Compiling Python functions for use with other languages#

Numba CUDA MLIR can compile Python code to PTX or LTO-IR so that Python functions can be incorporated into CUDA code written in other languages (e.g. C/C++). It is commonly used to support User-Defined Functions written in Python within the context of a library or application.

The compilation API can be used without a GPU present, as it uses no driver functions and avoids initializing CUDA in the process. It is invoked through the following function:

If a device is available and compiled code for the compute capability of the current device is required (for example when building a JIT compilation workflow using Numba CUDA MLIR), the compile_for_current_device function can be used:

Most users should use the two functions described above; for backwards compatibility with existing use cases, the following functions are also provided:

Using the C ABI#

Numba CUDA MLIR internally uses its own ABI - this is as described in Device function ABI, without the extern "C" modifier. Calling Numba CUDA MLIR ABI device functions requires three issues to be addressed:

  • The name of the function will be mangled according to Numba CUDA MLIR’s ABI rules - these are based on the Itanium C++ ABI rules, but are extended beyond its specifications.

  • The Python return value is expected to be stored into a pointer value passed in the first argument.

  • The return value of the compiled function will contain a status code, instead of the return value of the function. For use of Numba CUDA MLIR-compiled functions outside of Numba CUDA MLIR, this can generally be ignored.

A simple way to address all these issues is to compile device functions with the C ABI instead. This results in the following:

  • The name of the device function in the compiled code can be controlled. By default it will match the name of the function in Python, so it is easy to determine. This is the function’s __name__, rather than __qualname__, because __qualname__ encodes additional scoping information that would make the function name hard to predict, and in a lot of cases, an illegal identifier in C.

  • The returned value of the Python code is placed in the return value of the compiled function.

  • Status codes are ignored / unreported, so they do not need to be handled.

If the name of the compiled function needs to be specified, it can be controlled by passing the name in the abi_info dict, under the key 'abi_name'.

Compilation with the C ABI is the default when using the compile() and compile_for_current_device() functions. The compile_ptx() and compile_ptx_for_current_device() functions default to the Numba CUDA MLIR ABI in order to maintain compatibility with existing use cases.

C and Numba CUDA MLIR ABI examples#

The following function:

def add(x, y):
    return x + y

compiled for the Numba ABI using, for example:

ptx, resty = cuda.compile_ptx(add, int32(int32, int32), device=True)

results in PTX where the function prototype is:

.visible .func  (.param .b32 func_retval0) _ZN8__main__3addB2v1B94cw51cXTLSUwv1sCUt9Uw1VEw0NRRQPKzLTg4gaGKFsG2oMQGEYakJSQB1PQBk0Bynm21OiwU1a0UoLGhDpQE8oxrNQE_3dEii(
    .param .b64 _ZN8__main__3addB2v1B94cw51cXTLSUwv1sCUt9Uw1VEw0NRRQPKzLTg4gaGKFsG2oMQGEYakJSQB1PQBk0Bynm21OiwU1a0UoLGhDpQE8oxrNQE_3dEii_param_0,
    .param .b32 _ZN8__main__3addB2v1B94cw51cXTLSUwv1sCUt9Uw1VEw0NRRQPKzLTg4gaGKFsG2oMQGEYakJSQB1PQBk0Bynm21OiwU1a0UoLGhDpQE8oxrNQE_3dEii_param_1,
    .param .b32 _ZN8__main__3addB2v1B94cw51cXTLSUwv1sCUt9Uw1VEw0NRRQPKzLTg4gaGKFsG2oMQGEYakJSQB1PQBk0Bynm21OiwU1a0UoLGhDpQE8oxrNQE_3dEii_param_2
 )

Note that there are three parameters, for the pointer to the return value, x, and y. The name is mangled in a way that is hard to predict outside of Numba CUDA MLIR internals.

Compiling for the C ABI with:

ptx, resty = cuda.compile_ptx(add, int32(int32, int32), device=True, abi="c")

instead results in the following PTX prototype:

.visible .func  (.param .b32 func_retval0) add(
    .param .b32 add_param_0,
    .param .b32 add_param_1
)

The function name matches the Python source function name, and there are exactly two parameters, for x and y. The result of the function is directly placed in the return value:

add.s32        %r3, %r2, %r1;
st.param.b32   [func_retval0+0], %r3;

To distinguish one variant of the compiled add() function from another, the following example specifies its ABI name in the abi_info dict:

ptx, resty = cuda.compile_ptx(add, float32(float32, float32), device=True,
                              abi="c", abi_info={"abi_name": "add_f32"})

Resulting in the PTX prototype:

.visible .func  (.param .b32 func_retval0) add_f32(
    .param .b32 add_f32_param_0,
    .param .b32 add_f32_param_1
)

which will not clash with definitions by other names (e.g. the variant for int32 above).