From f9b2bee3d24f9a6dc99f9c20b48f4646463680a0 Mon Sep 17 00:00:00 2001
From: Miroslav Stoyanov <stoyanovmk@ornl.gov>
Date: Wed, 23 Oct 2024 14:56:35 -0400
Subject: [PATCH] added call to rocfft_setup()

---
 include/heffte_backend_rocm.h   | 25 +++++++++++++++++++++++--
 src/heffte_backend_rocm.hip.cpp |  2 ++
 2 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/include/heffte_backend_rocm.h b/include/heffte_backend_rocm.h
index 0787e94..a4906a9 100644
--- a/include/heffte_backend_rocm.h
+++ b/include/heffte_backend_rocm.h
@@ -180,6 +180,23 @@ namespace rocm {
             return 4 * (length-1);
         }
     };
+
+    /*!
+     * \ingroup heffterocm
+     * \brief Indicates whether the rocfft_setup() method has been called
+     */
+    struct initialize {
+        //! \brief Static (global) variable indicating if rocfft_setup() has been called
+        static bool is_initialized;
+        //! \brief Make initialize
+        static void make() {
+            if (not is_initialized) {
+                rocfft_setup();
+                is_initialized = true;
+            }
+
+        }
+    };
 }
 
 namespace backend{
@@ -549,7 +566,9 @@ class rocfft_executor : public executor_base{
         total_size(box.count()),
         embed({0, 0}),
         worksize(compute_workspace_size())
-    {}
+    {
+        rocm::initialize::make();
+    }
     //! \brief Merges two FFTs into one.
     template<typename index>
     rocfft_executor(hipStream_t active_stream, box3d<index> const box, int dir1, int dir2) :
@@ -740,7 +759,9 @@ class rocfft_executor_r2c : public executor_base{
         rsize(box.count()),
         csize(box.r2c(dimension).count()),
         worksize(compute_workspace_size())
-    {}
+    {
+        rocm::initialize::make();
+    }
 
     //! \brief Forward transform, single precision.
     template<typename precision_type>
diff --git a/src/heffte_backend_rocm.hip.cpp b/src/heffte_backend_rocm.hip.cpp
index e2ff591..22bd2ef 100644
--- a/src/heffte_backend_rocm.hip.cpp
+++ b/src/heffte_backend_rocm.hip.cpp
@@ -34,6 +34,8 @@ void synchronize_default_stream(){
 
 namespace rocm {
 
+bool initialize::is_initialized = false;
+
 /*
  * Launch with one thread per entry.
  *