Merge branch 'github_devel_staging' into 'github_devel_branch_tracking'

[LibFabric] Include stdio.h with libfabric and edit unit tests

See merge request nvshmem/nvshmem!1520

Author: benhg

CMakeLists.txt

@@ -48,7 +48,7 @@ endif()
 project(
   NVSHMEM
   LANGUAGES CUDA CXX C
-  VERSION 3.5.20.0
+  VERSION 3.5.19.1
 )
 
 if(NOT (CMAKE_C_COMPILER_ID STREQUAL "GNU" OR CMAKE_C_COMPILER_ID STREQUAL "Clang") OR

src/host/mem/mem_transport.cpp

@@ -257,7 +257,7 @@ nvshmemi_mem_p2p_transport::nvshmemi_mem_p2p_transport(int mype, int npes) {
        This is to avoid a case where some PEs on some node are P2P reachable and some PEs on some
        node are not, causing asymmetry
     */
-    peer_error_status = (bool *)std::calloc(sizeof(bool), npes);
+    peer_error_status = (bool *)std::calloc(npes, sizeof(*peer_error_status));
     nvshmemi_boot_handle.allgather((void *)(&errored_on_initialization_), peer_error_status,
                                    sizeof(bool), &nvshmemi_boot_handle);
     NVSHMEMU_FOR_EACH(i, npes) {

src/modules/transport/libfabric/libfabric.h

@@ -6,6 +6,7 @@
 
 #include <assert.h>
 #include <stdint.h>  // IWYU pragma: keep
+#include <stdio.h>
 #include <stddef.h>
 #include <string.h>
 #include <deque>

test/host/interop/CMakeLists.txt

@@ -39,6 +39,17 @@ target_link_libraries(simplelib2 PRIVATE nvshmem_test_helper)
 target_link_libraries(app PRIVATE simplelib1 simplelib2 nvshmem_test_helper)
 target_link_libraries(app_multi_init PRIVATE simplelib1 simplelib2 nvshmem_test_helper)
 
+#
+# Some toolchains/distros enable PIE by default for executables. NVCC-produced objects for these
+# CUDA test apps may not be compiled with -fPIE, which then fails to link with:
+#   relocation R_X86_64_32 ... can not be used when making a PIE object
+# Disable PIE for these interop test executables to keep the build portable.
+#
+if (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+    target_link_options(app PRIVATE -no-pie)
+    target_link_options(app_multi_init PRIVATE -no-pie)
+endif()
+
 install(TARGETS simplelib1 LIBRARY DESTINATION "${NVSHMEM_TEST_INSTALL_PREFIX}/host/interop")
 install(TARGETS simplelib2 LIBRARY DESTINATION "${NVSHMEM_TEST_INSTALL_PREFIX}/host/interop")
 install(TARGETS app RUNTIME DESTINATION "${NVSHMEM_TEST_INSTALL_PREFIX}/host/interop")

test/unit/mem/heap/internal/host/nvshmemi_mem_transport.hpp

@@ -59,21 +59,27 @@ class nvshmemi_mem_p2p_transport {
 
     void set_heap_size_(nvshmemi_symmetric_heap &heap, size_t size) { heap.heap_size_ = size; }
     void set_heap_base_(nvshmemi_symmetric_heap &heap) {
-        int heap_base = 1;
-        void *heap_base_ptr = &heap_base;
-        heap.heap_base_ = heap_base_ptr;
+        auto &storage = heap_base_storage_[&heap];
+        if (!storage) storage = std::make_unique<int>(1);
+        heap.heap_base_ = storage.get();
     }
 
     void set_peer_heap_base_p2p_(nvshmemi_symmetric_heap &heap) {
-        int value = 1;
-        void *ptr = &value;
-        heap.peer_heap_base_p2p_ = (void **)std::calloc(1, sizeof(void *));
+        auto &storage = peer_heap_base_p2p_storage_[&heap];
+        if (!storage) storage = std::make_unique<int>(1);
+        void *ptr = storage.get();
+
+        if (heap.peer_heap_base_p2p_ == nullptr) {
+            heap.peer_heap_base_p2p_ = (void **)std::calloc(1, sizeof(void *));
+        }
         heap.peer_heap_base_p2p_[0] = ptr;
     }
 
    private:
     static nvml_function_table nvml_ftable_;
     static void *nvml_handle_;
+    std::map<nvshmemi_symmetric_heap *, std::unique_ptr<int>> heap_base_storage_;
+    std::map<nvshmemi_symmetric_heap *, std::unique_ptr<int>> peer_heap_base_p2p_storage_;
     explicit nvshmemi_mem_p2p_transport(int mype, int npes){};
     static nvshmemi_mem_p2p_transport *p2p_objref_;  // singleton instance
 };