/***************************************************************************//**
* Parallel construction of Q using tile V (application to identity;
* reduction Householder) - dynamic scheduling
**/
void plasma_pzunglqrh_quark(PLASMA_desc A, PLASMA_desc Q, PLASMA_desc T, int BS,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int K, N, RD, lastRD;
int ldak;
int ldqm;
int tempkm, tempNn, tempnn, tempmm, tempNRDn, tempkmin;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
K = min(A.mt, A.nt);
for (k = K-1; k >= 0; k--) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
lastRD = 0;
for (RD = BS; RD < A.nt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; RD /= 2) {
for (N = k;
N+RD < A.nt-1; /* No reduction with rightmost single-column subdomain */
N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
for (m = 0; m < Q.mt; m++) {
tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
ldqm = BLKLDD(Q, m );
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempmm, Q.nb, tempmm, tempNRDn,
tempkm, ib, T.nb,
Q (m, N ), ldqm,
Q (m, N+RD), ldqm,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
for (N = k;
N < A.nt-1 || N == k; /* No rightmost single-column */
N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm, tempNn);
for (n = N+BS-1 == A.nt-2 ? A.nt-1 : min(N+BS-1, A.nt-1); /* Suck in rightmost single-column domain */
n >= N+1;
n--) {
tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
for (m = 0; m < Q.mt; m++) {
tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
ldqm = BLKLDD(Q, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempmm, Q.nb, tempmm, tempnn,
tempkm, ib, T.nb,
Q(m, N), ldqm,
Q(m, n), ldqm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
for (m = 0; m < Q.mt; m++) {
tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
ldqm = BLKLDD(Q, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempmm, tempNn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
Q(m, N), ldqm);
}
}
}
}
/***************************************************************************//**
* Parallel application of Q using tile V - LQ factorization (reduction
* Householder) - dynamic scheduling
**/
void plasma_pzunmlqrh_quark(PLASMA_enum side, PLASMA_enum trans,
PLASMA_desc A, PLASMA_desc B, PLASMA_desc T, int BS,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int K, N, RD, lastRD;
int ldaN, ldak;
int ldbN, ldbm, ldbNRD;
int tempNn, tempkm, tempnn, tempmm, tempNRDn, tempkmin;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
K = min(A.mt, A.nt);
if (side == PlasmaLeft ) {
if (trans == PlasmaNoTrans) {
/*
* PlasmaLeft / PlasmaNoTrans
*/
for (k = 0; k < K; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
ldbN = BLKLDD(B, N);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempNn, tempnn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(N, n), ldbN);
}
for (m = N+1; m < min(N+BS, A.nt); m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.nb, tempnn, tempmm, tempnn,
tempkm, ib, T.nb,
B(N, n), ldbN,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
}
for (RD = BS; RD < A.nt-k; RD *= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
ldbN = BLKLDD(B, N );
ldbNRD = BLKLDD(B, N+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempNRDn, tempnn,
tempkm, ib, T.nb,
B (N, n), ldbN,
B (N+RD, n), ldbNRD,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
}
} else {
/*
* PlasmaLeft / PlasmaConjTrans
*/
for (k = K-1; k >= 0; k--) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
lastRD = 0;
for (RD = BS; RD < A.nt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; RD /= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
ldbN = BLKLDD(B, N );
ldbNRD = BLKLDD(B, N+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
B.nb, tempnn, tempNRDn, tempnn,
tempkm, ib, T.nb,
B (N, n), ldbN,
B (N+RD, n), ldbNRD,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
ldbN = BLKLDD(B, N);
for (m = min(N+BS, A.nt)-1; m > N; m--) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn,
tempkm, ib, T.nb,
B(N, n), ldbN,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempNn, tempnn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(N, n), ldbN);
}
}
}
}
} else {
if (trans == PlasmaNoTrans) {
/*
* PlasmaRight / PlasmaNoTrans
*/
for (k = K-1; k >= 0; k--) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
lastRD = 0;
for (RD = BS; RD < A.nt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; RD /= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
for (m = 0; m < B.mt; m++) {
ldbm = BLKLDD(B, m);
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempNRDn,
tempkm, ib, T.nb,
B (m, N ), ldbm,
B (m, N+RD), ldbm,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
for (n = min(N+BS, A.nt)-1; n > N; n--) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn,
tempkm, ib, T.nb,
B(m, N), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(m, N), ldbm);
}
}
}
} else {
/*
* PlasmaRight / PlasmaConjTrans
*/
for (k = 0; k < K; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
for (m = 0; m < B.mt; m++) {
ldbm = BLKLDD(B, m);
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn,
tempkmin, ib, T.nb,
A(k, N), ldaN,
T(k, N), T.mb,
B(m, N), ldbm);
}
for (n = N+1; n < min(N+BS, A.nt); n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn, tempmm, tempnn,
tempkm, ib, T.nb,
B(m, N), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
for (RD = BS; RD < A.nt-k; RD *= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempNRDn,
tempkm, ib, T.nb,
B (m, N ), ldbm,
B (m, N+RD), ldbm,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile LQ factorization (reduction Householder) - dynamic scheduling
**/
void plasma_pzgelqfrh_quark(PLASMA_desc A, PLASMA_desc T, int BS,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int K, N, RD;
int ldak, ldam;
int tempkm, tempNn, tempmm, tempnn, tempNRDn;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
K = min(A.mt, A.nt);
for (k = 0; k < K; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
// for (N = k; N < A.nt; N += BS) {
for (N = k;
N < A.nt-1 || N == k; // No rightmost single-column subdomain
N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
QUARK_CORE_zgelqt(
plasma->quark, &task_flags,
tempkm, tempNn, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaConjTrans,
tempmm, tempNn, tempNn, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
A(m, N), ldam);
}
// for (n = N+1; n < N+BS && n < A.nt; n++) {
for (n = N+1;
(n < N+BS && n < A.nt) || n == A.nt-1; // Suck in rightmost single-column domain
n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_ztslqt(
plasma->quark, &task_flags,
tempkm, tempnn, ib, T.nb,
A(k, N), ldak,
A(k, n), ldak,
T(k, n), T.mb);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaConjTrans,
tempmm, A.nb, tempmm, tempnn, A.mb, ib, T.nb,
A(m, N), ldam,
A(m, n), ldam,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
for (RD = BS; RD < A.nt-k; RD *= 2) {
// for (N = k; N+RD < A.nt; N += 2*RD) {
for (N = k;
N+RD < A.nt-1; // No reduction with rightmost single-column subdomain
N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
QUARK_CORE_zttlqt(
plasma->quark, &task_flags,
tempkm, tempNRDn, ib, T.nb,
A (k, N ), ldak,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
PlasmaRight, PlasmaConjTrans,
tempmm, A.nb, tempmm, tempNRDn, A.mb, ib, T.nb,
A (m, N ), ldam,
A (m, N+RD), ldam,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
}
}
