From 24785df2747ee9f31f8b602885513b479fcca5ad Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=B0=88=E5=AE=87=E9=9A=86?= <tanyulong2@huawei.com>
Date: Wed, 10 Sep 2025 15:06:07 +0800
Subject: [PATCH 1/2] solve 1 process ineigbor_alltoall(v|w) result bug

---
 ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c | 34 ++++++++++++++++---
 .../mca/coll/libnbc/nbc_ineighbor_alltoallv.c | 34 ++++++++++++++++---
 .../mca/coll/libnbc/nbc_ineighbor_alltoallw.c | 34 ++++++++++++++++---
 3 files changed, 90 insertions(+), 12 deletions(-)

diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
index d57fcdf3470..8dea1efad1b 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
@@ -89,14 +89,40 @@ static int nbc_neighbor_alltoall_init(const void *sbuf, int scount, MPI_Datatype
 
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    for (int i = indegree - 1 ; i >= 0 ; --i) {
-      if (MPI_PROC_NULL != srcs[i]) {
-        res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * i * rcount, true, rcount, rtype, srcs[i], schedule, false);
-        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+    int flag = 0;
+    if (OMPI_COMM_IS_CART(comm)) {
+      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+        if (comm->c_topo->mtc.cart->dims[dim] == 1) {
+          flag = 1;
           break;
         }
       }
     }
+    if (flag != 0) {
+      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
+        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
+          res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * (2 * dim + 1) * rcount, true, rcount, rtype, srcs[2 * dim + 1], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+        if (MPI_PROC_NULL != srcs[2 * dim]) {
+          res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * (2 * dim) * rcount, true, rcount, rtype, srcs[2 * dim], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    } else {
+      for (int i = indegree - 1 ; i >= 0 ; --i) {
+        if (MPI_PROC_NULL != srcs[i]) {
+          res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * i * rcount, true, rcount, rtype, srcs[i], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    }
 
     free (srcs);
 
diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
index e871862c2bd..b886fc2cf2c 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
@@ -93,14 +93,40 @@ static int nbc_neighbor_alltoallv_init(const void *sbuf, const int *scounts, con
     /* simply loop over neighbors and post send/recv operations */
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    for (int i = indegree - 1 ; i >= 0 ; --i) {
-      if (srcs[i] != MPI_PROC_NULL) {
-        res = NBC_Sched_recv ((char *) rbuf + rdispls[i] * rcvext, false, rcounts[i], rtype, srcs[i], schedule, false);
-        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+    int flag = 0;
+    if (OMPI_COMM_IS_CART(comm)) {
+      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+        if (comm->c_topo->mtc.cart->dims[dim] == 1) {
+          flag = 1;
           break;
         }
       }
     }
+    if (flag != 0) {
+      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
+        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
+          res = NBC_Sched_recv ((char *) rbuf + rdispls[2 * dim + 1] * rcvext, false, rcounts[2 * dim + 1], rtype, srcs[2 * dim + 1], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+        if (MPI_PROC_NULL != srcs[2 * dim]) {
+          res = NBC_Sched_recv ((char *) rbuf + rdispls[2 * dim] * rcvext, false, rcounts[2 * dim], rtype, srcs[2 * dim], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    } else {
+      for (int i = indegree - 1 ; i >= 0 ; --i) {
+        if (srcs[i] != MPI_PROC_NULL) {
+          res = NBC_Sched_recv ((char *) rbuf + rdispls[i] * rcvext, false, rcounts[i], rtype, srcs[i], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    }
 
     free (srcs);
 
diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
index 8b7ee0f050c..3bae06ce5c0 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
@@ -79,14 +79,40 @@ static int nbc_neighbor_alltoallw_init(const void *sbuf, const int *scounts, con
     /* simply loop over neighbors and post send/recv operations */
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    for (int i = indegree - 1 ; i >= 0 ; --i) {
-      if (srcs[i] != MPI_PROC_NULL) {
-        res = NBC_Sched_recv ((char *) rbuf + rdisps[i], false, rcounts[i], rtypes[i], srcs[i], schedule, false);
-        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+    int flag = 0;
+    if (OMPI_COMM_IS_CART(comm)) {
+      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+        if (comm->c_topo->mtc.cart->dims[dim] == 1) {
+          flag = 1;
           break;
         }
       }
     }
+    if (flag != 0) {
+      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
+        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
+          res = NBC_Sched_recv ((char *) rbuf + rdisps[2 * dim + 1], false, rcounts[2 * dim + 1], rtypes[2 * dim + 1], srcs[2 * dim + 1], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+        if (MPI_PROC_NULL != srcs[2 * dim]) {
+          res = NBC_Sched_recv ((char *) rbuf + rdisps[2 * dim], false, rcounts[2 * dim], rtypes[2 * dim], srcs[2 * dim], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    } else {
+      for (int i = indegree - 1 ; i >= 0 ; --i) {
+        if (srcs[i] != MPI_PROC_NULL) {
+          res = NBC_Sched_recv ((char *) rbuf + rdisps[i], false, rcounts[i], rtypes[i], srcs[i], schedule, false);
+          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            break;
+          }
+        }
+      }
+    }
 
     free (srcs);
 
-- 
Gitee


From f469bf9486e3e9d662ff93ef87e7182be521f209 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=B0=88=E5=AE=87=E9=9A=86?= <tanyulong2@huawei.com>
Date: Wed, 10 Sep 2025 16:51:51 +0800
Subject: [PATCH 2/2] cleancode

---
 ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c | 29 ++++++++++---------
 .../mca/coll/libnbc/nbc_ineighbor_alltoallv.c | 29 ++++++++++---------
 .../mca/coll/libnbc/nbc_ineighbor_alltoallw.c | 29 ++++++++++---------
 3 files changed, 45 insertions(+), 42 deletions(-)

diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
index 8dea1efad1b..5323dea9f11 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoall.c
@@ -19,6 +19,7 @@
  *
  */
 #include "nbc_internal.h"
+#include "ompi/mca/topo/base/base.h"
 
 /* cannot cache schedules because one cannot check locally if the pattern is the same!! */
 #undef NBC_CACHE_SCHEDULE
@@ -89,28 +90,28 @@ static int nbc_neighbor_alltoall_init(const void *sbuf, int scount, MPI_Datatype
 
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    int flag = 0;
+    /* comm is cart and process of a certain dim is 1 need to special handle */
+    bool is_cart_dim_one = false;
     if (OMPI_COMM_IS_CART(comm)) {
-      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+      for(int dim = 0 ; dim <comm->c_topo->mtc.cart->ndims ; ++dim) {
         if (comm->c_topo->mtc.cart->dims[dim] == 1) {
-          flag = 1;
+          is_cart_dim_one = true;
           break;
         }
       }
     }
-    if (flag != 0) {
-      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
-        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
-          res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * (2 * dim + 1) * rcount, true, rcount, rtype, srcs[2 * dim + 1], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
+    if (is_cart_dim_one) {
+      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims ; ++dim) {
+        for (int i = 1 ; i >= 0 ; --i) {
+          if (MPI_PROC_NULL != srcs[2 * dim + i]) {
+            res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * (2 * dim + i) * rcount, true, rcount, rtype, srcs[2 * dim + i], schedule, false);
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+              break;
+            }
           }
         }
-        if (MPI_PROC_NULL != srcs[2 * dim]) {
-          res = NBC_Sched_recv ((char *) rbuf + (MPI_Aint) rcvext * (2 * dim) * rcount, true, rcount, rtype, srcs[2 * dim], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
-          }
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+          break;
         }
       }
     } else {
diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
index b886fc2cf2c..93c772ab8f0 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallv.c
@@ -19,6 +19,7 @@
  *
  */
 #include "nbc_internal.h"
+#include "ompi/mca/topo/base/base.h"
 
 /* cannot cache schedules because one cannot check locally if the pattern is the same!! */
 #undef NBC_CACHE_SCHEDULE
@@ -93,28 +94,28 @@ static int nbc_neighbor_alltoallv_init(const void *sbuf, const int *scounts, con
     /* simply loop over neighbors and post send/recv operations */
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    int flag = 0;
+    /* comm is cart and process of a certain dim is 1 need to special handle */
+    bool is_cart_dim_one = false;
     if (OMPI_COMM_IS_CART(comm)) {
-      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+      for(int dim = 0 ; dim <comm->c_topo->mtc.cart->ndims ; ++dim) {
         if (comm->c_topo->mtc.cart->dims[dim] == 1) {
-          flag = 1;
+          is_cart_dim_one = true;
           break;
         }
       }
     }
-    if (flag != 0) {
-      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
-        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
-          res = NBC_Sched_recv ((char *) rbuf + rdispls[2 * dim + 1] * rcvext, false, rcounts[2 * dim + 1], rtype, srcs[2 * dim + 1], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
+    if (is_cart_dim_one) {
+      for (int dim = 0 ; dim < comm->c_topo->mtc.cart->ndims ; ++dim) {
+        for (int i = 1 ; i >= 0 ; --i) {
+          if (MPI_PROC_NULL != srcs[2 * dim + i]) {
+            res = NBC_Sched_recv ((char *) rbuf + rdispls[2 * dim + i] * rcvext, false, rcounts[2 * dim + i], rtype, srcs[2 * dim + i], schedule, false);
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+              break;
+            }
           }
         }
-        if (MPI_PROC_NULL != srcs[2 * dim]) {
-          res = NBC_Sched_recv ((char *) rbuf + rdispls[2 * dim] * rcvext, false, rcounts[2 * dim], rtype, srcs[2 * dim], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
-          }
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+          break;
         }
       }
     } else {
diff --git a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
index 3bae06ce5c0..f012d58ccec 100644
--- a/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
+++ b/ompi/mca/coll/libnbc/nbc_ineighbor_alltoallw.c
@@ -19,6 +19,7 @@
  *
  */
 #include "nbc_internal.h"
+#include "ompi/mca/topo/base/base.h"
 
 /* cannot cache schedules because one cannot check locally if the pattern is the same!! */
 #undef NBC_CACHE_SCHEDULE
@@ -79,28 +80,28 @@ static int nbc_neighbor_alltoallw_init(const void *sbuf, const int *scounts, con
     /* simply loop over neighbors and post send/recv operations */
     /* change recv order to solve the problem of opposite results in loop neigbor under 2 processes */
     /* issue can see https://github.com/mpi-forum/mpi-issues/issues/153 */
-    int flag = 0;
+    /* comm is cart and process of a certain dim is 1 need to special handle */
+    bool is_cart_dim_one = false;
     if (OMPI_COMM_IS_CART(comm)) {
-      for(int dim = 0; dim <comm->c_topo->mtc.cart->ndims; dim++) {
+      for(int dim = 0 ; dim <comm->c_topo->mtc.cart->ndims ; ++dim) {
         if (comm->c_topo->mtc.cart->dims[dim] == 1) {
-          flag = 1;
+          is_cart_dim_one = true;
           break;
         }
       }
     }
-    if (flag != 0) {
-      for (int dim = 0; dim < comm->c_topo->mtc.cart->ndims; dim++) {
-        if (MPI_PROC_NULL != srcs[2 * dim + 1]) {
-          res = NBC_Sched_recv ((char *) rbuf + rdisps[2 * dim + 1], false, rcounts[2 * dim + 1], rtypes[2 * dim + 1], srcs[2 * dim + 1], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
+    if (is_cart_dim_one) {
+      for (int dim = 0 ; dim < comm->c_topo->mtc.cart->ndims ; ++dim) {
+        for (int i = 1 ; i >= 0 ; --i) {
+          if (MPI_PROC_NULL != srcs[2 * dim + i]) {
+            res = NBC_Sched_recv ((char *) rbuf + rdisps[2 * dim + i], false, rcounts[2 * dim + i], rtypes[2 * dim + i], srcs[2 * dim + i], schedule, false);
+            if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+              break;
+            }
           }
         }
-        if (MPI_PROC_NULL != srcs[2 * dim]) {
-          res = NBC_Sched_recv ((char *) rbuf + rdisps[2 * dim], false, rcounts[2 * dim], rtypes[2 * dim], srcs[2 * dim], schedule, false);
-          if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-            break;
-          }
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+          break;
         }
       }
     } else {
-- 
Gitee