Calling np.fft.irfft simultaneously in multiple threads prevents handling of SIGINT

When running np.fft.irfft or np.fft.irfft2 simultaneously in multiple threads, it messes up something with the handling of SIGINT for the entire Python process. I even tried using the signal module to register a SIGINT handler separately. After starting the threads, whether they complete successfully or not, the process can no longer handle SIGINT.

Reproducing code example:

Running forever, can’t CTRL-C:

import threading
import numpy as np
import time

def task(lock, done):
    while not done.is_set():
        # Try using the lock and see SIGINT handling restored
        #with lock:
        np.fft.irfft(np.array(range(10000)))
    print("Task done")

threads = []
done_event = threading.Event()
lock = threading.Lock()
for i in range(10):
    thread = threading.Thread(target=task, args=(lock, done_event))
    threads.append(thread)
    thread.start()

print("All tasks started")

try:
    while True:
        time.sleep(0.1)
except KeyboardInterrupt:
    print("Keyboard interrupt!")

print("Shutting down")
done_event.set()
for thread in threads:
    thread.join()
print("Shutdown successful")

Even weirder - can’t CTRL-C even after the threads are finished:

import threading
import numpy as np
import time

def task(lock, done):
    for i in range(100):
        if done.is_set():
            break
        # Try using the lock and see SIGINT handling restored
        #with lock:
        np.fft.irfft(np.array(range(10000)))
    print("Task done")

threads = []
done_event = threading.Event()
lock = threading.Lock()
for i in range(10):
    thread = threading.Thread(target=task, args=(lock, done_event))
    threads.append(thread)
    thread.start()

print("All tasks started")

try:
    while True:
        time.sleep(0.1)
except KeyboardInterrupt:
    print("Keyboard interrupt!")

print("Shutting down")
done_event.set()
for thread in threads:
    thread.join()
print("Shutdown successful")

Numpy/Python version information:

1.15.1 3.5.2 (default, Nov 23 2017, 16:37:01) [GCC 5.4.0 20160609]

EDIT: Running on Ubuntu 16.04

[seberg]

Closing, we removed the sigint handling from fft, its just not worth the trouble anymore.

[mreineck]

I’m not sure what the plans for future numpy releases are, but if inclusion of C++ modules is on the table, it may be worth switching numpy to scipy’s implementation of pocketfft. The speedups are considerable, and the interface code is simpler.