static void ucp_perf_test_destroy_eps(ucx_perf_context_t* perf,
unsigned group_size)
{
ucs_status_ptr_t *reqs;
ucp_tag_recv_info_t info;
ucs_status_t status;
unsigned i;
reqs = calloc(sizeof(*reqs), group_size);
for (i = 0; i < group_size; ++i) {
if (perf->ucp.peers[i].rkey != NULL) {
ucp_rkey_destroy(perf->ucp.peers[i].rkey);
}
if (perf->ucp.peers[i].ep != NULL) {
reqs[i] = ucp_disconnect_nb(perf->ucp.peers[i].ep);
}
}
for (i = 0; i < group_size; ++i) {
if (!UCS_PTR_IS_PTR(reqs[i])) {
continue;
}
do {
ucp_worker_progress(perf->ucp.worker);
status = ucp_request_test(reqs[i], &info);
} while (status == UCS_INPROGRESS);
ucp_request_release(reqs[i]);
}
free(reqs);
free(perf->ucp.peers);
}
int ompi_osc_ucx_compare_and_swap(const void *origin_addr, const void *compare_addr,
void *result_addr, struct ompi_datatype_t *dt,
int target, ptrdiff_t target_disp,
struct ompi_win_t *win) {
ompi_osc_ucx_module_t *module = (ompi_osc_ucx_module_t *)win->w_osc_module;
ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, target);
uint64_t remote_addr = (module->win_info_array[target]).addr + target_disp * OSC_UCX_GET_DISP(module, target);
ucp_rkey_h rkey;
size_t dt_bytes;
ompi_osc_ucx_internal_request_t *req = NULL;
int ret = OMPI_SUCCESS;
ucs_status_t status;
ret = check_sync_state(module, target, false);
if (ret != OMPI_SUCCESS) {
return ret;
}
ret = start_atomicity(module, ep, target);
if (ret != OMPI_SUCCESS) {
return ret;
}
if (module->flavor == MPI_WIN_FLAVOR_DYNAMIC) {
status = get_dynamic_win_info(remote_addr, module, ep, target);
if (status != UCS_OK) {
return OMPI_ERROR;
}
}
rkey = (module->win_info_array[target]).rkey;
ompi_datatype_type_size(dt, &dt_bytes);
memcpy(result_addr, origin_addr, dt_bytes);
req = ucp_atomic_fetch_nb(ep, UCP_ATOMIC_FETCH_OP_CSWAP, *(uint64_t *)compare_addr,
result_addr, dt_bytes, remote_addr, rkey, req_completion);
if (UCS_PTR_IS_PTR(req)) {
ucp_request_release(req);
}
ret = incr_and_check_ops_num(module, target, ep);
if (ret != OMPI_SUCCESS) {
return ret;
}
return end_atomicity(module, ep, target);
}
int ompi_osc_ucx_rget(void *origin_addr, int origin_count,
struct ompi_datatype_t *origin_dt,
int target, ptrdiff_t target_disp, int target_count,
struct ompi_datatype_t *target_dt, struct ompi_win_t *win,
struct ompi_request_t **request) {
ompi_osc_ucx_module_t *module = (ompi_osc_ucx_module_t*) win->w_osc_module;
ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, target);
uint64_t remote_addr = (module->state_info_array[target]).addr + OSC_UCX_STATE_REQ_FLAG_OFFSET;
ucp_rkey_h rkey;
ompi_osc_ucx_request_t *ucx_req = NULL;
ompi_osc_ucx_internal_request_t *internal_req = NULL;
ucs_status_t status;
int ret = OMPI_SUCCESS;
ret = check_sync_state(module, target, true);
if (ret != OMPI_SUCCESS) {
return ret;
}
if (module->flavor == MPI_WIN_FLAVOR_DYNAMIC) {
status = get_dynamic_win_info(remote_addr, module, ep, target);
if (status != UCS_OK) {
return OMPI_ERROR;
}
}
rkey = (module->win_info_array[target]).rkey;
OMPI_OSC_UCX_REQUEST_ALLOC(win, ucx_req);
if (NULL == ucx_req) {
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
ret = ompi_osc_ucx_get(origin_addr, origin_count, origin_dt, target, target_disp,
target_count, target_dt, win);
if (ret != OMPI_SUCCESS) {
return ret;
}
status = ucp_worker_fence(mca_osc_ucx_component.ucp_worker);
if (status != UCS_OK) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_worker_fence failed: %d\n",
__FILE__, __LINE__, status);
return OMPI_ERROR;
}
internal_req = ucp_atomic_fetch_nb(ep, UCP_ATOMIC_FETCH_OP_FADD, 0,
&(module->req_result), sizeof(uint64_t),
remote_addr, rkey, req_completion);
if (UCS_PTR_IS_PTR(internal_req)) {
internal_req->external_req = ucx_req;
mca_osc_ucx_component.num_incomplete_req_ops++;
} else {
ompi_request_complete(&ucx_req->super, true);
}
*request = &ucx_req->super;
return incr_and_check_ops_num(module, target, ep);
}
int ompi_osc_ucx_fetch_and_op(const void *origin_addr, void *result_addr,
struct ompi_datatype_t *dt, int target,
ptrdiff_t target_disp, struct ompi_op_t *op,
struct ompi_win_t *win) {
ompi_osc_ucx_module_t *module = (ompi_osc_ucx_module_t*) win->w_osc_module;
int ret = OMPI_SUCCESS;
ret = check_sync_state(module, target, false);
if (ret != OMPI_SUCCESS) {
return ret;
}
if (op == &ompi_mpi_op_no_op.op || op == &ompi_mpi_op_replace.op ||
op == &ompi_mpi_op_sum.op) {
ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, target);
uint64_t remote_addr = (module->win_info_array[target]).addr + target_disp * OSC_UCX_GET_DISP(module, target);
ucp_rkey_h rkey;
uint64_t value = *(uint64_t *)origin_addr;
ucp_atomic_fetch_op_t opcode;
size_t dt_bytes;
ompi_osc_ucx_internal_request_t *req = NULL;
ucs_status_t status;
ret = start_atomicity(module, ep, target);
if (ret != OMPI_SUCCESS) {
return ret;
}
if (module->flavor == MPI_WIN_FLAVOR_DYNAMIC) {
status = get_dynamic_win_info(remote_addr, module, ep, target);
if (status != UCS_OK) {
return OMPI_ERROR;
}
}
rkey = (module->win_info_array[target]).rkey;
ompi_datatype_type_size(dt, &dt_bytes);
if (op == &ompi_mpi_op_replace.op) {
opcode = UCP_ATOMIC_FETCH_OP_SWAP;
} else {
opcode = UCP_ATOMIC_FETCH_OP_FADD;
if (op == &ompi_mpi_op_no_op.op) {
value = 0;
}
}
req = ucp_atomic_fetch_nb(ep, opcode, value, result_addr,
dt_bytes, remote_addr, rkey, req_completion);
if (UCS_PTR_IS_PTR(req)) {
ucp_request_release(req);
}
ret = incr_and_check_ops_num(module, target, ep);
if (ret != OMPI_SUCCESS) {
return ret;
}
return end_atomicity(module, ep, target);
} else {
return ompi_osc_ucx_get_accumulate(origin_addr, 1, dt, result_addr, 1, dt,
target, target_disp, 1, dt, op, win);
}
}
void print_ucp_info(int print_opts, ucs_config_print_flags_t print_flags,
uint64_t ctx_features, const ucp_ep_params_t *base_ep_params,
size_t estimated_num_eps, unsigned dev_type_bitmap,
const char *mem_size)
{
ucp_config_t *config;
ucs_status_t status;
ucs_status_ptr_t status_ptr;
ucp_context_h context;
ucp_worker_h worker;
ucp_params_t params;
ucp_worker_params_t worker_params;
ucp_ep_params_t ep_params;
ucp_address_t *address;
size_t address_length;
resource_usage_t usage;
ucp_ep_h ep;
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
return;
}
memset(¶ms, 0, sizeof(params));
params.field_mask = UCP_PARAM_FIELD_FEATURES |
UCP_PARAM_FIELD_ESTIMATED_NUM_EPS;
params.features = ctx_features;
params.estimated_num_eps = estimated_num_eps;
get_resource_usage(&usage);
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SELF))) {
ucp_config_modify(config, "SELF_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SHM))) {
ucp_config_modify(config, "SHM_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_NET))) {
ucp_config_modify(config, "NET_DEVICES", "");
}
status = ucp_init(¶ms, config, &context);
if (status != UCS_OK) {
printf("<Failed to create UCP context>\n");
goto out_release_config;
}
if ((print_opts & PRINT_MEM_MAP) && (mem_size != NULL)) {
ucp_mem_print_info(mem_size, context, stdout);
}
if (print_opts & PRINT_UCP_CONTEXT) {
ucp_context_print_info(context, stdout);
print_resource_usage(&usage, "UCP context");
}
if (!(print_opts & (PRINT_UCP_WORKER|PRINT_UCP_EP))) {
goto out_cleanup_context;
}
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = UCS_THREAD_MODE_MULTI;
get_resource_usage(&usage);
status = ucp_worker_create(context, &worker_params, &worker);
if (status != UCS_OK) {
printf("<Failed to create UCP worker>\n");
goto out_cleanup_context;
}
if (print_opts & PRINT_UCP_WORKER) {
ucp_worker_print_info(worker, stdout);
print_resource_usage(&usage, "UCP worker");
}
if (print_opts & PRINT_UCP_EP) {
status = ucp_worker_get_address(worker, &address, &address_length);
if (status != UCS_OK) {
printf("<Failed to get UCP worker address>\n");
goto out_destroy_worker;
}
ep_params = *base_ep_params;
ep_params.field_mask |= UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = address;
status = ucp_ep_create(worker, &ep_params, &ep);
ucp_worker_release_address(worker, address);
if (status != UCS_OK) {
printf("<Failed to create UCP endpoint>\n");
goto out_destroy_worker;
}
ucp_ep_print_info(ep, stdout);
status_ptr = ucp_disconnect_nb(ep);
if (UCS_PTR_IS_PTR(status_ptr)) {
do {
ucp_worker_progress(worker);
status = ucp_request_test(status_ptr, NULL);
} while (status == UCS_INPROGRESS);
ucp_request_release(status_ptr);
}
}
out_destroy_worker:
ucp_worker_destroy(worker);
out_cleanup_context:
ucp_cleanup(context);
out_release_config:
ucp_config_release(config);
}