Fix CUDA pointer and coredump attribute types (#2929)

* Fix pointer device ordinal attribute result

* Allow pointer ordinal null sentinels

* Fix additional CUDA attribute result types

* Format CUDA attribute regression test

Author: kkraus14

cuda_bindings/cuda/bindings/_lib/utils.pxd

@@ -5,6 +5,7 @@ cimport cuda.bindings.driver as driver
 cimport cuda.bindings.cydriver as cydriver
 cimport cuda.bindings.cyruntime as cyruntime
 from libcpp.vector cimport vector
+from libcpp cimport bool as cpp_bool
 from cpython.buffer cimport PyBuffer_Release, Py_buffer
 
 cdef class _HelperKernelParams:
@@ -62,7 +63,7 @@ cdef class _HelperCUpointer_attribute:
     cdef unsigned int _uint
     cdef int _int
     cdef driver.CUdeviceptr _devptr
-    cdef void** _void
+    cdef void* _void
     cdef driver.CUDA_POINTER_ATTRIBUTE_P2P_TOKENS _token
     cdef bint _bool
     cdef unsigned long long _ull
@@ -165,5 +166,5 @@ cdef class _HelperCUcoredumpSettings:
     cdef object _references  # keeps caller bytes alive so _charstar stays valid
 
     # Return values
-    cdef bint _bool
+    cdef cpp_bool _bool
     cdef char* _charstar

cuda_bindings/cuda/bindings/_lib/utils.pxi

@@ -251,7 +251,6 @@ cdef class _HelperCUpointer_attribute:
             else:
                 self._cptr = <void*><void_ptr>init_value.getPtr()
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MEMORY_TYPE,
-                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ACCESS_FLAGS,):
             self._uint = init_value
@@ -267,7 +266,7 @@ cdef class _HelperCUpointer_attribute:
             else:
                 self._cptr = <void*><void_ptr>init_value.getPtr()
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_HOST_POINTER,):
-            self._void = <void**><void_ptr>init_value
+            self._void = <void*><void_ptr>init_value
             self._cptr = <void*>&self._void
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_P2P_TOKENS,):
             if self._is_getter:
@@ -281,7 +280,8 @@ cdef class _HelperCUpointer_attribute:
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MAPPED,):
             self._bool = init_value
             self._cptr = <void*>&self._bool
-        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_BUFFER_ID,):
+        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_BUFFER_ID,
+                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,):
             self._ull = init_value
             self._cptr = <void*>&self._ull
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_SIZE,):
@@ -308,11 +308,11 @@ cdef class _HelperCUpointer_attribute:
         if self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_CONTEXT,):
             return self._ctx
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MEMORY_TYPE,
-                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,
-                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ACCESS_FLAGS,):
             return self._uint
+        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,):
+            return self._int
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_POINTER,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,):
             return self._devptr
@@ -325,7 +325,8 @@ cdef class _HelperCUpointer_attribute:
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_LEGACY_CUDA_IPC_CAPABLE,
                             cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MAPPED,):
             return self._bool
-        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_BUFFER_ID,):
+        elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_BUFFER_ID,
+                            cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,):
             return self._ull
         elif self._attr in (cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_SIZE,):
             return self._size
@@ -680,7 +681,7 @@ cdef class _HelperCUcoredumpSettings:
                 self._bool = init_value
 
             self._cptr = <void*>&self._bool
-            self._size = 1
+            self._size = sizeof(cpp_bool)
         else:
             raise TypeError('Unsupported attribute: {}'.format(attr.name))
 

cuda_bindings/tests/test_cuda.py

@@ -379,12 +379,23 @@ def test_pointer_get_attributes_device_ordinal():
         cuda.CUpointer_attribute.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,
     ]
 
-    attrs = cuda.cuPointerGetAttributes(len(attributes), attributes, 0)
+    err, attrs = cuda.cuPointerGetAttributes(len(attributes), attributes, 0)
 
-    # device ordinals are always small numbers.  A large number would indicate
-    # an overflow error.
+    assert err == cuda.CUresult.CUDA_SUCCESS
+    # Drivers use either negative sentinel for a null pointer. Both ensure the
+    # signed device ordinal is returned rather than an unsigned value.
+    assert attrs[0] in (cuda.CU_DEVICE_CPU, cuda.CU_DEVICE_INVALID)
+
+
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_pointer_allowed_handle_types_preserves_64_bits():
+    helper = cuda._HelperCUpointer_attribute(
+        cuda.CUpointer_attribute.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES, 0, is_getter=True
+    )
+    expected = 1 << 40
+    ctypes.c_uint64.from_address(helper.cptr).value = expected
 
-    assert abs(attrs[1][0]) < 256
+    assert helper.pyObj() == expected
 
 
 @pytest.mark.skipif(not supportsManagedMemory(), reason="When new attributes were introduced")
@@ -546,6 +557,16 @@ def test_cuda_coredump_attr():
     assert attr_list[3] is True
 
 
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_coredump_bool_uses_single_byte_storage():
+    helper = cuda._HelperCUcoredumpSettings(cuda.CUcoredumpSettings.CU_COREDUMP_TRIGGER_HOST, 0, is_getter=True)
+    ctypes.c_uint32.from_address(helper.cptr).value = 0xFFFFFFFF
+    ctypes.c_uint8.from_address(helper.cptr).value = 0
+
+    assert helper.size() == ctypes.sizeof(ctypes.c_bool)
+    assert helper.pyObj() is False
+
+
 def test_get_error_name_and_string():
     err, device = cuda.cuDeviceGet(0)
     _, s = cuda.cuGetErrorString(err)