buffer don't use page locked memory for device to host transfers

using cudaMallocHost caused performance problems on Perlmutter.
LBL-EESA · Sep 9, 2023 · 3d84c83 · 3d84c83
1 parent c0b2167
commit 3d84c83
Showing 1 changed file with 4 additions and 2 deletions.
diff --git a/hamr_buffer_impl.h b/hamr_buffer_impl.h
@@ -1767,15 +1767,17 @@ std::shared_ptr<const T> buffer<T>::get_host_accessible() const
 #if defined(HAMR_ENABLE_CUDA)
     else if ((m_alloc == allocator::cuda) || (m_alloc == allocator::cuda_async))
     {
-        // make a copy on the host
+        // make a copy on the host.
+        std::shared_ptr<T> tmp = malloc_allocator<T>::allocate(m_size);
+        /*TODO:Using cudaMallocHost caused performance issues on Perlmutter
         std::shared_ptr<T> tmp = cuda_malloc_host_allocator<T>::allocate(m_size);
         if (!tmp)
         {
             std::cerr << "[" << __FILE__ << ":" << __LINE__ << "] ERROR:"
                 " CUDA failed to allocate host pinned memory, falling back"
                 " to the default system allocator." << std::endl;
             tmp = malloc_allocator<T>::allocate(m_size);
-        }
+        }*/
 
         activate_cuda_device dev(m_owner);