Closed
Description
Describe the issue:
With 2.0.0dev0 wheel on scientific-python-nightly-wheels there is a big slowdown for @ / dot for us. First noticed a ~1.5x slowdown in MNE-Python CIs, then went through a bunch of stuff with @seberg on Discord (thanks for your patience!). Finally I think I created a minimal example:
$ python3 -m venv ~/python/virtualenvs/npbad
$ source ~/python/virtualenvs/npbad/bin/activate
$ pip install --default-timeout=60 --extra-index-url "https://pypi.anaconda.org/scientific-python-nightly-wheels/simple" "numpy==1.25.0rc1+218.g0e5a362fd"
$ pip list
Package Version
---------- ------------------------
numpy 1.25.0rc1+218.g0e5a362fd
pip 23.0.1
setuptools 66.1.1
$ which python
/home/larsoner/python/virtualenvs/npbad/bin/python
$ python -m timeit -s "import numpy as np; x = np.random.RandomState(0).randn(300, 10000)" "x @ x.T"
20 loops, best of 5: 18.6 msec per loop
$ pip install --default-timeout=60 --upgrade --extra-index-url "https://pypi.anaconda.org/scientific-python-nightly-wheels/simple" "numpy==2.0.0dev0"
$ python -m timeit -s "import numpy as np; x = np.random.RandomState(0).randn(300, 10000)" "x @ x.T"
1 loop, best of 5: 911 msec per loopTL;DR: 18.6ms on "1.25.0rc1" from just under a month ago, 911ms on latest 2.0.0dev0 on my machine.
I've tried to reproduce this on main on my machine as well by building myself by setting dispatches, using meson or not, etc. but have only ever managed to get the good/fast time.
Reproduce the code example:
AboveError message:
N/ARuntime information:
Above
Context for the issue:
Above