/***************************************************************************//**
* Parallel tile LQ factorization - static scheduling
**/
void plasma_pcgelqf(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc T;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_k;
int next_m;
int next_n;
int ldak, ldam;
int tempkm, tempkn, tempmm, tempnn;
int ib = PLASMA_IB;
PLASMA_Complex32_t *work, *tau;
plasma_unpack_args_4(A, T, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
work = (PLASMA_Complex32_t*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
tau = (PLASMA_Complex32_t*)plasma_private_alloc(plasma, A.nb, A.dtyp);
ss_init(A.mt, A.nt, -1);
k = 0;
m = PLASMA_RANK;
while (m >= A.mt) {
k++;
m = m-A.mt+k;
}
n = k;
while (k < min(A.mt, A.nt) && m < A.mt) {
next_m = m;
next_n = n;
next_k = k;
next_n++;
if (next_n == A.nt) {
next_m += PLASMA_SIZE;
while (next_m >= A.mt && next_k < min(A.nt, A.mt)) {
next_k++;
next_m = next_m-A.mt+next_k;
}
next_n = next_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;
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldam = BLKLDD(A, m);
if (m == k) {
if (n == k) {
ss_cond_wait(k, k, k-1);
CORE_cgelqt(
tempkm, tempkn, ib,
A(k, k), ldak,
T(k, k), T.mb,
tau, work);
ss_cond_set(k, k, k);
}
else {
ss_cond_wait(k, n, k-1);
CORE_ctslqt(
tempkm, tempnn, ib,
A(k, k), ldak,
A(k, n), ldak,
T(k, n), T.mb,
tau, work);
ss_cond_set(k, n, k);
}
}
else {
if (n == k) {
ss_cond_wait(k, k, k);
ss_cond_wait(m, k, k-1);
CORE_cunmlq(
PlasmaRight, PlasmaConjTrans,
tempmm, tempkn, tempkn, ib,
A(k, k), ldak,
T(k, k), T.mb,
A(m, k), ldam,
work, T.nb);
}
else {
ss_cond_wait(k, n, k);
ss_cond_wait(m, n, k-1);
CORE_ctsmlq(
PlasmaRight, PlasmaConjTrans,
tempmm, A.nb, tempmm, tempnn, A.nb, ib,
A(m, k), ldam,
A(m, n), ldam,
A(k, n), ldak,
T(k, n), T.mb,
work, T.nb);
ss_cond_set(m, n, k);
}
}
m = next_m;
n = next_n;
k = next_k;
}
plasma_private_free(plasma, work);
plasma_private_free(plasma, tau);
ss_finalize();
}
/***************************************************************************//**
* Parallel tile LU factorization - dynamic scheduling
**/
void plasma_psgetrf_incpiv_quark(PLASMA_desc A, PLASMA_desc L, int *IPIV,
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_sgetrf_incpiv(
plasma->quark, &task_flags,
tempkm, tempkn, ib, L.nb,
A(k, k), ldak, IPIV(k, k),
sequence, request,
k == A.mt-1, A.nb*k);
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_sgessm(
plasma->quark, &task_flags,
tempkm, tempnn, tempkm, ib, L.nb,
IPIV(k, k),
A(k, k), ldak,
A(k, n), ldak);
}
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_ststrf(
plasma->quark, &task_flags,
tempmm, tempkn, ib, L.nb,
A(k, k), ldak,
A(m, k), ldam,
L(m, k), L.mb,
IPIV(m, k),
sequence, request,
m == A.mt-1, A.nb*k);
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_sssssm(
plasma->quark, &task_flags,
A.nb, tempnn, tempmm, tempnn, A.nb, ib, L.nb,
A(k, n), ldak,
A(m, n), ldam,
L(m, k), L.mb,
A(m, k), ldam,
IPIV(m, k));
}
}
}
}
/***************************************************************************//**
* Parallel tile LQ factorization (reduction Householder) - static / sequential
**/
void plasma_pzgelqfrh(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc T;
int BS;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int K, N, RD;
int ldak, ldam;
int tempkm, tempNn, tempmm, tempnn, tempNRDn;
int ib;
if (PLASMA_RANK != 0) return;
plasma_unpack_args_5(A, T, BS, sequence, request);
ib = PLASMA_IB;
PLASMA_Complex64_t *work, *tau;
work = (PLASMA_Complex64_t*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
tau = (PLASMA_Complex64_t*)plasma_private_alloc(plasma, A.nb, A.dtyp);
if (sequence->status != PLASMA_SUCCESS)
return;
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-1 || N == k; // No rightmost single-column subdomain
N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
CORE_zgelqt(
tempkm, tempNn, ib,
A(k, N), ldak,
T(k, N), T.mb,
tau, work);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
CORE_zunmlq(
PlasmaRight, PlasmaConjTrans,
tempmm, tempNn, tempNn, ib,
A(k, N), ldak,
T(k, N), T.mb,
A(m, N), ldam,
work , A.nb);
}
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;
CORE_ztslqt(
tempkm, tempnn, ib,
A(k, N), ldak,
A(k, n), ldak,
T(k, n), T.mb,
tau, work);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
CORE_ztsmlq(
PlasmaRight, PlasmaConjTrans,
tempmm, A.mb, tempmm, tempnn, A.mb, ib,
A(m, N), ldam,
A(m, n), ldam,
A(k, n), ldak,
T(k, n), T.mb,
work , A.nb);
}
}
}
for (RD = BS; RD < A.nt-k; 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;
CORE_zttlqt(
tempkm, tempNRDn, ib,
A (k, N ), ldak,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb,
tau, work);
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
CORE_zttmlq(
PlasmaRight, PlasmaConjTrans,
tempmm, A.nb, tempmm, tempNRDn, tempkm, ib,
A (m, N ), ldam,
A (m, N+RD), ldam,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb,
work , A.nb);
}
}
}
}
}
/***************************************************************************//**
* 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 QR factorization - dynamic scheduling
**/
void plasma_pdgeqrf_quark(PLASMA_desc A, PLASMA_desc T, int ib)
{
int k, m, n;
int ldak, ldam;
int tempkm, tempkn, tempnn, tempmm;
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);
double *dA = A(k, k);
double *dT = T(k, k);
#if defined(USE_OMPEXT)
omp_set_task_priority(1);
#endif
#pragma omp task depend(inout: dA[0:T.nb*T.nb]) depend(out:dT[0:ib*T.nb])
{
double tau[T.nb];
double work[ib * T.nb];
CORE_dgeqrt(tempkm, tempkn, ib, dA, ldak, dT, T.mb, &tau[0], &work[0]);
}
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
double *dA = A(k, k);
double *dT = T(k, k);
double *dC = A(k, n);
#pragma omp task depend(in: dA[0:T.nb*T.nb], dT[0:ib*T.nb]) depend(inout:dC[0:T.nb*T.nb])
{
double work[T.nb * ib];
CORE_dormqr(PlasmaLeft, PlasmaTrans,
tempkm, tempnn, tempkm, ib,
dA, ldak,
dT, T.mb,
dC, ldak,
&work[0], T.nb);
}
}
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
double *dA = A(k, k);
double *dB = A(m, k);
double *dT = T(m, k);
#pragma omp task depend(inout:dA[0:T.nb*T.nb], dB[0:T.nb*T.nb]) depend(out:dT[0:ib*T.nb])
{
double tau[T.nb];
double work[ib * T.nb];
CORE_dtsqrt(tempmm, tempkn, ib,
dA, ldak,
dB, ldam,
dT, T.mb, &tau[0], &work[0]);
}
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
double *dA = A(k, n);
double *dB = A(m, n);
double *dV = A(m, k);
double *dT = T(m, k);
#pragma omp task depend(inout:dA[0:T.nb*T.nb], dB[0:T.nb*T.nb]) depend(in:dV[0:T.nb*T.nb], dT[0:ib*T.nb])
{
double work[ib * T.nb];
CORE_dtsmqr(PlasmaLeft, PlasmaTrans,
A.mb, tempnn, tempmm, tempnn, A.nb, ib,
dA, ldak,
dB, ldam,
dV, ldam,
dT, T.mb, &work[0], ib);
}
}
}
}
}
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization - static scheduling
**/
void plasma_pcunmqr(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, ldam, ldbm;
int tempkm, tempnn, tempkmin, tempmm;
int minMT, minM;
int ib = PLASMA_IB;
PLASMA_Complex32_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_Complex32_t*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
ss_init(B.mt, B.nt, -1);
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
k = 0;
n = PLASMA_RANK;
while (n >= B.nt) {
k++;
n = n-B.nt;
}
m = k;
while (k < minMT && n < B.nt) {
next_n = n;
next_m = m;
next_k = k;
next_m++;
if (next_m == A.mt) {
next_n += PLASMA_SIZE;
while (next_n >= B.nt && next_k < minMT) {
next_k++;
next_n = next_n-B.nt;
}
next_m = next_k;
}
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);
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
if (m == k) {
ss_cond_wait(k, n, k-1);
CORE_cunmqr(
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_ctsmqr(
side, trans,
A.mb, tempnn, tempmm, tempnn, tempkmin, ib,
B(k, n), ldbk,
B(m, n), ldbm,
A(m, k), ldam,
T(m, k), T.mb,
work, ib);
ss_cond_set(m, n, k);
}
n = next_n;
m = next_m;
k = next_k;
}
plasma_private_free(plasma, work);
ss_finalize();
}
/***************************************************************************//**
* 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);
}
}
}
/***************************************************************************//**
* Parallel tile Hermitian rank-k update - dynamic scheduling
**/
void plasma_pcsyr2k_quark(PLASMA_enum uplo, PLASMA_enum trans,
PLASMA_Complex32_t alpha, PLASMA_desc A, PLASMA_desc B,
PLASMA_Complex32_t beta, PLASMA_desc C,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int m, n, k;
int ldak, ldam, ldan, ldcm, ldcn;
int ldbk, ldbm, ldbn;
int tempnn, tempmm, tempkn, tempkm;
PLASMA_Complex32_t zone = (PLASMA_Complex32_t)1.0;
PLASMA_Complex32_t zbeta;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
for (n = 0; n < C.nt; n++) {
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
ldan = BLKLDD(A, n);
ldbn = BLKLDD(B, n);
ldcn = BLKLDD(C, n);
/*
* PlasmaNoTrans
*/
if (trans == PlasmaNoTrans) {
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
QUARK_CORE_csyr2k(
plasma->quark, &task_flags,
uplo, trans,
tempnn, tempkn, A.mb,
alpha, A(n, k), ldan, /* ldan * K */
B(n, k), ldbn,
zbeta, C(n, n), ldcn); /* ldc * N */
}
/*
* PlasmaNoTrans / PlasmaLower
*/
if (uplo == PlasmaLower) {
for (m = n+1; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
ldcm = BLKLDD(C, m);
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, A(m, k), ldam, /* ldam * K */
B(n, k), ldbn, /* ldan * K */
zbeta, C(m, n), ldcm); /* ldc * N */
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, B(m, k), ldbm, /* ldam * K */
A(n, k), ldan, /* ldan * K */
zone, C(m, n), ldcm); /* ldc * N */
}
}
}
/*
* PlasmaNoTrans / PlasmaUpper
*/
else {
for (m = n+1; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaTrans,
tempnn, tempmm, tempkn, A.mb,
alpha, A(n, k), ldan, /* ldan * K */
B(m, k), ldbm, /* ldam * M */
zbeta, C(n, m), ldcn); /* ldc * M */
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaTrans,
tempnn, tempmm, tempkn, A.mb,
alpha, B(n, k), ldan, /* ldan * K */
A(m, k), ldam, /* ldam * M */
zone, C(n, m), ldcn); /* ldc * M */
}
}
}
}
/*
* Plasma[Conj]Trans
*/
else {
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_csyr2k(
plasma->quark, &task_flags,
uplo, trans,
tempnn, tempkm, A.mb,
alpha, A(k, n), ldak, /* lda * N */
B(k, n), ldbk,
zbeta, C(n, n), ldcn); /* ldc * N */
}
/*
* Plasma[Conj]Trans / PlasmaLower
*/
if (uplo == PlasmaLower) {
for (m = n+1; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldcm = BLKLDD(C, m);
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* lda * M */
B(k, n), ldbk, /* lda * N */
zbeta, C(m, n), ldcm); /* ldc * N */
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, B(k, m), ldbk, /* lda * M */
A(k, n), ldak, /* lda * N */
zone, C(m, n), ldcm); /* ldc * N */
}
}
}
/*
* Plasma[Conj]Trans / PlasmaUpper
*/
else {
for (m = n+1; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaNoTrans,
tempnn, tempmm, tempkm, A.mb,
alpha, A(k, n), ldak, /* lda * K */
B(k, m), ldbk, /* lda * M */
zbeta, C(n, m), ldcn); /* ldc * M */
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
trans, PlasmaNoTrans,
tempnn, tempmm, tempkm, A.mb,
alpha, B(k, n), ldbk, /* lda * K */
A(k, m), ldak, /* lda * M */
zone, C(n, m), ldcn); /* ldc * M */
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile BAND Tridiagonal Reduction - dynamic scheduler
**/
void plasma_pssyrbt_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_sgeqrt(
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_ssyrfb(
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_sormqr(
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_stsqrt(
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_stsmqr_sytra1(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaTrans,
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_stsmqr(
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_stsmqr_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_sgelqt(
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_ssyrfb(
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_sormlq(
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_stslqt(
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_stsmlq_sytra1(
plasma->quark, &task_flags,
PlasmaRight, PlasmaTrans,
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_stsmlq(
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_stsmlq_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 tile triangular matrix inverse - dynamic scheduling
**/
void plasma_pztrtri_quark(PLASMA_enum uplo, PLASMA_enum diag, PLASMA_desc A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldam, ldan;
int tempkn, tempmm, tempnn;
PLASMA_Complex64_t zone = (PLASMA_Complex64_t) 1.0;
PLASMA_Complex64_t mzone = (PLASMA_Complex64_t)-1.0;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
for (n = 0; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldan = BLKLDD(A, n);
for (m = n+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_ztrsm(
plasma->quark, &task_flags,
PlasmaRight, uplo, PlasmaNoTrans, diag,
tempmm, tempnn, A.mb,
mzone, A(n, n), ldan,
A(m, n), ldam);
}
for (m = n+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
for (k = 0; k < n; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
QUARK_CORE_zgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempkn, tempnn, A.mb,
zone, A(m, n), ldam,
A(n, k), ldan,
zone, A(m, k), ldam);
}
}
for (m = 0; m < n; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
QUARK_CORE_ztrsm(
plasma->quark, &task_flags,
PlasmaLeft, uplo, PlasmaNoTrans, diag,
tempnn, tempmm, A.mb,
zone, A(n, n), ldan,
A(n, m), ldan);
}
QUARK_CORE_ztrtri(
plasma->quark, &task_flags,
uplo, diag,
tempnn, A.mb,
A(n, n), ldan,
sequence, request, A.nb*n);
}
}
/*
* PlasmaUpper
*/
else {
for (m = 0; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
for (n = m+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_ztrsm(
plasma->quark, &task_flags,
PlasmaLeft, uplo, PlasmaNoTrans, diag,
tempmm, tempnn, A.mb,
mzone, A(m, m), ldam,
A(m, n), ldam);
}
for (n = 0; n < m; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldan = BLKLDD(A, n);
for (k = m+1; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
QUARK_CORE_zgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempnn, tempkn, tempmm, A.mb,
zone, A(n, m), ldan,
A(m, k), ldam,
zone, A(n, k), ldan);
}
QUARK_CORE_ztrsm(
plasma->quark, &task_flags,
PlasmaRight, uplo, PlasmaNoTrans, diag,
tempnn, tempmm, A.mb,
zone, A(m, m), ldam,
A(n, m), ldan);
}
QUARK_CORE_ztrtri(
plasma->quark, &task_flags,
uplo, diag,
tempmm, A.mb,
A(m, m), ldam,
sequence, request, A.mb*m);
}
}
}
/***************************************************************************//**
* Parallel tile symmetric matrix-matrix multiplication - dynamic scheduling
**/
void plasma_pdsymm_quark(PLASMA_enum side, PLASMA_enum uplo,
double alpha, PLASMA_desc A, PLASMA_desc B,
double beta, PLASMA_desc C,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int lda, ldak, ldb, ldc;
int tempmm, tempnn, tempkn, tempkm;
double zbeta;
double zone = (double)1.0;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
for (m = 0; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldc = BLKLDD(C, m);
for (n = 0; n < C.nt; n++) {
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
/*
* PlasmaLeft / PlasmaLower
*/
if (side == PlasmaLeft) {
lda = BLKLDD(A, m);
if (uplo == PlasmaLower) {
for (k = 0; k < C.mt; k++) {
tempkm = k == C.mt-1 ? C.m-k*C.mb : C.mb;
ldak = BLKLDD(A, k);
ldb = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
if (k < m) {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, A(m, k), lda, /* lda * K */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
if (k == m) {
QUARK_CORE_dsymm(
plasma->quark, &task_flags,
side, uplo,
tempmm, tempnn, A.mb,
alpha, A(k, k), ldak, /* ldak * X */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* ldak * X */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
}
}
}
/*
* PlasmaLeft / PlasmaUpper
*/
else {
for (k = 0; k < C.mt; k++) {
tempkm = k == C.mt-1 ? C.m-k*C.mb : C.mb;
ldak = BLKLDD(A, k);
ldb = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
if (k < m) {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* ldak * X */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
if (k == m) {
QUARK_CORE_dsymm(
plasma->quark, &task_flags,
side, uplo,
tempmm, tempnn, A.mb,
alpha, A(k, k), ldak, /* ldak * K */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm, A.mb,
alpha, A(m, k), lda, /* lda * K */
B(k, n), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
}
}
}
}
/*
* PlasmaRight / PlasmaLower
*/
else {
lda = BLKLDD(A, n);
ldb = BLKLDD(B, m);
if (uplo == PlasmaLower) {
for (k = 0; k < C.nt; k++) {
tempkn = k == C.nt-1 ? C.n-k*C.nb : C.nb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
if (k < n) {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, B(m, k), ldb, /* ldb * K */
A(n, k), lda, /* lda * K */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
if (k == n) {
QUARK_CORE_dsymm(
plasma->quark, &task_flags,
side, uplo,
tempmm, tempnn, A.mb,
alpha, A(k, k), ldak, /* ldak * Y */
B(m, k), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, B(m, k), ldb, /* ldb * K */
A(k, n), ldak, /* ldak * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
}
}
}
/*
* PlasmaRight / PlasmaUpper
*/
else {
for (k = 0; k < C.nt; k++) {
tempkn = k == C.nt-1 ? C.n-k*C.nb : C.nb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
if (k < n) {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, B(m, k), ldb, /* ldb * K */
A(k, n), ldak, /* ldak * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
if (k == n) {
QUARK_CORE_dsymm(
plasma->quark, &task_flags,
side, uplo,
tempmm, tempnn, A.mb,
alpha, A(k, k), ldak, /* ldak * Y */
B(m, k), ldb, /* ldb * Y */
zbeta, C(m, n), ldc); /* ldc * Y */
}
else {
QUARK_CORE_dgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaTrans,
tempmm, tempnn, tempkn, A.mb,
alpha, B(m, k), ldb, /* ldb * K */
A(n, k), lda, /* lda * K */
zbeta, C(m, n), ldc); /* ldc * Y */
}
}
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile symmetric matrix-matrix multiplication - static scheduling
**/
void plasma_pdsymm(plasma_context_t *plasma)
{
PLASMA_enum side;
PLASMA_enum uplo;
double alpha;
PLASMA_desc A;
PLASMA_desc B;
double beta;
PLASMA_desc C;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_m;
int next_n;
int lda, ldak, ldb, ldc;
int tempmm, tempnn, tempkn, tempkm;
double zbeta;
double zone = (double)1.0;
plasma_unpack_args_9(side, uplo, alpha, A, B, beta, C, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
n = 0;
m = PLASMA_RANK;
while (m >= C.mt && n < C.nt) {
n++;
m = m-C.mt;
}
while (n < C.nt) {
next_m = m;
next_n = n;
next_m += PLASMA_SIZE;
while (next_m >= C.mt && next_n < C.nt) {
next_n++;
next_m = next_m - C.mt;
}
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
ldc = BLKLDD(C, m);
/*
* PlasmaLeft / PlasmaLower
*/
if (side == PlasmaLeft) {
lda = BLKLDD(A, m);
if (uplo == PlasmaLower) {
for (k = 0; k < C.mt; k++) {
tempkm = k == C.mt-1 ? C.m-k*C.mb : C.mb;
ldak = BLKLDD(A, k);
ldb = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
if (k < m) {
CORE_dgemm(
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, A(m, k), lda,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
else {
if (k == m) {
CORE_dsymm(
side, uplo,
tempmm, tempnn,
alpha, A(k, k), ldak,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
else {
CORE_dgemm(
PlasmaTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, A(k, m), ldak,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
}
}
}
/*
* PlasmaLeft / PlasmaUpper
*/
else {
for (k = 0; k < C.mt; k++) {
tempkm = k == C.mt-1 ? C.m-k*C.mb : C.mb;
ldak = BLKLDD(A, k);
ldb = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
if (k < m) {
CORE_dgemm(
PlasmaTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, A(k, m), ldak,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
else {
if (k == m) {
CORE_dsymm(
side, uplo,
tempmm, tempnn,
alpha, A(k, k), ldak,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
else {
CORE_dgemm(
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, A(m, k), lda,
B(k, n), ldb,
zbeta, C(m, n), ldc);
}
}
}
}
}
/*
* PlasmaRight / PlasmaLower
*/
else {
lda = BLKLDD(A, n);
ldb = BLKLDD(B, m);
if (uplo == PlasmaLower) {
for (k = 0; k < C.nt; k++) {
tempkn = k == C.nt-1 ? C.n-k*C.nb : C.nb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
if (k < n) {
CORE_dgemm(
PlasmaNoTrans, PlasmaTrans,
tempmm, tempnn, tempkn,
alpha, B(m, k), ldb,
A(n, k), lda,
zbeta, C(m, n), ldc);
}
else {
if (n == k) {
CORE_dsymm(
side, uplo,
tempmm, tempnn,
alpha, A(k, k), ldak,
B(m, k), ldb,
zbeta, C(m, n), ldc);
}
else {
CORE_dgemm(
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkn,
alpha, B(m, k), ldb,
A(k, n), ldak,
zbeta, C(m, n), ldc);
}
}
}
}
/*
* PlasmaRight / PlasmaUpper
*/
else {
for (k = 0; k < C.nt; k++) {
tempkn = k == C.nt-1 ? C.n-k*C.nb : C.nb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
if (k < n) {
CORE_dgemm(
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, tempkn,
alpha, B(m, k), ldb,
A(k, n), ldak,
zbeta, C(m, n), ldc);
}
else {
if (n == k) {
CORE_dsymm(
side, uplo,
tempmm, tempnn,
alpha, A(k, k), ldak,
B(m, k), ldb,
zbeta, C(m, n), ldc);
}
else {
CORE_dgemm(
PlasmaNoTrans, PlasmaTrans,
tempmm, tempnn, tempkn,
alpha, B(m, k), ldb,
A(n, k), lda,
zbeta, C(m, n), ldc);
}
}
}
}
}
m = next_m;
n = next_n;
}
}
/***************************************************************************//**
* Parallel tile LU factorization - static scheduling
**/
void plasma_psgetrf_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;
float *work;
plasma_unpack_args_5(A, L, IPIV, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
work = (float*)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_sgetrf_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_ststrf(
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_sgessm(
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_sssssm(
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 forward substitution for tile LU - static scheduling
**/
void plasma_pztrsmpl(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc B;
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, ldbk, ldam, ldbm;
int tempkm, tempnn, tempkmin, tempmm, tempkn;
int ib;
plasma_unpack_args_6(A, B, L, IPIV, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
ss_init(B.mt, B.nt, -1);
ib = PLASMA_IB;
k = 0;
n = PLASMA_RANK;
while (n >= B.nt) {
k++;
n = n-B.nt;
}
m = k;
while (k < min(A.mt, A.nt) && n < B.nt) {
next_n = n;
next_m = m;
next_k = k;
next_m++;
if (next_m == A.mt) {
next_n += PLASMA_SIZE;
while (next_n >= B.nt && next_k < min(A.mt, A.nt)) {
next_k++;
next_n = next_n-B.nt;
}
next_m = next_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;
tempkmin = k == min(A.mt, A.nt)-1 ? min(A.m, A.n)-k*A.mb : A.mb;
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
if (m == k) {
ss_cond_wait(k, n, k-1);
CORE_zgessm(
tempkm, tempnn, tempkmin, ib,
IPIV(k, k),
A(k, k), ldak,
B(k, n), ldbk);
ss_cond_set(k, n, k);
}
else {
ss_cond_wait(m, n, k-1);
CORE_zssssm(
A.nb, tempnn, tempmm, tempnn, tempkn, ib,
B(k, n), ldbk,
B(m, n), ldbm,
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;
}
ss_finalize();
}
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization - dynamic scheduling
**/
void plasma_pdormqr_quark(PLASMA_enum side, PLASMA_enum trans,
PLASMA_desc A, PLASMA_desc B, PLASMA_desc T, int ib)
{
int k, m, n;
int ldak, ldbk, ldam, ldan, ldbm;
int tempkm, tempnn, tempkmin, tempmm, tempkn;
int minMT, minM;
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
double *work = (double *)alloca(sizeof(double) * T.nb * ib);
/*
* PlasmaLeft / PlasmaTrans
*/
if (side == PlasmaLeft ) {
if (trans == PlasmaTrans) {
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;
double *dA = A(k, k);
double *dT = T(k, k);
double *dB = B(k, n);
{
CORE_dormqr(side, trans, tempkm, tempnn, tempkmin, ib,
dA, ldak,
dT, T.mb,
dB, ldbk, work, T.nb);
}
}
for (m = k+1; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
double *dA = B(k, n);
double *dB = B(m, n);
double *dV = A(m, k);
double *dT = T(m, k);
{
CORE_dtsmqr(side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib,
dA, ldbk,
dB, ldbm,
dV, ldam,
dT, T.mb, work, (side == PlasmaLeft)?ib:T.nb);
}
}
}
}
}
/*
* PlasmaLeft / PlasmaNoTrans
*/
else {
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;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
double *dA = B(k, n);
double *dB = B(m, n);
double *dV = A(m, k);
double *dT = T(m, k);
{
CORE_dtsmqr(side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib,
dA, ldbk,
dB, ldbm,
dV, ldam,
dT, T.mb, work, (side == PlasmaLeft)?ib:T.nb);
}
}
}
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
double *dA = A(k, k);
double *dT = T(k, k);
double *dB = B(k, n);
{
CORE_dormqr(side, trans, tempkm, tempnn, tempkmin, ib,
dA, ldak,
dT, T.mb,
dB, ldbk, work, T.nb);
}
}
}
}
}
/*
* PlasmaRight / PlasmaTrans
*/
else {
if (trans == PlasmaTrans) {
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);
ldbk = BLKLDD(B, 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);
double *dA = B(m, k);
double *dB = B(m, n);
double *dV = A(n, k);
double *dT = T(n, k);
{
CORE_dtsmqr(side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib,
dA, ldbm,
dB, ldbm,
dV, ldan,
dT, T.mb, work, (side == PlasmaLeft)?ib:T.nb);
}
}
}
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
double *dA = A(k, k);
double *dT = T(k, k);
double *dB = B(m, k);
{
CORE_dormqr(side, trans, tempmm, tempkn, tempkmin, ib,
dA, ldak,
dT, T.mb,
dB, ldbm, work, T.nb);
}
}
}
}
/*
* PlasmaRight / PlasmaNoTrans
*/
else {
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.nb : A.nb;
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);
double *dA = A(k, k);
double *dT = T(k, k);
double *dB = B(m, k);
{
CORE_dormqr(side, trans, tempmm, tempkn, tempkmin, ib,
dA, ldak,
dT, T.mb,
dB, ldbm, work, T.nb);
}
}
for (n = k+1; n < B.nt; 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);
double *dA = B(m, k);
double *dB = B(m, n);
double *dV = A(n, k);
double *dT = T(n, k);
{
CORE_dtsmqr(side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib,
dA, ldbm,
dB, ldbm,
dV, ldan,
dT, T.mb, work, (side == PlasmaLeft)?ib:T.nb);
}
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile QR factorization - static scheduling
**/
void plasma_pdgeqrf(plasma_context_t *plasma)
{
PLASMA_desc A;
PLASMA_desc T;
PLASMA_sequence *sequence;
PLASMA_request *request;
int k, m, n;
int next_k;
int next_m;
int next_n;
int ldak, ldam;
int tempkm, tempkn, tempnn, tempmm;
int ib = PLASMA_IB;
double *work, *tau;
plasma_unpack_args_4(A, T, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
work = (double*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
tau = (double*)plasma_private_alloc(plasma, A.nb, A.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) {
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;
}
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;
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldam = BLKLDD(A, m);
if (n == k) {
if (m == k) {
ss_cond_wait(k, k, k-1);
CORE_dgeqrt(
tempkm, tempkn, ib,
A(k, k), ldak,
T(k, k), T.mb,
tau, work);
ss_cond_set(k, k, k);
}
else {
ss_cond_wait(m, k, k-1);
CORE_dtsqrt(
tempmm, tempkn, ib,
A(k, k), ldak,
A(m, k), ldam,
T(m, k), T.mb,
tau, work);
ss_cond_set(m, k, k);
}
}
else {
if (m == k) {
ss_cond_wait(k, k, k);
ss_cond_wait(k, n, k-1);
CORE_dormqr(
PlasmaLeft, PlasmaTrans,
tempkm, tempnn, tempkm, ib,
A(k, k), ldak,
T(k, k), T.mb,
A(k, n), ldak,
work, T.nb);
}
else {
ss_cond_wait(m, k, k);
ss_cond_wait(m, n, k-1);
CORE_dtsmqr(
PlasmaLeft, PlasmaTrans,
A.nb, tempnn, tempmm, tempnn, A.nb, ib,
A(k, n), ldak,
A(m, n), ldam,
A(m, k), ldam,
T(m, k), T.mb,
work, ib);
ss_cond_set(m, n, k);
}
}
n = next_n;
m = next_m;
k = next_k;
}
plasma_private_free(plasma, work);
plasma_private_free(plasma, tau);
ss_finalize();
}
/***************************************************************************//**
* Parallel tile QR factorization (reduction Householder) - dynamic scheduling
**/
void plasma_psgeqrfrh_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, M, RD;
int ldaM, ldam, ldaMRD;
int tempkn, tempMm, tempnn, tempmm, tempMRDm;
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++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
for (M = k;
M < A.mt-1 || M == k; /* No bottom single-row subdomain */
M += BS) {
tempMm = M == A.mt-1 ? A.m-M*A.mb : A.mb;
ldaM = BLKLDD(A, M);
QUARK_CORE_sgeqrt(
plasma->quark, &task_flags,
tempMm, tempkn, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb);
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_sormqr(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaTrans,
tempMm, tempnn, tempMm, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb,
A(M, n), ldaM);
}
for (m = M+1;
(m < M+BS && m < A.mt) || m == A.mt-1; /* Suck in bottom single-row domain */
m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_stsqrt(
plasma->quark, &task_flags,
tempmm, tempkn, ib, T.nb,
A(M, k), ldaM,
A(m, k), ldam,
T(m, k), T.mb);
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_stsmqr(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaTrans,
A.nb, tempnn, tempmm, tempnn, A.nb, ib, T.nb,
A(M, n), ldaM,
A(m, n), ldam,
A(m, k), ldam,
T(m, k), T.mb);
}
}
}
for (RD = BS; RD < A.mt-k; RD *= 2) {
for (M = k;
M+RD < A.mt-1; /* No reduction with bottom single-row subdomain */
M += 2*RD) {
tempMRDm = M+RD == A.mt-1 ? A.m-(M+RD)*A.mb : A.mb;
ldaM = BLKLDD(A, M );
ldaMRD = BLKLDD(A, M+RD);
QUARK_CORE_sttqrt(
plasma->quark, &task_flags,
tempMRDm, tempkn, ib, T.nb,
A (M , k), ldaM,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
for (n = k+1; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_sttmqr(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaTrans,
A.nb, tempnn, tempMRDm, tempnn, A.nb, ib, T.nb,
A (M, n), ldaM,
A (M+RD, n), ldaMRD,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
}
}
}
}
}
/***************************************************************************//**
* Parallel tile Hermitian rank-k update - static scheduling
**/
void plasma_pcsyr2k(plasma_context_t *plasma)
{
PLASMA_enum uplo;
PLASMA_enum trans;
PLASMA_Complex32_t alpha;
PLASMA_desc A;
PLASMA_desc B;
PLASMA_Complex32_t beta;
PLASMA_desc C;
PLASMA_sequence *sequence;
PLASMA_request *request;
int m, n, k;
int next_m;
int next_n;
int ldam, ldan, ldak;
int ldbm, ldbn, ldbk;
int ldcm, ldcn;
int tempkn, tempkm, tempmm, tempnn;
PLASMA_Complex32_t zone = (PLASMA_Complex32_t)1.0;
PLASMA_Complex32_t zbeta;
plasma_unpack_args_9(uplo, trans, alpha, A, B, beta, C, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
n = 0;
m = PLASMA_RANK;
while (m >= C.mt && n < C.nt) {
n++;
m = m-C.mt+n;
}
while (n < C.nt) {
next_n = n;
next_m = m + PLASMA_SIZE;
while (next_m >= C.mt && next_n < C.nt) {
next_n++;
next_m = next_m - C.mt + next_n;
}
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
ldcn = BLKLDD(C, n);
ldcm = BLKLDD(C, m);
if (m == n) {
/*
* PlasmaNoTrans
*/
if (trans == PlasmaNoTrans) {
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
CORE_csyr2k(
uplo, trans,
tempnn, tempkn,
alpha, A(m, k), ldam,
B(m, k), ldbm,
zbeta, C(m, m), ldcm);
}
}
/*
* Plasma[Conj]Trans
*/
else {
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
CORE_csyr2k(
uplo, trans,
tempnn, tempkm,
alpha, A(k, m), ldak,
B(k, m), ldbk,
zbeta, C(m, m), ldcm);
}
}
}
else {
if (trans == PlasmaNoTrans) {
ldam = BLKLDD(A, m);
ldan = BLKLDD(A, n);
ldbm = BLKLDD(B, m);
ldbn = BLKLDD(B, n);
/*
* PlasmaNoTrans / PlasmaLower
*/
if (uplo == PlasmaLower) {
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
CORE_cgemm(
trans, PlasmaTrans,
tempmm, tempnn, tempkn,
alpha, A(m, k), ldam,
B(n, k), ldbn,
zbeta, C(m, n), ldcm);
CORE_cgemm(
trans, PlasmaTrans,
tempmm, tempnn, tempkn,
alpha, B(m, k), ldbm,
A(n, k), ldan,
zone, C(m, n), ldcm);
}
}
/*
* PlasmaNoTrans / PlasmaUpper
*/
else {
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
zbeta = k == 0 ? beta : zone;
CORE_cgemm(
trans, PlasmaTrans,
tempnn, tempmm, tempkn,
alpha, A(n, k), ldan,
B(m, k), ldbm,
zbeta, C(n, m), ldcn);
CORE_cgemm(
trans, PlasmaTrans,
tempnn, tempmm, tempkn,
alpha, B(n, k), ldbn,
A(m, k), ldam,
zone, C(n, m), ldcn);
}
}
}
else {
/*
* Plasma[Conj]Trans / PlasmaLower
*/
if (uplo == PlasmaLower) {
for (k = 0; k < A.mt; k++) {
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
zbeta = k == 0 ? beta : zone;
CORE_cgemm(
trans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, A(k, m), ldak,
B(k, n), ldbk,
zbeta, C(m, n), ldcm);
CORE_cgemm(
trans, PlasmaNoTrans,
tempmm, tempnn, tempkm,
alpha, B(k, m), ldbk,
A(k, n), ldak,
zone, C(m, n), ldcm);
}
}
/*
* Plasma[Conj]Trans / PlasmaUpper
*/
else {
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
CORE_cgemm(
trans, PlasmaNoTrans,
tempnn, tempmm, tempkm,
alpha, A(k, n), ldak,
B(k, m), ldbk,
zbeta, C(n, m), ldcm);
CORE_cgemm(
trans, PlasmaNoTrans,
tempnn, tempmm, tempkm,
alpha, B(k, n), ldbk,
A(k, m), ldak,
zone, C(n, m), ldcn);
}
}
}
}
m = next_m;
n = next_n;
}
}
/***************************************************************************//**
* Parallel initializztion a 2-D array A to
* ALPHA on the offdiagonals.
**/
void plasma_pzlaset2_quark(PLASMA_enum uplo, PLASMA_Complex64_t alpha,
PLASMA_desc A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int i, j;
int ldai, ldaj;
int tempim;
int tempjm, tempjn;
int minmn = min(A.mt, A.nt);
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
if (uplo == PlasmaLower) {
for (j = 0; j < minmn; j++){
tempjm = j == A.mt-1 ? A.m-j*A.mb : A.mb;
tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
ldaj = BLKLDD(A, j);
QUARK_CORE_zlaset2(
plasma->quark, &task_flags,
PlasmaLower, tempjm, tempjn, alpha,
A(j, j), ldaj);
for (i = j+1; i < A.mt; i++){
tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
ldai = BLKLDD(A, i);
QUARK_CORE_zlaset2(
plasma->quark, &task_flags,
PlasmaUpperLower, tempim, tempjn, alpha,
A(i, j), ldai);
}
}
}
else if (uplo == PlasmaUpper) {
for (j = 1; j < A.nt; j++){
tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
for (i = 0; i < min(j, A.mt); i++){
tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
ldai = BLKLDD(A, i);
QUARK_CORE_zlaset2(
plasma->quark, &task_flags,
PlasmaUpperLower, tempim, tempjn, alpha,
A(i, j), ldai);
}
}
for (j = 0; j < minmn; j++){
tempjm = j == A.mt-1 ? A.m-j*A.mb : A.mb;
tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
ldaj = BLKLDD(A, j);
QUARK_CORE_zlaset2(
plasma->quark, &task_flags,
PlasmaUpper, tempjm, tempjn, alpha,
A(j, j), ldaj);
}
}
else {
for (i = 0; i < A.mt; i++){
tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
ldai = BLKLDD(A, i);
for (j = 0; j < A.nt; j++){
tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
QUARK_CORE_zlaset2(
plasma->quark, &task_flags,
PlasmaUpperLower, tempim, tempjn, alpha,
A(i, j), ldai);
}
}
}
}
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization - dynamic scheduling
**/
void plasma_pcunmqr_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, ldbk, ldam, ldan, ldbm;
int tempkm, tempnn, tempkmin, tempmm, tempkn;
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;
}
/*
* PlasmaLeft / PlasmaConjTrans
*/
if (side == PlasmaLeft ) {
if (trans == PlasmaConjTrans) {
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_cunmqr(
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;
ldam = BLKLDD(A, m);
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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib, T.nb,
B(k, n), ldbk,
B(m, n), ldbm,
A(m, k), ldam,
T(m, k), T.mb);
}
}
}
}
/*
* PlasmaLeft / PlasmaNoTrans
*/
else {
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;
ldam = BLKLDD(A, m);
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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib, T.nb,
B(k, n), ldbk,
B(m, n), ldbm,
A(m, k), ldam,
T(m, k), T.mb);
}
}
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_cunmqr(
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);
}
}
}
}
/*
* PlasmaRight / PlasmaConjTrans
*/
else {
if (trans == PlasmaConjTrans) {
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);
ldbk = BLKLDD(B, 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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib, T.nb,
B(m, k), ldbm,
B(m, n), ldbm,
A(n, k), ldan,
T(n, k), 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_cunmqr(
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);
}
}
}
/*
* PlasmaRight / PlasmaNoTrans
*/
else {
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.nb : A.nb;
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_cunmqr(
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;
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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib, T.nb,
B(m, k), ldbm,
B(m, n), ldbm,
A(n, k), ldan,
T(n, k), T.mb);
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile Cholesky factorization - dynamic scheduling
**/
void plasma_pspotrf_quark(PLASMA_enum uplo, PLASMA_desc A,
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, tempmm;
float zone = (float) 1.0;
float mzone = (float)-1.0;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
QUARK_CORE_spotrf(
plasma->quark, &task_flags,
PlasmaLower, tempkm, A.mb,
A(k, k), ldak,
sequence, request, A.nb*k);
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_strsm(
plasma->quark, &task_flags,
PlasmaRight, PlasmaLower, PlasmaTrans, PlasmaNonUnit,
tempmm, A.mb, A.mb,
zone, A(k, k), ldak,
A(m, k), ldam);
}
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_ssyrk(
plasma->quark, &task_flags,
PlasmaLower, PlasmaNoTrans,
tempmm, A.mb, A.mb,
-1.0, A(m, k), ldam,
1.0, A(m, m), ldam);
for (n = k+1; n < m; n++) {
QUARK_CORE_sgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaTrans,
tempmm, A.mb, A.mb, A.mb,
mzone, A(m, k), ldam,
A(n, k), A.mb,
zone, A(m, n), ldam);
}
}
}
}
/*
* PlasmaUpper
*/
else {
for (k = 0; k < A.nt; k++) {
tempkm = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
QUARK_CORE_spotrf(
plasma->quark, &task_flags,
PlasmaUpper,
tempkm, A.mb,
A(k, k), ldak,
sequence, request, A.nb*k);
for (m = k+1; m < A.nt; m++) {
tempmm = m == A.nt-1 ? A.n-m*A.nb : A.nb;
QUARK_CORE_strsm(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaUpper, PlasmaTrans, PlasmaNonUnit,
A.nb, tempmm, A.mb,
zone, A(k, k), ldak,
A(k, m), ldak);
}
for (m = k+1; m < A.nt; m++) {
tempmm = m == A.nt-1 ? A.n-m*A.nb : A.nb;
ldam = BLKLDD(A, m);
QUARK_CORE_ssyrk(
plasma->quark, &task_flags,
PlasmaUpper, PlasmaTrans,
tempmm, A.mb, A.mb,
-1.0, A(k, m), ldak,
1.0, A(m, m), ldam);
for (n = k+1; n < m; n++) {
QUARK_CORE_sgemm(
plasma->quark, &task_flags,
PlasmaTrans, PlasmaNoTrans,
A.mb, tempmm, A.mb, A.mb,
mzone, A(k, n), ldak,
A(k, m), ldak,
zone, A(n, m), A.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);
}
}
}
}
}
/***************************************************************************//**
* 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();
}
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization (reduction Householder)
* - dynamic scheduling
**/
void plasma_psormlqrh_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, ldam, 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-1 || N == k; /* No rightmost single-column subdomain */
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_sormlq(
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 < N+BS && m < A.nt) || m == A.nt-1; /* Suck in rightmost single-column domain */
m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
ldam = BLKLDD(A, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_stsmlq(
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-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;
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_sttmlq(
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 / PlasmaTrans
*/
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;
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_sttmlq(
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-1 || N == k; /* No rightmost single-column subdomain */
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 = N+BS-1 == A.nt-2 ? A.nt-1 : min(N+BS-1, A.nt-1); /* Suck in rightmost single-column domain */
m >= N+1;
m--) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
ldam = BLKLDD(A, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_stsmlq(
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_sormlq(
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-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 < B.mt; m++) {
ldbm = BLKLDD(B, m);
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
QUARK_CORE_sttmlq(
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-1 || N == k; /* No rightmost single-column subdomain */
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 == 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_stsmlq(
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_sormlq(
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 / PlasmaTrans
*/
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-1 || N == k; /* No rightmost single-column subdomain */
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_sormlq(
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 < N+BS && n < A.nt) || n == A.nt-1; /* Suck in rightmost single-column domain */
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_stsmlq(
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-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 < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_sttmlq(
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);
}
}
}
}
}
}
}
/***************************************************************************//**
*
**/
void plasma_pdlacpy_quark(PLASMA_enum uplo, PLASMA_desc A, PLASMA_desc B,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int X, Y;
int m, n;
int ldam, ldbm;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
switch (uplo) {
/*
* PlasmaUpper
*/
case PlasmaUpper:
for (m = 0; m < A.mt; m++) {
X = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
if (m < A.nt) {
Y = m == A.nt-1 ? A.n-m*A.nb : A.nb;
QUARK_CORE_dlacpy(
plasma->quark, &task_flags,
PlasmaUpper,
X, Y, A.mb,
A(m, m), ldam,
B(m, m), ldbm);
}
for (n = m+1; n < A.nt; n++) {
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_dlacpy(
plasma->quark, &task_flags,
PlasmaUpperLower,
X, Y, A.mb,
A(m, n), ldam,
B(m, n), ldbm);
}
}
break;
/*
* PlasmaLower
*/
case PlasmaLower:
for (m = 0; m < A.mt; m++) {
X = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
if (m < A.nt) {
Y = m == A.nt-1 ? A.n-m*A.nb : A.nb;
QUARK_CORE_dlacpy(
plasma->quark, &task_flags,
PlasmaLower,
X, Y, A.mb,
A(m, m), ldam,
B(m, m), ldbm);
}
for (n = 0; n < min(m, A.nt); n++) {
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_dlacpy(
plasma->quark, &task_flags,
PlasmaUpperLower,
X, Y, A.mb,
A(m, n), ldam,
B(m, n), ldbm);
}
}
break;
/*
* PlasmaUpperLower
*/
case PlasmaUpperLower:
default:
for (m = 0; m < A.mt; m++) {
X = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
for (n = 0; n < A.nt; n++) {
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_dlacpy(
plasma->quark, &task_flags,
PlasmaUpperLower,
X, Y, A.mb,
A(m, n), ldam,
B(m, n), ldbm);
}
}
}
}
/***************************************************************************//**
* Parallel tile LQ factorization - dynamic scheduling
**/
void plasma_pcgelqf_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_cgelqt(
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_cunmlq(
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_ctslqt(
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_ctsmlq(
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);
}
}
}
}
/***************************************************************************//**
*
**/
void plasma_pdlacpy(plasma_context_t *plasma)
{
PLASMA_enum uplo;
PLASMA_desc A;
PLASMA_desc B;
PLASMA_sequence *sequence;
PLASMA_request *request;
int X, Y;
int m, n;
int next_m;
int next_n;
int ldam, ldbm;
plasma_unpack_args_5(uplo, A, B, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
switch (uplo) {
/*
* PlasmaUpper
*/
case PlasmaUpper:
m = 0;
n = PLASMA_RANK;
while (n >= A.nt) {
m++;
n = n - A.nt + m;
}
while (m < A.mt) {
next_m = m;
next_n = n;
next_n += PLASMA_SIZE;
while (next_n >= A.nt && next_m < A.mt) {
next_m++;
next_n = next_n - A.nt + next_m;
}
X = m == A.mt-1 ? A.m-m*A.mb : A.mb;
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
CORE_dlacpy(
m == n ? uplo : PlasmaUpperLower,
X, Y,
A(m, n), ldam,
B(m, n), ldbm);
n = next_n;
m = next_m;
}
break;
/*
* PlasmaLower
*/
case PlasmaLower:
n = 0;
m = PLASMA_RANK;
while (m >= A.mt) {
n++;
m = m - A.mt + n;
}
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 + next_n;
}
X = m == A.mt-1 ? A.m-m*A.mb : A.mb;
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
CORE_dlacpy(
m == n ? uplo : PlasmaUpperLower,
X, Y,
A(m, n), ldam,
B(m, n), ldbm);
n = next_n;
m = next_m;
}
break;
/*
* PlasmaUpperLower
*/
case PlasmaUpperLower:
default:
n = 0;
m = PLASMA_RANK;
while (m >= A.mt) {
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-m*A.mb : A.mb;
Y = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldam = BLKLDD(A, m);
ldbm = BLKLDD(B, m);
CORE_dlacpy(
PlasmaUpperLower,
X, Y,
A(m, n), ldam,
B(m, n), ldbm);
n = next_n;
m = next_m;
}
break;
}
}
/***************************************************************************//**
* Parallel UU' or L'L operation - dynamic scheduling
**/
void plasma_pclauum_quark(PLASMA_enum uplo, PLASMA_desc A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldam;
int tempkm, tempmm, tempnn;
PLASMA_Complex32_t zone = (PLASMA_Complex32_t)1.0;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
/*
* PlasmaLower
*/
if (uplo == PlasmaLower) {
for (m = 0; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
for(n = 0; n < m; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_cherk(
plasma->quark, &task_flags,
uplo, PlasmaConjTrans,
tempnn, tempmm, A.mb,
1.0, A(m, n), ldam,
1.0, A(n, n), A.mb);
for(k = n+1; k < m; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
PlasmaConjTrans, PlasmaNoTrans,
tempkm, tempnn, tempmm, A.mb,
zone, A(m, k), ldam,
A(m, n), ldam,
zone, A(k, n), A.mb);
}
}
for (n = 0; n < m; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_ctrmm(
plasma->quark, &task_flags,
PlasmaLeft, uplo, PlasmaConjTrans, PlasmaNonUnit,
tempmm, tempnn, A.mb,
zone, A(m, m), ldam,
A(m, n), ldam);
}
QUARK_CORE_clauum(
plasma->quark, &task_flags,
uplo,
tempmm,
A.mb, A(m, m), ldam);
}
}
/*
* PlasmaUpper
*/
else {
for (m = 0; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
for (n = 0; n < m; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_cherk(
plasma->quark, &task_flags,
uplo, PlasmaNoTrans,
tempnn, tempmm, A.mb,
1.0, A(n, m), A.mb,
1.0, A(n, n), A.mb);
for (k = n+1; k < m; k++){
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
QUARK_CORE_cgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaConjTrans,
tempnn, tempkm, tempmm, A.mb,
zone, A(n, m), A.mb,
A(k, m), A.mb,
zone, A(n, k), A.mb);
}
}
for (n = 0; n < m; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_ctrmm(
plasma->quark, &task_flags,
PlasmaRight, uplo, PlasmaConjTrans, PlasmaNonUnit,
tempnn, tempmm, A.mb,
zone, A(m, m), ldam,
A(n, m), A.mb);
}
QUARK_CORE_clauum(
plasma->quark, &task_flags,
uplo,
tempmm,
A.mb, A(m, m), ldam);
}
}
}
/***************************************************************************//**
* Parallel tile LU factorization with no pivoting - dynamic scheduling
**/
void plasma_pzgetrf_nopiv_quark(PLASMA_desc A,
PLASMA_sequence *sequence,
PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n, ib;
int ldak, ldam;
int tempkm, tempkn, tempmm, tempnn;
PLASMA_Complex64_t zone = (PLASMA_Complex64_t) 1.0;
PLASMA_Complex64_t mzone = (PLASMA_Complex64_t)-1.0;
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_zgetrf_nopiv(
plasma->quark, &task_flags,
tempkm, tempkn, ib, A.mb,
A(k, k), ldak,
sequence, request, A.mb*k);
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_ztrsm(
plasma->quark, &task_flags,
PlasmaRight, PlasmaUpper, PlasmaNoTrans, PlasmaNonUnit,
tempmm, tempkn, A.mb,
zone, A(k, k), ldak,
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_ztrsm(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaLower, PlasmaNoTrans, PlasmaUnit,
tempkm, tempnn, A.mb,
zone, A(k, k), ldak,
A(k, n), ldak);
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_zgemm(
plasma->quark, &task_flags,
PlasmaNoTrans, PlasmaNoTrans,
tempmm, tempnn, A.mb, A.mb,
mzone, A(m, k), ldam,
A(k, n), ldak,
zone, A(m, n), ldam);
}
}
}
}