UCT/CUDA_IPC: Enforce host memory support for mem_ttpe EP

Author: shasson5

src/ucp/core/ucp_ep.c

@@ -684,12 +684,18 @@ ucs_status_t ucp_worker_mem_type_eps_create(ucp_worker_h worker)
     ucs_status_t status;
     void *address_buffer;
     size_t address_length;
-    ucp_tl_bitmap_t mem_access_tls;
+    ucp_tl_bitmap_t mem_access_tls, host_mem_access_tls;
     char ep_name[UCP_WORKER_ADDRESS_NAME_MAX];
     unsigned addr_indices[UCP_MAX_LANES];
+    ucp_lane_index_t num_lanes;
 
     ucs_memory_type_for_each(mem_type) {
         ucp_context_memaccess_tl_bitmap(context, mem_type, 0, &mem_access_tls);
+        /* Mem type EP requires host memory support */
+        ucp_context_memaccess_tl_bitmap(context, UCS_MEMORY_TYPE_HOST, 0,
+                                        &host_mem_access_tls);
+        UCS_STATIC_BITMAP_AND_INPLACE(&mem_access_tls, host_mem_access_tls);
+
         if (UCP_MEM_IS_HOST(mem_type) ||
             UCS_STATIC_BITMAP_IS_ZERO(mem_access_tls)) {
             continue;
@@ -725,6 +731,9 @@ ucs_status_t ucp_worker_mem_type_eps_create(ucp_worker_h worker)
             goto err_free_address_list;
         }
 
+        /* Mem type EP cannot have more than one lane */
+        num_lanes = ucp_ep_num_lanes(worker->mem_type_ep[mem_type]);
+        ucs_assertv_always(num_lanes == 1, "num_lanes=%u", num_lanes);
         UCS_ASYNC_UNBLOCK(&worker->async);
 
         ucs_free(local_address.address_list);

src/uct/cuda/cuda_ipc/cuda_ipc_iface.c

@@ -74,10 +74,6 @@ static ucs_config_field_t uct_cuda_ipc_iface_config_table[] = {
      "Estimated CPU overhead for transferring GPU memory",
      ucs_offsetof(uct_cuda_ipc_iface_config_t, params.overhead), UCS_CONFIG_TYPE_TIME},
 
-    {"ENABLE_SAME_PROCESS", "n",
-     "Enable same process same device communication for cuda_ipc",
-     ucs_offsetof(uct_cuda_ipc_iface_config_t, params.enable_same_process), UCS_CONFIG_TYPE_BOOL},
-
     {NULL}
 };
 
@@ -146,8 +142,7 @@ uct_cuda_ipc_iface_is_reachable_v2(const uct_iface_h tl_iface,
     dev_addr     = (const uct_cuda_ipc_device_addr_t *)params->device_addr;
     same_uuid    = (ucs_get_system_id() == dev_addr->system_uuid);
 
-    if ((getpid() == *(pid_t*)params->iface_addr) && same_uuid &&
-        !iface->config.enable_same_process) {
+    if ((getpid() == *(pid_t*)params->iface_addr) && same_uuid) {
         uct_iface_fill_info_str_buf(params, "same process");
         return 0;
     }

src/uct/cuda/cuda_ipc/cuda_ipc_iface.h

@@ -28,7 +28,6 @@ typedef struct {
     double                  bandwidth;           /* estimated bandwidth */
     double                  latency;             /* estimated latency */
     double                  overhead;            /* estimated CPU overhead */
-    int                     enable_same_process; /* enable cuda_ipc for same pid same device */
 } uct_cuda_ipc_iface_config_params_t;