/***************************************************************************//**
* Parallel tile Cholesky factorization - dynamic scheduling
**/
void plasma_pdplrnt_quark( PLASMA_desc A, unsigned long long int seed,
PLASMA_sequence *sequence, PLASMA_request *request )
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int m, n;
int ldam;
int tempmm, tempnn;
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 < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
for (n = 0; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_dplrnt(
plasma->quark, &task_flags,
tempmm, tempnn, A(m, n), ldam,
A.m, m*A.mb, n*A.nb, seed );
}
}
}
void warmup(Quark *q){
int NB = 200;
double *H = (double*) malloc(NB*NB*OOC_NTHREADS*sizeof(double));
double *D = (double*) offload_Alloc(NB*NB*OOC_NTHREADS*sizeof(double), 0);
{
Quark_Task_Flags tflags = Quark_Task_Flags_Initializer;
// for(int r = 0; r < OOC_NTHREADS; r++){
for(int r = 0; r < 2; r++){
QUARK_Task_Flag_Set(&tflags, TASK_LOCK_TO_THREAD, r);
// QUARK_Task_Flag_Set(&tflags, THREAD_SET_TO_MANUAL_SCHEDULING, (r==0)||(r==1));
QUARK_Insert_Task(q, CORE_H2D, &tflags,
sizeof(int), &NB, VALUE,
sizeof(int), &NB, VALUE,
sizeof(double), H+r*NB*NB, INPUT,
sizeof(int), &NB, VALUE,
sizeof(double), D+r*NB*NB, OUTPUT,
sizeof(int), &NB, VALUE,
0);
QUARK_Insert_Task(q, CORE_D2H, &tflags,
sizeof(int), &NB, VALUE,
sizeof(int), &NB, VALUE,
sizeof(double), D+r*NB*NB, INPUT,
sizeof(int), &NB, VALUE,
sizeof(double), H+r*NB*NB, OUTPUT,
sizeof(int), &NB, VALUE,
0);
}
}
QUARK_Barrier(q);
offload_Free(D, 0);
free(H);
}
/***************************************************************************//**
*
**/
void plasma_pslag2d_quark(PLASMA_desc SA, 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);
for(m = 0; m < SA.mt; m++) {
X = m == SA.mt-1 ? SA.m-m*SA.mb : SA.mb;
ldam = BLKLDD(SA, m);
ldbm = BLKLDD(B, m);
for(n = 0; n < SA.nt; n++) {
Y = n == SA.nt-1 ? SA.n-n*SA.nb : SA.nb;
QUARK_CORE_slag2d(
plasma->quark, &task_flags,
X, Y, SA.mb,
SA(m, n), ldam,
B(m, n), ldbm);
}
}
}
/***************************************************************************//**
*
**/
void plasma_pdaxpy_quark(double alpha, 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);
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_daxpy(
plasma->quark, &task_flags,
X, Y, A.mb,
alpha, A(m, n), ldam,
B(m, n), ldbm);
}
}
}
/***************************************************************************//**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
**/
void plasma_pdorgqr_quark(PLASMA_desc A, PLASMA_desc Q, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldak, ldqk, ldam, ldqm;
int tempmm, tempnn, tempkmin, tempkm;
int tempAkm, tempAkn;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
for (k = min(A.mt, A.nt)-1; k >= 0; k--) {
tempAkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
tempAkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
tempkmin = min( tempAkn, tempAkm );
tempkm = k == Q.mt-1 ? Q.m-k*Q.mb : Q.mb;
ldak = BLKLDD(A, k);
ldqk = BLKLDD(Q, k);
for (m = Q.mt - 1; m > k; m--) {
tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
ldam = BLKLDD(A, m);
ldqm = BLKLDD(Q, m);
for (n = 0; n < Q.nt; n++) {
tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
QUARK_CORE_dtsmqr(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
Q.mb, tempnn, tempmm, tempnn, tempAkn, ib, T.nb,
Q(k, n), ldqk,
Q(m, n), ldqm,
A(m, k), ldam,
T(m, k), T.mb);
}
}
for (n = 0; n < Q.nt; n++) {
tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
QUARK_CORE_dormqr(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
tempkm, tempnn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
Q(k, n), ldqk);
}
}
}
/***************************************************************************//**
* Parallel forward substitution for tile LU - dynamic scheduling
**/
void plasma_pztrsmpl_quark(PLASMA_desc A, PLASMA_desc B, 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, ldbk, ldbm;
int tempkm, tempnn, tempkmin, tempmm, tempkn;
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;
tempkmin = k == min(A.mt, A.nt)-1 ? min(A.m, A.n)-k*A.mb : A.mb;
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_zgessm(
plasma->quark, &task_flags,
tempkm, tempnn, tempkmin, ib, L.nb,
IPIV(k, k),
A(k, k), ldak,
B(k, n), ldbk);
}
for (m = k+1; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.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_zssssm(
plasma->quark, &task_flags,
A.nb, tempnn, tempmm, tempnn, tempkn, ib, L.nb,
B(k, n), ldbk,
B(m, n), ldbm,
L(m, k), L.mb,
A(m, k), ldam,
IPIV(m, k));
}
}
}
}
/***************************************************************************//**
* Parallel tile row interchanges - dynamic scheduling
**/
void plasma_pclaswp_quark(PLASMA_desc B, int *IPIV, int inc,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int m, n;
int tempi, tempm, tempmm, tempnn;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
if ( inc > 0 )
{
for (m = 0; m < B.mt; m++) {
tempi = m * B.mb;
tempm = B.m - tempi;
tempmm = m == B.mt-1 ? tempm : B.mb;
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n - n * B.nb : B.nb;
QUARK_CORE_claswp_ontile(
plasma->quark, &task_flags,
plasma_desc_submatrix(B, tempi, n*B.nb, tempm, tempnn),
B(m, n), 1, tempmm, IPIV(m), inc, B(B.mt-1, n) );
}
}
}
else
{
for (m = B.mt-1; m > -1; m--) {
tempi = m * B.mb;
tempm = B.m - tempi;
tempmm = m == B.mt-1 ? tempm : B.mb;
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n - n * B.nb : B.nb;
QUARK_CORE_claswp_ontile(
plasma->quark, &task_flags,
plasma_desc_submatrix(B, tempi, n*B.nb, tempm, tempnn),
B(m, n), 1, tempmm, IPIV(m), inc, B(0, n) );
}
}
}
}
/***************************************************************************//**
* Zeroes a submatrix in tile layout - dynamic scheduling
**/
void plasma_pztile_zero_quark(PLASMA_desc A, PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_Complex64_t *bdl;
plasma_context_t *plasma;
int X1, Y1;
int X2, Y2;
int n, m, ldt;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
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 < A.mt; m++)
{
ldt = BLKLDD(A, m);
for (n = 0; n < A.nt; n++)
{
X1 = n == 0 ? A.j%A.nb : 0;
Y1 = m == 0 ? A.i%A.mb : 0;
X2 = n == A.nt-1 ? (A.j+A.n-1)%A.nb+1 : A.nb;
Y2 = m == A.mt-1 ? (A.i+A.m-1)%A.mb+1 : A.mb;
bdl = ABDL(m, n);
QUARK_Insert_Task(plasma->quark, CORE_ztile_zero_quark, &task_flags,
sizeof(int), &X1, VALUE,
sizeof(int), &X2, VALUE,
sizeof(int), &Y1, VALUE,
sizeof(int), &Y2, VALUE,
sizeof(PLASMA_Complex64_t)*A.bsiz, bdl, OUTPUT | LOCALITY,
sizeof(int), &ldt, VALUE,
0);
}
}
}
/***************************************************************************//**
* Conversion from LAPACK F77 matrix layout to tile layout - dynamic scheduling
**/
void plasma_pzlapack_to_tile_quark(PLASMA_Complex64_t *Af77, int lda, PLASMA_desc A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
PLASMA_Complex64_t *f77;
PLASMA_Complex64_t *bdl;
plasma_context_t *plasma;
int X1, Y1;
int X2, Y2;
int n, m, ldt;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
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 < A.mt; m++)
{
ldt = BLKLDD(A, m);
for (n = 0; n < A.nt; n++)
{
X1 = n == 0 ? A.j%A.nb : 0;
Y1 = m == 0 ? A.i%A.mb : 0;
X2 = n == A.nt-1 ? (A.j+A.n-1)%A.nb+1 : A.nb;
Y2 = m == A.mt-1 ? (A.i+A.m-1)%A.mb+1 : A.mb;
f77 = AF77(m, n);
bdl = ABDL(m, n);
QUARK_CORE_zlacpy(
plasma->quark, &task_flags,
PlasmaUpperLower, (Y2-Y1), (X2-X1), A.mb,
&(f77[X1*lda+Y1]), lda,
&(bdl[X1*lda+Y1]), ldt);
}
}
}
/***************************************************************************//**
* 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 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);
}
}
}
/** ****************************************************************************
*
* @ingroup InPlaceTransformation
*
* plasma_pcgetmi2_quark - realises nprob independant transpositions. Each
* subproblem is a tile of mb-by-nb elements.
* This function use an extra space of PLASMA_SIZE*(mb*nb). This is a
* maximum in case of dynamic scheduling.
*
*
*******************************************************************************
*
* @param[in] idep
* PlasmaIPT_Nodep: No fake dependencies are added.
* PlasmaIPT_Panel: A gatherv is added on each panel and panel size is m*nb.
* PlasmaIPT_All: A gatherv is added on the whole matrix.
*
* @param[in] odep
* PlasmaIPT_Nodep: No fake dependencies are added.
* PlasmaIPT_Panel: A gatherv is added on each panel and panel size is m*nb.
* PlasmaIPT_All: A gatherv is added on the whole matrix.
*
* @param[in] storev
* PlasmaColumnWise: Data stored in column major.
* PlasmaRowWise: Data stored in row major.
*
* @param[in] m
* Number of row of A if tiles are sorted in column major format,
* number of columns otherwise.
*
* @param[in] n
* Number of columns of A if tiles are sorted in column major format,
* number of rows otherwise.
*
* @param[in] mb
* Number of rows in each individual subproblem if storev == PlasmaColumnWise,
* number of columns otherwise. m%mb must be 0.
*
* @param[in] nb
* Number of columns in each individual subproblem if storev == PlasmaColumnWise,
* number of rows otherwise. n%nb must be 0.
*
* @param[in,out] A
* Matrix of size m*n.
*
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
*
* @param[out] request
* Identifies this function call (for exception handling purposes).
*
*******************************************************************************
*
* @sa plasma_pcgetmi2
*
******************************************************************************/
void plasma_pcgetmi2_quark(PLASMA_enum idep, PLASMA_enum odep, PLASMA_enum storev,
int m, int n, int mb, int nb, PLASMA_Complex32_t *A,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
PLASMA_Complex32_t *Al, *Ap;
int i, j, nprob, mt, nt;
int bsiz, psiz, size;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
/* quick return */
if( (mb < 2) || (nb < 2) ) {
return ;
}
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
bsiz = mb*nb;
if ( storev == PlasmaColumnwise ) {
psiz = m*nb;
mt = ( m / mb );
nt = ( n / nb );
} else {
psiz = n*mb;
mt = ( n / nb );
nt = ( m / mb );
}
size = m*n;
switch ( idep ) {
/*
* Dependencies on each panel as input
*/
case PlasmaIPT_Panel:
switch ( odep ) {
case PlasmaIPT_Panel:
for (j=0; j<nt; j++) {
Ap = A + (psiz*j);
for (i=0; i<mt; i++) {
Al = Ap + i*bsiz;
QUARK_CORE_cgetrip_f1(
plasma->quark, &task_flags,
mb, nb, Al, bsiz,
Ap, psiz, INOUT|GATHERV);
}
}
break;
case PlasmaIPT_All:
for (j=0; j<nt; j++) {
Ap = A + (psiz*j);
for (i=0; i<mt; i++) {
Al = Ap + i*bsiz;
QUARK_CORE_cgetrip_f2(plasma->quark, &task_flags,
mb, nb, Al, bsiz,
Ap, size, INPUT,
A, size, INOUT|GATHERV);
}
}
break;
case PlasmaIPT_NoDep:
default:
for (j=0; j<nt; j++) {
Ap = A + (psiz*j);
for (i=0; i<mt; i++) {
Al = Ap + i*bsiz;
QUARK_CORE_cgetrip_f1(
plasma->quark, &task_flags,
mb, nb, Al, bsiz,
Ap, psiz, INPUT);
}
}
}
break;
/*
* Dependency on all the matrix as input
*/
case PlasmaIPT_All:
switch ( odep ) {
case PlasmaIPT_Panel:
for (j=0; j<nt; j++) {
Ap = A + (psiz*j);
for (i=0; i<mt; i++) {
Al = Ap + i*bsiz;
QUARK_CORE_cgetrip_f2(
plasma->quark, &task_flags,
mb, nb, Al, bsiz,
A, size, INPUT,
Ap, psiz, INOUT|GATHERV);
}
}
break;
case PlasmaIPT_All:
nprob = mt*nt;
for (i=0; i<nprob; i++) {
QUARK_CORE_cgetrip_f1(plasma->quark, &task_flags,
mb, nb, &(A[ i*bsiz ]), bsiz,
A, size, INOUT|GATHERV);
}
break;
case PlasmaIPT_NoDep:
default:
nprob = mt*nt;
for (i=0; i<nprob; i++) {
QUARK_CORE_cgetrip_f1(plasma->quark, &task_flags,
mb, nb, &(A[ i*bsiz ]), bsiz,
A, size, INPUT);
}
}
break;
/*
* No Dependencies as input
*/
case PlasmaIPT_NoDep:
default:
switch ( odep ) {
case PlasmaIPT_Panel:
for (j=0; j<nt; j++) {
Ap = A + (psiz*j);
for (i=0; i<mt; i++) {
Al = Ap + i*bsiz;
QUARK_CORE_cgetrip_f1(
plasma->quark, &task_flags,
mb, nb, Al, bsiz,
Ap, psiz, INOUT|GATHERV);
}
}
break;
case PlasmaIPT_All:
nprob = mt*nt;
for (i=0; i<nprob; i++) {
QUARK_CORE_cgetrip_f1(plasma->quark, &task_flags,
mb, nb, &(A[ i*bsiz ]), bsiz,
A, size, INOUT|GATHERV);
}
break;
case PlasmaIPT_NoDep:
default:
nprob = mt*nt;
for (i=0; i<nprob; i++) {
QUARK_CORE_cgetrip(plasma->quark, &task_flags,
mb, nb, &(A[ i*bsiz ]), bsiz);
}
}
}
}
/***************************************************************************//**
* Parallel application of Q using tile V - LQ factorization (reduction
* Householder) - dynamic scheduling
**/
void plasma_pzunmlqrh_quark(PLASMA_enum side, PLASMA_enum trans,
PLASMA_desc A, PLASMA_desc B, PLASMA_desc T, int BS,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int K, N, RD, lastRD;
int ldaN, ldak;
int ldbN, ldbm, ldbNRD;
int tempNn, tempkm, tempnn, tempmm, tempNRDn, tempkmin;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
K = min(A.mt, A.nt);
if (side == PlasmaLeft ) {
if (trans == PlasmaNoTrans) {
/*
* PlasmaLeft / PlasmaNoTrans
*/
for (k = 0; k < K; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
ldbN = BLKLDD(B, N);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempNn, tempnn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(N, n), ldbN);
}
for (m = N+1; m < min(N+BS, A.nt); m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.nb, tempnn, tempmm, tempnn,
tempkm, ib, T.nb,
B(N, n), ldbN,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
}
for (RD = BS; RD < A.nt-k; RD *= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
ldbN = BLKLDD(B, N );
ldbNRD = BLKLDD(B, N+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempNRDn, tempnn,
tempkm, ib, T.nb,
B (N, n), ldbN,
B (N+RD, n), ldbNRD,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
}
} else {
/*
* PlasmaLeft / PlasmaConjTrans
*/
for (k = K-1; k >= 0; k--) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
lastRD = 0;
for (RD = BS; RD < A.nt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; RD /= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
ldbN = BLKLDD(B, N );
ldbNRD = BLKLDD(B, N+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
B.nb, tempnn, tempNRDn, tempnn,
tempkm, ib, T.nb,
B (N, n), ldbN,
B (N+RD, n), ldbNRD,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
ldbN = BLKLDD(B, N);
for (m = min(N+BS, A.nt)-1; m > N; m--) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn,
tempkm, ib, T.nb,
B(N, n), ldbN,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempNn, tempnn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(N, n), ldbN);
}
}
}
}
} else {
if (trans == PlasmaNoTrans) {
/*
* PlasmaRight / PlasmaNoTrans
*/
for (k = K-1; k >= 0; k--) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
lastRD = 0;
for (RD = BS; RD < A.nt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; RD /= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
for (m = 0; m < B.mt; m++) {
ldbm = BLKLDD(B, m);
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempNRDn,
tempkm, ib, T.nb,
B (m, N ), ldbm,
B (m, N+RD), ldbm,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
for (n = min(N+BS, A.nt)-1; n > N; n--) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn,
tempkm, ib, T.nb,
B(m, N), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn,
tempkmin, ib, T.nb,
A(k, N), ldak,
T(k, N), T.mb,
B(m, N), ldbm);
}
}
}
} else {
/*
* PlasmaRight / PlasmaConjTrans
*/
for (k = 0; k < K; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
for (N = k; N < A.nt; N += BS) {
tempNn = N == A.nt-1 ? A.n-N*A.nb : A.nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
for (m = 0; m < B.mt; m++) {
ldbm = BLKLDD(B, m);
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn,
tempkmin, ib, T.nb,
A(k, N), ldaN,
T(k, N), T.mb,
B(m, N), ldbm);
}
for (n = N+1; n < min(N+BS, A.nt); n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempNn, tempmm, tempnn,
tempkm, ib, T.nb,
B(m, N), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
for (RD = BS; RD < A.nt-k; RD *= 2) {
for (N = k; N+RD < A.nt; N += 2*RD) {
tempNRDn = N+RD == A.nt-1 ? A.n-(N+RD)*A.nb : A.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zttmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempNRDn,
tempkm, ib, T.nb,
B (m, N ), ldbm,
B (m, N+RD), ldbm,
A (k, N+RD), ldak,
T2(k, N+RD), T.mb);
}
}
}
}
}
}
}
/***************************************************************************//**
* Parallel 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 - LQ factorization - dynamic scheduling
**/
void plasma_pzunmlq_quark(PLASMA_enum side, PLASMA_enum trans,
PLASMA_desc A, PLASMA_desc B, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n;
int ldak, ldan, ldbk, ldbm;
int tempmm, tempnn, tempkn, tempkm, tempkmin;
int ib, minMT, minM;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
if (A.m > A.n) {
minM = A.n;
minMT = A.nt;
} else {
minM = A.m;
minMT = A.mt;
}
if (side == PlasmaLeft ) {
if (trans == PlasmaNoTrans) {
/*
* PlasmaLeft / PlasmaNoTrans
*/
for (k = 0; k < minMT; k++) {
tempkm = k == B.mt -1 ? B.m -k*B.mb : B.mb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempkm, tempnn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(k, n), ldbk);
}
for (m = k+1; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib, T.nb,
B(k, n), ldbk,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
}
}
else {
/*
* PlasmaLeft / PlasmaConjTrans
*/
for (k = minMT-1; k >= 0; k--) {
tempkm = k == B.mt -1 ? B.m -k*B.mb : B.mb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
for (m = B.mt-1; m > k; m--) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
B.mb, tempnn, tempmm, tempnn, tempkmin, ib, T.nb,
B(k, n), ldbk,
B(m, n), ldbm,
A(k, m), ldak,
T(k, m), T.mb);
}
}
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempkm, tempnn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(k, n), ldbk);
}
}
}
}
else {
if (trans == PlasmaNoTrans) {
/*
* PlasmaRight / PlasmaNoTrans
*/
for (k = minMT-1; k >= 0; k--) {
tempkn = k == B.nt -1 ? B.n -k*B.nb : B.nb;
tempkmin = k == minMT-1 ? minM-k*A.nb : A.nb;
ldak = BLKLDD(A, k);
for (n = B.nt-1; n > k; n--) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
ldan = BLKLDD(A, n);
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib, T.nb,
B(m, k), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempkn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(m, k), ldbm);
}
}
}
else {
/*
* PlasmaRight / PlasmaConjTrans
*/
for (k = 0; k < minMT; k++) {
tempkn = k == B.nt -1 ? B.n -k*B.nb : B.nb;
tempkmin = k == minMT-1 ? minM-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, tempkn, tempkmin, ib, T.nb,
A(k, k), ldak,
T(k, k), T.mb,
B(m, k), ldbm);
}
for (n = k+1; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
for (m = 0; m < B.mt; m++) {
tempmm = m == B.mt-1 ? B.m-m*B.mb : B.mb;
ldbm = BLKLDD(B, m);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempnn, tempkmin, ib, T.nb,
B(m, k), ldbm,
B(m, n), ldbm,
A(k, n), ldak,
T(k, n), T.mb);
}
}
}
}
}
}
/***************************************************************************//**
* Parallel tile BAND Tridiagonal Reduction - dynamic scheduler
**/
void plasma_pzherbt_quark(PLASMA_enum uplo,
PLASMA_desc A, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
int k, m, n, i, j;
int ldak, ldam, ldan, ldaj, ldai;
int tempkn, tempmm, tempnn, tempjj;
int ib;
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
ib = PLASMA_IB;
if (uplo == PlasmaLower) {
for (k = 0; k < A.nt-1; k++){
tempkn = k+1 == A.nt-1 ? A.n-(k+1)*A.nb : A.nb;
ldak = BLKLDD(A, k+1);
QUARK_CORE_zgeqrt(
plasma->quark, &task_flags,
tempkn, A.nb, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb);
/* LEFT and RIGHT on the symmetric diagonal block */
QUARK_CORE_zherfb(
plasma->quark, &task_flags,
PlasmaLower,
tempkn, tempkn, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb,
A(k+1, k+1), ldak);
/* RIGHT on the remaining tiles until the bottom */
for (m = k+2; m < A.mt ; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_zunmqr(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempmm, A.nb, tempkn, ib, T.nb,
A(k+1, k), ldak,
T(k+1, k), T.mb,
A(m , k+1), ldam);
}
for (m = k+2; m < A.mt; m++) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
ldam = BLKLDD(A, m);
QUARK_CORE_ztsqrt(
plasma->quark, &task_flags,
tempmm, A.nb, ib, T.nb,
A(k+1, k), ldak,
A(m , k), ldam,
T(m , k), T.mb);
/* LEFT */
for (i = k+2; i < m; i++) {
ldai = BLKLDD(A, i);
QUARK_CORE_ztsmqr_hetra1(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaConjTrans,
A.mb, A.nb, tempmm, A.nb, A.nb, ib, T.nb,
A(i, k+1), ldai,
A(m, i), ldam,
A(m, k), ldam,
T(m, k), T.mb);
}
/* RIGHT */
for (j = m+1; j < A.mt ; j++) {
tempjj = j == A.mt-1 ? A.m-j*A.mb : A.mb;
ldaj = BLKLDD(A, j);
QUARK_CORE_ztsmqr(
plasma->quark, &task_flags,
PlasmaRight, PlasmaNoTrans,
tempjj, A.nb, tempjj, tempmm, A.nb, ib, T.nb,
A(j, k+1), ldaj,
A(j, m), ldaj,
A(m, k), ldam,
T(m, k), T.mb);
}
/* LEFT->RIGHT */
QUARK_CORE_ztsmqr_corner(
plasma->quark, &task_flags,
A.nb, A.nb, tempmm, A.nb, tempmm, tempmm, A.nb, ib, T.nb,
A(k+1, k+1), ldak,
A(m , k+1), ldam,
A(m , m), ldam,
A(m , k), ldam,
T(m , k), T.mb);
}
}
}
else {
for (k = 0; k < A.nt-1; k++){
tempkn = k+1 == A.nt-1 ? A.n-(k+1)*A.nb : A.nb;
ldak = BLKLDD(A, k+1);
QUARK_CORE_zgelqt(
plasma->quark, &task_flags,
A.nb, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb);
/* RIGHT and LEFT on the symmetric diagonal block */
QUARK_CORE_zherfb(
plasma->quark, &task_flags,
PlasmaUpper,
tempkn, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb,
A(k+1, k+1), ldak);
/* LEFT on the remaining tiles until the left side */
for (n = k+2; n < A.nt ; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
QUARK_CORE_zunmlq(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
A.nb, tempnn, tempkn, ib, T.nb,
A(k, k+1), A.nb,
T(k, k+1), T.mb,
A(k+1, n), ldak);
}
for (n = k+2; n < A.nt; n++) {
tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
ldan = BLKLDD(A, n);
QUARK_CORE_ztslqt(
plasma->quark, &task_flags,
A.nb, tempnn, ib, T.nb,
A(k, k+1), A.nb,
A(k, n), A.nb,
T(k, n), T.mb);
/* RIGHT */
for (i = k+2; i < n; i++) {
ldai = BLKLDD(A, i);
QUARK_CORE_ztsmlq_hetra1(
plasma->quark, &task_flags,
PlasmaRight, PlasmaConjTrans,
A.mb, A.nb, A.nb, tempnn, A.nb, ib, T.nb,
A(k+1, i), ldak,
A(i, n), ldai,
A(k, n), A.nb,
T(k, n), T.mb);
}
/* LEFT */
for (j = n+1; j < A.nt ; j++) {
tempjj = j == A.nt-1 ? A.n-j*A.nb : A.nb;
ldaj = BLKLDD(A, j);
QUARK_CORE_ztsmlq(
plasma->quark, &task_flags,
PlasmaLeft, PlasmaNoTrans,
A.nb, tempjj, tempnn, tempjj, A.nb, ib, T.nb,
A(k+1, j), ldak,
A(n, j), ldan,
A(k, n), A.nb,
T(k, n), T.mb);
}
/* RIGHT->LEFT */
QUARK_CORE_ztsmlq_corner(
plasma->quark, &task_flags,
A.nb, A.nb, A.nb, tempnn, tempnn, tempnn, A.nb, ib, T.nb,
A(k+1, k+1), ldak,
A(k+1, n), ldak,
A(n , n), ldan,
A(k , n), A.nb,
T(k , n), T.mb);
}
}
}
}
/***************************************************************************//**
* Parallel 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 Reduction from BAND tridiagonal to the final condensed form - dynamic scheduler
**/
void plasma_pzhbrdt_quark(PLASMA_enum uplo,
PLASMA_desc A, double *D, double *E, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
#ifdef COMPLEX
static PLASMA_Complex64_t zone = (PLASMA_Complex64_t) 1.0;
static double dzero = (double) 0.0;
PLASMA_Complex64_t ztmp;
double absztmp;
#endif
PLASMA_Complex64_t *C, *S;
int blksweep, lcsweep, blkid, lcNB;
int N, NB, NT, grsiz, lcgrsiz;
int i;
size_t eltsize = plasma_element_size(A.dtyp);
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
NT = A.nt;
N = A.m;
NB = A.mb;
/* Quick return */
if (N == 0){
return;
}
if (NB == 0) {
memset(D, 0, N*sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
#ifdef COMPLEX
for (i=0; i<N; i++)
D[i] = cabs(*A(i,i));
#else
for (i=0; i<N; i++)
D[i] = *A(i,i);
#endif
return;
}
/*
* Barrier is used because the bulge have to wait until
* the reduction to band has been finish.
* otherwise, I can remove this BARRIER when I integrate
* the function dependencies link inside the reduction to
* band. Keep in min the case when NB=1, where no bulge-chasing.
*/
/***************************************************************/
QUARK_Barrier(plasma->quark);
tblg = -Wtimming();
/***************************************************************/
/*
* Case NB=1 ==> matrix is already Bidiagonal. no need to bulge.
* Make diagonal and superdiagonal elements real, storing them in
* D and E. if PlasmaLower, first transform lower bidiagonal form
* to upper bidiagonal by applying plane rotations/ Householder
* from the left, overwriting superdiagonal elements then make
* elements real of the resulting upper Bidiagonal. if PlasmaUpper
* then make its elements real. For Q, PT: ZSCAL should be done
* in case of WANTQ.
*/
if (NB == 1){
memset(D, 0, N *sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
#ifdef COMPLEX
if(uplo==PlasmaLower){
for (i=0; i<N; i++)
{
D[i] = creal( *A(i, i) ); /* diag value */
if( i < (N-1)) { /* lower off-diag value */
ztmp = *A((i+1),i);
absztmp = cabs(ztmp);
*A((i+1),i) = absztmp;
E[i] = absztmp;
if(absztmp != dzero)
ztmp = (PLASMA_Complex64_t) (ztmp / absztmp);
else
ztmp = zone;
if(i<(N-2)) *A((i+2),(i+1)) = *A((i+2),(i+1)) * ztmp;
/* for Q: ZSCAL should be done in case of WANTQ */
}
}
} else { /* PlasmaUpper */
for (i=0; i<N; i++)
{
D[i] = creal( *A(i,i) ); /* diag value*/
if(i<(N-1)) { /* lower off-diag value */
ztmp = *A(i, (i+1));
absztmp = cabs(ztmp);
*A(i,(i+1)) = absztmp;
E[i] = absztmp;
if(absztmp != dzero)
ztmp = (PLASMA_Complex64_t) (ztmp / absztmp);
else
ztmp = zone;
if(i<(N-2)) *A((i+1),(i+2)) = *A((i+1),(i+2)) * ztmp;
/* for Q: ZSCAL should be done in case of WANTQ. HERE NEED THE multiply by CONJ(T) */
}
}
} /* end PlasmaUpper*/
#else
if( uplo == PlasmaLower ){
for (i=0; i < N-1; i++) {
D[i] = *A(i, i);
E[i] = *A(i+1, i);
}
D[i] = *A(i, i);
} else {
for (i=0; i < N-1; i++) {
D[i] = *A(i, i );
E[i] = *A(i, i+1);
}
D[i] = *A(i, i);
}
#endif
return;
}
/* Case N<NB ==> matrix is very small and better to call lapack XHETRD. */
if( N <= 0 ) /* this will be removed we don t need it. */
{
PLASMA_Complex64_t *work, *TTau;
int info, ldwork = N*N;
work = (PLASMA_Complex64_t *) plasma_shared_alloc(plasma, ldwork, PlasmaComplexDouble);
TTau = (PLASMA_Complex64_t *) plasma_shared_alloc(plasma, N, PlasmaComplexDouble);
info = LAPACKE_zhetrd_work(LAPACK_COL_MAJOR, lapack_const(uplo), N,
A(0,0), A.lm, D, E, TTau, work, ldwork);
plasma_shared_free(plasma, (void*) work);
plasma_shared_free(plasma, (void*) TTau);
if( info == 0 )
sequence->status = PLASMA_SUCCESS;
else
plasma_sequence_flush(plasma->quark, sequence, request, info);
return;
}
/* General case NB > 1 && N > NB */
C = (PLASMA_Complex64_t *) plasma_shared_alloc(plasma, N, PlasmaComplexDouble);
S = (PLASMA_Complex64_t *) plasma_shared_alloc(plasma, N, PlasmaComplexDouble);
/***************************************************************************
* START BULGE CHASING CODE
**************************************************************************/
/*
* Initialisation of local parameter. those parameter should be
* input or tuned parameter.
*/
grsiz = 1;
if( NB > 160 ) {
grsiz = 1;
}
else if( NB > 100 ) {
grsiz = 1;
/*
if( N < 5000 )
grsiz = 1;
else
grsiz = 2;
*/
} else {
grsiz = 2;
}
grsiz = max(1, grsiz);
/*grsiz=1;*/
/*printf(" Version -dp- N %5d NB %5d lcNB %5d grsiz %5d A.ln %5d A.nb %5d \n",N,NB,lcNB,grsiz,A.ln,A.nb);*/
for (blksweep = 0; blksweep<NT; blksweep++){
lcNB = blksweep == NT-1 ? A.n-blksweep*A.nb : A.nb;
/*printf(" Version -dp- N %5d NB %5d lcNB %5d grsiz %5d blksweep%5d NT %5d \n",N,NB,lcNB,grsiz,blksweep,NT);*/
for (lcsweep = 0; lcsweep<lcNB; lcsweep++){
for (blkid = blksweep; blkid<NT; blkid=blkid+grsiz){
lcgrsiz = (blkid+1) < NT ? grsiz : NT-blkid;
/*printf(" Version -dp- N %5d NB %5d lcNB %5d grsiz %5d lcgrsiz %5d blkid %5d \n",N,NB,lcNB,grsiz,lcgrsiz,blkid);*/
QUARK_CORE_ztrdalg_v2(
plasma->quark, &task_flags,
uplo, &A, C, S,
lcgrsiz, lcsweep, blkid, blksweep);
}
}
}
/*
* Barrier used only for now, to be sure that everything
* is done before copying the D and E and free workspace.
* this will be removed later when D and E are directly filled
* during the bulge process.
*/
QUARK_Barrier(plasma->quark);
tblg += Wtimming();
printf(" done with bulge %lf \n\n\n",tblg);
plasma_shared_free(plasma, (void*) C);
plasma_shared_free(plasma, (void*) S);
/*
* STORE THE RESULTING diagonal/off-diagonal in D AND E
*/
memset(D, 0, N *sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
/* Make diagonal and superdiagonal elements real,
* storing them in D and E
*/
/* In complex case, the off diagonal element are
* not necessary real. we have to make off-diagonal
* elements real and copy them to E.
* When using HouseHolder elimination,
* the ZLARFG give us a real as output so, all the
* diagonal/off-diagonal element except the last one are already
* real and thus we need only to take the abs of the last
* one.
* */
#ifdef COMPLEX
if(uplo==PlasmaLower){
for (i=0; i < N-1 ; i++)
{
D[i] = creal( *A(i,i) );
/*
* Alternative for Householder case, all off-diag
* are real except the last off-diag, where we
* have to take the abs
*/
if(i<(N-2))
E[i] = creal(*A(i+1, i));
else
E[i] = cabs( *A(i+1, i));
}
D[i] = creal( *A(i, i) );
} else { /* PlasmaUpper */
for (i=0; i<N-1; i++)
{
D[i] = creal( *A(i,i) );
/*
* Alternative for Householder case, all off-diag
* are real except the last off-diag, where we
* have to take the abs
*/
if( i < (N-2) )
E[i] = creal(*A(i, (i+1)));
else
E[i] = cabs(*A(i, (i+1)));
}
D[i] = creal( *A(i, i) );
} /* end PlasmaUpper */
#else
if( uplo == PlasmaLower ){
for (i=0; i < N-1; i++) {
D[i] = *A(i, i);
E[i] = *A(i+1, i);
}
D[i] = *A(i, i);
} else {
for (i=0; i < N-1; i++) {
D[i] = *A(i, i );
E[i] = *A(i, i+1);
}
D[i] = *A(i, i);
}
#endif
} /* END FUNCTION */
/***************************************************************************//**
* 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 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 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);
}
}
}
}
/***************************************************************************//**
* 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 Reduction from BAND tridiagonal to the final condensed form - dynamic scheduler
**/
void plasma_pdsbrdt_quark(PLASMA_enum uplo,
PLASMA_desc A, double *D, double *E, PLASMA_desc T,
PLASMA_sequence *sequence, PLASMA_request *request)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
#ifdef COMPLEX
static double zone = (double) 1.0;
static double dzero = (double) 0.0;
double ztmp;
double absztmp;
#endif
double *C, *S;
int N, NB, INgrsiz, INthgrsiz, BAND;
int myid, grsiz, shift=3, stt, st, ed, stind, edind;
int blklastind, colpt, PCOL, ACOL, MCOL;
int stepercol, mylastid, grnb, grid;
int *DEP,*MAXID;
int i, j, m;
int thgrsiz, thgrnb, thgrid, thed;
size_t eltsize = plasma_element_size(A.dtyp);
plasma = plasma_context_self();
if (sequence->status != PLASMA_SUCCESS)
return;
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
N = A.m;
NB = A.mb;
/* Quick return */
if (N == 0){
return;
}
if (NB == 0) {
memset(D, 0, N*sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
#ifdef COMPLEX
for (i=0; i<N; i++)
D[i] = fabs(*A(i,i));
#else
for (i=0; i<N; i++)
D[i] = *A(i,i);
#endif
return;
}
/*
* Barrier is used because the bulge have to wait until
* the reduction to band has been finish.
* otherwise, I can remove this BARRIER when I integrate
* the function dependencies link inside the reduction to
* band. Keep in min the case when NB=1, where no bulge-chasing.
*/
/***************************************************************/
QUARK_Barrier(plasma->quark);
tblg = -Wtimming();
/***************************************************************/
/*
* Case NB=1 ==> matrix is already Bidiagonal. no need to bulge.
* Make diagonal and superdiagonal elements real, storing them in
* D and E. if PlasmaLower, first transform lower bidiagonal form
* to upper bidiagonal by applying plane rotations/ Householder
* from the left, overwriting superdiagonal elements then make
* elements real of the resulting upper Bidiagonal. if PlasmaUpper
* then make its elements real. For Q, PT: ZSCAL should be done
* in case of WANTQ.
*/
if (NB == 1){
memset(D, 0, N *sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
#ifdef COMPLEX
if(uplo==PlasmaLower){
for (i=0; i<N; i++)
{
D[i] = ( *A(i, i) ); /* diag value */
if( i < (N-1)) { /* lower off-diag value */
ztmp = *A((i+1),i);
absztmp = fabs(ztmp);
*A((i+1),i) = absztmp;
E[i] = absztmp;
if(absztmp != dzero)
ztmp = (double) (ztmp / absztmp);
else
ztmp = zone;
if(i<(N-2)) *A((i+2),(i+1)) = *A((i+2),(i+1)) * ztmp;
/* for Q: ZSCAL should be done in case of WANTQ */
}
}
} else { /* PlasmaUpper */
for (i=0; i<N; i++)
{
D[i] = ( *A(i,i) ); /* diag value*/
if(i<(N-1)) { /* lower off-diag value */
ztmp = *A(i, (i+1));
absztmp = fabs(ztmp);
*A(i,(i+1)) = absztmp;
E[i] = absztmp;
if(absztmp != dzero)
ztmp = (double) (ztmp / absztmp);
else
ztmp = zone;
if(i<(N-2)) *A((i+1),(i+2)) = *A((i+1),(i+2)) * ztmp;
/* for Q: ZSCAL should be done in case of WANTQ. HERE NEED THE multiply by CONJ(T) */
}
}
} /* end PlasmaUpper*/
#else
if( uplo == PlasmaLower ){
for (i=0; i < N-1; i++) {
D[i] = *A(i, i);
E[i] = *A(i+1, i);
}
D[i] = *A(i, i);
} else {
for (i=0; i < N-1; i++) {
D[i] = *A(i, i );
E[i] = *A(i, i+1);
}
D[i] = *A(i, i);
}
#endif
return;
}
/* Case N<NB ==> matrix is very small and better to call lapack XHETRD. */
if( N <= 0 ) /* this will be removed we don t need it. */
{
double *work, *TTau;
int info, ldwork = N*N;
work = (double *) plasma_shared_alloc(plasma, ldwork, PlasmaRealDouble);
TTau = (double *) plasma_shared_alloc(plasma, N, PlasmaRealDouble);
info = LAPACKE_dsytrd_work(LAPACK_COL_MAJOR, lapack_const(uplo), N,
A(0,0), A.lm, D, E, TTau, work, ldwork);
plasma_shared_free(plasma, (void*) work);
plasma_shared_free(plasma, (void*) TTau);
if( info == 0 )
sequence->status = PLASMA_SUCCESS;
else
plasma_sequence_flush(plasma->quark, sequence, request, info);
return;
}
/* General case NB > 1 && N > NB */
DEP = (int *) plasma_shared_alloc(plasma, N+1, PlasmaInteger );
MAXID = (int *) plasma_shared_alloc(plasma, N+1, PlasmaInteger );
C = (double *) plasma_shared_alloc(plasma, N, PlasmaRealDouble);
S = (double *) plasma_shared_alloc(plasma, N, PlasmaRealDouble);
memset(MAXID,0,(N+1)*sizeof(int));
/***************************************************************************
* START BULGE CHASING CODE
**************************************************************************/
/*
* Initialisation of local parameter. those parameter should be
* input or tuned parameter.
*/
INgrsiz = 1;
if( NB > 160 ) {
INgrsiz = 2;
}
else if( NB > 100 ) {
if( N < 5000 )
INgrsiz = 2;
else
INgrsiz = 4;
} else {
INgrsiz = 6;
}
INthgrsiz = N;
BAND = 0;
grsiz = INgrsiz;
thgrsiz = INthgrsiz;
if( grsiz == 0 ) grsiz = 6;
if( thgrsiz == 0 ) thgrsiz = N;
i = shift/grsiz;
stepercol = i*grsiz == shift ? i:i+1;
i = (N-2)/thgrsiz;
thgrnb = i*thgrsiz == (N-2) ? i:i+1;
for (thgrid = 1; thgrid<=thgrnb; thgrid++){
stt = (thgrid-1)*thgrsiz+1;
thed = min( (stt + thgrsiz -1), (N-2));
for (i = stt; i <= N-2; i++){
ed=min(i,thed);
if(stt>ed)break;
for (m = 1; m <=stepercol; m++){
st=stt;
for (j = st; j <=ed; j++){
/* PCOL: dependency on the ID of the master of the group of the previous column. (Previous Column:PCOL). */
/* ACOL: dependency on the ID of the master of the previous group of my column. (Acctual Column:ACOL). (it is 0(NULL) for myid=1) */
/* MCOL: OUTPUT dependency on the my ID, to be used by the next ID. (My Column: MCOL). I am the master of this group. */
myid = (i-j)*(stepercol*grsiz) +(m-1)*grsiz + 1;
mylastid = myid+grsiz-1;
PCOL = mylastid+shift-1; /* to know the dependent ID of the previous column. need to know the master of its group */
MAXID[j] = myid;
PCOL = min(PCOL,MAXID[j-1]); /* for the last columns, we might do only 1 or 2 kernel, so the PCOL will be wrong. this is to force it to the last ID of the previous col.*/
grnb = PCOL/grsiz;
grid = grnb*grsiz == PCOL ? grnb:grnb+1;
PCOL = (grid-1)*grsiz +1; /* give me the ID of the master of the group of the previous column. */
ACOL = myid-grsiz;
if(myid==1)ACOL=0;
MCOL = myid;
QUARK_CORE_dtrdalg(
plasma->quark, &task_flags,
uplo, N, NB,
&A, C, S, i, j, m, grsiz, BAND,
DEP(PCOL), DEP(ACOL), DEP(MCOL) );
if(mylastid%2 ==0){
blklastind = (mylastid/2)*NB+1+j-1;
}else{
colpt = ((mylastid+1)/2)*NB + 1 +j -1 ;
stind = colpt-NB+1;
edind = min(colpt,N);
if( (stind>=edind-1) && (edind==N) )
blklastind=N;
else
blklastind=0;
}
if(blklastind >= (N-1)) stt=stt+1;
} /* END for j=st:ed */
} /* END for m=1:stepercol */
} /* END for i=1:MINMN-2 */
} /* END for thgrid=1:thgrnb */
/*
* Barrier used only for now, to be sure that everything
* is done before copying the D and E and free workspace.
* this will be removed later when D and E are directly filled
* during the bulge process.
*/
QUARK_Barrier(plasma->quark);
tblg += Wtimming();
//printf(" done with bulge %lf \n\n\n",tblg);
plasma_shared_free(plasma, (void*) DEP);
plasma_shared_free(plasma, (void*) MAXID);
plasma_shared_free(plasma, (void*) C);
plasma_shared_free(plasma, (void*) S);
/*
* STORE THE RESULTING diagonal/off-diagonal in D AND E
*/
memset(D, 0, N *sizeof(double));
memset(E, 0, (N-1)*sizeof(double));
/* Make diagonal and superdiagonal elements real,
* storing them in D and E
*/
/* In complex case, the off diagonal element are
* not necessary real. we have to make off-diagonal
* elements real and copy them to E.
* When using HouseHolder elimination,
* the ZLARFG give us a real as output so, all the
* diagonal/off-diagonal element except the last one are already
* real and thus we need only to take the abs of the last
* one.
* */
#ifdef COMPLEX
if(uplo==PlasmaLower){
for (i=0; i < N-1 ; i++)
{
D[i] = ( *A(i,i) );
/*
* Alternative for Householder case, all off-diag
* are real except the last off-diag, where we
* have to take the abs
*/
if(i<(N-2))
E[i] = (*A(i+1, i));
else
E[i] = fabs( *A(i+1, i));
}
D[i] = ( *A(i, i) );
} else { /* PlasmaUpper */
for (i=0; i<N-1; i++)
{
D[i] = ( *A(i,i) );
/*
* Alternative for Householder case, all off-diag
* are real except the last off-diag, where we
* have to take the abs
*/
if( i < (N-2) )
E[i] = (*A(i, (i+1)));
else
E[i] = fabs(*A(i, (i+1)));
}
D[i] = ( *A(i, i) );
} /* end PlasmaUpper */
#else
if( uplo == PlasmaLower ){
for (i=0; i < N-1; i++) {
D[i] = *A(i, i);
E[i] = *A(i+1, i);
}
D[i] = *A(i, i);
} else {
for (i=0; i < N-1; i++) {
D[i] = *A(i, i );
E[i] = *A(i, i+1);
}
D[i] = *A(i, i);
}
#endif
} /* END FUNCTION */
/***************************************************************************//**
* 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 application of Q using tile V - QR factorization (reduction Householder)
* - dynamic scheduling
**/
void plasma_pcunmqrrh_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, M, RD, lastRD;
int ldaM, ldam, ldan, ldaMRD;
int ldbM, ldbm, ldbMRD;
int tempMm, tempkn, tempnn, tempmm, tempMRDm, 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 == PlasmaConjTrans) {
/*
* PlasmaLeft / PlasmaConjTrans
*/
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;
tempkmin = min(tempMm, tempkn);
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_cunmqr(
plasma->quark, &task_flags,
side, trans,
tempMm, tempnn,
tempkmin, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb,
B(M, n), ldbM);
}
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;
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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
A.nb, tempnn, tempmm, tempnn,
tempkn, ib, T.nb,
B(M, n), ldbM,
B(m, n), ldbm,
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;
ldbM = BLKLDD(B, M );
ldbMRD = BLKLDD(B, M+RD);
ldaMRD = BLKLDD(A, M+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_cttmqr(
plasma->quark, &task_flags,
side, trans,
A.nb, tempnn, tempMRDm, tempnn,
tempkn, ib, T.nb,
B (M, n), ldbM,
B (M+RD, n), ldbMRD,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
}
}
}
}
} else {
/*
* PlasmaLeft / PlasmaNoTrans
*/
for (k = K-1; k >= 0; k--) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
lastRD = 0;
for (RD = BS; RD < A.mt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; 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;
ldbM = BLKLDD(B, M );
ldbMRD = BLKLDD(B, M+RD);
ldaMRD = BLKLDD(A, M+RD);
for (n = 0; n < B.nt; n++) {
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.nb;
QUARK_CORE_cttmqr(
plasma->quark, &task_flags,
side, trans,
A.nb, tempnn, tempMRDm, tempnn,
tempkn, ib, T.nb,
B (M, n), ldbM,
B (M+RD, n), ldbMRD,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
}
}
}
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;
tempkmin = min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
ldbM = BLKLDD(B, M);
for (m = M+BS-1 == A.mt-2 ? A.mt-1 : min(M+BS-1, A.mt-1); /* Suck in bottom single-row domain */
m >= M+1;
m--) {
tempmm = m == A.mt-1 ? A.m-m*A.mb : A.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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
A.nb, tempnn, tempmm, tempnn,
tempkn, ib, T.nb,
B(M, n), ldbM,
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,
tempMm, tempnn,
tempkmin, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb,
B(M, n), ldbM);
}
}
}
}
} else {
if (trans == PlasmaConjTrans) {
/*
* PlasmaRight / PlasmaConjTrans
*/
for (k = K-1; k >= 0; k--) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
lastRD = 0;
for (RD = BS; RD < A.mt-k; RD *= 2)
lastRD = RD;
for (RD = lastRD; RD >= BS; 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;
ldaMRD = BLKLDD(A, M+RD);
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_cttmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempMRDm,
tempkn, ib, T.nb,
B (m, M), ldbm,
B (m, M+RD), ldbm,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
}
}
}
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;
tempkmin = min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
ldbM = BLKLDD(B, M);
for (n = M+BS-1 == A.mt-2 ? A.mt-1 : min(M+BS-1, A.mt-1); /* Suck in bottom single-row domain */
n >= M+1;
n--) {
ldan = BLKLDD(A, n);
tempnn = n == B.nt-1 ? B.n-n*B.nb : B.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_ctsmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, tempMm, tempmm, tempnn,
tempkn, ib, T.nb,
B(m, M), ldbm,
B(m, n), ldbm,
A(n, k), ldan,
T(n, k), T.mb);
}
}
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_cunmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, tempMm,
tempkmin, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb,
B(m, M), ldbm);
}
}
}
} else {
/*
* PlasmaRight / PlasmaNoTrans
*/
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;
tempkmin = min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
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_cunmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, tempMm,
tempkmin, ib, T.nb,
A(M, k), ldaM,
T(M, k), T.mb,
B(m, M), ldbm);
}
for (n = M+1;
(n < M+BS && n < A.mt) || n == A.mt-1; /* Suck in bottom single-row domain */
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, tempMm, tempmm, tempnn,
tempkn, ib, T.nb,
B(m, M), ldbm,
B(m, n), ldbm,
A(n, k), ldan,
T(n, 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;
ldaMRD = BLKLDD(A, M+RD);
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_cttmqr(
plasma->quark, &task_flags,
side, trans,
tempmm, B.nb, tempmm, tempMRDm,
tempkn, ib, T.nb,
B (m, M ), ldbm,
B (m, M+RD), ldbm,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T.mb);
}
}
}
}
}
}
}
/***************************************************************************//**
* 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 */
}
}
}
}
}
}
static int
RunTest(int *iparam, double *dparam, real_Double_t *t_)
{
plasma_context_t *plasma;
Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
PLASMA_Complex64_t *A, *A2 = NULL;
real_Double_t t;
int *ipiv, *ipiv2 = NULL;
int i;
int m = iparam[TIMING_N];
int n = iparam[TIMING_NRHS];
int check = iparam[TIMING_CHECK];
int lda = m;
PLASMA_sequence *sequence = NULL;
PLASMA_request request = PLASMA_REQUEST_INITIALIZER;
/* Initialize Plasma */
PLASMA_Init( iparam[TIMING_THRDNBR] );
PLASMA_Set(PLASMA_SCHEDULING_MODE, PLASMA_DYNAMIC_SCHEDULING );
PLASMA_Disable(PLASMA_AUTOTUNING);
PLASMA_Set(PLASMA_TILE_SIZE, iparam[TIMING_NB] );
PLASMA_Set(PLASMA_INNER_BLOCK_SIZE, iparam[TIMING_IB] );
/* Allocate Data */
A = (PLASMA_Complex64_t *)malloc(lda*n*sizeof(PLASMA_Complex64_t));
/* Check if unable to allocate memory */
if ( (! A) ) {
printf("Out of Memory \n ");
return -1;
}
/* Initialiaze Data */
LAPACKE_zlarnv_work(1, ISEED, lda*n, A);
/* Allocate Workspace */
ipiv = (int *)malloc( n*sizeof(int) );
/* Save A in lapack layout for check */
if ( check ) {
A2 = (PLASMA_Complex64_t *)malloc(lda*n*sizeof(PLASMA_Complex64_t));
ipiv2 = (int *)malloc( n*sizeof(int) );
LAPACKE_zlacpy_work(LAPACK_COL_MAJOR,' ', m, n, A, lda, A2, lda);
LAPACKE_zgetrf_work(LAPACK_COL_MAJOR, m, n, A2, lda, ipiv2 );
}
plasma = plasma_context_self();
PLASMA_Sequence_Create(&sequence);
QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
QUARK_Task_Flag_Set(&task_flags, TASK_THREAD_COUNT, iparam[TIMING_THRDNBR] );
plasma_dynamic_spawn();
CORE_zgetrf_reclap_init();
t = -cWtime();
QUARK_CORE_zgetrf_reclap(plasma->quark, &task_flags,
m, n, n,
A, lda, ipiv,
sequence, &request,
0, 0,
iparam[TIMING_THRDNBR]);
PLASMA_Sequence_Wait(sequence);
t += cWtime();
*t_ = t;
PLASMA_Sequence_Destroy(sequence);
/* Check the solution */
if ( check )
{
double *work = (double *)malloc(max(m,n)*sizeof(double));
/* Check ipiv */
for(i=0; i<n; i++)
{
if( ipiv[i] != ipiv2[i] ) {
fprintf(stderr, "\nPLASMA (ipiv[%d] = %d, A[%d] = %e) / LAPACK (ipiv[%d] = %d, A[%d] = [%e])\n",
i, ipiv[i], i, creal(A[ i * lda + i ]),
i, ipiv2[i], i, creal(A2[ i * lda + i ]));
break;
}
}
dparam[TIMING_ANORM] = LAPACKE_zlange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A, lda, work);
dparam[TIMING_XNORM] = LAPACKE_zlange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A2, lda, work);
dparam[TIMING_BNORM] = 0.0;
CORE_zaxpy( m, n, -1.0, A, lda, A2, lda);
dparam[TIMING_RES] = LAPACKE_zlange_work(LAPACK_COL_MAJOR, lapack_const(PlasmaMaxNorm),
m, n, A2, lda, work);
free( A2 );
free( ipiv2 );
free( work );
}
free( A );
free( ipiv );
PLASMA_Finalize();
return 0;
}
/***************************************************************************//**
*
**/
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 matrix-matrix multiplication - dynamic scheduling
**/
void plasma_pzgemm_quark(PLASMA_enum transA, PLASMA_enum transB,
PLASMA_Complex64_t alpha, PLASMA_desc A, PLASMA_desc B,
PLASMA_Complex64_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 ldam, ldak, ldbn, ldbk, ldcm;
int tempmm, tempnn, tempkn, tempkm;
PLASMA_Complex64_t zbeta;
PLASMA_Complex64_t zone = (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);
for (m = 0; m < C.mt; m++) {
tempmm = m == C.mt-1 ? C.m-m*C.mb : C.mb;
ldcm = BLKLDD(C, m);
for (n = 0; n < C.nt; n++) {
tempnn = n == C.nt-1 ? C.n-n*C.nb : C.nb;
/*
* A: PlasmaNoTrans / B: PlasmaNoTrans
*/
if (transA == PlasmaNoTrans) {
ldam = BLKLDD(A, m);
if (transB == PlasmaNoTrans) {
for (k = 0; k < A.nt; k++) {
tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
ldbk = BLKLDD(B, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_zgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkn, A.mb,
alpha, A(m, k), ldam, /* lda * Z */
B(k, n), ldbk, /* ldb * Y */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
/*
* A: PlasmaNoTrans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
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_zgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkn, A.mb,
alpha, A(m, k), ldam, /* lda * Z */
B(n, k), ldbn, /* ldb * Z */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
}
/*
* A: Plasma[Conj]Trans / B: PlasmaNoTrans
*/
else {
if (transB == PlasmaNoTrans) {
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_zgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* lda * X */
B(k, n), ldbk, /* ldb * Y */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
/*
* A: Plasma[Conj]Trans / B: Plasma[Conj]Trans
*/
else {
ldbn = BLKLDD(B, n);
for (k = 0; k < A.mt; k++) {
tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
ldak = BLKLDD(A, k);
zbeta = k == 0 ? beta : zone;
QUARK_CORE_zgemm(
plasma->quark, &task_flags,
transA, transB,
tempmm, tempnn, tempkm, A.mb,
alpha, A(k, m), ldak, /* lda * X */
B(n, k), ldbn, /* ldb * Z */
zbeta, C(m, n), ldcm); /* ldc * Y */
}
}
}
}
}
}
/***************************************************************************//**
* 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));
}
}
}
}