From 66122818b7dbed561cdef6c23b4ab9a49d4f0985 Mon Sep 17 00:00:00 2001
From: Joongi Kim <joongi@an.kaist.ac.kr>
Date: Thu, 14 Jan 2016 00:54:53 +0900
Subject: [PATCH] refs #6: Revive no-opped kernel execution and more comments.

 * Now the performance gets affected by computation.
---
 src/lib/elementgraph.cc |  3 ++-
 src/lib/io.cc           |  3 ++-
 src/lib/offloadtask.cc  | 12 ++++++------
 3 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/src/lib/elementgraph.cc b/src/lib/elementgraph.cc
index f97f726..efb938f 100644
--- a/src/lib/elementgraph.cc
+++ b/src/lib/elementgraph.cc
@@ -110,12 +110,13 @@ void ElementGraph::send_offload_task_to_device(OffloadTask *task)
         task->prepare_read_buffer();
         task->prepare_write_buffer();
         task->state = TASK_PREPARED;
-    } /* endif(!task.prepared) */
+    }
 
     /* Send the offload task to device thread. */
     assert(task->state == TASK_PREPARED);
     int ret = rte_ring_enqueue(ctx->offload_input_queues[dev_idx], (void*) task);
     if (ret == -ENOBUFS) {
+        /* The input queue is full.  Delay the task. */
         enqueue_offload_task(task, task->tracker.element, task->tracker.input_port);
     } else {
         /* It may return -EDQUOT, but here we ignore this HWM signal.
diff --git a/src/lib/io.cc b/src/lib/io.cc
index 54d50f0..53f5fa9 100644
--- a/src/lib/io.cc
+++ b/src/lib/io.cc
@@ -154,7 +154,8 @@ static void comp_offload_task_completion_cb(struct ev_loop *loop, struct ev_asyn
                         ((float) task_cycles / total_batch_size
                          - ((float) task->batches[b]->delay_time / task->batches[b]->count));
             }
-            /* We need to rewind the state so that it gets executed by ElemGraph. */
+            /* Rewind the state so that it gets "prepared" by ElemGraph.
+             * (e.g., update datablock list used by next element) */
             task->state = TASK_INITIALIZED;
             ctx->elem_graph->enqueue_offload_task(task,
                                                   ctx->elem_graph->get_first_next(task->elem),
diff --git a/src/lib/offloadtask.cc b/src/lib/offloadtask.cc
index 3851a0a..4ddaf04 100644
--- a/src/lib/offloadtask.cc
+++ b/src/lib/offloadtask.cc
@@ -327,13 +327,13 @@ void OffloadTask::execute()
         arg = {(void *) &checkbits_d, sizeof(void *), alignof(void *)};
         cctx->push_kernel_arg(arg);
 
-        //offload_compute_handler &handler = elem->offload_compute_handlers[cctx->type_name];
-        //handler(cctx, &res);
+        offload_compute_handler &handler = elem->offload_compute_handlers[cctx->type_name];
+        handler(cctx, &res);
         /* Skip kernel execution. */
-        res.num_workitems = 0;
-        res.num_threads_per_workgroup = 1;
-        res.num_workgroups = 1;
-        cctx->get_host_checkbits()[0] = 1;
+        //res.num_workitems = 0;
+        //res.num_threads_per_workgroup = 1;
+        //res.num_workgroups = 1;
+        //cctx->get_host_checkbits()[0] = 1;
 
     } else {