avoid extra copy + ensure compiled objcode loadable

Author: leofang

cuda_core/cuda/core/experimental/_module.py

@@ -11,7 +11,7 @@
 from cuda.core.experimental._stream import Stream
 from cuda.core.experimental._utils.clear_error_support import (
     assert_type,
-    assert_type_str_or_bytes,
+    assert_type_str_or_bytes_like,
     raise_code_path_meant_to_be_unreachable,
 )
 from cuda.core.experimental._utils.cuda_utils import driver, get_binding_version, handle_return, precondition
@@ -615,14 +615,14 @@ def _lazy_load_module(self, *args, **kwargs):
         if self._handle is not None:
             return
         module = self._module
-        assert_type_str_or_bytes(module)
+        assert_type_str_or_bytes_like(module)
         if isinstance(module, str):
             if self._backend_version == "new":
                 self._handle = handle_return(self._loader["file"](module.encode(), [], [], 0, [], [], 0))
             else:  # "old" backend
                 self._handle = handle_return(self._loader["file"](module.encode()))
             return
-        if isinstance(module, bytes):
+        if isinstance(module, (bytes, bytearray)):
             if self._backend_version == "new":
                 self._handle = handle_return(self._loader["data"](module, [], [], 0, [], [], 0))
             else:  # "old" backend

cuda_core/cuda/core/experimental/_program.py

@@ -657,8 +657,7 @@ def compile(self, target_type, name_expressions=(), logs=None):
                     nvvm.get_program_log(self._mnff.handle, log)
                     logs.write(log.decode("utf-8", errors="backslashreplace"))
 
-            data_bytes = bytes(data)
-            return ObjectCode._init(data_bytes, target_type, name=self._options.name)
+            return ObjectCode._init(data, target_type, name=self._options.name)
 
         supported_backends = ("nvJitLink", "driver")
         if self._backend not in supported_backends:

cuda_core/cuda/core/experimental/_utils/clear_error_support.py

@@ -9,10 +9,10 @@ def assert_type(obj, expected_type):
         raise TypeError(f"Expected type {expected_type.__name__}, but got {type(obj).__name__}")
 
 
-def assert_type_str_or_bytes(obj):
+def assert_type_str_or_bytes_like(obj):
     """Ensure obj is of type str or bytes, else raise AssertionError with a clear message."""
-    if not isinstance(obj, (str, bytes)):
-        raise TypeError(f"Expected type str or bytes, but got {type(obj).__name__}")
+    if not isinstance(obj, (str, bytes, bytearray)):
+        raise TypeError(f"Expected type str or bytes or bytearray, but got {type(obj).__name__}")
 
 
 def raise_code_path_meant_to_be_unreachable():

cuda_core/tests/test_program.py

@@ -312,11 +312,13 @@ def test_nvvm_deferred_import():
 
 
 @nvvm_available
-def test_nvvm_program_creation(nvvm_ir):
+def test_nvvm_program_creation_compilation(nvvm_ir):
     """Test basic NVVM program creation"""
     program = Program(nvvm_ir, "nvvm")
     assert program.backend == "NVVM"
     assert program.handle is not None
+    obj = program.compile("ptx")
+    ker = obj.get_kernel("simple")  # noqa: F841
     program.close()