/***************************************************************************//**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
**/
void plasma_pzunglq_quark(PLASMA_desc A, PLASMA_desc Q, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldak, ldqm;
int tempnn, tempmm, tempkmin, tempkn;
int tempAkm, tempAkn;
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;
for (k = min(A.mt, A.nt)-1; k >= 0; k--) {
tempAkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
tempAkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
tempkmin = min( tempAkn, tempAkm );
tempkn = k == Q.nt-1 ? Q.n-k*Q.nb : Q.nb;
ldak = BLKLDD(A, k);
for (n = Q.nt-1; n > k; 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, tempAkm, ib, T.nb,
Q(m, k), 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, tempkn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
Q(m, k), ldqm);
}
}
}
/***************************************************************************//**
* Parallel tile BAND Tridiagonal Reduction - dynamic scheduler
**/
void plasma_pzherbt_quark(PLASMA_enum uplo,
PLASMA_desc A, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n, i, j;
int ldak, ldam, ldan, ldaj, ldai;
int tempkn, tempmm, tempnn, tempjj;
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;
if (uplo == PlasmaLower) {
for (k = 0; k < A.nt-1; k++){
tempkn = k+1 == A.nt-1 ? A.n-(k+1)*A.nb : A.nb;
ldak = BLKLDD(A, k+1);
QUARK_CORE_zgeqrt(
plasma->quark, &task_flags,
tempkn, A.nb, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb);
/* LEFT and RIGHT on the symmetric diagonal block */
QUARK_CORE_zherfb(
plasma->quark, &task_flags,
PlasmaLower,
tempkn, tempkn, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb,
A(k+1, k+1), ldak);
/* RIGHT on the remaining tiles until the bottom */
for (m = k+2; m < A.mt ; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_zunmqr(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempmm, A.nb, tempkn, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb,
A(m , k+1), 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);
QUARK_CORE_ztsqrt(
plasma->quark, &task_flags,
tempmm, A.nb, ib, T.nb,
A(k+1, k), ldak,
A(m , k), ldam,
T(m , k), T.mb);
/* LEFT */
for (i = k+2; i < m; i++) {
ldai = BLKLDD(A, i);
QUARK_CORE_ztsmqr_hetra1(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaConjTrans,
A.mb, A.nb, tempmm, A.nb, A.nb, ib, T.nb,
A(i, k+1), ldai,
A(m, i), ldam,
A(m, k), ldam,
T(m, k), T.mb);
}
/* RIGHT */
for (j = m+1; j < A.mt ; j++) {
tempjj = j == A.mt-1 ? A.m-j*A.mb : A.mb;
ldaj = BLKLDD(A, j);
QUARK_CORE_ztsmqr(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempjj, A.nb, tempjj, tempmm, A.nb, ib, T.nb,
A(j, k+1), ldaj,
A(j, m), ldaj,
A(m, k), ldam,
T(m, k), T.mb);
}
/* LEFT->RIGHT */
QUARK_CORE_ztsmqr_corner(
plasma->quark, &task_flags,
A.nb, A.nb, tempmm, A.nb, tempmm, tempmm, A.nb, ib, T.nb,
A(k+1, k+1), ldak,
A(m , k+1), ldam,
A(m , m), ldam,
A(m , k), ldam,
T(m , k), T.mb);
}
}
}
else {
for (k = 0; k < A.nt-1; k++){
tempkn = k+1 == A.nt-1 ? A.n-(k+1)*A.nb : A.nb;
ldak = BLKLDD(A, k+1);
QUARK_CORE_zgelqt(
plasma->quark, &task_flags,
A.nb, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb);
/* RIGHT and LEFT on the symmetric diagonal block */
QUARK_CORE_zherfb(
plasma->quark, &task_flags,
PlasmaUpper,
tempkn, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb,
A(k+1, k+1), ldak);
/* LEFT on the remaining tiles until the left side */
for (n = k+2; n < A.nt ; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
A.nb, tempnn, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb,
A(k+1, n), ldak);
}
for (n = k+2; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldan = BLKLDD(A, n);
QUARK_CORE_ztslqt(
plasma->quark, &task_flags,
A.nb, tempnn, ib, T.nb,
A(k, k+1), A.nb,
A(k, n), A.nb,
T(k, n), T.mb);
/* RIGHT */
for (i = k+2; i < n; i++) {
ldai = BLKLDD(A, i);
QUARK_CORE_ztsmlq_hetra1(
plasma->quark, &task_flags,
PlasmaRight, PlasmaConjTrans,
A.mb, A.nb, A.nb, tempnn, A.nb, ib, T.nb,
A(k+1, i), ldak,
A(i, n), ldai,
A(k, n), A.nb,
T(k, n), T.mb);
}
/* LEFT */
for (j = n+1; j < A.nt ; j++) {
tempjj = j == A.nt-1 ? A.n-j*A.nb : A.nb;
ldaj = BLKLDD(A, j);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
A.nb, tempjj, tempnn, tempjj, A.nb, ib, T.nb,
A(k+1, j), ldak,
A(n, j), ldan,
A(k, n), A.nb,
T(k, n), T.mb);
}
/* RIGHT->LEFT */
QUARK_CORE_ztsmlq_corner(
plasma->quark, &task_flags,
A.nb, A.nb, A.nb, tempnn, tempnn, tempnn, A.nb, ib, T.nb,
A(k+1, k+1), ldak,
A(k+1, n), ldak,
A(n , n), ldan,
A(k , n), A.nb,
T(k , n), T.mb);
}
}
}
}
/***************************************************************************//**
* 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 application of Q using tile V - LQ factorization - dynamic scheduling
**/
void plasma_pzunmlq_quark(PLASMA_enum side, PLASMA_enum trans,
PLASMA_desc A, PLASMA_desc B, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldak, ldan, ldbk, ldbm;
int tempmm, tempnn, tempkn, tempkm, tempkmin;
int ib, minMT, minM;
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;
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
if (side == PlasmaLeft ) {
if (trans == PlasmaNoTrans) {
/*
* PlasmaLeft / PlasmaNoTrans
*/
for (k = 0; k < minMT; k++) {
tempkm = k == B.mt -1 ? B.m -k*B.mb : B.mb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
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,
tempkm, tempnn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(k, n), ldbk);
}
for (m = k+1; m < B.mt; 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, tempkmin, ib, T.nb,
B(k, n), ldbk,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
}
}
else {
/*
* PlasmaLeft / PlasmaConjTrans
*/
for (k = minMT-1; k >= 0; k--) {
tempkm = k == B.mt -1 ? B.m -k*B.mb : B.mb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
for (m = B.mt-1; m > k; 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, tempkmin, ib, T.nb,
B(k, n), ldbk,
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,
tempkm, tempnn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(k, n), ldbk);
}
}
}
}
else {
if (trans == PlasmaNoTrans) {
/*
* PlasmaRight / PlasmaNoTrans
*/
for (k = minMT-1; k >= 0; k--) {
tempkn = k == B.nt -1 ? B.n -k*B.nb : B.nb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
for (n = B.nt-1; n > k; n--) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
ldan = BLKLDD(A, n);
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, tempkmin, ib, T.nb,
B(m, k), 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, tempkn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(m, k), ldbm);
}
}
}
else {
/*
* PlasmaRight / PlasmaConjTrans
*/
for (k = 0; k < minMT; k++) {
tempkn = k == B.nt -1 ? B.n -k*B.nb : B.nb;
tempkmin = k == minMT-1 ? minM-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
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, tempkn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(m, k), ldbm);
}
for (n = k+1; n < B.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, B.nb, tempmm, tempnn, tempkmin, ib, T.nb,
B(m, k), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
}
}
}
/***************************************************************************//**
* 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 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);
}
}
}
}
}
/***************************************************************************//**
* Parallel tile LQ factorization - dynamic scheduling
**/
void plasma_pzgelqf_quark(PLASMA_desc A, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldak, ldam;
int tempkm, tempkn, tempmm, tempnn;
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;
for (k = 0; k < min(A.mt, A.nt); k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
QUARK_CORE_zgelqt(
plasma->quark, &task_flags,
tempkm, tempkn, ib, T.nb,
A(k, k), ldak,
T(k, k), 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, tempkn, tempkn, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
A(m, k), ldam);
}
for (n = k+1; n < A.nt; 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, k), 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, k), ldam,
A(m, n), ldam,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
}