onnx-extended: extensions for onnx and onnxruntime

onnx-extended extends the list of supported operators in onnx reference implementation and onnxruntime, or implements faster versions in C++. Documentation onnx-extended. Source are available on github/onnx-extended.

Use a C++ implementation of existing operators

import timeit
import numpy as np
from onnx import TensorProto
from onnx.helper import (
    make_graph,
    make_model,
    make_node,
    make_opsetid,
    make_tensor_value_info,
)
from onnx.reference import ReferenceEvaluator
from onnxruntime import InferenceSession
from onnx_extended.ext_test_case import measure_time
from onnx_extended.reference import CReferenceEvaluator


X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None])
Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None])
B = make_tensor_value_info("B", TensorProto.FLOAT, [None, None, None, None])
W = make_tensor_value_info("W", TensorProto.FLOAT, [None, None, None, None])
node = make_node(
    "Conv",
    ["X", "W", "B"],
    ["Y"],
    pads=[1, 1, 1, 1],
    dilations=[1, 1],
    strides=[2, 2],
)
graph = make_graph([node], "g", [X, W, B], [Y])
onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)])

sH, sW = 64, 64
X = np.arange(sW * sH).reshape((1, 1, sH, sW)).astype(np.float32)
W = np.ones((1, 1, 3, 3), dtype=np.float32)
B = np.array([[[[0]]]], dtype=np.float32)

sess1 = ReferenceEvaluator(onnx_model)
sess2 = CReferenceEvaluator(onnx_model)  # 100 times faster

expected = sess1.run(None, {"X": X, "W": W, "B": B})[0]
got = sess2.run(None, {"X": X, "W": W, "B": B})[0]
diff = np.abs(expected - got).max()
print(f"difference: {diff}")

f1 = lambda: sess1.run(None, {"X": X, "W": W, "B": B})[0]
f2 = lambda: sess2.run(None, {"X": X, "W": W, "B": B})[0]
print("onnx:", timeit.timeit(f1, globals=globals(), number=5))
print("onnx-extended:", timeit.timeit(f2, globals=globals(), number=5))

difference: 0.0
onnx: 0.024006774998269975
onnx-extended: 0.0002316169993719086

Build with CUDA, openmp, eigen, onnxruntime

The package also contains some dummy examples on how to build with C++ functions (pybind11, cython), with openmp, eigen with or without CUDA. It also shows how to create a custom operator for onnxruntime in C++.

The version released on pypi/onnx-extended only works on CPU. It needs to be manually built to enable the code using CUDA. The build will automatically link with CUDA if it is found. If not, some extensions might not be available.

python setup.py build_ext --inplace
# pip install -e .

It is possible to use a specific version of CUDA:

python setup.py build_ext --inplace --cuda-version=11.8
# or (not working yet)
# pip install -e . --config-settings="--cuda-version=11.8"
# pip install -e . --global-option="--cuda-version=11.8"
export USE_CUDA=11.8
pip install -e .

NVTX can be enabled with the following command:

python setup.py build_ext --inplace --use_nvtx 1
# or (not working yet)
# pip install -e . --config-settings="--use_nvtx=1"
pip install -e . --global-option "--use_nvtx=1"

Experimental cython binding for onnxruntime

The python onnxruntime package relies on pybind11 to expose its functionalities. onnx-extended tries to build a cython wrapper around the C/C++ API of onnxruntime. cython relies on python C API and is faster than pybind11. This different may be significant when onnxruntime is used on small graphs and tensors.

Custom kernels for onnxruntime

onnxruntime provides an API to add custom implementation for existing or new onnx operators. An example for CPU.

from onnxruntime import InferenceSession, SessionOptions
from onnx_extended.ortops.optim.cpu import get_ort_ext_libs

r = get_ort_ext_libs()
opts = SessionOptions()
if r is not None:
    opts.register_custom_ops_library(r[0])

sess_cus = InferenceSession(
    onx_modified.SerializeToString(), opts, providers=["CPUExecutionProvider"]
)