Intel(R) oneAPI Collective Communications Library (oneCCL) 2021.15.2 (#171)

Author: nikitaxgusev

CMakeLists.txt

@@ -335,7 +335,7 @@ endif()
 
 set(CCL_MAJOR_VERSION     "2021")
 set(CCL_MINOR_VERSION     "15")
-set(CCL_UPDATE_VERSION    "1")
+set(CCL_UPDATE_VERSION    "2")
 set(CCL_PRODUCT_STATUS    "Gold")
 string(TIMESTAMP CCL_PRODUCT_BUILD_DATE "%Y-%m-%dT %H:%M:%SZ")
 get_vcs_properties("git")

include/oneapi/ccl/config.h

@@ -32,10 +32,10 @@
 
 #define CCL_MAJOR_VERSION           2021
 #define CCL_MINOR_VERSION           15
-#define CCL_UPDATE_VERSION          1
+#define CCL_UPDATE_VERSION          2
 #define CCL_PRODUCT_STATUS     "Gold"
-#define CCL_PRODUCT_BUILD_DATE "2025-05-05T 03:34:33Z"
-#define CCL_PRODUCT_FULL       "Gold-2021.15.1 2025-05-05T 03:34:33Z (master/0b49b8f)"
+#define CCL_PRODUCT_BUILD_DATE "2025-05-22T 03:58:44Z"
+#define CCL_PRODUCT_FULL       "Gold-2021.15.2 2025-05-22T 03:58:44Z (master/9bd0c7a)"
 
 #if defined(SYCL_LANGUAGE_VERSION) && defined (__INTEL_LLVM_COMPILER)
 #define CCL_ENABLE_SYCL

man/doxconfig

@@ -1,5 +1,5 @@
 PROJECT_NAME           = "Intel® oneAPI Collective Communications Library"
-PROJECT_NUMBER         = "2021.15.1"
+PROJECT_NUMBER         = "2021.15.2"
 
 INPUT = ../src/common/env/vars.hpp ../src/common/env/vars_experimental.hpp
 

man/man3/OneCCL.3

@@ -1,4 +1,4 @@
-.TH "OneCCLvars" 3 "Tue Mar 18 2025" "Version 2021.15.1" "Intel® oneAPI Collective Communications Library" \" -*- nroff -*-
+.TH "OneCCLvars" 3 "Tue Mar 18 2025" "Version 2021.15.2" "Intel® oneAPI Collective Communications Library" \" -*- nroff -*-
 .ad l
 .nh
 .SH NAME
@@ -1051,7 +1051,7 @@ By-default: '536870912'
 .SH "Author"
 .PP 
 Generated automatically by Doxygen for Intel® oneAPI Collective Communications Library from the source code\&.
-.TH "ExpOneCCLvars" 3 "Tue Mar 18 2025" "Version 2021.15.1" "Intel® oneAPI Collective Communications Library" \" -*- nroff -*-
+.TH "ExpOneCCLvars" 3 "Tue Mar 18 2025" "Version 2021.15.2" "Intel® oneAPI Collective Communications Library" \" -*- nroff -*-
 .ad l
 .nh
 .SH NAME

src/common/api_wrapper/ofi_api_wrapper.cpp

@@ -67,7 +67,7 @@ std::string get_ofi_lib_path() {
     return ofi_lib_path;
 }
 
-std::string get_relative_ofi_lib_path() {
+static std::string get_relative_ccl_root_path() {
     Dl_info info;
 
     if (dladdr((void*)ccl::get_library_version, &info)) {
@@ -84,16 +84,15 @@ std::string get_relative_ofi_lib_path() {
             char* ccl_root_cstr = dirname(libccl_dir);
 
             auto ccl_root = std::string(ccl_root_cstr);
-            setenv("FI_PROVIDER_PATH", (ccl_root + "/lib/libfabric/prov").c_str(), 0);
-            return ccl_root + "/lib/libfabric/libfabric.so";
+
+            return ccl_root;
         }
     }
 
-    LOG_DEBUG("Could not fetch relative path to libfabric. Fallback to `libfabric.so`");
-    return "libfabric.so";
+    return {};
 }
 
-bool ofi_api_init() {
+static bool load_libfabric() {
     ofi_lib_info.ops = &ofi_lib_ops;
     ofi_lib_info.fn_names = ofi_fn_names;
     ofi_lib_info.path = get_ofi_lib_path();
@@ -106,7 +105,20 @@ bool ofi_api_init() {
     print_error(error, ofi_lib_info);
     LOG_INFO("Retrying to load libfabric.so using relative path");
 
-    ofi_lib_info.path = get_relative_ofi_lib_path();
+    auto realtive_root = get_relative_ccl_root_path();
+    if (realtive_root.empty()) {
+        return false;
+    }
+
+    // Path up to IMPI 2021.14
+    ofi_lib_info.path = realtive_root + "/lib/libfabric/libfabric.so";
+    error = load_library(ofi_lib_info);
+    if (error == CCL_LOAD_LB_SUCCESS) {
+        return true;
+    }
+
+    // Path in IMPI 2021.15
+    ofi_lib_info.path = realtive_root + "/lib/libfabric.so";
     error = load_library(ofi_lib_info);
     if (error == CCL_LOAD_LB_SUCCESS) {
         return true;
@@ -116,6 +128,46 @@ bool ofi_api_init() {
     return false;
 }
 
+static void setup_providers() {
+    const char* fi_provider_path = getenv("FI_PROVIDER_PATH");
+    if (fi_provider_path != nullptr) {
+        LOG_DEBUG("FI_PROVIDER_PATH is already set to: ", fi_provider_path);
+        return;
+    }
+
+    char libfabric_path[PATH_MAX];
+    dlinfo(ofi_lib_info.handle, RTLD_DI_ORIGIN, &libfabric_path);
+
+    // Add realpath to resolve any symlinks and get the absolute path
+    char real_libfabric_path[PATH_MAX];
+    if (!realpath(libfabric_path, real_libfabric_path)) {
+        LOG_ERROR("Failed to resolve libfabric realpath: ", strerror(errno));
+        return;
+    }
+
+    std::string primary_path = std::string(real_libfabric_path);
+    std::string secondary_path = primary_path + "/prov";
+
+    // Construct the full provider path with colon separator
+    std::string full_provider_path = primary_path + ":" + secondary_path;
+
+    if (setenv("FI_PROVIDER_PATH", full_provider_path.c_str(), 1) != 0) {
+        LOG_ERROR("Failed to set FI_PROVIDER_PATH with error: ", strerror(errno));
+        return;
+    }
+
+    LOG_DEBUG("FI_PROVIDER_PATH set to: ", full_provider_path);
+}
+
+bool ofi_api_init() {
+    if (load_libfabric() == false) {
+        return false;
+    }
+
+    setup_providers();
+    return true;
+}
+
 void ofi_api_fini() {
     LOG_DEBUG("close OFI lib: handle: ", ofi_lib_info.handle);
     close_library(ofi_lib_info);

third-party-programs.txt

@@ -1,5 +1,5 @@
 Intel(R) oneAPI Collective Communications Library (oneCCL) 
-2021.15.1 Third Party Programs File
+2021.15.2 Third Party Programs File
 
 This file is the "third-party-programs.txt" file specified in the associated 
 Intel end user license agreement for the Intel software you are licensing.