int main(int argc, char *argv[]) {
double t1, t2;
Node root;
StealStack *ss;
/* initialize stealstacks and comm. layer */
ss = ss_init(&argc, &argv);
/* determine benchmark parameters */
uts_parseParams(argc, argv);
/* Initialize trace collection structures */
ss_initStats(ss);
/* show parameter settings */
if (ss_get_thread_num() == 0) {
uts_printParams();
}
fflush(NULL);
// Workers will return 1 from ss_start(), all others (managers)
// will return 0 here once the computation ends
if (ss_start(sizeof(Node), chunkSize)) {
/* initialize root node and push on thread 0 stack */
if (ss_get_thread_num() == 0) {
uts_initRoot(&root, type);
#ifdef TRACE
ss_markSteal(ss, 0); // first session is own "parent session"
#endif
ss_put_work(ss, &root);
}
/* time parallel search */
t1 = uts_wctime();
parTreeSearch(ss);
t2 = uts_wctime();
ss->walltime = t2 - t1;
#ifdef TRACE
ss->startTime = t1;
ss->sessionRecords[SS_IDLE][ss->entries[SS_IDLE] - 1].endTime = t2;
#endif
}
ss_stop();
/* display results */
showStats();
ss_finalize();
return 0;
}
/***************************************************************************//**
* Parallel tile LU factorization - static scheduling
**/
void plasma_pzgetrf_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;
PLASMA_Complex64_t *work;
plasma_unpack_args_5(A, L, IPIV, sequence, request);
if (sequence->status != PLASMA_SUCCESS)
return;
work = (PLASMA_Complex64_t*)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_zgetrf_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_ztstrf(
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_zgessm(
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_zssssm(
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 application of Q using tile V - LQ factorization - static scheduling
**/
void plasma_pzunmlq(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, ldbm;
int tempmm, tempnn, tempkm, tempkmin;
int minMT, minM;
int ib = PLASMA_IB;
PLASMA_Complex64_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_Complex64_t*)plasma_private_alloc(plasma, ib*T.nb, T.dtyp);
ss_init(B.mt, B.nt, min(A.mt, A.nt));
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
k = minMT-1;
n = PLASMA_RANK;
while (n >= B.nt) {
k--;
n = n-B.nt;
}
m = B.mt-1;
while (k >= 0 && n < B.nt) {
next_n = n;
next_m = m;
next_k = k;
next_m--;
if (next_m == k-1) {
next_n += PLASMA_SIZE;
while (next_n >= B.nt && next_k >= 0) {
next_k--;
next_n = next_n-B.nt;
}
next_m = B.mt-1;
}
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);
ldbm = BLKLDD(B, m);
if (m == k) {
CORE_zunmlq(
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_ztsmlq(
side, trans,
A.mb, tempnn, tempmm, tempnn, tempkmin, ib,
B(k, n), ldbk,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb,
work, ib);
ss_cond_set(m, n, k);
}
m = next_m;
n = next_n;
k = next_k;
}
plasma_private_free(plasma, work);
ss_finalize();
}
/***************************************************************************//**
* 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 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 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 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();
}
