Fix the double loading problem of libamdhip64.so.

PyTorch wheel has its own bundled libamdhip64.so, and the original logic has problems
detecting this shared object file because it can be installed at any place with
venv/DT_RUNPATH/LD_LIBRARY_PATH.

Now Triton enumerates the address space with dl_iterate_phdr, and prefer the
libamdhip64.so that is already loaded into address space over everything other
than TRITON_LIBHIP_PATH.

third_party/amd/backend/driver.py

@@ -13,6 +13,53 @@
 include_dir = [os.path.join(dirname, "include")]
 
 
+def _linux_find_already_mmapped_dylib():
+    import platform;
+    if platform.system() != 'Linux':
+        return ''
+    from ctypes import (
+        Structure,
+        c_char,
+        c_void_p,
+        c_char_p,
+        c_int,
+        c_size_t,
+        POINTER,
+        CFUNCTYPE,
+        CDLL,
+        memmove,
+    )
+    from ctypes.util import find_library
+    import ctypes
+    class dl_phdr_info(Structure):
+        _fields_ = [
+                ('dlpi_addr', c_void_p),
+                ('dlpi_name', c_char_p),
+                # Don't care the remaining fields
+        ]
+    # callback_t must use POINTER(c_char) to avoid copying
+    callback_t = CFUNCTYPE(c_int, POINTER(dl_phdr_info), POINTER(c_size_t), POINTER(c_char))
+    dl_iterate_phdr = CDLL('libc.so.6').dl_iterate_phdr
+    # argtypes must use c_char_p to accept ctypes.create_string_buffer
+    dl_iterate_phdr.argtypes = [callback_t, c_char_p]
+    dl_iterate_phdr.restype = c_int
+
+    MAX_PATH = 4096
+    path = ctypes.create_string_buffer(MAX_PATH + 1)
+    def callback(info, size, data):
+        dlpi_name = info.contents.dlpi_name
+        p = Path(os.fsdecode(dlpi_name))
+        if 'libamdhip64.so' in p.name:
+            memmove(data, dlpi_name, min(MAX_PATH, len(dlpi_name)))
+            return 1
+        return 0
+    found = dl_iterate_phdr(callback_t(callback), path)
+    if found:
+        filepath = os.fsdecode(ctypes.string_at(path))
+        print(f'_linux_find_already_mmapped_dylib {filepath=}')
+        return filepath
+    return ''
+
 @functools.lru_cache()
 def _get_path_to_hip_runtime_dylib():
     lib_name = "libamdhip64.so"
@@ -24,6 +71,13 @@ def _get_path_to_hip_runtime_dylib():
             return env_libhip_path
         raise RuntimeError(f"TRITON_LIBHIP_PATH '{env_libhip_path}' does not point to a valid {lib_name}")
 
+    # If the shared object is already mmapped to address space, use it.
+    mmapped_libhip = _linux_find_already_mmapped_dylib()
+    if mmapped_libhip:
+        if os.path.exists(mmapped_libhip):
+            return mmapped_libhip
+        raise RuntimeError(f"Memory Mapped '{mmapped_libhip}' does not point to a valid {lib_name}")
+
     paths = []
 
     import site