU(i).
@ingroup magma_dgelqf_comp
********************************************************************/
extern "C" magma_int_t
magma_dgelqf_gpu(
magma_int_t m, magma_int_t n,
magmaDouble_ptr dA, magma_int_t ldda,
double *tau,
double *work, magma_int_t lwork,
magma_int_t *info)
{
const double c_one = MAGMA_D_ONE;
const magma_int_t ione = 1;
MAGMA_UNUSED( ione ); // used only for real
double *dAT;
magma_int_t min_mn, maxm, maxn, nb;
magma_int_t iinfo;
int lquery;
*info = 0;
nb = magma_get_dgelqf_nb(m);
min_mn = min(m,n);
work[0] = MAGMA_D_MAKE( (double)(m*nb), 0 );
lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,m) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
magma_int_t lddat = maxn;
dAT = dA;
if ( m == n ) {
lddat = ldda;
magmablas_dtranspose_inplace( m, dAT, ldda );
}
else {
if (MAGMA_SUCCESS != magma_dmalloc( &dAT, maxm*maxn ) ) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_dtranspose( m, n, dA, ldda, dAT, lddat );
}
magma_dgeqrf2_gpu( n, m, dAT, lddat, tau, &iinfo );
assert( iinfo >= 0 );
if ( iinfo > 0 ) {
*info = iinfo;
}
// conjugate tau
#ifdef COMPLEX
lapackf77_dlacgv( &min_mn, tau, &ione );
#endif
if ( m == n ) {
magmablas_dtranspose_inplace( m, dAT, lddat );
}
else {
magmablas_dtranspose( n, m, dAT, lddat, dA, ldda );
magma_free( dAT );
}
return *info;
} /* magma_dgelqf_gpu */
/**
Purpose
-------
DGELQF computes an LQ factorization of a DOUBLE_PRECISION M-by-N matrix dA:
dA = L * Q.
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 DOUBLE_PRECISION array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and below the diagonal of the array
contain the m-by-min(m,n) lower trapezoidal matrix L (L is
lower triangular if m <= n); the elements above the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors (see Further Details).
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
@param[out]
tau DOUBLE_PRECISION array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) DOUBLE_PRECISION 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(1,M).
For optimum performance LWORK >= M*NB, where NB is the
optimal blocksize.
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -10 internal GPU memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(k) . . . H(2) H(1), 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:n) is stored on exit in A(i,i+1:n),
and tau in TAU(i).
@ingroup magma_dgelqf_comp
********************************************************************/
extern "C" magma_int_t
magma_dgelqf_gpu(
magma_int_t m, magma_int_t n,
magmaDouble_ptr dA, magma_int_t ldda,
double *tau,
double *work, magma_int_t lwork,
magma_int_t *info)
{
double *dAT;
double c_one = MAGMA_D_ONE;
magma_int_t maxm, maxn, maxdim, nb;
magma_int_t iinfo;
int lquery;
*info = 0;
nb = magma_get_dgelqf_nb(m);
work[0] = MAGMA_D_MAKE( (double)(m*nb), 0 );
lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,m) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (min(m, n) == 0) {
work[0] = c_one;
return *info;
}
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
magma_int_t lddat = maxn;
dAT = dA;
if ( m == n ) {
lddat = ldda;
magmablas_dtranspose_inplace( m, dAT, ldda );
}
else {
if (MAGMA_SUCCESS != magma_dmalloc( &dAT, maxm*maxn ) ) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_dtranspose( m, n, dA, ldda, dAT, lddat );
}
magma_dgeqrf2_gpu(n, m, dAT, lddat, tau, &iinfo);
if ( m == n ) {
magmablas_dtranspose_inplace( m, dAT, lddat );
}
else {
magmablas_dtranspose( n, m, dAT, lddat, dA, ldda );
magma_free( dAT );
}
return *info;
} /* magma_dgelqf_gpu */
extern "C" magma_int_t
magma_dgetrf_gpu(
magma_int_t m, magma_int_t n,
magmaDouble_ptr dA, size_t dA_offset, magma_int_t ldda,
magma_int_t *ipiv,
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
=======
DGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
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.
A (input/output) DOUBLE_PRECISION array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDDA (input) INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
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.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
#define dA(i_, j_) dA, dA_offset + (i_)*nb + (j_)*nb*ldda
#define dAT(i_, j_) dAT, dAT_offset + (i_)*nb*lddat + (j_)*nb
#define dAP(i_, j_) dAP, (i_) + (j_)*maxm
#define work(i_) (work + (i_))
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, j, rows, s, lddat, ldwork;
magmaDouble_ptr dAT, dAP;
double *work;
size_t dAT_offset;
/* 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;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
mindim = min(m, n);
nb = magma_get_dgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
if ( MAGMA_SUCCESS != magma_dmalloc_cpu( &work, m*n )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_dgetmatrix( m, n, dA(0,0), ldda, work(0), m, queue );
lapackf77_dgetrf( &m, &n, work, &m, ipiv, info );
magma_dsetmatrix( m, n, work(0), m, dA(0,0), ldda, queue );
magma_free_cpu( work );
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
if ( MAGMA_SUCCESS != magma_dmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// square matrices can be done in place;
// rectangular requires copy to transpose
if ( m == n ) {
dAT = dA;
dAT_offset = dA_offset;
lddat = ldda;
magmablas_dtranspose_inplace( m, dAT(0,0), lddat, queue );
}
else {
lddat = maxn; // N-by-M
dAT_offset = 0;
if ( MAGMA_SUCCESS != magma_dmalloc( &dAT, lddat*maxm )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_dtranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queue );
}
ldwork = maxm;
if ( MAGMA_SUCCESS != magma_dmalloc_cpu( &work, ldwork*nb )) {
magma_free( dAP );
if ( dA != dAT )
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
for( j=0; j < s; j++ ) {
// download j-th panel
magmablas_dtranspose( nb, m-j*nb, dAT(j,j), lddat, dAP(0,0), maxm, queue );
magma_dgetmatrix( m-j*nb, nb, dAP(0,0), maxm, work(0), ldwork, queue );
if ( j > 0 ){
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queue );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queue );
}
// do the cpu part
rows = m - j*nb;
lapackf77_dgetrf( &rows, &nb, work, &ldwork, ipiv+j*nb, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_dlaswp( n, dAT(0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queue );
// upload j-th panel
magma_dsetmatrix( m-j*nb, nb, work(0), ldwork, dAP(0,0), maxm, queue );
magmablas_dtranspose( m-j*nb, nb, dAP(0,0), maxm, dAT(j,j), lddat, queue );
// do the small non-parallel computations (next panel update)
if ( s > (j+1) ) {
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queue );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queue );
}
else {
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queue );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queue );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
magmablas_dtranspose( nb0, rows, dAT(s,s), lddat, dAP(0,0), maxm, queue );
magma_dgetmatrix( rows, nb0, dAP(0,0), maxm, work(0), ldwork, queue );
// do the cpu part
lapackf77_dgetrf( &rows, &nb0, work, &ldwork, ipiv+s*nb, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_dlaswp( n, dAT(0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queue );
// upload j-th panel
magma_dsetmatrix( rows, nb0, work(0), ldwork, dAP(0,0), maxm, queue );
magmablas_dtranspose( rows, nb0, dAP(0,0), maxm, dAT(s,s), lddat, queue );
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat, queue );
}
// undo transpose
if ( dA == dAT ) {
magmablas_dtranspose_inplace( m, dAT(0,0), lddat, queue );
}
else {
magmablas_dtranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queue );
magma_free( dAT );
}
magma_free( dAP );
magma_free_cpu( work );
}
return *info;
} /* magma_dgetrf_gpu */
extern "C" magma_int_t
magma_dgetrf_gpu_work_amc(
magma_int_t m, magma_int_t n,
double *dA, magma_int_t dA_LD,
magma_int_t *ipiv, magma_int_t *info,
/*workspace on the cpu side*/
double *AWORK, magma_int_t AWORK_LD, magma_int_t AWORK_n
)
{
/* -- MAGMA (version 1.5.0-beta3) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2011
Purpose
=======
DGETRF_GPU_WORK_AMC computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. The technique used for the panel factorization
is the parallel recursif LU (see lawn 259).
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
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.
A (input/output) DOUBLE_PRECISION array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDDA (input) INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
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.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
int ONE = 1;
magma_int_t iinfo, nb;
magma_int_t mindim;
magma_int_t nrows, ncols;
//double *work;
magma_int_t dm_max, dn_max;
magma_int_t I, J, K, M, N, U_K;
//magma_int_t A_K;
double *dAT;
magma_int_t dAT_LD;
double *dAP_set,*dAP_get;
magma_int_t dAP_LD;
//magma_int_t nrows, ncols;
magma_int_t gpu_nrows, gpu_ncols;
int nbcores; /*Number of cores available for the whole factorization*/
int panel_num_threads; /*Number of threads for the panel*/
double dcpu; /*percentage of the matrix to allocate on the CPUs*/
int B_rows;
double t1;
/* Recommanded dimension in the workspace*/
int A_m, A_n, A_N, A_NMAX, A_LD;
double *A;
#ifdef USE_CALU
int i_nrows;
#endif
amc_args_t *args;
/*magma_event_t *A_event;*/ /*Control bucket*/
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (dA_LD < max(1,m))
*info = -4;
else if (AWORK_LD < max(1,m))
*info = -5;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/*Get parameters*/
args = magma_amc_args_get_default();
nb= args->nb;
nbcores = args->P;
panel_num_threads = args->Pr;
dcpu = args->dcpu;
/* Check and fix parameters */
if(nb==0)
nb = magma_get_dgetrf_nb(m) ;/*magma dgetrf block size*/
else
nb = args->nb;
if(nb>n) nb = n;
if(panel_num_threads>nbcores) panel_num_threads = nbcores;
/* Compute the maximum number of panels we can store in the workspace*/
A_NMAX = (int) (AWORK_n/ nb);
/* Compute the recommanded number of columns for the cpu part*/
A_n = (int) ceil(n*dcpu);
/*Make sure we work with multiple of 32*/
/*
if(A_n%32!=0) {
A_n = ((A_n + 31)/32)*32;
}
*/
/* Compute the recommanded number of panels for the cpu part*/
A_N = (int) (A_n/ nb);
/* Check if there are enough workspace. In case the user gave a workspace lower than the optimal*/
/* NOTE: using small workspace may reduce performance*/
if(A_N>A_NMAX){
#if (dbglevel >=1)
printf("[DBG_WARNING] Resizing buffer to feet user preferences. Recommanded:%d, Max given:%d\n",A_N, A_NMAX);
#endif
A_N = A_NMAX;
}
A = AWORK;
A_m = m;
A_LD = AWORK_LD;
#if (dbglevel >=1)
/* Initialize the tracing*/
ca_dbg_trace_init(nbcores,1); //nbcores + 1 GPU
#endif
#if (dbglevel >=1)
t1 = magma_wtime();
#endif
/*Transfer the first column block of the matrix from the GPU to the CPUs*/
magma_dgetmatrix(A_m, A_n, dA, dA_LD, A, A_LD);
#if (dbglevel >=1)
printf("[DBG] Time First getmatrix: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat(m, A_n, A, A_LD,"A after first getMatrix");
#endif
/*Allocate a workspace for the panels transposition*/
dAP_LD = m;
if(dAP_LD%32!=0) dAP_LD = ((dAP_LD + 31)/32)*32;/*Make dAP_LD multiple of 32*/
if (MAGMA_SUCCESS != magma_dmalloc(&dAP_set, dAP_LD*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_dmalloc(&dAP_get, dAP_LD*nb)) {
magma_free(dAP_set);
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
#if (dbglevel >=1)
printf("[DBG] Time workspace memory alloc (dAP): %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
/*Transpose the gpu part of the matrix in/out of place*/
if ((m == n) ){ //&& (m % 32 == 0) && (dA_LD%32 == 0)
dAT = dA;
dAT_LD= dA_LD;
magmablas_dtranspose_inplace(m, dAT, dAT_LD);
}
else {
dm_max = m;
dn_max = n;
/*Make sure m and n are multiple of 32*/
if(dm_max%32!=0) dm_max = ((dm_max + 31)/32)*32;
if(dn_max%32!=0) dn_max = ((dn_max + 31)/32)*32;
if (MAGMA_SUCCESS != magma_dmalloc(&dAT, dm_max*dn_max )) {
magma_free(dAP_set);
magma_free(dAP_get);
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dAT_LD = dn_max;
magmablas_dtranspose2( dAT, dAT_LD, dA, dA_LD, m, n );
}
#if (dbglevel >=1)
printf("[DBG] Time First transposition: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_LD,"matrix dAT to factorize");
#endif
/* Compute the maximun number of steps*/
mindim = min(m, n);
M = (int) ceil( (double) m / nb);
N = (int) ceil( (double) mindim / nb); /*N = n/nb*/
/*Let the asynchronous algorithm begin*/
#if (dbglevel >=1)
printf("Starting recursif code ... m:%d, n:%d, nb:%d, nbcores:%d, N:%d, A_N:%d\n", m, n, nb, nbcores, N, A_N); //Summary
#endif
/*Initialize the scheduler*/
magma_schedule_init(nbcores, 1);
K = 0;
#ifdef USE_CALU
/*initialize calu environment*/
core_dtslu_alloc(panel_num_threads, A_m, nb);
core_dtslu_init(panel_num_threads);
/*Initialize rows indice: required*/
for(I=0;I<A_m;I++) ipiv[I]=I;
#else
/*initialize parallel recursif panel environment*/
CORE_zgetrf_reclap_init();
#endif
magma_schedule_set_task_priority(INT_MAX-1);
/*Schedule the first panel factorization*/
#ifdef USE_CALU
magma_insert_core_dtslu(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, panel_num_threads, colptr(K));
B_rows = (int) ceil((double) (M-K-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+1; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, nb, c_one, A(0,K), A_LD, A(I,K), A_LD, colptr(K));
}
#else
magma_insert_core_dgetrf_rec(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, panel_num_threads, colptr(K));
#endif
/*Transfer the factorized panel to the GPU (transposition included)*/
magma_insert_dsetmatrix_transpose(A_m, nb, A(0,K), A_LD, dAT(0,K), dAT_LD, dAP_set, dAP_LD, colptr(K), dAT(K,K));
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, nb, A(0,0), A_LD,"A(0,0)");
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_LD,"dA");
#endif
for(K=0;K<=N-1;K++){
/*insert the coarse update of the trailing submatrix corresponding to panel K to the GPU, that is submatrix A[K+1:M, K+1+d-1:N]*/
gpu_nrows = m - (K+1)*nb;
gpu_ncols = n - (K+1+A_N-1)*nb;
if(gpu_ncols >0)
{
/*NOTE: Here we work on the matrix transpose*/
/*Set the priority max for the GPU computations*/
magma_schedule_set_task_priority(INT_MAX);
//// magma_schedule_set_task_priority(INT_MAX - N*K);
/*schedule a swap of the trailing submatrix in the gpu using ipiv[K]*/
/*dependency dAT((K+1)-1, (K+A_N)-1) = dAT(K, K+A_N-1) with previous dgemm*/
magma_insert_dlaswp(gpu_ncols, dAT(K, K+A_N), dAT_LD, ONE, nb, ipiv(K), ONE, dAT(K, K+A_N-1)); /*non blocking*/
//printf("debug barrier\n");
//magma_schedule_barrier();
magma_insert_dtrsm(MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit, gpu_ncols, nb, c_one, dAT(K,K), dAT_LD, dAT(K,K+A_N), dAT_LD);/*non blocking*/
/* aij^T = aij^T - (lik.ukj)^T = aij^T - ukj^T.lik^T*/
magma_insert_dgemm(MagmaNoTrans,MagmaNoTrans, gpu_ncols, gpu_nrows, nb, c_neg_one, dAT(K,K+A_N), dAT_LD, dAT(K+1,K), dAT_LD, c_one, dAT(K+1,K+A_N), dAT_LD);/*non blocking*/
}
/*iterate over the rest of the columns to update the trailing submatrix on the cpu*/
for(J=K+1;J<=min(K+A_N-1, N-1);J++){
ncols = min(nb, n - J*nb);
/*Set the priority max for column having the next panel (look ahead of deep 1),
and process the rest of the update in a right looking way*/
if(J==K+1)
magma_schedule_set_task_priority(INT_MAX -2 );
//// magma_schedule_set_task_priority(INT_MAX - N*K -1);
else
magma_schedule_set_task_priority(INT_MAX -3 - J );//- N*K
/*dependency colptr(J): make sure column J is sent from GPU, and all previous update was done*/
magma_insert_core_dlaswp(ncols, A(K,J), A_LD, ONE, nb, ipiv(K), ONE, colptr(J));
magma_insert_core_dtrsm('L', 'L', 'N', 'U', nb, ncols, c_one, A(K,K), A_LD, A(K,J), A_LD, colptr(J));
/*Compute the number of blocs rows to group together before the update. To avoid scheduling overhead.*/
B_rows = (int) ceil((double) (M-K-1)/panel_num_threads);
//B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+1; I<=M-1; I+=B_rows){
nrows = min(B_rows*nb, m-I*nb);
/*dep colptr(K):make sure the panel is not overwritten or swapped since dgemm use A[I,K]*/
/*dep colptr(J): Gather all dgemm on one column and create dependencies with previous dgemm and the next panel*/
magma_insert_core_dgemm('N','N', nrows, ncols, nb, c_neg_one, A(I,K), A_LD, A(K,J), A_LD, c_one, A(I,J), A_LD, colptr(K), colptr(J));
}
if(J==K+1)
{
/*Look ahead and insert the next panel*/
nrows = m - (K+1)*nb;
ncols = min(nb, n - (K+1)*nb);
/*Schedule the next panel factorization with maximum priority*/
magma_schedule_set_task_priority(INT_MAX -1);
#ifdef USE_CALU
magma_insert_core_dtslu(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
B_rows = (int) ceil((double) (M-(K+1)-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+2; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, ncols, c_one, A(K+1,K+1), A_LD, A(I,K+1), A_LD, colptr(K+1));
//dtrsm("R", "U", "N", "N", &nrowPblock, &panel_NB, &dONE, &(A[M*pos+pos]), &LDA, &(A[lpos]), &LDA); //
}
#else
magma_insert_core_dgetrf_rec(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
#endif
/*Determine the upper part of the matrix done by the CPU on that column and send it to the GPU with the panel*/
U_K = max(0, K+1 - A_N +1);
nrows = m - U_K*nb;
/*Transfer the upper part of the matrix for that column and the factorized panel to the GPU*/
magma_insert_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
}
}
/*Transfer asynchronously one column (column K+A_N) from the GPU to the CPU to balance work*/
/*Make sure this is inserted after all dgemm before it schedules to replace a current panel in case A_N< N*/
if(K+A_N<N) {
ncols = min(nb, gpu_ncols);
magma_schedule_set_task_priority(INT_MAX);
magma_insert_dgetmatrix_transpose(gpu_nrows, ncols, dAT(K+1,K+A_N), dAT_LD, A(K+1,K+A_N), A_LD, dAP_get, dAP_LD, colptr(K+A_N)); //blocking
/*if A_N==1 there is no look-ahead, so insert the panel here*/
if(A_N==1){
/*Look ahead and insert the next panel*/
nrows = m - (K+1)*nb;
ncols = min(nb, n - (K+1)*nb);
/*Schedule the next panel factorization with maximum priority*/
magma_schedule_set_task_priority(INT_MAX -1);
#ifdef USE_CALU
magma_insert_core_dtslu(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
B_rows = (int) ceil((double) (M-(K+1)-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+2; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, ncols, c_one, A(K+1,K+1), A_LD, A(I,K+1), A_LD, colptr(K+1));
//dtrsm("R", "U", "N", "N", &nrowPblock, &panel_NB, &dONE, &(A[M*pos+pos]), &LDA, &(A[lpos]), &LDA); //
}
#else
magma_insert_core_dgetrf_rec(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
//magma_insert_core_dgetrf(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, colptr(K+1));
#endif
/*Determine the upper part of the matrix done by the CPU on that column and send it to the GPU with the panel*/
U_K = max(0, K+1 - A_N +1);
nrows = m - U_K*nb;
///magma_schedule_set_task_priority(INT_MAX);
/*Transfer the upper part of the matrix for that column and the factorized panel to the GPU*/
magma_insert_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
}
}
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, A_n, A, A_LD,"A");
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_LD,"dA");
#endif
} //Step K done
/*Wait for all thread termination*/
magma_schedule_barrier();
/*TODO: don't need quark here*/
/*Perform a sequence of left swap on the matrix corresponding to the different panel*/
for(K=1;K<=N-1;K++){
#if (dbglevel >=1)
ca_trace_start();
#endif
nrows = min(nb,m - K*nb);
ncols = min(K*nb,n);
/*dep dAT(K-1): Make sure the last swap is completed, and also the dgemm using the panel*/
// magma_insert_dlaswp(ncols, dAT(K, 0), dAT_LD, ONE, nrows, ipiv(K), ONE, dAT(K-1,0));
magmablas_dlaswp(ncols, dAT(K, 0), dAT_LD, ONE, nrows, ipiv(K), ONE);
#if (dbglevel >=1)
ca_trace_end_1gpu('W');
#endif
}
/*Shutdown the scheduler*/
magma_schedule_delete();
/*update permutation vector indexes*/
for(K=1;K<=N-1;K++){
nrows = min(nb, n-K*nb);
for(J=0;J<=nrows-1;J++){
ipiv[K*nb+J] += K*nb;
}
}
#if dbglevel>=1
printf("[DBG] Time Factorization:%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
/*No need for synchro, since dtranspose is blocking*/
if (m == n) {
magmablas_dtranspose_inplace(m, dAT, dAT_LD); //( m, dAT, dAT_LD );
dA = dAT;
}
else {
magmablas_dtranspose2( dA, dA_LD, dAT, dAT_LD, n, m );
magma_free( dAT );
}
#if dbglevel>=1
printf("[DBG] Time Final in/out of place transpose:%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#ifdef USE_CALU
core_dtslu_free();
#endif
magma_free( dAP_set );
magma_free( dAP_get );
#if dbglevel>=1
printf("[DBG] Time memory free (dAP):%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat_transpose_gpu(m, n, dA, dA_LD,"dA = LU");
#endif
#if dbglevel>=1
/*Finalize the tracing*/
ca_dbg_trace_finalize();
printf("[DBG] Time llog:%f\n",magma_wtime()-t1);
#endif
return *info;
} /* End of MAGMA_DGETRF_REC_ASYNC_WORK_GPU */
extern "C" magma_int_t
magma_dgetrf(magma_int_t m, magma_int_t n, double *a, magma_int_t lda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
DGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
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.
A (input/output) DOUBLE_PRECISION array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
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.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
#define dAT(i,j) (dAT + (i)*nb*ldda + (j)*nb)
double *dAT, *dA, *da, *work;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t iinfo, nb;
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
nb = magma_get_dgetrf_nb(m);
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_dgetrf(&m, &n, a, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, maxdim;
magma_int_t i, rows, cols, s = min(m, n)/nb;
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
/* set number of GPUs */
magma_int_t num_gpus = magma_num_gpus();
if ( num_gpus > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
/* explicitly checking the memory requirement */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(double);
int h = 1+(2+num_gpus), num_gpus2 = num_gpus;
int NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
char * ngr_nb_char = getenv("MAGMA_NGR_NB");
if( ngr_nb_char != NULL ) NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
if( num_gpus > ceil((double)NB/nb) ) {
num_gpus2 = (int)ceil((double)NB/nb);
h = 1+(2+num_gpus2);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
}
if( num_gpus2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
ldda = maxn;
work = a;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
if (MAGMA_SUCCESS != magma_dmalloc( &dA, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
ldda = maxdim;
magma_dsetmatrix( m, n, a, lda, da, ldda );
dAT = da;
magmablas_dtranspose_inplace( ldda, dAT, ldda );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
if (MAGMA_SUCCESS != magma_dmalloc( &dA, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
magma_dsetmatrix( m, n, a, lda, da, maxm );
if (MAGMA_SUCCESS != magma_dmalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dA );
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
magmablas_dtranspose2( dAT, ldda, da, maxm, m, n );
}
lapackf77_dgetrf( &m, &nb, work, &lda, ipiv, &iinfo);
/* Define user stream if current stream is NULL */
cudaStream_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;
for( i = 0; i < s; i++ )
{
// download i-th panel
cols = maxm - i*nb;
if (i>0){
// download i-th panel
magmablas_dtranspose( dA, cols, dAT(i,i), ldda, nb, cols );
// make sure that gpu queue is empty
magma_device_sync();
magma_dgetmatrix_async( m-i*nb, nb, dA, cols, work, lda,
stream[0]);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (i+1)*nb, nb,
c_one, dAT(i-1,i-1), ldda,
dAT(i-1,i+1), ldda );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-i*nb, nb,
c_neg_one, dAT(i-1,i+1), ldda,
dAT(i, i-1), ldda,
c_one, dAT(i, i+1), ldda );
// do the cpu part
rows = m - i*nb;
magma_queue_sync( stream[0] );
lapackf77_dgetrf( &rows, &nb, work, &lda, ipiv+i*nb, &iinfo);
}
if (*info == 0 && iinfo > 0)
*info = iinfo + i*nb;
// upload i-th panel
magma_dsetmatrix_async( m-i*nb, nb, work, lda, dA, cols,
stream[0]);
magmablas_dpermute_long2( ldda, dAT, ldda, ipiv, nb, i*nb );
magma_queue_sync( stream[0] );
magmablas_dtranspose( dAT(i,i), ldda, dA, cols, cols, nb);
// do the small non-parallel computations
if (s > (i+1)){
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
else{
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_dtranspose2( dA, cols, dAT(s,s), ldda, nb0, rows);
magma_dgetmatrix( rows, nb0, dA, cols, work, lda );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
lapackf77_dgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo);
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
magmablas_dpermute_long2( ldda, dAT, ldda, ipiv, nb0, s*nb );
magma_dsetmatrix( rows, nb0, work, lda, dA, cols );
magmablas_dtranspose2( dAT(s,s), ldda, dA, cols, rows, nb0);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), ldda,
dAT(s, s)+nb0, ldda);
}
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_dtranspose_inplace( ldda, dAT, ldda );
magma_dgetmatrix( m, n, da, ldda, a, lda );
} else {
magmablas_dtranspose2( da, maxm, dAT, ldda, n, m );
magma_dgetmatrix( m, n, da, maxm, a, lda );
magma_free( dAT );
}
magma_free( dA );
magma_queue_destroy( stream[0] );
if (current_stream == NULL) {
magma_queue_destroy( stream[1] );
magmablasSetKernelStream(NULL);
}
}
return *info;
} /* magma_dgetrf */
