CUDA Array Interface (Version 2)¶
The cuda array interface is created for interoperability between different implementation of GPU array-like objects in various projects. The idea is borrowed from the numpy array interface.
Note
Currently, we only define the Python-side interface. In the future, we may add a C-side interface for efficient exchange of the information in compiled code.
Python Interface Specification¶
Note
Experimental feature. Specification may change.
The __cuda_array_interface__
attribute returns a dictionary (dict
)
that must contain the following entries:
shape:
(integer, ...)
A tuple of
int
(orlong
) representing the size of each dimension.typestr:
str
The type string. This has the same definition as
typestr
in the numpy array interface.data:
(integer, boolean)
The data is a 2-tuple. The first element is the data pointer as a Python
int
(orlong
). The data must be device-accessible. For zero-size arrays, use0
here. The second element is the read-only flag as a Pythonbool
.Because the user of the interface may or may not be in the same context, the most common case is to use
cuPointerGetAttribute
withCU_POINTER_ATTRIBUTE_DEVICE_POINTER
in the CUDA driver API (or the equivalent CUDA Runtime API) to retrieve a device pointer that is usable in the currently active context.version:
integer
An integer for the version of the interface being exported. The current version is 2.
The following are optional entries:
strides:
None
or(integer, ...)
If strides is not given, or it is
None
, the array is in C-contiguous layout. Otherwise, a tuple ofint
(orlong
) is explicitly given for representing the number of bytes to skip to access the next element at each dimension.descr
This is for describing more complicated types. This follows the same specification as in the numpy array interface.
mask:
None
or object exposing the__cuda_array_interface__
If
None
then all values in data are valid. All elements of the mask array should be interpreted only as true or not true indicating which elements of this array are valid. This has the same definition asmask
in the numpy array interface.Note
Numba does not currently support working with masked CUDA arrays and will raise a
NotImplementedError
exception if one is passed to a GPU function.
Lifetime management¶
Obtaining the value of the __cuda_array_interface__
property of any object
has no effect on the lifetime of the object from which it was created. In
particular, note that the interface has no slot for the owner of the data.
It is therefore imperative for a consumer to retain a reference to the object owning the data for as long as they make use of the data.
Lifetime management in Numba¶
Numba provides two mechanisms for creating device arrays. Which to use depends on whether the created device array should maintain the life of the object from which it is created:
as_cuda_array
: This creates a device array that holds a reference to the owning object. As long as a reference to the device array is held, its underlying data will also be kept alive, even if all other references to the original owning object have been dropped.from_cuda_array_interface
: This creates a device array with no reference to the owning object by default. The owning object, or some other object to be considered the owner can be passed in theowner
parameter.
The interfaces of these functions are:
-
cuda.
as_cuda_array
(obj)¶ Create a DeviceNDArray from any object that implements the cuda array interface.
A view of the underlying GPU buffer is created. No copying of the data is done. The resulting DeviceNDArray will acquire a reference from obj.
-
cuda.
from_cuda_array_interface
(desc, owner=None)¶ Create a DeviceNDArray from a cuda-array-interface description. The owner is the owner of the underlying memory. The resulting DeviceNDArray will acquire a reference from it.
Pointer Attributes¶
Additional information about the data pointer can be retrieved using
cuPointerGetAttribute
or cudaPointerGetAttributes
. Such information
include:
the CUDA context that owns the pointer;
is the pointer host-accessible?
is the pointer a managed memory?
Differences with CUDA Array Interface (Version 0)¶
Version 0 of the CUDA Array Interface did not have the optional mask attribute to support masked arrays.
Differences with CUDA Array Interface (Version 1)¶
Versions 0 and 1 of the CUDA Array Interface neither clarified the strides attribute for C-contiguous arrays nor specified the treatment for zero-size arrays.