/***************************************************************************//**
*
* @ingroup PLASMA_Complex64_t_Tile_Async
*
* PLASMA_zlaswp_Tile_Async - performs a series of row interchanges
* on the matrix A. One row interchange is initiated for each of
* rows K1 through K2 of A.
* Non-blocking equivalent of PLASMA_zlaswp_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_zlaswp
* @sa PLASMA_zlaswp_Tile
* @sa PLASMA_claswp_Tile_Async
* @sa PLASMA_dlaswp_Tile_Async
* @sa PLASMA_slaswp_Tile_Async
* @sa PLASMA_zgetrf_Tile_Async
*
******************************************************************************/
int PLASMA_zlaswp_Tile_Async(PLASMA_desc *A, int K1, int K2, int *IPIV, int INCX,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA = *A;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_zlaswp_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_zlaswp_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_zlaswp_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_zlaswp_Tile", "invalid first descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if ( (K1 != 1) || (K2 != descA.m) ) {
plasma_error("PLASMA_zlaswp_Tile", "invalid K1 or K2 (1..M is the only interval supported right now)");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
plasma_dynamic_call_3(
plasma_pzbarrier_tl2pnl,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
/* swap */
plasma_dynamic_call_5(
plasma_pzlaswp,
PLASMA_desc, descA,
int *, IPIV,
int, INCX,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
plasma_dynamic_call_3(
plasma_pzbarrier_pnl2tl,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
* @ingroup double_Tile_Async
*
* PLASMA_dsygst_Tile_Async - reduces a complex Hermitian-definite
* generalized eigenproblem to standard form.
* If PlasmaItype == 1, the problem is A*x = lambda*B*x, and A is
* overwritten by inv(U**T)*A*inv(U) or inv(L)*A*inv(L**T)
* If PlasmaItype == 2 or 3, the problem is A*B*x = lambda*x or B*A*x
* = lambda*x, and A is overwritten by U*A*U**T or L**T*A*L. B must
* have been previously factorized as U**T*U or L*L**T by
* PLASMA_DPOTRF.
* ONLY PlasmaItype == 1 and PlasmaLower supported!
* Non-blocking equivalent of PLASMA_dsygst_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_dsygst
* @sa PLASMA_dsygst_Tile
* @sa PLASMA_chegst_Tile_Async
* @sa PLASMA_dsygst_Tile_Async
* @sa PLASMA_ssygst_Tile_Async
* @sa PLASMA_dsygv_Tile_Async
*
******************************************************************************/
int PLASMA_dsygst_Tile_Async(PLASMA_enum itype, PLASMA_enum uplo,
PLASMA_desc *A,
PLASMA_desc *B,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA = *A;
PLASMA_desc descB = *B;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_dsygst_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_dsygst_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_dsygst_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_dsygst_Tile", "invalid first descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (plasma_desc_check(&descB) != PLASMA_SUCCESS) {
plasma_error("PLASMA_dsygst_Tile", "invalid second descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_dsygst_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/*
* Transform Hermitian-definite generalized eigenproblem
* to standard form
*/
plasma_dynamic_call_6(plasma_pdsygst,
PLASMA_enum, itype,
PLASMA_enum, uplo,
PLASMA_desc, descA,
PLASMA_desc, descB,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
* @ingroup PLASMA_Complex32_t_Tile_Async
*
* PLASMA_cpotrf_Tile_Async - Computes the Cholesky factorization of a symmetric
* positive definite or Hermitian positive definite matrix.
* Non-blocking equivalent of PLASMA_cpotrf_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_cpotrf
* @sa PLASMA_cpotrf_Tile
* @sa PLASMA_cpotrf_Tile_Async
* @sa PLASMA_dpotrf_Tile_Async
* @sa PLASMA_spotrf_Tile_Async
* @sa PLASMA_cpotrs_Tile_Async
*
******************************************************************************/
int PLASMA_cpotrf_Tile_Async(PLASMA_enum uplo, PLASMA_desc *A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA = *A;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_cpotrf_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_cpotrf_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_cpotrf_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_cpotrf_Tile", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_cpotrf_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (uplo != PlasmaUpper && uplo != PlasmaLower) {
plasma_error("PLASMA_cpotrf_Tile", "illegal value of uplo");
return plasma_request_fail(sequence, request, -1);
}
/* Quick return */
/*
if (max(N, 0) == 0)
return PLASMA_SUCCESS;
*/
plasma_parallel_call_4(plasma_pcpotrf,
PLASMA_enum, uplo,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
* @ingroup float_Tile_Async
*
* PLASMA_splgsy_Tile_Async - Generate a random hermitian matrix by tiles.
* Non-blocking equivalent of PLASMA_splgsy_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_splgsy
* @sa PLASMA_splgsy_Tile
* @sa PLASMA_cplgsy_Tile_Async
* @sa PLASMA_dplgsy_Tile_Async
* @sa PLASMA_splgsy_Tile_Async
* @sa PLASMA_splgsy_Tile_Async
* @sa PLASMA_splgsy_Tile_Async
*
******************************************************************************/
int PLASMA_splgsy_Tile_Async( float bump,
PLASMA_desc *A,
unsigned long long int seed,
PLASMA_sequence *sequence,
PLASMA_request *request)
{
PLASMA_desc descA = *A;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_splgsy_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_splgsy_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_splgsy_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_splgsy_Tile", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_splgsy_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Quick return */
if (min( descA.m, descA.n ) == 0)
return PLASMA_SUCCESS;
plasma_parallel_call_5(plasma_psplgsy,
float, bump,
PLASMA_desc, descA,
unsigned long long int, seed,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
* @ingroup PLASMA_Complex64_t_Tile_Async
*
* PLASMA_zgetrf_nopiv_Tile_Async - Computes the tile LU factorization of a
* matrix. Non-blocking equivalent of PLASMA_zgetrf_nopiv_Tile(). May return
* before the computation is finished. Allows for pipelining of operations ar
* runtime.
*
*******************************************************************************
*
* @param[in,out] A
* On entry, the M-by-N matrix to be factored.
* On exit, the tile factors L and U from the factorization.
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_zgetrf_nopiv
* @sa PLASMA_zgetrf_nopiv_Tile
* @sa PLASMA_cgetrf_nopiv_Tile_Async
* @sa PLASMA_dgetrf_nopiv_Tile_Async
* @sa PLASMA_sgetrf_nopiv_Tile_Async
* @sa PLASMA_zgetrs_Tile_Async
*
******************************************************************************/
int PLASMA_zgetrf_nopiv_Tile_Async(PLASMA_desc *A,
PLASMA_sequence *sequence,
PLASMA_request *request)
{
PLASMA_desc descA;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_zgetrf_nopiv_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_zgetrf_nopiv_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_zgetrf_nopiv_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(A) != PLASMA_SUCCESS) {
plasma_error("PLASMA_zgetrf_nopiv_Tile", "invalid first descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
} else {
descA = *A;
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_zgetrf_nopiv_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
plasma_dynamic_call_3(plasma_pzgetrf_nopiv,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
**/
void plasma_pdlag2s(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc SB;
PLASMA_sequence *sequence;
PLASMA_request *request;
int X, Y;
int m, n;
int next_m;
int next_n;
int ldam, ldbm;
int info = PLASMA_SUCCESS;
plasma_unpack_args_4(A, SB, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
n = 0;
m = PLASMA_RANK;
while (m >= A.mt && n < A.nt) {
n++;
m = m-A.mt;
}
while (n < A.nt) {
next_m = m;
next_n = n;
next_m += PLASMA_SIZE;
while (next_m >= A.mt && next_n < A.nt) {
next_n++;
next_m = next_m-A.mt;
}
X = m == A.mt-1 ? A.m-A.mb*m : A.nb;
Y = n == A.nt-1 ? A.n-A.nb*n : A.nb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(SB, m);
CORE_dlag2s(X, Y, A(m, n), ldam, SB(m, n), ldbm, &info);
if (info != 0)
plasma_request_fail(sequence, request, info);
m = next_m;
n = next_n;
}
}
/***************************************************************************//**
*
* @ingroup float_Tile_Async
*
* PLASMA_ssytrd_Tile_Async - Computes all eigenvalues and,
* optionally, eigenvectors of a complex Hermitian matrix A using a
* two-stage approach:
* First stage: reduction to band tridiagonal form;
* Second stage: reduction from band to tridiagonal form.
*
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_ssytrd
* @sa PLASMA_ssytrd_Tile
* @sa PLASMA_chetrd_Tile_Async
* @sa PLASMA_dsytrd_Tile_Async
* @sa PLASMA_ssytrd_Tile_Async
*
******************************************************************************/
int PLASMA_ssytrd_Tile_Async(PLASMA_enum jobz, PLASMA_enum uplo,
PLASMA_desc *A,
float *D,
float *E,
PLASMA_desc *T,
PLASMA_desc *Q,
PLASMA_sequence *sequence, PLASMA_request *request)
{
int NB, IB, IBNB, NT;
PLASMA_desc descA = *A;
PLASMA_desc descT = *T;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_ssytrd_Tile_Async", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_ssytrd_Tile_Async", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_ssytrd_Tile_Async", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Set NT & NTRHS */
NB = PLASMA_NB;
IB = PLASMA_IB;
IBNB = IB*NB;
NT = (descA.ln%NB==0) ? (descA.ln/NB) : (descA.ln/NB+1);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_ssytrd_Tile_Async", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (plasma_desc_check(&descT) != PLASMA_SUCCESS) {
plasma_error("PLASMA_ssytrd_Tile_Async", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if ( (jobz == PlasmaVec) && (plasma_desc_check(Q) != PLASMA_SUCCESS) ) {
plasma_error("PLASMA_ssytrd_Tile_Async", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Check input arguments */
if (jobz != PlasmaNoVec && jobz != PlasmaVec) {
plasma_error("PLASMA_ssytrd_Tile_Async", "illegal value of jobz");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (descA.m != descA.n) {
plasma_error("PLASMA_ssytrd_Tile_Async", "matrix need to be square");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (descA.nb != descA.mb) {
plasma_error("PLASMA_ssytrd_Tile_Async", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (jobz == PlasmaVec) {
plasma_error("PLASMA_ssytrd_Tile_Async", "computing the eigenvectors is not supported in this version");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if ( (jobz == PlasmaVec) && (Q->nb != Q->mb) ) {
plasma_error("PLASMA_ssytrd_Tile_Async", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Reduction to tridiagonal form
* with a two-stage approach.
*/
/* Reduction to BAND tridiagonal form
*/
plasma_dynamic_call_5(plasma_pssyrbt,
PLASMA_enum, uplo,
PLASMA_desc, descA,
PLASMA_desc, descT,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
/*
* Build the Q of the first stage
*/
/* if (jobz == PlasmaVec){ */
/* /\* Initialize Q to Identity *\/ */
/* plasma_dynamic_call_6(plasma_pslaset, */
/* PLASMA_enum, PlasmaUpperLower, */
/* float, 0.0, */
/* float, 1.0, */
/* PLASMA_desc, descQ, */
/* PLASMA_sequence*, sequence, */
/* PLASMA_request*, request); */
/* /\* Accumulate the transformations from the first stage*\/ */
/* plasma_dynamic_call_6(plasma_psorgtr, */
/* PLASMA_enum, uplo, */
/* PLASMA_desc, descA, */
/* PLASMA_desc, descQ, */
/* PLASMA_desc, descT, */
/* PLASMA_sequence*, sequence, */
/* PLASMA_request*, request); */
/* } */
/* Set the V's to zero before the 2nd stage (bulge chasing) */
/*
*/
plasma_dynamic_call_5(plasma_pslaset2,
PLASMA_enum, uplo,
float, 0.0,
PLASMA_desc, uplo == PlasmaLower ? plasma_desc_submatrix(descA, descA.mb, 0, descA.m-descA.mb, descA.n-descA.nb)
: plasma_desc_submatrix(descA, 0, descA.nb, descA.m-descA.mb, descA.n-descA.nb),
PLASMA_sequence*, sequence,
PLASMA_request*, request);
/* Reduction from BAND tridiagonal to the final condensed form
*/
plasma_dynamic_call_7(plasma_pssbrdt,
PLASMA_enum, uplo,
PLASMA_desc, descA,
float*, D,
float*, E,
PLASMA_desc, descT,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
* Parallel tile LU factorization - static scheduling
**/
void plasma_pzgetrf_incpiv(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc L;
int *IPIV;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_k;
int next_m;
int next_n;
int ldak, ldam;
int info;
int tempkn, tempkm, tempmm, tempnn;
int ib = PLASMA_IB;
PLASMA_Complex64_t *work;
plasma_unpack_args_5(A, L, IPIV, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
work = (PLASMA_Complex64_t*)plasma_private_alloc(plasma, ib*L.nb, L.dtyp);
ss_init(A.mt, A.nt, -1);
k = 0;
n = PLASMA_RANK;
while (n >= A.nt) {
k++;
n = n-A.nt+k;
}
m = k;
while (k < min(A.mt, A.nt) && n < A.nt && !ss_aborted()) {
next_n = n;
next_m = m;
next_k = k;
next_m++;
if (next_m == A.mt) {
next_n += PLASMA_SIZE;
while (next_n >= A.nt && next_k < min(A.mt, A.nt)) {
next_k++;
next_n = next_n-A.nt+next_k;
}
next_m = next_k;
}
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldam = BLKLDD(A, m);
if (n == k) {
if (m == k) {
ss_cond_wait(k, k, k-1);
CORE_zgetrf_incpiv(
tempkm, tempkn, ib,
A(k, k), ldak,
IPIV(k, k), &info);
if (info != 0 && m == A.mt-1) {
plasma_request_fail(sequence, request, info + A.nb*k);
ss_abort();
}
ss_cond_set(k, k, k);
}
else {
ss_cond_wait(m, k, k-1);
CORE_ztstrf(
tempmm, tempkn, ib, A.nb,
A(k, k), ldak,
A(m, k), ldam,
L(m, k), L.mb,
IPIV(m, k),
work, L.nb, &info);
if (info != 0 && m == A.mt-1) {
plasma_request_fail(sequence, request, info + A.nb*k);
ss_abort();
}
ss_cond_set(m, k, k);
}
}
else {
if (m == k) {
ss_cond_wait(k, k, k);
ss_cond_wait(k, n, k-1);
CORE_zgessm(
tempkm, tempnn, tempkm, ib,
IPIV(k, k),
A(k, k), ldak,
A(k, n), ldak);
}
else {
ss_cond_wait(m, k, k);
ss_cond_wait(m, n, k-1);
CORE_zssssm(
A.nb, tempnn, tempmm, tempnn, A.nb, ib,
A(k, n), ldak,
A(m, n), ldam,
L(m, k), L.mb,
A(m, k), ldam,
IPIV(m, k));
ss_cond_set(m, n, k);
}
}
n = next_n;
m = next_m;
k = next_k;
}
plasma_private_free(plasma, work);
ss_finalize();
}
/***************************************************************************//**
* Parallel application of Q using tile V - LQ factorization - static scheduling
**/
void plasma_pzunmlq(plasma_context_t *plasma)
{
PLASMA_enum side;
PLASMA_enum trans;
PLASMA_desc A;
PLASMA_desc B;
PLASMA_desc T;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_k;
int next_m;
int next_n;
int ldak, ldbk, ldbm;
int tempmm, tempnn, tempkm, tempkmin;
int minMT, minM;
int ib = PLASMA_IB;
PLASMA_Complex64_t *work;
plasma_unpack_args_7(side, trans, A, B, T, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
if (side != PlasmaLeft) {
plasma_request_fail(sequence, request, PLASMA_ERR_NOT_SUPPORTED);
return;
}
if (trans != PlasmaConjTrans) {
plasma_request_fail(sequence, request, PLASMA_ERR_NOT_SUPPORTED);
return;
}
work = (PLASMA_Complex64_t*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
ss_init(B.mt, B.nt, min(A.mt, A.nt));
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
k = minMT-1;
n = PLASMA_RANK;
while (n >= B.nt) {
k--;
n = n-B.nt;
}
m = B.mt-1;
while (k >= 0 && n < B.nt) {
next_n = n;
next_m = m;
next_k = k;
next_m--;
if (next_m == k-1) {
next_n += PLASMA_SIZE;
while (next_n >= B.nt && next_k >= 0) {
next_k--;
next_n = next_n-B.nt;
}
next_m = B.mt-1;
}
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
tempkm = k == B.mt-1 ? B.m-k*B.mb : B.mb;
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
ldbm = BLKLDD(B, m);
if (m == k) {
CORE_zunmlq(
side, trans,
tempkm, tempnn, tempkmin, ib,
A(k, k), ldak,
T(k, k), T.mb,
B(k, n), ldbk,
work, T.nb);
ss_cond_set(k, n, k);
}
else {
ss_cond_wait(m, n, k+1);
CORE_ztsmlq(
side, trans,
A.mb, tempnn, tempmm, tempnn, tempkmin, ib,
B(k, n), ldbk,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb,
work, ib);
ss_cond_set(m, n, k);
}
m = next_m;
n = next_n;
k = next_k;
}
plasma_private_free(plasma, work);
ss_finalize();
}
/***************************************************************************//**
*
* @ingroup PLASMA_Complex64_t_Tile_Async
*
* PLASMA_zgetri_Tile_Async - Computes the inverse of a matrix using the LU
* factorization computed by PLASMA_zgetrf.
* This method inverts U and then computes inv(A) by solving the system
* inv(A)*L = inv(U) for inv(A).
* Non-blocking equivalent of PLASMA_zgetri_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations at runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_zgetri
* @sa PLASMA_zgetri_Tile
* @sa PLASMA_cgetri_Tile_Async
* @sa PLASMA_dgetri_Tile_Async
* @sa PLASMA_sgetri_Tile_Async
* @sa PLASMA_zgetrf_Tile_Async
*
******************************************************************************/
int PLASMA_zgetri_Tile_Async(PLASMA_desc *A, int *IPIV, PLASMA_desc *W,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA;
PLASMA_desc descW;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_zgetri_Tile_Async", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_zgetri_Tile_Async", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_zgetri_Tile_Async", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(A) != PLASMA_SUCCESS) {
plasma_error("PLASMA_zgetri_Tile_Async", "invalid A descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
} else {
descA = *A;
}
/* Check descriptors for correctness */
if (plasma_desc_check(W) != PLASMA_SUCCESS) {
plasma_error("PLASMA_zgetri_Tile_Async", "invalid W descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
} else {
descW = *W;
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_zgetri_Tile_Async", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Quick return */
if (max(descA.m, 0) == 0)
return PLASMA_SUCCESS;
plasma_dynamic_call_5(plasma_pztrtri,
PLASMA_enum, PlasmaUpper,
PLASMA_enum, PlasmaNonUnit,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
plasma_dynamic_call_9(plasma_pztrsmrv,
PLASMA_enum, PlasmaRight,
PLASMA_enum, PlasmaLower,
PLASMA_enum, PlasmaNoTrans,
PLASMA_enum, PlasmaUnit,
PLASMA_Complex64_t, (PLASMA_Complex64_t) 1.0,
PLASMA_desc, descA,
PLASMA_desc, descW,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
/* No need for barrier tile2row because of previous dependencies */
/* swap */
plasma_dynamic_call_5(
plasma_pzlaswpc,
PLASMA_desc, descA,
int *, IPIV,
int, -1,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
plasma_dynamic_call_3(
plasma_pzbarrier_row2tl,
PLASMA_desc, descA,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
return PLASMA_SUCCESS;
}
/***************************************************************************//**
*
* @ingroup PLASMA_Complex64_t_Tile_Async
*
* PLASMA_zlansy_Tile_Async - Non-blocking equivalent of PLASMA_zlansy_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations at runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_zlansy
* @sa PLASMA_zlansy_Tile
* @sa PLASMA_clansy_Tile_Async
* @sa PLASMA_dlansy_Tile_Async
* @sa PLASMA_slansy_Tile_Async
*
******************************************************************************/
int PLASMA_zlansy_Tile_Async(PLASMA_enum norm, PLASMA_enum uplo, PLASMA_desc *A, double *value,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA;
double *work = NULL;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_zlansy_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_zlansy_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_zlansy_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(A) != PLASMA_SUCCESS) {
plasma_error("PLASMA_zlansy_Tile", "invalid descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
} else {
descA = *A;
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_zlansy_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if ( (norm != PlasmaMaxNorm) && (norm != PlasmaOneNorm)
&& (norm != PlasmaInfNorm) && (norm != PlasmaFrobeniusNorm) ) {
plasma_error("PLASMA_zlansy_Tile", "illegal value of norm");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if ( (uplo != PlasmaUpper) && (uplo != PlasmaLower) ) {
plasma_error("PLASMA_zlansy_Tile", "illegal value of uplo");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Quick return */
if ( descA.m == 0) {
*value = 0.0;
return PLASMA_SUCCESS;
}
if (PLASMA_SCHEDULING == PLASMA_STATIC_SCHEDULING) {
if (norm == PlasmaFrobeniusNorm) {
work = plasma_shared_alloc(plasma, 2*PLASMA_SIZE, PlasmaRealDouble );
} else {
work = plasma_shared_alloc(plasma, PLASMA_SIZE, PlasmaRealDouble );
}
}
plasma_parallel_call_7(plasma_pzlansy,
PLASMA_enum, norm,
PLASMA_enum, uplo,
PLASMA_desc, descA,
double*, work,
double*, value,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
if (work != NULL)
plasma_shared_free( plasma, work );
return PLASMA_SUCCESS;
}
/** ****************************************************************************
*
* @ingroup InPlaceTransformation
*
* plasma_dgetmi2 Implementation of inplace transposition
* based on the GKK algorithm by Gustavson, Karlsson, Kagstrom.
* This algorithm shift some cycles to transpose the matrix.
*
*******************************************************************************
*
* @param[in] m
* Number of rows of matrix A
*
* @param[in] n
* Number of columns of matrix A
*
* @param[in,out] A
* Matrix of size L*m*n
*
* @param[in] nprob
* Number of parallel and independant problems
*
* @param[in] me
* Number of rows of the problem
*
* @param[in] ne
* Number of columns in the problem
*
* @param[in] L
* Size of chunk to use for transformation
*
******************************************************************************/
int plasma_dshift(plasma_context_t *plasma, int m, int n, double *A,
int nprob, int me, int ne, int L,
PLASMA_sequence *sequence, PLASMA_request *request)
{
int *leaders = NULL;
int ngrp, thrdbypb, thrdtot, nleaders;
/* Check Plasma context */
thrdtot = PLASMA_SIZE;
thrdbypb = PLASMA_GRPSIZE;
ngrp = thrdtot/thrdbypb;
/* check input */
if( (nprob * me * ne * L) != (m * n) ) {
plasma_error(__func__, "problem size does not match matrix size");
/*printf("m=%d, n=%d, nprob=%d, me=%d, ne=%d, L=%d\n", m, n, nprob, me, ne, L);*/
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if( thrdbypb > thrdtot ) {
plasma_error(__func__, "number of thread per problem must be less or equal to total number of threads");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if( (thrdtot % thrdbypb) != 0 ) {
plasma_error(__func__, "number of thread per problem must divide the total number of thread");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* quick return */
if( (me < 2) || (ne < 2) || (nprob < 1) ) {
return PLASMA_SUCCESS;
}
GKK_getLeaderNbr(me, ne, &nleaders, &leaders);
nleaders *= 3;
if (PLASMA_SCHEDULING == PLASMA_STATIC_SCHEDULING) {
int *Tp = NULL;
int i, ipb;
int q, owner;
q = me*ne - 1;
Tp = (int *)plasma_shared_alloc(plasma, thrdtot, PlasmaInteger);
for (i=0; i<thrdtot; i++)
Tp[i] = 0;
ipb = 0;
/* First part with coarse parallelism */
if (nprob > ngrp) {
ipb = (nprob / ngrp)*ngrp;
/* loop over leader */
if (thrdbypb > 1) {
for (i=0; i<nleaders; i+=3) {
/* assign this cycle to a thread */
owner = minloc(thrdbypb, Tp);
/* assign it to owner */
Tp[owner] = Tp[owner] + leaders[i+1] * L;
leaders[i+2] = owner;
}
GKK_BalanceLoad(thrdbypb, Tp, leaders, nleaders, L);
}
else {
for (i=0; i<nleaders; i+=3) {
Tp[0] = Tp[0] + leaders[i+1] * L;
leaders[i+2] = 0;
}
}
/* shift in parallel */
for (i=0; i< (nprob/ngrp); i++) {
plasma_static_call_9(plasma_pdshift,
int, me,
int, ne,
int, L,
double*, &(A[i*ngrp*me*ne*L]),
int *, leaders,
int, nleaders,
int, thrdbypb,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
}
}
/***************************************************************************//**
* Parallel tile Cholesky factorization - static scheduling
**/
void plasma_pspotrf(plasma_context_t *plasma)
{
PLASMA_enum uplo;
PLASMA_desc A;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_k;
int next_m;
int next_n;
int ldak, ldam, ldan;
int info;
int tempkn, tempmn;
float zone = (float) 1.0;
float mzone = (float)-1.0;
plasma_unpack_args_4(uplo, A, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
ss_init(A.nt, A.nt, 0);
k = 0;
m = PLASMA_RANK;
while (m >= A.nt) {
k++;
m = m-A.nt+k;
}
n = 0;
while (k < A.nt && m < A.nt && !ss_aborted()) {
next_n = n;
next_m = m;
next_k = k;
next_n++;
if (next_n > next_k) {
next_m += PLASMA_SIZE;
while (next_m >= A.nt && next_k < A.nt) {
next_k++;
next_m = next_m-A.nt+next_k;
}
next_n = 0;
}
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
tempmn = m == A.nt-1 ? A.n-m*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldan = BLKLDD(A, n);
ldam = BLKLDD(A, m);
if (m == k) {
if (n == k) {
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
CORE_spotrf(
PlasmaLower,
tempkn,
A(k, k), ldak,
&info);
}
/*
* PlasmaUpper
*/
else {
CORE_spotrf(
PlasmaUpper,
tempkn,
A(k, k), ldak,
&info);
}
if (info != 0) {
plasma_request_fail(sequence, request, info + A.nb*k);
ss_abort();
}
ss_cond_set(k, k, 1);
}
else {
ss_cond_wait(k, n, 1);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
CORE_ssyrk(
PlasmaLower, PlasmaNoTrans,
tempkn, A.nb,
-1.0, A(k, n), ldak,
1.0, A(k, k), ldak);
}
/*
* PlasmaUpper
*/
else {
CORE_ssyrk(
PlasmaUpper, PlasmaTrans,
tempkn, A.nb,
-1.0, A(n, k), ldan,
1.0, A(k, k), ldak);
}
}
}
else {
if (n == k) {
ss_cond_wait(k, k, 1);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
CORE_strsm(
PlasmaRight, PlasmaLower, PlasmaTrans, PlasmaNonUnit,
tempmn, A.nb,
zone, A(k, k), ldak,
A(m, k), ldam);
}
/*
* PlasmaUpper
*/
else {
CORE_strsm(
PlasmaLeft, PlasmaUpper, PlasmaTrans, PlasmaNonUnit,
A.nb, tempmn,
zone, A(k, k), ldak,
A(k, m), ldak);
}
ss_cond_set(m, k, 1);
}
else {
ss_cond_wait(k, n, 1);
ss_cond_wait(m, n, 1);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
CORE_sgemm(
PlasmaNoTrans, PlasmaTrans,
tempmn, A.nb, A.nb,
mzone, A(m, n), ldam,
A(k, n), ldak,
zone, A(m, k), ldam);
}
/*
* PlasmaUpper
*/
else {
CORE_sgemm(
PlasmaTrans, PlasmaNoTrans,
A.nb, tempmn, A.nb,
mzone, A(n, k), ldan,
A(n, m), ldan,
zone, A(k, m), ldak);
}
}
}
n = next_n;
m = next_m;
k = next_k;
}
ss_finalize();
}
/***************************************************************************//**
*
* @ingroup float_Tile_Async
*
* PLASMA_sgelqf_Tile_Async - Computes the tile LQ factorization of a matrix.
* Non-blocking equivalent of PLASMA_sgelqf_Tile().
* May return before the computation is finished.
* Allows for pipelining of operations ar runtime.
*
*******************************************************************************
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa PLASMA_sgelqf
* @sa PLASMA_sgelqf_Tile
* @sa PLASMA_cgelqf_Tile_Async
* @sa PLASMA_dgelqf_Tile_Async
* @sa PLASMA_sgelqf_Tile_Async
* @sa PLASMA_sgelqs_Tile_Async
*
******************************************************************************/
int PLASMA_sgelqf_Tile_Async(PLASMA_desc *A, PLASMA_desc *T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_desc descA = *A;
PLASMA_desc descT = *T;
plasma_context_t *plasma;
plasma = plasma_context_self();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_sgelqf_Tile", "PLASMA not initialized");
return PLASMA_ERR_NOT_INITIALIZED;
}
if (sequence == NULL) {
plasma_fatal_error("PLASMA_sgelqf_Tile", "NULL sequence");
return PLASMA_ERR_UNALLOCATED;
}
if (request == NULL) {
plasma_fatal_error("PLASMA_sgelqf_Tile", "NULL request");
return PLASMA_ERR_UNALLOCATED;
}
/* Check sequence status */
if (sequence->status == PLASMA_SUCCESS)
request->status = PLASMA_SUCCESS;
else
return plasma_request_fail(sequence, request, PLASMA_ERR_SEQUENCE_FLUSHED);
/* Check descriptors for correctness */
if (plasma_desc_check(&descA) != PLASMA_SUCCESS) {
plasma_error("PLASMA_sgelqf_Tile", "invalid first descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
if (plasma_desc_check(&descT) != PLASMA_SUCCESS) {
plasma_error("PLASMA_sgelqf_Tile", "invalid second descriptor");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Check input arguments */
if (descA.nb != descA.mb) {
plasma_error("PLASMA_sgelqf_Tile", "only square tiles supported");
return plasma_request_fail(sequence, request, PLASMA_ERR_ILLEGAL_VALUE);
}
/* Quick return */
/*
if (min(M, N) == 0)
return PLASMA_SUCCESS;
*/
if (plasma->householder == PLASMA_FLAT_HOUSEHOLDER) {
plasma_parallel_call_4(plasma_psgelqf,
PLASMA_desc, descA,
PLASMA_desc, descT,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
}
else {
plasma_dynamic_call_5(plasma_psgelqfrh,
PLASMA_desc, descA,
PLASMA_desc, descT,
PLASMA_enum, PLASMA_RHBLK,
PLASMA_sequence*, sequence,
PLASMA_request*, request);
}
return PLASMA_SUCCESS;
}
