int CORE_dswptr_ontile(PLASMA_desc descA, int i1, int i2, const int *ipiv, int inc,
const double *Akk, int ldak)
{
double zone = 1.0;
int lda;
int m = descA.mt == 1 ? descA.m : descA.mb;
if ( descA.nt > 1 ) {
coreblas_error(1, "Illegal value of descA.nt");
return -1;
}
if ( i1 < 1 ) {
coreblas_error(2, "Illegal value of i1");
return -2;
}
if ( (i2 < i1) || (i2 > m) ) {
coreblas_error(3, "Illegal value of i2");
return -3;
}
CORE_dlaswp_ontile(descA, i1, i2, ipiv, inc);
lda = BLKLDD(descA, 0);
cblas_dtrsm( CblasColMajor, CblasLeft, CblasLower,
CblasNoTrans, CblasUnit,
m, descA.n, (zone),
Akk, ldak,
A(0, 0), lda );
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);
}
}
}
