AMD CPU limitation

Describe the bug AMD CPU produces significantly worse performance than Intel CPUs in the “Simple PYSR Example” from pysr_demo.ipynb. AMD appears limited to utilize one core while Intel can utilize all cores.

Additional context Tested CPU / OS:

• Ryzen 5900x (12 cores / 24 threads): 6% CPU usage (using 1 core, 1x baseline)
• i5 2020 macbook pro (4 cores / 8 threads): 60% CPU usage (using all cores, 2.1 x faster than the AMD)
• i5 6600k windows 10 desktop (4 cores / 4 threads): 80% CPU usage (using all cores,1.2 x faster than the AMD)

Running the PyTorch trainer In [12], with gpus=0 utilizes all CPU cores on all systems tested.

Version (please complete the following information):

• OS: macOS 12.4, intel (OK), Windows 10, intel (OK), Windows 11, AMD (Problem)

• Version: PYSR 0.9.3

• Python 3.9 venv in pycharm 2022.1.3

• Does the bug appear with the latest version of PySR? Yes

Configuration

• What are your PySR settings? Exactly the Simple PySR example found in pysr_demo.ipynb

default_pysr_params = dict(
    populations=30,
    procs=4,
    model_selection="best",
)
# Learn equations
model = PySRRegressor(
    niterations=30,
    binary_operators=["plus", "mult"],
    unary_operators=["cos", "exp", "sin"],
    **default_pysr_params
)

model.fit(X, y)

• What dataset are you running on?

# Dataset
np.random.seed(0)
X = 2 * np.random.randn(100, 5)
y = 2.5382 * np.cos(X[:, 3]) + X[:, 0] ** 2 - 2

Error message No message, just performance loss on the AMD chip.

[mtaylorDW]

Thanks @MilesCranmer , multithreading=False gave a 2-3x increase in cpu utilization on the AMD machine, utilizing 20% of the CPU resources and doubles the MBP’s speed in cycles per second. I was then able to saturate the CPU by cranking up the population_size parameter, which previously had no effect.

[MilesCranmer]

As a quick fix, if you set multithreading=False, it will use multiprocessing instead, which is almost guaranteed to fix this, and give you the same performance as on other machines. Windows seems to have issues with the multithreading backend of PySR for some reason.