Different path result under same random seed. Sionna v1.2

[WindMelon]

Hi,
I am using Sionna v1.2 to do some work on Channel Simulation and I found I get different result when I enable diffuse_reflection
Here is my code, using example scene.

import random

import numpy as np
from sionna.rt import load_scene, Camera, Transmitter, Receiver, PlanarArray, PathSolver, RadioMapSolver
import mitsuba as mi

import os
try:
    import sionna
except ImportError as e:
    # Install Sionna if package is not already installed
    os.system("pip install sionna")
    import sionna

import tensorflow as tf

tf.get_logger().setLevel('ERROR')

tf.random.set_seed(128)  # Set global random seed for reproducibility
np.random.seed(128)
random.seed(128)

# Load example scene
scene = load_scene(sionna.rt.scene.munich)
# scene = load_scene(sionna.rt.scene.simple_street_canyon_with_cars)

# Configure antenna array for all transmitters
scene.tx_array = PlanarArray(num_rows=1,
                             num_cols=1,
                             vertical_spacing=0.5,
                             horizontal_spacing=0.5,
                             pattern="iso",
                             polarization="V")

# Configure antenna array for all receivers
scene.rx_array = PlanarArray(num_rows=1,
                             num_cols=1,
                             vertical_spacing=0.5,
                             horizontal_spacing=0.5,
                             pattern="iso",
                             polarization="V")
my_cam = Camera(position=mi.Point3f(-250, 250, 150), look_at=mi.Point3f(-15, 30, 28))

tx = Transmitter(name="tx", position=mi.Point3f([8.5, 21, 27]))
scene.add(tx)
rx = Receiver(name="rx", position=mi.Point3f(45, 90, 1.5), orientation=mi.Point3f(0, 0, 0))
scene.add(rx)

tx.look_at(rx)

scene.frequency = 4.9e9
refl = scene.get('ground')
refl.radio_material.scattering_coefficient = 0.1
solver = PathSolver()
paths = solver(scene, max_depth=3, los=True,
               specular_reflection=True,
               diffuse_reflection=True,
               refraction=True, seed=128)

path_idx = 0
print(f"\n--- Detailed results for path {path_idx} ---")
print(
    f"Channel coefficient: {paths.a[0].numpy()[0, 0, 0, 0, path_idx] + 1j * paths.a[1].numpy()[0, 0, 0, 0, path_idx]}")
print(f"Propagation delay: {paths.tau[0, 0, path_idx].numpy() * 1e6:.5f} us")
print(f"Zenith angle of departure: {paths.theta_t.numpy()[0, 0, path_idx]:.4f} rad")
print(f"Azimuth angle of departure: {paths.phi_t.numpy()[0, 0, path_idx]:.4f} rad")
print(f"Zenith angle of arrival: {paths.theta_r.numpy()[0, 0, path_idx]:.4f} rad")
print(f"Azimuth angle of arrival: {paths.phi_r.numpy()[0, 0, path_idx]:.4f} rad")
print(f"Doppler shift: {paths.doppler.numpy()[0, 0, path_idx]:.4f} Hz")
print(paths.valid.numpy().shape)

output 1:

--- Detailed results for path 0 ---
Channel coefficient: (5.9288719057803974e-05+0j)
Propagation delay: 0.27392 us
Zenith angle of departure: 1.8865 rad
Azimuth angle of departure: 1.0842 rad
Zenith angle of arrival: 1.2551 rad
Azimuth angle of arrival: -2.0574 rad
Doppler shift: 0.0000 Hz
(1, 1, 1008)

output 2:

--- Detailed results for path 0 ---
Channel coefficient: (5.836293226479938e-09+0j)
Propagation delay: 0.30180 us
Zenith angle of departure: 1.9982 rad
Azimuth angle of departure: 1.3340 rad
Zenith angle of arrival: 1.6298 rad
Azimuth angle of arrival: -2.6752 rad
Doppler shift: 0.0000 Hz
(1, 1, 1010)

output 3:

--- Detailed results for path 0 ---
Channel coefficient: (6.1929532613191896e-09+0j)
Propagation delay: 0.61432 us
Zenith angle of departure: 1.8070 rad
Azimuth angle of departure: 1.7342 rad
Zenith angle of arrival: 1.5924 rad
Azimuth angle of arrival: 2.4911 rad
Doppler shift: 0.0000 Hz
(1, 1, 1009)

I set all seeds as 128, and still get different result. Is there any way to ensure the determinacy? Thanks.

[WindMelon]

It's running on CPU

[jhoydis]

Hi @WindMelon,

There are two issues: (1) The order of returned paths is not deterministic and you should check the full channel impulse response (or order path by their delay). (2) I believe that the way diffuse reflections are computed is non-deterministic due to importance sampling (However, I need to check this with @faycalaa). Is there a particular reason why you need deterministic diffuse reflections? Since there are generally so many, it should not impact the channel quality in any way.

[WindMelon]

Hi @jhoydis,
Thanks for your reply!
The reason I want to make path result deterministic is that I'm working on NR system simulation, I wonder if I can use sionna to generate stable Channels to make sure the system simulation deterministic.
Now I found a way to eliminate randomness by setting drjit thread count.

import drjit as dr
dr.set_thread_count(1)

Although this might be much slower. :)

[jhoydis]

Hi @WindMelon,

I am inclined to believe that deterministic vs random diffuse scattering will not make a huge impact on observed system level performance.

[WindMelon]

@jhoydis
Thank you for your answer and suggestions.