world.apply_settings() slow in 0.9.11 synchronous mode

[syveqc]

Hi,

I just upgraded to Carla 0.9.11 and now my code runs slow. (Ubuntu 20.04). After some digging I found that world.apply_settings() needs significantly more time than before.
Code example:

import carla
import time

#load carla and apply basic settings
client = carla.Client("localhost", 2000)
world = client.load_world("Town04")
settings = world.get_settings()
settings.no_rendering_mode = True
settings.fixed_delta_seconds = 0.1
settings.synchronous_mode = True
world.apply_settings(settings)

#needed because first tick takes longer than subsequent ones sometimes
world.tick(1)

#tick without applying settings
start = time.time()
world.tick(1)
print(f"normal step needed {int((time.time()-start)*1000)}ms")

#tick after applying settings
start = time.time()
world.apply_settings(settings)
print(f"apply settings needed {int((time.time()-start)*1000)}ms")
world.tick(1)
print(f"apply settings and step needed {int((time.time()-start)*1000)}ms")

I tested it in opengl and vulkan mode, the error exists in both. (in vulkan the non-apply_settings() steps are a lot quicker, the apply_settings() steps need nearly identical time)

Output in 0.9.10 on my machine (time may vary a couple ms)

normal step needed 6ms
apply settings needed 0ms
apply settings and step needed 9ms

Output in 0.9.11 on my machine (time may vary a couple ms)

normal step needed 8ms
apply settings needed 1002ms
apply settings and step needed 1003ms

For completion's sake here is my hardware, but the behavior was also confirmed on another machine, so I do not think the hardware actually matters:

• CPU: AMD Ryzen 5950x
• GPU: Nvidia RTX 3090
• RAM: 32GB 3600MHz C16

Thanks in advance for your help!

[hh0rva1h]

I can confirm this.
@germanros1987 could you please triage this issue?

[corkyw10]

@bernatx could you follow up on this please?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[hh0rva1h]

go away stale bot

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[hh0rva1h]

@bernatx Please follow up, Code example is given, should be quick to look into.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[hh0rva1h]

@bernatx Still an issue with 0.9.13, why is it so much slower? This is actually relevant for people doing reinforcement learning and applying settings after each episode: 1 second wait time after each episode can amount to much training time being wasted, especially at the begin of the training when episodes are very short!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.