From b3c2e80ee68523da768ac880a365655ddc4e27e7 Mon Sep 17 00:00:00 2001
From: Yael Yacobovich <yyacobovich@nvidia.com>
Date: Wed, 30 Oct 2024 16:37:11 +0200
Subject: [PATCH] change bounds in ring

---
 src/components/ec/cuda/ec_cuda_executor.c     |  1 +
 .../tl/ucp/allgather/allgather_ring.c         | 29 +++++++++++--------
 2 files changed, 18 insertions(+), 12 deletions(-)

diff --git a/src/components/ec/cuda/ec_cuda_executor.c b/src/components/ec/cuda/ec_cuda_executor.c
index 50a8f6f24c..5d51cf924c 100644
--- a/src/components/ec/cuda/ec_cuda_executor.c
+++ b/src/components/ec/cuda/ec_cuda_executor.c
@@ -110,6 +110,7 @@ ucc_status_t ucc_cuda_executor_task_finalize(ucc_ee_executor_task_t *task)
 ucc_status_t ucc_cuda_executor_start(ucc_ee_executor_t *executor,
                                      void *ee_context)
 {
+    printf("======== in executer start =========");
     ucc_ec_cuda_executor_t *eee = ucc_derived_of(executor,
                                                  ucc_ec_cuda_executor_t);
 
diff --git a/src/components/tl/ucp/allgather/allgather_ring.c b/src/components/tl/ucp/allgather/allgather_ring.c
index 9a95d374ee..ad68a73828 100644
--- a/src/components/tl/ucp/allgather/allgather_ring.c
+++ b/src/components/tl/ucp/allgather/allgather_ring.c
@@ -43,29 +43,31 @@ void ucc_tl_ucp_allgather_ring_progress(ucc_coll_task_t *coll_task)
     ucc_rank_t         sendto, recvfrom, sblock, rblock;
     int                step;
     void              *buf;
-
+    
 
     if (UCC_INPROGRESS == ucc_tl_ucp_test(task)) {
         return;
     }
-    printf("after prog\n");
-    
-    /*int k = 1;
-    int a = 1;
-    while(k){
-        a = 1;
-    }*/
+    printf("in progress top, rank = %d task->tagged.send_posted = %d task->tagged.recv_posted = %d \n",(int)trank, (int)task->tagged.send_posted,
+                (int)task->tagged.recv_posted);
+
     sendto   = ucc_ep_map_eval(task->subset.map, (trank + 1) % tsize);
     recvfrom = ucc_ep_map_eval(task->subset.map, (trank - 1 + tsize) % tsize);
-    printf("in prog,tsize = %d, trank = %d, sndto = %d, recv = %d\n", (int)tsize, (int)trank, (int)sendto, (int)recvfrom);
+    
+    
+    uint32_t USE_CUDA = UCC_TL_UCP_TEAM_LIB(team)->cfg.allgather_use_cuda;
+    int iter = USE_CUDA ? tsize - 1 : tsize;
+
+    while (task->tagged.send_posted < iter) {
+        
+        step = USE_CUDA ? task->tagged.send_posted : task->tagged.send_posted - 1;
 
-    while (task->tagged.send_posted < tsize - 1) {
-        step = task->tagged.send_posted;
         sblock = task->allgather_ring.get_send_block(&task->subset, trank,
                                                      tsize, step);
         rblock = task->allgather_ring.get_recv_block(&task->subset, trank,
                                                      tsize, step);
         buf = PTR_OFFSET(rbuf, sblock * data_size);
+
         UCPCHECK_GOTO(
             ucc_tl_ucp_send_nb(buf, data_size, rmem, sendto, team, task),
             task, out);
@@ -73,8 +75,8 @@ void ucc_tl_ucp_allgather_ring_progress(ucc_coll_task_t *coll_task)
         UCPCHECK_GOTO(
             ucc_tl_ucp_recv_nb(buf, data_size, rmem, recvfrom, team, task),
             task, out);
+        
         if (UCC_INPROGRESS == ucc_tl_ucp_test(task)) {
-            printf("in while loop\ntask->tagged.send_posted = %d\n", (int)task->tagged.send_posted);
             return;
         }
     }
@@ -110,13 +112,16 @@ ucc_status_t ucc_tl_ucp_allgather_ring_start(ucc_coll_task_t *coll_task)
         if(USE_CUDA){
             status = ucc_mc_memcpy(PTR_OFFSET(rbuf, data_size * block),
                                sbuf, data_size, rmem, smem);
+            //printf("in use cuda: task->tagged.send_posted = %d \n", (int)task->tagged.send_posted);
             if (ucc_unlikely(UCC_OK != status)) {
                 return status;
             }
         } else {
             /* Loopback */
             UCPCHECK_GOTO(ucc_tl_ucp_send_nb(sbuf, data_size, smem, trank, team, task),task, out);
+            //printf("start: trank = %d send task->tagged.send_posted = %d\n",(int)trank, (int)task->tagged.send_posted);
             UCPCHECK_GOTO(ucc_tl_ucp_recv_nb(PTR_OFFSET(rbuf, data_size * block), data_size, rmem, trank, team, task),task, out);
+            //printf("start: trank = %d recv task->tagged.recv_posted = %d\n",(int)trank, (int)task->tagged.recv_posted);
         }
     }
     return ucc_progress_queue_enqueue(UCC_TL_CORE_CTX(team)->pq, &task->super);