Add LIBNAMES_REQUIRING_RTLD_DEEPBIND feature (for cufftMp)

Author: rwgk

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/load_dl_linux.py

@@ -8,7 +8,10 @@
 from typing import Optional, cast
 
 from cuda.pathfinder._dynamic_libs.load_dl_common import LoadedDL
-from cuda.pathfinder._dynamic_libs.supported_nvidia_libs import SUPPORTED_LINUX_SONAMES
+from cuda.pathfinder._dynamic_libs.supported_nvidia_libs import (
+    LIBNAMES_REQUIRING_RTLD_DEEPBIND,
+    SUPPORTED_LINUX_SONAMES,
+)
 
 CDLL_MODE = os.RTLD_NOW | os.RTLD_GLOBAL
 
@@ -138,6 +141,13 @@ def check_if_already_loaded_from_elsewhere(libname: str, _have_abs_path: bool) -
     return None
 
 
+def _load_lib(libname: str, filename: str) -> ctypes.CDLL:
+    cdll_mode = CDLL_MODE
+    if libname in LIBNAMES_REQUIRING_RTLD_DEEPBIND:
+        cdll_mode |= os.RTLD_DEEPBIND
+    return ctypes.CDLL(filename, cdll_mode)
+
+
 def load_with_system_search(libname: str) -> Optional[LoadedDL]:
     """Try to load a library using system search paths.
 
@@ -152,13 +162,14 @@ def load_with_system_search(libname: str) -> Optional[LoadedDL]:
     """
     for soname in get_candidate_sonames(libname):
         try:
-            handle = ctypes.CDLL(soname, CDLL_MODE)
+            handle = _load_lib(libname, soname)
+        except OSError:
+            pass
+        else:
             abs_path = abs_path_for_dynamic_library(libname, handle)
             if abs_path is None:
                 raise RuntimeError(f"No expected symbol for {libname=!r}")
             return LoadedDL(abs_path, False, handle._handle)
-        except OSError:
-            pass
     return None
 
 
@@ -196,7 +207,7 @@ def load_with_abs_path(libname: str, found_path: str) -> LoadedDL:
     """
     _work_around_known_bugs(libname, found_path)
     try:
-        handle = ctypes.CDLL(found_path, CDLL_MODE)
+        handle = _load_lib(libname, found_path)
     except OSError as e:
         raise RuntimeError(f"Failed to dlopen {found_path}: {e}") from e
     return LoadedDL(found_path, False, handle._handle)

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/supported_nvidia_libs.py

@@ -394,6 +394,8 @@
     "nvrtc",
 )
 
+LIBNAMES_REQUIRING_RTLD_DEEPBIND = ("cufftMp",)
+
 
 def is_suppressed_dll_file(path_basename: str) -> bool:
     if path_basename.startswith("nvrtc"):