Memory access problem with 1.2.1 with Python in Linux

[Johndeep]

Hi everyone,

after roughly checking this issue for duplicates or similarities, I came to the conclusion to try to post this.

Just to be clear, I work with Python in Ubuntu 22.04 LTS in Wayland. But I also tested this issue in Xorg. Unfortunately same result.

After updating imgui-bundle from 1.1.0 to 1.2.1 via pip, I noticed that the line
imgui.create_context()
leads to a crash of my app. Now the curious part.

• I've tried this with sdl2 first, since it is my main backend in use currently. It doesn't work. That's how I noticed it.
• Then I tried the same with glfw3. It also crashes.
• After that, I then tried the example backend files from the repository. Neither do they work.
• Desperately I thought, "Maybe I did something wrong? Maybe the backends doesn't work anymore.", thus, I tried the "demo_hello_world.py" (with immapp) demo code, which can be found in the doc. Also no luck.

The app starts briefly in all cases. It shows no content. However, it either stops without any terminal output or outputs a memory access error line, which is not helpful.

I'm working on a system with Intel core i7-1365U, which comes with Iris Xe and thus in Ubuntu with Mesa Intel Graphics (RPL-P) driver. I think this is quite strange, since imgui-bundle 1.1.0 works fine.

Currently, I'm working with version 1.1.0 until this is resolved.

This is quite a bummer actually, since I really like to use the new functionality of ChildFlags to make the child area re-sizable.

[pthom]

Hello,

Thanks for reaching out. I cannot reproduce your issue here, so I will need your help to investigate.

Python side

Could you confirm that this fails:

from imgui_bundle import imgui
imgui.create_context()

If so, this is probably something inside imgui (or in the bindings around it).

Can you please:

• reboot and check whether the problem persists
• try to install from source (see https://pthom.github.io/imgui_bundle/install-instructions.html#_install_for_python) and see if that solves the issue

C++ side

Could you clone https://github.com/pthom/imgui_bundle_template and build it (it is a simple hello world with imgui bundle in C++) and tell me if it works?

# clone
https://github.com/pthom/imgui_bundle_template.git
cd imgui_bundle_template
# build
mkdir -p build && cd build
cmake ..
make -j 4
# Run
./imgui_bundle_example_integration

Follow up

If the problem persists, since I cannot reproduce it on my side, please contact me by email at pthomet at gmail dot com, I'll try to help by email.

[Johndeep]

Thank you for the very fast reply! Thus, I'll do my best to answer the effort.

Python side

1. The following code (pip version) fails with memory access error..

from imgui_bundle import imgui
imgui.create_context()

2. Reboot doesn't work either. Since it was the first thing I did after updating. Sorry for not mentioned it.
3. The installation from source does work. (Btw. did version 1.2.1 added impl.process_event? I wondered that clicking didn't work until I've added that. Is that something ocornut added/separated from process_inputs? Sorry for not being up-to-date)

C++ side

I did all those steps and it also seems to work. To make sure it is the right version, I've added the child windows snippet with the variable resizing. It works as intended.

Preliminary Conclusion

It seems the imgui-bundle version 1.2.1 on pip is currently faulty, since the compiled version directly from the source does work. I don't need to mention that the C++ version also has no problems.

[pthom]

I cannot replicate your issue on Ubuntu22.04 (but I'm using arm64 on a Mac).
I'm afraid the wheel build might be faulty, although I don't know why.

If you have time could you try to install from source at tag v1.2.1:

cd imgui_bundle
git checkout v1.2.1
git submodule update
pip install -v .

If it also works, I'll have to suspect the wheels built by github's CI...

Btw. did version 1.2.1 added impl.process_event? I wondered that clicking didn't work until I've added that. Is that something ocornut added/separated from process_inputs? Sorry for not being up-to-date

I did not understand what you are referring to. Can you clarify?

Child windows snippet with the variable resizing.

Is this a new flag in ImGui?

[Johndeep]

Now I had the chance to test this issue on another notebook, which is a Intel core i5-7200U, which entails an older Intel HD Graphics.

I tested it again with the pip version 1.2.1, but I also had the same issues. Namely the same memory access error as on the other machine..
Of course I also tried to compile everything from source (older CPU, more tea time I guess), which was again the same. I successfully got imgui-bundle to work with the main branch.

I tried what you suggested and planned to build against v1.2.1 by checking out that branch (or tag, whatever). But I could not. Rather, an error just popped up after trying to pip install -v .

  error: subprocess-exited-with-error
  
  × Building wheel for imgui-bundle (pyproject.toml) did not run successfully.
  │ exit code: 1
  ╰─> See above for output.
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  full command: /usr/bin/python3 /usr/lib/python3/dist-packages/pip/_vendor/pep517/in_process/_in_process.py build_wheel /tmp/tmpy0okxmmq
  cwd: [..]/imgui_bundle
  Building wheel for imgui-bundle (pyproject.toml) ... error
  ERROR: Failed building wheel for imgui-bundle
Failed to build imgui-bundle
ERROR: Could not build wheels for imgui-bundle, which is required to install pyproject.toml-based projects

You may be right about the wheels.. and just for the record, I successfully build the entire thing with the main branch.

---

About the impl.process_event. I use the sdl2-backend combination. In sdl2 impl.process_inputs() needs to be called for any inputs (mouse and key) to be recognized by imgui at all. But after updating to v1.2.1, I need to call impl.process_event(*event) instead. But since you're not aware of this, I assume that this is rather a change done in the main imgui branch I guess.

About Child Flags. Yes, this seems to be rather new, since ocornut implemented it just last year in summer it seems. With this, the parameter for imgui.begin_child changes slightly. Previously, one parameter was to determine, if it has a drawn border or not. Now it is integrated into the flags. It can be set by imgui.ChildFlags_.border. The new thing is, that you can also set imgui.ChildFlags_.resize_x for example. This allows you to resize the child directly by dragging the border on the right.

child_border_drag.webm

[pthom]

Now I had the chance to test this issue on another notebook, which is a Intel core i5-7200U, which entails an older Intel HD Graphics. I tested it again with the pip version 1.2.1, but I also had the same issues. Namely the same memory access error as on the other machine..

Hum, this is getting worrying.

I tried what you suggested and planned to build against v1.2.1 by checking out that branch (or tag, whatever). But I could not. Rather, an error just popped up after trying to pip install -v .

If a source build fails, this is also an issue... The log extract you inserted is not long enough and does not contain the root cause of the error. Can you please attach the full log, or a longer extract where the root cause is displayed.

About the impl.process_event. I use the sdl2-backend combination. In sdl2 impl.process_inputs() needs to be called for any inputs (mouse and key) to be recognized by imgui at all. But after updating to v1.2.1, I need to call impl.process_event(*event) instead. But since you're not aware of this, I assume that this is rather a change done in the main imgui branch I guess.

Do you mean this, when using a pure python backend?

---

As far as wheels built by github are concerned, I'd like to make a test: each time a modification is made on the main branch wheels are built by github. They contain an artifact which you can download, for example here, and this artifact contains wheels for all platforms.

Could you tell me if the wheels in this artifact work for you?

[Johndeep]

Hi there again.

I tried to reproduce the installation of the git repo with v1.2.1 again. On the newer machine today, strangely enough, it could compile without problems. So for verification, I plan to do the same with the other machine again later today, since I don't have access to it right now. But since both machine are similar (Intel 64bit etc.), I suspect being sloppy myself yesterday for not cleaning up the build. But that would not explain why, without cleanup, a git checkout to the latest branch in the very same folder worked fine again though.

Sorry for currently not being able to provide the full log from yesterday, since I just copied it from terminal. (Forgot the proper manner of full logs)

I assume there weren't any changes since yesterday with the v1.2.1 tag?

Anyway, after the successful installation (demo says version 1.19.0), imgui-bundle works just fine. There I've wondered if I did something wrong yesterday and tried to install imgui-bundle from pip directly again. But sadly, it still crashes like yesterday.. I could reproduce the error from yesterday..

---

Yes exactly, I mean that pure python backend snippet! Line 40 and 41. It is strange though. I guess line 41 impl.process_inputs() could be obsolete now? I did tested it with the show_demo_window line, but every input works just fine without line 41.

---

About the wheels, I downloaded the example artifact from the link you've provided (#586). And just as I'm writing this line, the compilation has finished. All the example codes works just fine.

[pthom]

...the compilation has finished. All the example codes works just fine.

Actually, I would like you to try without compiling, by using the compiled wheel that corresponds to your python version.
i.e. you could use either:

artifact/imgui_bundle-1.2.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
artifact/imgui_bundle-1.2.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
artifact/imgui_bundle-1.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

On the newer machine today, strangely enough, it could compile without problems

Sometimes, a previous compilation left temp files inside _skbuild, and removing this folder before recompiling sometimes avoids errors.

---

Yes exactly, I mean that pure python backend snippet! Line 40 and 41. It is strange though. I guess line 41 impl.process_inputs() could be obsolete now? I did tested it with the show_demo_window line, but every input works just fine without line 41.

Well, the python backends are almost direct adaptions from pyimgui bindings.

The sdl backend for pyimgui also provide process_inputs, and inside imgui_bundle process_inputs was adapted from it.

I removed the mouse handling from it, since it was done in process_events. However it still has some impact on imgui (set io.display_size, io.display_framebuffer_scale and io.delta_time)

[Johndeep]

Okay, without compilation. No problem.
I still have Ubuntu 22.04 with Python 3.10, I went by cp310.

Now I understand what you want from me. So I went on the journey, went by divide and conquer and downloaded around 11 artifacts in total. I could pin down the problem to be in between artifact 528 and 531. As you can see, 529 and 530 were not build due to errors.

So artifact 528 is the last package, where imgui-bundle actually works (for Ubuntu). From 531 until the latest (which is 586 as of now), I get the memory access error. So I tried to compare the source folders to see what differences happened, and the following might be the cause:

That is what I noticed. You can try to analyze it yourself to what to do about that. I have no knowledge about the automatic builds in github (I should invest some time for that though), but I would be glad, if you figure it out and could explain what the problem was.

[pthom]

Woo, many thanks for this analysis.

So, after analysing the artifacts numbers, the possible commits that cause the error are:

14faa56 update hello_imgui [Pascal Thomet]
([hello_imgui] updated from e8ec91 to 0b5027)

b722ca9 Update ImGui bindings: publish ImGuiInputTextCallback and ImGuiSizeCallback as std::function [Pascal Thomet]

191532f imgui: publish IniFileName and LogFilename as string [Pascal Thomet]

I would suspect 14faa56, but I cannot find the related history in hello_imgui. Something strange...

Anyhow, I will have to publish an update on pypi, since the latest is broken for ubuntu users. Many thanks to you!

As I cannot reliably test this on my side (as it seems), can you please check wait until the latest wheels build is finished (now compiling, and should be ready in 1 hour), and tell me if the binary wheels work for you.

If so, I'll try to prepare a release based on this.

[pthom]

PS: if you have something to show, I'm always pleased to see what users are doing with imgui_bundle in the gallery

[Johndeep]

Okay, just downloaded the newest artifact 587 and tried it out. But unfortunately, the binary wheel didn't work. Same error (memory access error, or if executed from vscodium, no output at all, silent exit so to say). And as always for control, either self compiled or version 1.1.0 do work.

PS: Currently not yet, but if my experimenting phase is done and I've something developed, I could show. Since this is academic work, it might be possible to let you have a glance on it in the future. Anyway, I should be very thankful that you bundled up imgui and some really useful "batteries" like ImPlot or Node-Editor into this for python! Else I would consider to develop in C++ or something.. or not using any GUI at all.

[Johndeep]

I know this little post is not really related to this issue, but the reason for me using imgui-bundle has something to do due to pyimgui not working on Wayland at all. Due to OpenGL.error.Error: Attempt to retrieve context when no valid context error pyimgui #318. None of the suggested workarounds there work for me.

Also, they are a tad behind the updates from ocornut directly (would not be a big deal since I didn't use the new features very much). Their version 2.0.0 is currently based on 1.82. And there are currently no sign of any updates since 9 month ago (as of writing now).

[pthom]

Hello,

Are you available for a session where we could work together on this (we would setup a conf call with screen sharing, so that I can try to diagnose more in details).

I could be available at 4PM paris time today, i.e 3PM UTC.

[pthom]

If not, I'll need to setup an Intel Linux machine somewhere