Skip to content

Instantly share code, notes, and snippets.

@mcollinswisc
Last active August 7, 2026 18:09
Show Gist options
  • Select an option

  • Save mcollinswisc/7cc504381d6c30ee50ef72745f47c6e5 to your computer and use it in GitHub Desktop.

Select an option

Save mcollinswisc/7cc504381d6c30ee50ef72745f47c6e5 to your computer and use it in GitHub Desktop.
Comparison of bit shift implementations (ONNX, numpy, torch, jax, mindspore)
"""Compare signed BitShift across NN frameworks.
Comparison is to NumPy, PyTorch, JAX, TensorFlow, MindSpore and the ONNX
reference evaluator.
Covers the cases checked in as ONNX node tests, plus the edge and ambiguous
cases deliberately left out of them (for now): negative shift amounts, shifts at
or past the bit width on types the node tests do not exercise, and the extremes
of each type.
Setup
-----
python -m venv /tmp/bitshift-venv
. /tmp/bitshift-venv/bin/activate
pip install numpy
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install 'jax[cpu]'
pip install -e /path/to/onnx/repo # builds C++, takes a few minutes
python bitshift_compare.py
The editable install is what puts the signed-BitShift branch under test. Against
released ONNX or the main branch, BitShift is still opset 11 and every signed
case is rejected by the type constraint.
TensorFlow is optional and is skipped when it is not installed. It publishes no
wheels for 3.14 yet. On a newer interpreter the rest of the comparison still
runs, just without TF.
pip install tensorflow-cpu
MindSpore is optional and is skipped when it is not installed. It requires
numpy < 2, which the others have long since moved past, so it needs a venv of
its own rather than a place alongside them:
pip install mindspore
"""
from __future__ import annotations
import jax
import numpy as np
import torch
# See the module docstring: MindSpore's version bounds keep it out of the venv
# the other implementations run in, so it is reported only where it is present.
# It must precede onnx, because it sets PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION
# on import, which has no effect once onnx has loaded protobuf's C++ backend,
# and MindSpore's own generated protos fail to load under that backend.
try:
import mindspore
except ImportError:
mindspore = None
import onnx
import onnx.reference
# TensorFlow ships no wheels for the newest Python releases, so it is reported
# only where it is present. It must follow onnx, because it links its own copy
# of protobuf: imported first, onnx's extension binds against that copy instead
# of the one it was built against and the checker segfaults.
try:
import tensorflow as tf
except ImportError:
tf = None
# JAX computes in 32 bits by default, downcasting the int64 and uint64 cases.
jax.config.update("jax_enable_x64", True)
BITSHIFT_OPSET = onnx.defs.get_schema("BitShift").since_version
UNSIGNED = [np.uint8, np.uint16, np.uint32, np.uint64]
SIGNED = [np.int8, np.int16, np.int32, np.int64]
DTYPES = UNSIGNED + SIGNED
INT8_MIN, INT8_MAX = -128, 127
INT32_MIN, INT32_MAX = -(2**31), 2**31 - 1
INT64_MIN = -(2**63)
def width(dtype):
"""Bit width of an integer type: 8 for int8, 32 for uint32, and so on."""
return np.dtype(dtype).itemsize * 8
class Case:
def __init__(self, name, dtype, direction, x, y):
self.direction = direction
self.x = np.array(x).astype(dtype)
self.y = np.array(y).astype(dtype)
# The dtype is on the arrays already, so the caller names only the
# behavior and the table gets the type appended.
self.name = f"{name}_{self.x.dtype.name}"
def numpy_result(self):
fn = np.left_shift if self.direction == "LEFT" else np.right_shift
return fn(self.x, self.y)
def torch_result(self):
fn = (
torch.bitwise_left_shift
if self.direction == "LEFT"
else torch.bitwise_right_shift
)
# from_numpy preserves the dtype exactly, and raises for dtypes this
# torch build has no equivalent for, which is reported per case.
return fn(torch.from_numpy(self.x), torch.from_numpy(self.y)).numpy()
def jax_result(self):
fn = (
jax.numpy.left_shift
if self.direction == "LEFT"
else jax.numpy.right_shift
)
# jax.numpy.right_shift picks an arithmetic or a logical shift from the
# dtype, which is the rule this operator change adopts.
return np.asarray(fn(self.x, self.y))
def tensorflow_result(self):
fn = (
tf.bitwise.left_shift
if self.direction == "LEFT"
else tf.bitwise.right_shift
)
# TensorFlow clamps the shift count into [0, bit width - 1] instead of
# letting it run off the end, so it parts company with the others
# wherever Y is negative or reaches the width.
return fn(self.x, self.y).numpy()
def mindspore_result(self):
fn = (
mindspore.ops.bitwise_left_shift
if self.direction == "LEFT"
else mindspore.ops.bitwise_right_shift
)
return fn(mindspore.Tensor(self.x), mindspore.Tensor(self.y)).asnumpy()
def _model(self):
def value_info(name, shape):
elem_type = onnx.helper.np_dtype_to_tensor_dtype(self.x.dtype)
return onnx.helper.make_tensor_value_info(
name, elem_type, list(shape)
)
node = onnx.helper.make_node(
"BitShift", ["x", "y"], ["z"], direction=self.direction
)
graph = onnx.helper.make_graph(
[node],
"bitshift",
[value_info("x", self.x.shape), value_info("y", self.y.shape)],
[value_info("z", np.broadcast_shapes(self.x.shape, self.y.shape))],
)
model = onnx.helper.make_model(
graph,
opset_imports=[onnx.helper.make_operatorsetid("", BITSHIFT_OPSET)],
)
# full_check runs shape inference, which is what enforces the type
# constraint; without it a disallowed dtype would slip through.
onnx.checker.check_model(model, full_check=True)
return model
def onnx_result(self):
evaluator = onnx.reference.ReferenceEvaluator(self._model())
return evaluator.run(None, {"x": self.x, "y": self.y})[0]
def run(self):
implementations = [
("numpy", self.numpy_result),
("torch", self.torch_result),
("jax", self.jax_result),
]
if tf is not None:
implementations.append(("tensorflow", self.tensorflow_result))
if mindspore is not None:
implementations.append(("mindspore", self.mindspore_result))
implementations.append(("onnx", self.onnx_result))
results = {}
for label, method in implementations:
try:
results[label] = method()
except Exception as exc: # noqa: BLE001
# A failure is a result too: torch has no shift for the wider
# unsigned types, and ONNX rejects signed ones before opset 28.
results[label] = exc
return results
def all_cases():
both = ("RIGHT", "LEFT")
# The two types the signed node tests exercise, and the six they do not.
tested = [np.int8, np.int32]
untested = [np.uint8, np.uint16, np.uint32, np.uint64, np.int16, np.int64]
# Everything down to the negative shift amounts exists as an ONNX node test
# on this branch; everything after that is deliberately not checked in (for
# now).
cases = [
Case("basic", dt, direction, [16, 4, 1], [1, 2, 3])
for dt in DTYPES
for direction in both
]
for name, dt, direction, x, y in [
("negative_input", np.int8, "RIGHT", [-8, -1, INT8_MIN], [1, 1, 1]),
("negative_input", np.int32, "RIGHT", [-8, -1, INT32_MIN], [1, 1, 1]),
("overflow", np.int8, "LEFT", [64, 1, -64], [1, 7, 1]),
("overflow", np.int32, "LEFT", [2**30, 1, -(2**30)], [1, 31, 1]),
("shift_ge_width", np.int8, "RIGHT", [-8, 4, -1], [8, 9, 127]),
("shift_ge_width", np.int8, "LEFT", [-8, 4, -1], [8, 9, 127]),
("shift_ge_width", np.int32, "RIGHT", [-8, 4, -1], [32, 33, 100]),
("shift_ge_width", np.int32, "LEFT", [-8, 4, -1], [32, 33, 100]),
]:
cases.append(Case(name, dt, direction, x, y))
# Negative shift amounts. The schema requires Y to be non-negative and
# leaves this undefined, mirroring what numpy.left_shift documents, so any
# result here conforms.
cases += [
Case(
"negative_shift",
dt,
direction,
[-8, 4, -1],
[-1, -width(dt), -width(dt) * 2],
)
for dt in tested
for direction in both
]
# Shift at or past the bit width, on the types the node tests skip.
cases += [
Case(
"shift_ge_width",
dt,
direction,
[-8, 4, -1] if dt in SIGNED else [8, 4, 1],
[width(dt), width(dt) + 1, width(dt) * 2],
)
for dt in untested
for direction in both
]
# Either side of the boundary, where masking the shift count would show up.
cases += [
Case(
"boundary",
dt,
direction,
[-8, -8, -8],
[width(dt) - 2, width(dt) - 1, width(dt)],
)
for dt in tested
for direction in both
]
for name, dt, x, y in [
("extremes", np.int8, [INT8_MIN, INT8_MAX, -1], [1, 1, 1]),
("extremes", np.int32, [INT32_MIN, INT32_MAX, -1], [1, 1, 1]),
("min_input", np.int64, [INT64_MIN, -1, 1], [1, 1, 1]),
("shift_by_zero", np.int8, [-8, 4, INT8_MIN], [0, 0, 0]),
]:
cases += [Case(name, dt, d, x, y) for d in both]
# Broadcasting with a signed dtype, which the node tests do not cover.
cases += [
Case("bcast_scalar_y", np.int32, "RIGHT", [-64, -8, 8], [2]),
Case("bcast_scalar_x", np.int32, "LEFT", [-1], [0, 1, 2]),
]
return cases
def fmt(value):
if isinstance(value, Exception):
return f"! {type(value).__name__}: {value}".split("\n")[0]
return "[" + ", ".join(str(v) for v in np.atleast_1d(value)) + "]"
def verdict(results):
values = [v for v in results.values() if not isinstance(v, Exception)]
# Disagreement is checked first: a case where one implementation is missing
# and the rest disagree is still a disagreement, not merely a gap.
if not all(np.array_equal(values[0], v) for v in values[1:]):
return "DIFFER"
if len(values) < len(results):
return "partial" # at least one implementation could not run it
return "same"
def main():
print(f"onnx {onnx.__version__} (BitShift opset {BITSHIFT_OPSET})")
print(f"numpy {np.__version__}")
print(f"torch {torch.__version__}")
print(f"jax {jax.__version__}")
if tf is not None:
print(f"tensorflow {tf.__version__}")
if mindspore is not None:
print(f"mindspore {mindspore.__version__}")
verdicts = []
for case in all_cases():
results = case.run()
verdicts.append(verdict(results))
print()
print(case.name)
print(f"x: {fmt(case.x)}")
print(f"y: {fmt(case.y)}")
print(f"dir: {case.direction}")
for label, value in results.items():
print(f"{label}: {fmt(value)}")
print(
f"\n{len(verdicts)} cases: {verdicts.count('DIFFER')} differ in value, "
f"{verdicts.count('partial')} ran in only some implementations."
)
if __name__ == "__main__":
main()
onnx 1.23.0 (BitShift opset 28)
numpy 2.5.1
torch 2.13.0+cpu
jax 0.11.0
basic_uint8
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint8
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint16
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt16'
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint16
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt16'
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint32
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt32'
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint32
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt32'
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint64
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt64'
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint64
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt64'
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_int8
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_int8
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_int16
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_int16
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_int32
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_int32
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
onnx: [32, 16, 8]
basic_int64
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
onnx: [8, 1, 0]
basic_int64
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
onnx: [32, 16, 8]
negative_input_int8
x: [-8, -1, -128]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4, -1, -64]
torch: [-4, -1, -64]
jax: [-4, -1, -64]
onnx: [-4, -1, -64]
negative_input_int32
x: [-8, -1, -2147483648]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4, -1, -1073741824]
torch: [-4, -1, -1073741824]
jax: [-4, -1, -1073741824]
onnx: [-4, -1, -1073741824]
overflow_int8
x: [64, 1, -64]
y: [1, 7, 1]
dir: LEFT
numpy: [-128, -128, -128]
torch: [-128, -128, -128]
jax: [-128, -128, -128]
onnx: [-128, -128, -128]
overflow_int32
x: [1073741824, 1, -1073741824]
y: [1, 31, 1]
dir: LEFT
numpy: [-2147483648, -2147483648, -2147483648]
torch: [-2147483648, -2147483648, -2147483648]
jax: [-2147483648, -2147483648, -2147483648]
onnx: [-2147483648, -2147483648, -2147483648]
shift_ge_width_int8
x: [-8, 4, -1]
y: [8, 9, 127]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int8
x: [-8, 4, -1]
y: [8, 9, 127]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_int32
x: [-8, 4, -1]
y: [32, 33, 100]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int32
x: [-8, 4, -1]
y: [32, 33, 100]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
negative_shift_int8
x: [-8, 4, -1]
y: [-1, -8, -16]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
negative_shift_int8
x: [-8, 4, -1]
y: [-1, -8, -16]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
negative_shift_int32
x: [-8, 4, -1]
y: [-1, -32, -64]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
negative_shift_int32
x: [-8, 4, -1]
y: [-1, -32, -64]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint8
x: [8, 4, 1]
y: [8, 9, 16]
dir: RIGHT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint8
x: [8, 4, 1]
y: [8, 9, 16]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint16
x: [8, 4, 1]
y: [16, 17, 32]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt16'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint16
x: [8, 4, 1]
y: [16, 17, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt16'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint32
x: [8, 4, 1]
y: [32, 33, 64]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt32'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint32
x: [8, 4, 1]
y: [32, 33, 64]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt32'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint64
x: [8, 4, 1]
y: [64, 65, 128]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt64'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint64
x: [8, 4, 1]
y: [64, 65, 128]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt64'
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_int16
x: [-8, 4, -1]
y: [16, 17, 32]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int16
x: [-8, 4, -1]
y: [16, 17, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_int64
x: [-8, 4, -1]
y: [64, 65, 128]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int64
x: [-8, 4, -1]
y: [64, 65, 128]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
boundary_int8
x: [-8, -8, -8]
y: [6, 7, 8]
dir: RIGHT
numpy: [-1, -1, -1]
torch: [-1, -1, -1]
jax: [-1, -1, -1]
onnx: [-1, -1, -1]
boundary_int8
x: [-8, -8, -8]
y: [6, 7, 8]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
boundary_int32
x: [-8, -8, -8]
y: [30, 31, 32]
dir: RIGHT
numpy: [-1, -1, -1]
torch: [-1, -1, -1]
jax: [-1, -1, -1]
onnx: [-1, -1, -1]
boundary_int32
x: [-8, -8, -8]
y: [30, 31, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
onnx: [0, 0, 0]
extremes_int8
x: [-128, 127, -1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-64, 63, -1]
torch: [-64, 63, -1]
jax: [-64, 63, -1]
onnx: [-64, 63, -1]
extremes_int8
x: [-128, 127, -1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, -2]
torch: [0, -2, -2]
jax: [0, -2, -2]
onnx: [0, -2, -2]
extremes_int32
x: [-2147483648, 2147483647, -1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-1073741824, 1073741823, -1]
torch: [-1073741824, 1073741823, -1]
jax: [-1073741824, 1073741823, -1]
onnx: [-1073741824, 1073741823, -1]
extremes_int32
x: [-2147483648, 2147483647, -1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, -2]
torch: [0, -2, -2]
jax: [0, -2, -2]
onnx: [0, -2, -2]
min_input_int64
x: [-9223372036854775808, -1, 1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4611686018427387904, -1, 0]
torch: [-4611686018427387904, -1, 0]
jax: [-4611686018427387904, -1, 0]
onnx: [-4611686018427387904, -1, 0]
min_input_int64
x: [-9223372036854775808, -1, 1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, 2]
torch: [0, -2, 2]
jax: [0, -2, 2]
onnx: [0, -2, 2]
shift_by_zero_int8
x: [-8, 4, -128]
y: [0, 0, 0]
dir: RIGHT
numpy: [-8, 4, -128]
torch: [-8, 4, -128]
jax: [-8, 4, -128]
onnx: [-8, 4, -128]
shift_by_zero_int8
x: [-8, 4, -128]
y: [0, 0, 0]
dir: LEFT
numpy: [-8, 4, -128]
torch: [-8, 4, -128]
jax: [-8, 4, -128]
onnx: [-8, 4, -128]
bcast_scalar_y_int32
x: [-64, -8, 8]
y: [2]
dir: RIGHT
numpy: [-16, -2, 2]
torch: [-16, -2, 2]
jax: [-16, -2, 2]
onnx: [-16, -2, 2]
bcast_scalar_x_int32
x: [-1]
y: [0, 1, 2]
dir: LEFT
numpy: [-1, -2, -4]
torch: [-1, -2, -4]
jax: [-1, -2, -4]
onnx: [-1, -2, -4]
54 cases: 0 differ in value, 12 ran in only some implementations.
onnx 1.23.0 (BitShift opset 28)
numpy 1.26.4
torch 2.13.0+cpu
jax 0.7.1
tensorflow 2.21.0
mindspore 2.10.0
basic_uint8
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint8
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint16
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt16'
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint16
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt16'
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint32
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt32'
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint32
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt32'
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_uint64
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt64'
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_uint64
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt64'
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_int8
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_int8
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_int16
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_int16
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_int32
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_int32
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
basic_int64
x: [16, 4, 1]
y: [1, 2, 3]
dir: RIGHT
numpy: [8, 1, 0]
torch: [8, 1, 0]
jax: [8, 1, 0]
tensorflow: [8, 1, 0]
mindspore: [8, 1, 0]
onnx: [8, 1, 0]
basic_int64
x: [16, 4, 1]
y: [1, 2, 3]
dir: LEFT
numpy: [32, 16, 8]
torch: [32, 16, 8]
jax: [32, 16, 8]
tensorflow: [32, 16, 8]
mindspore: [32, 16, 8]
onnx: [32, 16, 8]
negative_input_int8
x: [-8, -1, -128]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4, -1, -64]
torch: [-4, -1, -64]
jax: [-4, -1, -64]
tensorflow: [-4, -1, -64]
mindspore: [-4, -1, -64]
onnx: [-4, -1, -64]
negative_input_int32
x: [-8, -1, -2147483648]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4, -1, -1073741824]
torch: [-4, -1, -1073741824]
jax: [-4, -1, -1073741824]
tensorflow: [-4, -1, -1073741824]
mindspore: [-4, -1, -1073741824]
onnx: [-4, -1, -1073741824]
overflow_int8
x: [64, 1, -64]
y: [1, 7, 1]
dir: LEFT
numpy: [-128, -128, -128]
torch: [-128, -128, -128]
jax: [-128, -128, -128]
tensorflow: [-128, -128, -128]
mindspore: [-128, -128, -128]
onnx: [-128, -128, -128]
overflow_int32
x: [1073741824, 1, -1073741824]
y: [1, 31, 1]
dir: LEFT
numpy: [-2147483648, -2147483648, -2147483648]
torch: [-2147483648, -2147483648, -2147483648]
jax: [-2147483648, -2147483648, -2147483648]
tensorflow: [-2147483648, -2147483648, -2147483648]
mindspore: [-2147483648, -2147483648, -2147483648]
onnx: [-2147483648, -2147483648, -2147483648]
shift_ge_width_int8
x: [-8, 4, -1]
y: [8, 9, 127]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-1, 0, -1]
mindspore: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int8
x: [-8, 4, -1]
y: [8, 9, 127]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, -128]
mindspore: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_int32
x: [-8, 4, -1]
y: [32, 33, 100]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-1, 0, -1]
mindspore: [-8, 2, -1]
onnx: [-1, 0, -1]
shift_ge_width_int32
x: [-8, 4, -1]
y: [32, 33, 100]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, -2147483648]
mindspore: [-8, 8, -16]
onnx: [0, 0, 0]
negative_shift_int8
x: [-8, 4, -1]
y: [-1, -8, -16]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-8, 4, -1]
mindspore: [-1, 0, -1]
onnx: [-1, 0, -1]
negative_shift_int8
x: [-8, 4, -1]
y: [-1, -8, -16]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [-8, 4, -1]
mindspore: [0, 0, 0]
onnx: [0, 0, 0]
negative_shift_int32
x: [-8, 4, -1]
y: [-1, -32, -64]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-8, 4, -1]
mindspore: [-1, 4, -1]
onnx: [-1, 0, -1]
negative_shift_int32
x: [-8, 4, -1]
y: [-1, -32, -64]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [-8, 4, -1]
mindspore: [0, 4, -1]
onnx: [0, 0, 0]
shift_ge_width_uint8
x: [8, 4, 1]
y: [8, 9, 16]
dir: RIGHT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint8
x: [8, 4, 1]
y: [8, 9, 16]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, 128]
mindspore: [0, 0, 0]
onnx: [0, 0, 0]
shift_ge_width_uint16
x: [8, 4, 1]
y: [16, 17, 32]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt16'
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [0, 0, 1]
onnx: [0, 0, 0]
shift_ge_width_uint16
x: [8, 4, 1]
y: [16, 17, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt16'
jax: [0, 0, 0]
tensorflow: [0, 0, 32768]
mindspore: [0, 0, 1]
onnx: [0, 0, 0]
shift_ge_width_uint32
x: [8, 4, 1]
y: [32, 33, 64]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt32'
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [8, 2, 1]
onnx: [0, 0, 0]
shift_ge_width_uint32
x: [8, 4, 1]
y: [32, 33, 64]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt32'
jax: [0, 0, 0]
tensorflow: [0, 0, 2147483648]
mindspore: [8, 8, 1]
onnx: [0, 0, 0]
shift_ge_width_uint64
x: [8, 4, 1]
y: [64, 65, 128]
dir: RIGHT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "rshift_cpu" not implemented for 'UInt64'
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [8, 2, 1]
onnx: [0, 0, 0]
shift_ge_width_uint64
x: [8, 4, 1]
y: [64, 65, 128]
dir: LEFT
numpy: [0, 0, 0]
torch: ! NotImplementedError: "lshift_cpu" not implemented for 'UInt64'
jax: [0, 0, 0]
tensorflow: [0, 0, 9223372036854775808]
mindspore: [8, 8, 1]
onnx: [0, 0, 0]
shift_ge_width_int16
x: [-8, 4, -1]
y: [16, 17, 32]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-1, 0, -1]
mindspore: [-1, 0, -1]
onnx: [-1, 0, -1]
shift_ge_width_int16
x: [-8, 4, -1]
y: [16, 17, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, -32768]
mindspore: [0, 0, -1]
onnx: [0, 0, 0]
shift_ge_width_int64
x: [-8, 4, -1]
y: [64, 65, 128]
dir: RIGHT
numpy: [-1, 0, -1]
torch: [-1, 0, -1]
jax: [-1, 0, -1]
tensorflow: [-1, 0, -1]
mindspore: [-8, 2, -1]
onnx: [-1, 0, -1]
shift_ge_width_int64
x: [-8, 4, -1]
y: [64, 65, 128]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, -9223372036854775808]
mindspore: [-8, 8, -1]
onnx: [0, 0, 0]
boundary_int8
x: [-8, -8, -8]
y: [6, 7, 8]
dir: RIGHT
numpy: [-1, -1, -1]
torch: [-1, -1, -1]
jax: [-1, -1, -1]
tensorflow: [-1, -1, -1]
mindspore: [-1, -1, -1]
onnx: [-1, -1, -1]
boundary_int8
x: [-8, -8, -8]
y: [6, 7, 8]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [0, 0, 0]
onnx: [0, 0, 0]
boundary_int32
x: [-8, -8, -8]
y: [30, 31, 32]
dir: RIGHT
numpy: [-1, -1, -1]
torch: [-1, -1, -1]
jax: [-1, -1, -1]
tensorflow: [-1, -1, -1]
mindspore: [-1, -1, -8]
onnx: [-1, -1, -1]
boundary_int32
x: [-8, -8, -8]
y: [30, 31, 32]
dir: LEFT
numpy: [0, 0, 0]
torch: [0, 0, 0]
jax: [0, 0, 0]
tensorflow: [0, 0, 0]
mindspore: [0, 0, -8]
onnx: [0, 0, 0]
extremes_int8
x: [-128, 127, -1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-64, 63, -1]
torch: [-64, 63, -1]
jax: [-64, 63, -1]
tensorflow: [-64, 63, -1]
mindspore: [-64, 63, -1]
onnx: [-64, 63, -1]
extremes_int8
x: [-128, 127, -1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, -2]
torch: [0, -2, -2]
jax: [0, -2, -2]
tensorflow: [0, -2, -2]
mindspore: [0, -2, -2]
onnx: [0, -2, -2]
extremes_int32
x: [-2147483648, 2147483647, -1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-1073741824, 1073741823, -1]
torch: [-1073741824, 1073741823, -1]
jax: [-1073741824, 1073741823, -1]
tensorflow: [-1073741824, 1073741823, -1]
mindspore: [-1073741824, 1073741823, -1]
onnx: [-1073741824, 1073741823, -1]
extremes_int32
x: [-2147483648, 2147483647, -1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, -2]
torch: [0, -2, -2]
jax: [0, -2, -2]
tensorflow: [0, -2, -2]
mindspore: [0, -2, -2]
onnx: [0, -2, -2]
min_input_int64
x: [-9223372036854775808, -1, 1]
y: [1, 1, 1]
dir: RIGHT
numpy: [-4611686018427387904, -1, 0]
torch: [-4611686018427387904, -1, 0]
jax: [-4611686018427387904, -1, 0]
tensorflow: [-4611686018427387904, -1, 0]
mindspore: [-4611686018427387904, -1, 0]
onnx: [-4611686018427387904, -1, 0]
min_input_int64
x: [-9223372036854775808, -1, 1]
y: [1, 1, 1]
dir: LEFT
numpy: [0, -2, 2]
torch: [0, -2, 2]
jax: [0, -2, 2]
tensorflow: [0, -2, 2]
mindspore: [0, -2, 2]
onnx: [0, -2, 2]
shift_by_zero_int8
x: [-8, 4, -128]
y: [0, 0, 0]
dir: RIGHT
numpy: [-8, 4, -128]
torch: [-8, 4, -128]
jax: [-8, 4, -128]
tensorflow: [-8, 4, -128]
mindspore: [-8, 4, -128]
onnx: [-8, 4, -128]
shift_by_zero_int8
x: [-8, 4, -128]
y: [0, 0, 0]
dir: LEFT
numpy: [-8, 4, -128]
torch: [-8, 4, -128]
jax: [-8, 4, -128]
tensorflow: [-8, 4, -128]
mindspore: [-8, 4, -128]
onnx: [-8, 4, -128]
bcast_scalar_y_int32
x: [-64, -8, 8]
y: [2]
dir: RIGHT
numpy: [-16, -2, 2]
torch: [-16, -2, 2]
jax: [-16, -2, 2]
tensorflow: [-16, -2, 2]
mindspore: [-16, -2, 2]
onnx: [-16, -2, 2]
bcast_scalar_x_int32
x: [-1]
y: [0, 1, 2]
dir: LEFT
numpy: [-1, -2, -4]
torch: [-1, -2, -4]
jax: [-1, -2, -4]
tensorflow: [-1, -2, -4]
mindspore: [-1, -2, -4]
onnx: [-1, -2, -4]
54 cases: 19 differ in value, 6 ran in only some implementations.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment