25 #include "sopalin/sopalin_data.h"
56 if ( cblk->
cblktype & CBLK_TASKS_2D ) {
60 for ( blok = cblk->
fblokptr + 1, m = 0; blok < lblk; blok++, m++ ) {
67 while ( ( blok < lblk ) &&
68 ( blok[0].fcblknm == blok[1].fcblknm ) &&
69 ( blok[0].lcblknm == blok[1].lcblknm ) )
113 if ( cblk->
cblktype & CBLK_TASKS_2D ) {
117 for ( blokA = blokB; blokA < lblk; blokA++ ) {
119 cblk, fcblk, blokA, blokB, prio );
122 while ( ( blokA < lblk ) &&
123 ( blokA[0].fcblknm == blokA[1].fcblknm ) &&
124 ( blokA[0].lcblknm == blokA[1].lcblknm ) )
131 starpu_data_unregister_submit( blokB->
handler[1] );
136 cblk, blokB, fcblk, prio );
173 for (k=0; k<solvmtx->
cblknbr; k++, cblk++){
176 assert( cblk->
cblktype & CBLK_LAYOUT_2D );
178 if ( cblk->
cblktype & CBLK_IN_SCHUR ) {
189 for(m=0; blok < lblk; blok++, m++ )
192 cblk_n = fcblk - solvmtx->
cblktab;
197 cblknbr - pastix_imin( k + m, cblk_n ) );
201 starpu_data_unregister_submit( cblk->
handler[1] );
242 for (k=0; k<solvmtx->
cblknbr; k++, cblk++){
245 assert( cblk->
cblktype & CBLK_LAYOUT_2D );
247 for ( m = cblk[0].brownum; m < cblk[1].
brownum; m++ ) {
251 if ( lcblk->
cblktype & CBLK_IN_SCHUR ) {
256 cblk_n = fcblk - solvmtx->
cblktab;
258 assert( fcblk == cblk );
263 cblknbr - pastix_imin( k + m, cblk_n ) );
266 if ( cblk->
cblktype & CBLK_IN_SCHUR ) {
275 for ( k = 0; k < solvmtx->
cblknbr; k++, cblk++ ) {
277 starpu_data_unregister_submit( cblk->
handler[1] );
320 for ( k = 0; k <= solvmtx->
cblkmax1d; k++, cblk++ ) {
323 assert( cblk->
cblktype & CBLK_LAYOUT_2D );
325 if ( cblk->
cblktype & CBLK_IN_SCHUR ) {
329 if ( cblk->
cblktype & CBLK_TASKS_2D ) {
339 for(m=0; blok < lblk; blok++, m++ )
342 cblk_n = fcblk - solvmtx->
cblktab;
345 cblknbr - pastix_imin( k + m, cblk_n ) );
349 starpu_data_unregister_submit( cblk->
handler[1] );
357 for ( k = solvmtx->
cblkmin2d; k < solvmtx->cblknbr; k++, cblk++ ) {
359 if ( cblk->
cblktype & CBLK_IN_SCHUR ) {
363 if ( ! ( cblk->
cblktype & CBLK_TASKS_2D ) ) {
371 for ( blok = cblk->
fblokptr + 1, m = 0; blok < lblk; blok++, m++ ) {
376 cblknbr - pastix_imin( k + m, cblk_n ) );
379 while( ( blok < lblk ) &&
381 ( blok[0].lcblknm == blok[1].lcblknm ) )
428 for ( k = 0; k < cblknbr; k++, fcblk++ ) {
430 for ( m = fcblk[0].brownum; m < fcblk[1].
brownum; m++ ) {
435 if ( cblk->
cblktype & CBLK_IN_SCHUR ) {
439 if( ( cblk->
cblktype & CBLK_TASKS_2D ) &&
440 ( blok_prev != NULL ) &&
451 if ( fcblk->
cblktype & CBLK_IN_SCHUR ) {
459 for ( k = 0; k < solvmtx->
cblknbr; k++, cblk++ ) {
461 if ( cblk->
handler[1] != NULL ) {
462 starpu_data_unregister_submit( cblk->
handler[1] );
500 sopalin_data_t *sopalin_data )
509 if (pastix_data->
starpu == NULL) {
514 if ( sdesc == NULL ) {
521 sdesc = sopalin_data->
solvmtx->starpu_desc;
524 starpu_profiling_status_set(STARPU_PROFILING_ENABLE);
525 #if defined(STARPU_USE_FXT)
527 starpu_fxt_start_profiling();
530 #if defined(PASTIX_STARPU_STATS)
558 #if defined(PASTIX_STARPU_STATS)
563 starpu_task_wait_for_all();
564 #if defined(PASTIX_WITH_MPI)
565 starpu_mpi_wait_for_all( pastix_data->
pastix_comm );
569 #if defined(STARPU_USE_FXT)
571 starpu_fxt_stop_profiling();
574 starpu_profiling_status_set(STARPU_PROFILING_DISABLE);
575 #if defined(PASTIX_STARPU_STATS)
577 print_stats( sub, com, pastix_data->
solvmatr );
BEGIN_C_DECLS typedef int pastix_int_t
void starpu_sparse_matrix_getoncpu(starpu_sparse_matrix_desc_t *desc)
Submit asynchronous calls to retrieve the data on main memory.
void starpu_task_blok_sscalo(sopalin_data_t *sopalin_data, pastix_trans_t trans, const SolverCblk *cblk, SolverBlok *blok, int prio)
TODO.
void starpu_sparse_matrix_init(SolverMatrix *solvmtx, pastix_mtxtype_t mtxtype, int nodes, int myrank, pastix_coeftype_t flttype)
Generate the StarPU descriptor of the sparse matrix.
void pastix_starpu_init(pastix_data_t *pastix, int *argc, char **argv[], const int *bindtab)
Startup the StarPU runtime system.
void starpu_task_blok_sgemmsp(sopalin_data_t *sopalin_data, pastix_coefside_t sideA, pastix_coefside_t sideB, pastix_trans_t trans, SolverCblk *cblk, SolverCblk *fcblk, const SolverBlok *blokA, const SolverBlok *blokB, int prio)
StarPU GPU implementation.
void starpu_task_blok_ssytrf(sopalin_data_t *sopalin_data, SolverCblk *cblk, int prio)
TODO.
void starpu_task_blok_strsmsp(sopalin_data_t *sopalin_data, pastix_coefside_t coef, pastix_side_t side, pastix_uplo_t uplo, pastix_trans_t trans, pastix_diag_t diag, const SolverCblk *cblk, SolverBlok *blok, int prio)
StarPU GPU implementation.
void starpu_task_cblk_ssytrfsp(sopalin_data_t *sopalin_data, SolverCblk *cblk, int prio)
TODO.
void starpu_sparse_cblk_wont_use(pastix_coefside_t coef, SolverCblk *cblk)
Submit asynchronous calls to retrieve the data on main memory.
void starpu_task_cblk_sgemmsp(sopalin_data_t *sopalin_data, pastix_coefside_t sideA, pastix_coefside_t sideB, pastix_trans_t trans, const SolverCblk *cblk, const SolverBlok *blok, SolverCblk *fcblk, int prio)
StarPU GPU implementation.
StarPU descriptor stucture for the sparse matrix.
PASTIX_Comm inter_node_comm
Main PaStiX data structure.
void starpu_ssytrf_sp1dplus_ll(sopalin_data_t *sopalin_data, starpu_sparse_matrix_desc_t *desc)
Perform a sparse LDL^t factorization with 1D kernels.
void starpu_task_ssytrfsp(sopalin_data_t *sopalin_data, SolverCblk *cblk, int prio)
TODO.
void starpu_task_sytrf_sgemmsp(sopalin_data_t *sopalin_data, SolverCblk *cblk, SolverBlok *blokB, SolverCblk *fcblk, int prio)
TODO.
void starpu_ssytrf(pastix_data_t *pastix_data, sopalin_data_t *sopalin_data)
Perform a sparse LDL^t factorization using StarPU runtime.
void starpu_ssytrf_sp2d_ll(sopalin_data_t *sopalin_data, starpu_sparse_matrix_desc_t *desc)
Perform a sparse LDL^t factorization with 1D and 2D kernels.
void starpu_ssytrf_sp1dplus_rl(sopalin_data_t *sopalin_data, starpu_sparse_matrix_desc_t *desc)
Perform a sparse LDL^t factorization with 1D kernels.
void starpu_ssytrf_sp2d_rl(sopalin_data_t *sopalin_data, starpu_sparse_matrix_desc_t *desc)
Perform a sparse LDL^t factorization with 1D and 2D kernels.
SolverBlok *restrict bloktab
pastix_int_t *restrict browtab
SolverCblk *restrict cblktab
Solver column block structure.
Solver column block structure.