Add a configure option to enable GPU-aware communications.

Signed-off-by: George Bosilca <gbosilca@nvidia.com>

Author: bosilca

CMakeLists.txt

@@ -165,6 +165,9 @@ option(PARSEC_DIST_WITH_MPI
 if(PARSEC_DIST_WITH_MPI AND 0)
   message(FATAL_ERROR "PARSEC_DIST_WITH_MPI and PARSEC_DIST_WITH_OTHER are mutually exclusive, please select only one")
 endif()
+option(PARSEC_MPI_IS_GPU_AWARE
+       "Build PaRSEC assuming the MPI library is GPU-aware, aka. can move data directly to and from GPU memory.\
+       As of today (mid 2024) while most MPI support such an option, they require a single process per GPU" ON)
 option(PARSEC_DIST_THREAD
   "Use an extra thread to progress the data movements" ON)
 option(PARSEC_DIST_PRIORITIES

parsec/include/parsec/parsec_options.h.in

@@ -71,6 +71,7 @@
 
 /* Communication engine */
 #cmakedefine PARSEC_DIST_WITH_MPI
+#cmakedefine PARSEC_MPI_IS_GPU_AWARE
 #cmakedefine PARSEC_DIST_THREAD
 #cmakedefine PARSEC_DIST_PRIORITIES
 #cmakedefine PARSEC_DIST_COLLECTIVES

parsec/mca/device/device_gpu.c

@@ -953,9 +953,9 @@ parsec_device_data_reserve_space( parsec_device_gpu_module_t* gpu_device,
 
         /* Skip CTL flows only */
         if(PARSEC_FLOW_ACCESS_NONE == (PARSEC_FLOW_ACCESS_MASK & flow->flow_flags)) {
-	        gpu_task->flow_nb_elts[i] = 0;  /* assume there is nothing to transfer to the GPU */
+            gpu_task->flow_nb_elts[i] = 0;  /* assume there is nothing to transfer to the GPU */
             continue;
-	    }
+        }
 
         PARSEC_DEBUG_VERBOSE(20, parsec_gpu_output_stream,
                              "GPU[%d:%s]:%s: Investigating flow %s:%d",
@@ -970,7 +970,7 @@ parsec_device_data_reserve_space( parsec_device_gpu_module_t* gpu_device,
                                  gpu_device->super.device_index, gpu_device->super.name, task_name,
                                  flow->name, i, gpu_elem,
                                  this_task->data[i].data_in->data_transfer_status == PARSEC_DATA_STATUS_UNDER_TRANSFER ? " [in transfer]" : "");
-	        this_task->data[i].data_out = this_task->data[i].data_in;
+            this_task->data[i].data_out = this_task->data[i].data_in;
             continue;
         }
         master   = this_task->data[i].data_in->original;
@@ -2476,7 +2476,10 @@ parsec_device_kernel_epilog( parsec_device_gpu_module_t *gpu_device,
             gpu_copy->coherency_state = PARSEC_DATA_COHERENCY_SHARED;
             assert(PARSEC_DATA_STATUS_UNDER_TRANSFER == cpu_copy->data_transfer_status);
             cpu_copy->data_transfer_status = PARSEC_DATA_STATUS_COMPLETE_TRANSFER;
-
+            if( 0 == (parsec_mpi_allow_gpu_memory_communications & PARSEC_RUNTIME_SEND_FROM_GPU_MEMORY) ) {
+                /* Report the CPU copy as the output of the task. */
+                this_task->data[i].data_out = cpu_copy;
+            }
             PARSEC_DEBUG_VERBOSE(20, parsec_gpu_output_stream,
                                  "GPU copy %p [ref_count %d] moved to the read LRU in %s",
                                  gpu_copy, gpu_copy->super.super.obj_reference_count, __func__);

parsec/parsec_internal.h

@@ -200,6 +200,9 @@ PARSEC_DECLSPEC PARSEC_OBJ_CLASS_DECLARATION(parsec_taskpool_t);
 #define PARSEC_DEPENDENCIES_STARTUP_TASK   ((parsec_dependency_t)(1<<29))
 #define PARSEC_DEPENDENCIES_BITMASK        (~(PARSEC_DEPENDENCIES_TASK_DONE|PARSEC_DEPENDENCIES_IN_DONE|PARSEC_DEPENDENCIES_STARTUP_TASK))
 
+#define PARSEC_RUNTIME_SEND_FROM_GPU_MEMORY 0x00000002
+#define PARSEC_RUNTIME_RECV_FROM_GPU_MEMORY 0x00000001
+
 /**
  * This structure is used internally by the parsec_dependencies_t structures
  */
@@ -491,6 +494,12 @@ PARSEC_DECLSPEC extern int parsec_slow_bind_warning;
  * the scheduler, but can provide a better cache reuse.
  */
 PARSEC_DECLSPEC extern int parsec_runtime_keep_highest_priority_task;
+/**
+ * Global configuration mask allowing or not for the data to be sent or received,
+ * from or to, GPU memory. It can be an OR between PARSEC_RUNTIME_SEND_FROM_GPU_MEMORY
+ * and PARSEC_RUNTIME_RECV_FROM_GPU_MEMORY.
+ */
+PARSEC_DECLSPEC extern int parsec_mpi_allow_gpu_memory_communications;
 
 /**
  * Description of the state of the task. It indicates what will be the next

parsec/parsec_mpi_funnelled.c

@@ -201,6 +201,8 @@ parsec_list_t mpi_funnelled_dynamic_sendreq_fifo; /* ordered non threaded fifo *
 parsec_list_t mpi_funnelled_dynamic_recvreq_fifo; /* ordered non threaded fifo */
 parsec_mempool_t *mpi_funnelled_dynamic_req_mempool = NULL;
 
+int parsec_mpi_allow_gpu_memory_communications = 3;
+
 /* This structure is used to save all the information necessary to
  * invoke a callback after a MPI_Request is satisfied
  */
@@ -506,6 +508,14 @@ static int mpi_funneled_init_once(parsec_context_t* context)
                              MAX_MPI_TAG, (unsigned int)MAX_MPI_TAG, MAX_MPI_TAG / MAX_DEP_OUT_COUNT);
     }
 
+#if !defined(PARSEC_MPI_IS_GPU_AWARE)
+    parsec_mpi_allow_gpu_memory_communications = 0;
+#endif
+    parsec_mca_param_reg_int_name("mpi", "gpu_aware",
+                                  "Enabled if PaRSEC should allow MPI to move data directly from or to GPU memory. Otherwise, all data"
+                                  " movements will transit through CPU memory, and will always have a backup copy there. Accepted values "
+                                  "are ORed between 1 for receiving into GPU memory and 2 for sending from GPU memory",
+                                  false, false, parsec_mpi_allow_gpu_memory_communications, &parsec_mpi_allow_gpu_memory_communications);
     (void)context;
     return 0;
 }

parsec/remote_dep_mpi.c

@@ -2110,8 +2110,8 @@ static void remote_dep_mpi_get_start(parsec_execution_stream_t* es,
         /* prepare the local receiving data */
         assert(NULL == deps->output[k].data.data); /* we do not support in-place tiles now, make sure it doesn't happen yet */
         if(NULL == deps->output[k].data.data) {
-            deps->output[k].data.data = remote_dep_copy_allocate(&deps->output[k].data.remote,
-                                                                 deps->output[k].data.preferred_device);
+            int best_device = (parsec_mpi_allow_gpu_memory_communications & PARSEC_RUNTIME_RECV_FROM_GPU_MEMORY) ? deps->output[k].data.preferred_device : 0;
+            deps->output[k].data.data = remote_dep_copy_allocate(&deps->output[k].data.remote, best_device);
         }
         /* Mark the data under tranfer */
         deps->output[k].data.data->data_transfer_status = PARSEC_DATA_STATUS_UNDER_TRANSFER;