/*
* Initialize module on the communicator
*/
static int mca_coll_hcoll_module_enable(mca_coll_base_module_t *module,
struct ompi_communicator_t *comm)
{
mca_coll_hcoll_module_t *hcoll_module = (mca_coll_hcoll_module_t*) module;
hcoll_module->comm = comm;
if (OMPI_SUCCESS != __save_coll_handlers(hcoll_module)){
HCOL_ERROR("coll_hcol: __save_coll_handlers failed");
return OMPI_ERROR;
}
hcoll_set_runtime_tag_offset(-100,mca_pml.pml_max_tag);
hcoll_module->hcoll_context =
hcoll_create_context((rte_grp_handle_t)comm);
if (NULL == hcoll_module->hcoll_context){
HCOL_VERBOSE(1,"hcoll_create_context returned NULL");
return OMPI_ERROR;
}
if (comm != &ompi_mpi_comm_world.comm){
mca_coll_hcoll_module_list_item_wrapper_t *mw =
OBJ_NEW(mca_coll_hcoll_module_list_item_wrapper_t);
mw->module = hcoll_module;
OBJ_RETAIN(hcoll_module->comm);
opal_list_append(&mca_coll_hcoll_component.active_modules,
(opal_list_item_t*)mw);
}
return OMPI_SUCCESS;
}
static void* get_coll_handle(void)
{
ompi_request_t *ompi_req;
ompi_free_list_item_t *item;
OMPI_FREE_LIST_WAIT_MT(&(mca_coll_hcoll_component.requests),item);
if (OPAL_UNLIKELY(NULL == item)) {
HCOL_ERROR("Wait for free list failed.\n");
return NULL;
}
ompi_req = (ompi_request_t *)item;
OMPI_REQUEST_INIT(ompi_req,false);
return (void *)ompi_req;
}
static void* get_coll_handle(void)
{
ompi_request_t *ompi_req;
opal_free_list_item_t *item;
item = opal_free_list_wait (&(mca_coll_hcoll_component.requests));
if (OPAL_UNLIKELY(NULL == item)) {
HCOL_ERROR("Wait for free list failed.\n");
return NULL;
}
ompi_req = (ompi_request_t *)item;
OMPI_REQUEST_INIT(ompi_req,false);
ompi_req->req_complete_cb = NULL;
ompi_req->req_status.MPI_ERROR = MPI_SUCCESS;
ompi_req->req_free = request_free;
return (void *)ompi_req;
}
/*
* Initialize module on the communicator
*/
static int mca_coll_hcoll_module_enable(mca_coll_base_module_t *module,
struct ompi_communicator_t *comm)
{
int ret;
if (OMPI_SUCCESS != mca_coll_hcoll_save_coll_handlers((mca_coll_hcoll_module_t *)module)) {
HCOL_ERROR("coll_hcol: mca_coll_hcoll_save_coll_handlers failed");
return OMPI_ERROR;
}
ret = ompi_attr_set_c(COMM_ATTR, comm, &comm->c_keyhash, hcoll_comm_attr_keyval, (void *)module, false);
if (OMPI_SUCCESS != ret) {
HCOL_VERBOSE(1,"hcoll ompi_attr_set_c failed");
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
/*
* Initialize module on the communicator
*/
static int mca_coll_hcoll_module_enable(mca_coll_base_module_t *module,
struct ompi_communicator_t *comm)
{
mca_coll_hcoll_module_t *hcoll_module = (mca_coll_hcoll_module_t*) module;
hcoll_module->comm = comm;
if (OMPI_SUCCESS != __save_coll_handlers(hcoll_module)){
HCOL_ERROR("coll_hcol: __save_coll_handlers failed");
return OMPI_ERROR;
}
hcoll_set_runtime_tag_offset(-100,mca_pml.pml_max_tag);
hcoll_set_rte_halt_flag_address(&ompi_mpi_finalized);
hcoll_set_rte_halt_flag_size(sizeof(ompi_mpi_finalized));
hcoll_module->hcoll_context =
hcoll_create_context((rte_grp_handle_t)comm);
if (NULL == hcoll_module->hcoll_context){
HCOL_VERBOSE(1,"hcoll_create_context returned NULL");
return OMPI_ERROR;
}
#if 0
{
printf("\033[33mrank %d: DOING EXTRA TEST\033[0m\n",ompi_comm_rank(comm));
fflush(stdout);
sleep(1);
rte_ec_handle_t handle;
rte_grp_handle_t world_group = hcoll_rte_functions.rte_world_group_fn();
int peer;
const int max_count = 10000000;
const int step = max_count/100;
int buf = 0;
int i;
rte_request_handle_t req;
peer = (ompi_comm_rank(comm)+1)%2;
hcoll_rte_functions.get_ec_handles_fn(1,&peer,world_group,&handle);
for (i=1; i<max_count+1; i++){
if (0 == ompi_comm_rank(comm)){
if (i/step*step == i){
printf("%d %% done...\n",i/step);fflush(stdout);
}
buf = 1;
hcoll_rte_functions.send_fn(DTE_INT32,1,&buf,handle,world_group,0,&req);
} else {
hcoll_rte_functions.recv_fn(DTE_INT32,1,&buf,handle,world_group,0,&req);
}
int completed = 0;
hcoll_rte_functions.test_fn(&req,&completed);
while(!completed){
hcoll_rte_functions.test_fn(&req,&completed);
/*hcoll_rte_functions.rte_progress_fn();*/
opal_progress();
}
}
printf("\033[32mrank %d: EXTRA TEST PASS\033[0m\n",ompi_comm_rank(comm));
fflush(stdout);
sleep(1);
}
#endif
return OMPI_SUCCESS;
}
/*
* Invoked when there's a new communicator that has been created.
* Look at the communicator and decide which set of functions and
* priority we want to return.
*/
mca_coll_base_module_t *
mca_coll_hcoll_comm_query(struct ompi_communicator_t *comm, int *priority)
{
mca_coll_base_module_t *module;
mca_coll_hcoll_module_t *hcoll_module;
ompi_attribute_fn_ptr_union_t del_fn;
ompi_attribute_fn_ptr_union_t copy_fn;
mca_coll_hcoll_component_t *cm;
int err;
int rc;
cm = &mca_coll_hcoll_component;
*priority = 0;
module = NULL;
if (!cm->hcoll_enable) {
return NULL;
}
if (OMPI_COMM_IS_INTER(comm) || ompi_comm_size(comm) < cm->hcoll_np
|| ompi_comm_size(comm) < 2) {
return NULL;
}
if (!cm->libhcoll_initialized)
{
/* libhcoll should be initialized here since current implmentation of
mxm bcol in libhcoll needs world_group fully functional during init
world_group, i.e. ompi_comm_world, is not ready at hcoll component open
call */
opal_progress_register(mca_coll_hcoll_progress);
HCOL_VERBOSE(10,"Calling hcoll_init();");
#if HCOLL_API >= HCOLL_VERSION(3,2)
hcoll_read_init_opts(&cm->init_opts);
cm->init_opts->base_tag = MCA_COLL_BASE_TAG_HCOLL_BASE;
cm->init_opts->max_tag = mca_pml.pml_max_tag;
cm->init_opts->enable_thread_support = ompi_mpi_thread_multiple;
rc = hcoll_init_with_opts(&cm->init_opts);
#else
hcoll_set_runtime_tag_offset(MCA_COLL_BASE_TAG_HCOLL_BASE, mca_pml.pml_max_tag);
rc = hcoll_init();
#endif
if (HCOLL_SUCCESS != rc) {
cm->hcoll_enable = 0;
opal_progress_unregister(mca_coll_hcoll_progress);
HCOL_ERROR("Hcol library init failed");
return NULL;
}
#if HCOLL_API >= HCOLL_VERSION(3,2)
if (cm->using_mem_hooks && cm->init_opts->mem_hook_needed) {
#else
if (cm->using_mem_hooks && hcoll_check_mem_release_cb_needed()) {
#endif
opal_mem_hooks_register_release(mca_coll_hcoll_mem_release_cb, NULL);
} else {
cm->using_mem_hooks = 0;
}
copy_fn.attr_communicator_copy_fn = (MPI_Comm_internal_copy_attr_function*) MPI_COMM_NULL_COPY_FN;
del_fn.attr_communicator_delete_fn = hcoll_comm_attr_del_fn;
err = ompi_attr_create_keyval(COMM_ATTR, copy_fn, del_fn, &hcoll_comm_attr_keyval, NULL ,0, NULL);
if (OMPI_SUCCESS != err) {
cm->hcoll_enable = 0;
hcoll_finalize();
opal_progress_unregister(mca_coll_hcoll_progress);
HCOL_ERROR("Hcol comm keyval create failed");
return NULL;
}
if (mca_coll_hcoll_component.derived_types_support_enabled) {
copy_fn.attr_datatype_copy_fn = (MPI_Type_internal_copy_attr_function *) MPI_TYPE_NULL_COPY_FN;
del_fn.attr_datatype_delete_fn = hcoll_type_attr_del_fn;
err = ompi_attr_create_keyval(TYPE_ATTR, copy_fn, del_fn, &hcoll_type_attr_keyval, NULL ,0, NULL);
if (OMPI_SUCCESS != err) {
cm->hcoll_enable = 0;
hcoll_finalize();
opal_progress_unregister(mca_coll_hcoll_progress);
HCOL_ERROR("Hcol type keyval create failed");
return NULL;
}
}
OBJ_CONSTRUCT(&cm->dtypes, opal_free_list_t);
opal_free_list_init(&cm->dtypes, sizeof(mca_coll_hcoll_dtype_t),
8, OBJ_CLASS(mca_coll_hcoll_dtype_t), 0, 0,
32, -1, 32, NULL, 0, NULL, NULL, NULL);
}
hcoll_module = OBJ_NEW(mca_coll_hcoll_module_t);
if (!hcoll_module) {
if (!cm->libhcoll_initialized) {
cm->hcoll_enable = 0;
hcoll_finalize();
opal_progress_unregister(mca_coll_hcoll_progress);
}
return NULL;
}
hcoll_module->comm = comm;
HCOL_VERBOSE(10,"Creating hcoll_context for comm %p, comm_id %d, comm_size %d",
(void*)comm,comm->c_contextid,ompi_comm_size(comm));
hcoll_module->hcoll_context =
hcoll_create_context((rte_grp_handle_t)comm);
if (NULL == hcoll_module->hcoll_context) {
HCOL_VERBOSE(1,"hcoll_create_context returned NULL");
OBJ_RELEASE(hcoll_module);
if (!cm->libhcoll_initialized) {
cm->hcoll_enable = 0;
hcoll_finalize();
opal_progress_unregister(mca_coll_hcoll_progress);
}
return NULL;
}
hcoll_module->super.coll_module_enable = mca_coll_hcoll_module_enable;
hcoll_module->super.coll_barrier = hcoll_collectives.coll_barrier ? mca_coll_hcoll_barrier : NULL;
hcoll_module->super.coll_bcast = hcoll_collectives.coll_bcast ? mca_coll_hcoll_bcast : NULL;
hcoll_module->super.coll_allgather = hcoll_collectives.coll_allgather ? mca_coll_hcoll_allgather : NULL;
hcoll_module->super.coll_allgatherv = hcoll_collectives.coll_allgatherv ? mca_coll_hcoll_allgatherv : NULL;
hcoll_module->super.coll_allreduce = hcoll_collectives.coll_allreduce ? mca_coll_hcoll_allreduce : NULL;
hcoll_module->super.coll_alltoall = hcoll_collectives.coll_alltoall ? mca_coll_hcoll_alltoall : NULL;
hcoll_module->super.coll_alltoallv = hcoll_collectives.coll_alltoallv ? mca_coll_hcoll_alltoallv : NULL;
hcoll_module->super.coll_gatherv = hcoll_collectives.coll_gatherv ? mca_coll_hcoll_gatherv : NULL;
hcoll_module->super.coll_reduce = hcoll_collectives.coll_reduce ? mca_coll_hcoll_reduce : NULL;
hcoll_module->super.coll_ibarrier = hcoll_collectives.coll_ibarrier ? mca_coll_hcoll_ibarrier : NULL;
hcoll_module->super.coll_ibcast = hcoll_collectives.coll_ibcast ? mca_coll_hcoll_ibcast : NULL;
hcoll_module->super.coll_iallgather = hcoll_collectives.coll_iallgather ? mca_coll_hcoll_iallgather : NULL;
#if HCOLL_API >= HCOLL_VERSION(3,5)
hcoll_module->super.coll_iallgatherv = hcoll_collectives.coll_iallgatherv ? mca_coll_hcoll_iallgatherv : NULL;
#else
hcoll_module->super.coll_iallgatherv = NULL;
#endif
hcoll_module->super.coll_iallreduce = hcoll_collectives.coll_iallreduce ? mca_coll_hcoll_iallreduce : NULL;
#if HCOLL_API >= HCOLL_VERSION(3,5)
hcoll_module->super.coll_ireduce = hcoll_collectives.coll_ireduce ? mca_coll_hcoll_ireduce : NULL;
#else
hcoll_module->super.coll_ireduce = NULL;
#endif
hcoll_module->super.coll_gather = /*hcoll_collectives.coll_gather ? mca_coll_hcoll_gather :*/ NULL;
hcoll_module->super.coll_igatherv = hcoll_collectives.coll_igatherv ? mca_coll_hcoll_igatherv : NULL;
hcoll_module->super.coll_ialltoall = /*hcoll_collectives.coll_ialltoall ? mca_coll_hcoll_ialltoall : */ NULL;
#if HCOLL_API >= HCOLL_VERSION(3,7)
hcoll_module->super.coll_ialltoallv = hcoll_collectives.coll_ialltoallv ? mca_coll_hcoll_ialltoallv : NULL;
#else
hcoll_module->super.coll_ialltoallv = NULL;
#endif
*priority = cm->hcoll_priority;
module = &hcoll_module->super;
if (!cm->libhcoll_initialized) {
cm->libhcoll_initialized = true;
}
return module;
}
OBJ_CLASS_INSTANCE(mca_coll_hcoll_module_t,
mca_coll_base_module_t,
mca_coll_hcoll_module_construct,
mca_coll_hcoll_module_destruct);
