/***************************************************************************//**
* Parallel tile row interchanges - dynamic scheduling
**/
void plasma_pclaswp_quark(PLASMA_desc B, int *IPIV, int inc,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int m, n;
int tempi, tempm, tempmm, tempnn;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
if ( inc > 0 )
{
for (m = 0; m < B.mt; m++) {
tempi = m * B.mb;
tempm = B.m - tempi;
tempmm = m == B.mt-1 ? tempm : B.mb;
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n - n * B.nb : B.nb;
QUARK_CORE_claswp_ontile(
plasma->quark, &task_flags,
plasma_desc_submatrix(B, tempi, n*B.nb, tempm, tempnn),
B(m, n), 1, tempmm, IPIV(m), inc, B(B.mt-1, n) );
}
}
}
else
{
for (m = B.mt-1; m > -1; m--) {
tempi = m * B.mb;
tempm = B.m - tempi;
tempmm = m == B.mt-1 ? tempm : B.mb;
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n - n * B.nb : B.nb;
QUARK_CORE_claswp_ontile(
plasma->quark, &task_flags,
plasma_desc_submatrix(B, tempi, n*B.nb, tempm, tempnn),
B(m, n), 1, tempmm, IPIV(m), inc, B(0, 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 - dynamic scheduling - Right looking
**/
void plasma_pdgetrf_rectil_quark(PLASMA_desc A, int *IPIV)
{
int k, m, n;
int tempk, tempm, tempkm, tempkn, tempmm, tempnn;
int ldak, ldam;
double zone = (double)1.0;
double mzone = (double)-1.0;
void * fakedep;
/* How many threads per panel? Probably needs to be adjusted during factorization. */
CORE_dgetrf_rectil_init();
for (k = 0; k < min(A.mt, A.nt); k++)
{
tempk = k * A.mb;
tempm = A.m - tempk;
tempkm = k == A.mt-1 ? tempm : A.mb;
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
double *dA = A(k, k);
int *dB = IPIV(k);
PLASMA_desc pDesc = plasma_desc_submatrix(A, tempk, k*A.nb, tempm, tempkn);
hclib_pragma_marker("omp", "task depend(inout:dA[0:A.mb*A.nb]) depend(out:dB[0:pDesc.n])", "pragma59_omp_task");
{
int info[3];
info[1] = 0;
info[2] = 1;
CORE_dgetrf_rectil( pDesc, dB, info );
}
/*
* Update the trailing submatrix
*/
fakedep = (void *)(intptr_t)(k+1);
for (n = k+1; n < A.nt; n++)
{
/*
* Apply row interchange after the panel (work on the panel)
*/
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
PLASMA_desc descA = plasma_desc_submatrix(A, tempk, n*A.nb, tempm, tempnn);
double *dA = A(k, n);
double *dB = A(k, k);
int *dipiv = IPIV(k);
hclib_pragma_marker("omp", "task depend(inout:dA[0:1]) depend(in:dB[0:ldak], dipiv[0:tempkm])", "pragma82_omp_task");
CORE_dswptr_ontile(descA, 1, tempkm, dipiv, 1, dB, ldak);
m = k+1;
if ( m < A.mt ) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
double *dA = A(m , k);
double *dB = A(k , n);
double *dC = A(m , n);
hclib_pragma_marker("omp", "task depend(in:dA[0:A.mb*A.mb], dB[0:A.mb*A.mb]) depend(inout:dC[0:A.mb*A.mb])", "pragma93_omp_task");
cblas_dgemm(CblasColMajor, (CBLAS_TRANSPOSE)PlasmaNoTrans, (CBLAS_TRANSPOSE)PlasmaNoTrans,
tempmm, tempnn, A.nb,
mzone, dA, ldam,
dB, ldak,
zone, dC, ldam);
for (m = k+2; m < A.mt; m++)
{
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
double *dA = A(m , k);
double *dB = A(k , n);
double *dC = A(m , n);
double *fake1 = A(k+1, n);
double *fake2 = (double *)fakedep;
hclib_pragma_marker("omp", "task depend(in:dA[0:A.mb*A.mb], dB[0:A.mb*A.mb], fake2[0:1]) depend(inout:dC[0:A.mb*A.mb], fake1[0:A.mb*A.nb])", "pragma110_omp_task");
cblas_dgemm(CblasColMajor, (CBLAS_TRANSPOSE)PlasmaNoTrans, (CBLAS_TRANSPOSE)PlasmaNoTrans,
tempmm, tempnn, A.nb,
mzone, dA, ldam,
dB, ldak,
zone, dC, ldam);
}
}
}
}
for (k = 0; k < min(A.mt, A.nt); k++)
{
int mintmp;
tempk = k * A.mb;
tempm = A.m - tempk;
tempkm = k == A.mt-1 ? tempm : A.mb;
tempkn = k == A.nt-1 ? A.n - k * A.nb : A.nb;
mintmp = min(tempkm, tempkn);
ldak = BLKLDD(A, k);
/*
* Apply row interchange behind the panel (work on the panel)
*/
fakedep = (void*)(intptr_t)k;
for (n = 0; n < k; n++)
{
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
double *Aij = A(k, n);
double *prevSwap = A(k-1, n);
int *dipiv = IPIV(k);
PLASMA_desc descA = plasma_desc_submatrix(A, tempk, n*A.nb, tempm, tempnn);
hclib_pragma_marker("omp", "task depend(inout:Aij[0:1],fakedep) depend(in:dipiv[0:mintmp], prevSwap[0:A.lm*A.nb])", "pragma142_omp_task");
CORE_dlaswp_ontile(descA, 1, mintmp, dipiv, 1);
}
}
}
