From b138a67d4201d9bc57f1737cfbeb263b1d1fb5db Mon Sep 17 00:00:00 2001
From: Eric Phipps <etphipp@sandia.gov>
Date: Fri, 14 Jun 2024 16:39:30 -0600
Subject: [PATCH 1/2] Fix Cuda warning in debug builds

---
 src/Genten_DistTensorContext.hpp | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/src/Genten_DistTensorContext.hpp b/src/Genten_DistTensorContext.hpp
index 30d01f64af..79e721bbe6 100644
--- a/src/Genten_DistTensorContext.hpp
+++ b/src/Genten_DistTensorContext.hpp
@@ -314,12 +314,15 @@ bool weightsAreSame(const KtensorT<ExecSpace>&) {
 #else
 template <typename ExecSpace>
 bool weightsAreSame(const KtensorT<ExecSpace> &u) {
-  const ttb_indx wspan = u.weights().values().span();
+  auto w = u.weights();
+  const ttb_indx wspan = w.values().span();
   if (!isValueSame(wspan)) {
     return false;
   }
+  auto w_h = create_mirror_view(w);
+  deep_copy(w_h, w);
   for (std::size_t i = 0; i < wspan; ++i) {
-    if (!isValueSame(u.weights(i))) {
+    if (!isValueSame(w_h[i])) {
       return false;
     }
   }

From 9523cdbbac099982e596b9c39a58a080bb390226 Mon Sep 17 00:00:00 2001
From: Eric Phipps <etphipp@sandia.gov>
Date: Fri, 14 Jun 2024 16:40:07 -0600
Subject: [PATCH 2/2] Fix MPI error in some CUDA-aware MPI implementations

---
 src/Genten_Pmap.hpp | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/src/Genten_Pmap.hpp b/src/Genten_Pmap.hpp
index 87eeb059b3..b47c49795d 100644
--- a/src/Genten_Pmap.hpp
+++ b/src/Genten_Pmap.hpp
@@ -164,7 +164,10 @@ class ProcessorMap {
     static_assert(std::is_arithmetic<scalar_type2>::value,
                   "subGridAllReduce requires something like a double, or int");
 
-    MPI_Reduce(send.data(), recv.data(), send.span(),
+    scalar_type2* recv_data = nullptr;
+    if (sub_grid_rank_[i] == root)
+      recv_data = recv.data();
+    MPI_Reduce(send.data(), recv_data, send.span(),
                DistContext::toMpiType<scalar_type1>(),
                convertOp(op), root, sub_maps_[i]);
   }