Backport bytes compiler options support to CUDA 12

Carry forward the runtime behavior from #1665/#1698 with its nvJitLink regression and NVVM bytes-options case. Extracted from #3173 so source integration can be reviewed independently.

Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>

Author: rwgk

cuda_bindings_12/cuda/bindings/_internal/utils.pyx

@@ -98,7 +98,14 @@ cdef int get_nested_resource_ptr(nested_resource[ResT] &in_out_ptr, object obj,
         nested_ptr.reset(nested_vec, True)
         for i, obj_i in enumerate(obj):
             if ResT is char:
-                obj_i_bytes = (<str?>(obj_i)).encode()
+                obj_i_type = type(obj_i)
+                if obj_i_type is str:
+                    obj_i_bytes = obj_i.encode("utf-8")
+                elif obj_i_type is bytes:
+                    obj_i_bytes = obj_i
+                else:
+                    raise TypeError(
+                        f"Expected str or bytes, got {obj_i_type.__name__}")
                 str_len = <size_t>(len(obj_i_bytes)) + 1  # including null termination
                 deref(nested_res_vec)[i].resize(str_len)
                 obj_i_ptr = <char*>(obj_i_bytes)

cuda_bindings_12/tests/test_nvjitlink.py

@@ -100,6 +100,13 @@ def test_create_and_destroy(option):
     nvjitlink.destroy(handle)
 
 
+@pytest.mark.human_authored
+def test_create_and_destroy_bytes_options():
+    handle = nvjitlink.create(1, [b"-arch=sm_80"])
+    assert handle != 0
+    nvjitlink.destroy(handle)
+
+
 @pytest.mark.parametrize("option", ARCHITECTURES)
 def test_complete_empty(option):
     handle = nvjitlink.create(1, [f"-arch={option}"])

cuda_bindings_12/tests/test_nvvm.py

@@ -246,7 +246,7 @@ def test_get_buffer_empty(get_size, get_buffer):
         assert buffer == b"\x00"
 
 
-@pytest.mark.parametrize("options", [[], ["-opt=0"], ["-opt=3", "-g"]])
+@pytest.mark.parametrize("options", [[], ["-opt=0"], ["-opt=3", "-g"], [b"-opt=0"]])
 def test_compile_program_with_minimal_nvvm_ir(minimal_nvvmir, options):
     with nvvm_program() as prog:
         nvvm.add_module_to_program(prog, minimal_nvvmir, len(minimal_nvvmir), "FileNameHere.ll")