Bug: Requesting hardware-accelerated video encoders crashes GitHub CI

[Vanilagy]

Describe the Bug

I noticed that querying the system for hardware-accelerated video encoder support crashes inside of GitHub Action's ubuntu-latest environment.

Here's the most minimal reproduction I was able to come up with:
https://github.com/Vanilagy/node-av-hardware-ci-crash

The relevant code is this:

import { AV_CODEC_ID_H264, HardwareContext } from 'node-av';

const hw = HardwareContext.auto();
const codec = hw.getEncoderCodec(AV_CODEC_ID_H264);

console.log(codec.name);

Which, before reaching the log statement, crashes with:

[CUDA @ 0x13f2c7c0] Cannot load libcuda.so.1
[CUDA @ 0x13f2c7c0] Could not dynamically load CUDA
implib-gen: libva-drm.so.2: failed to load library 'libva-drm.so.2' via dlopen: libva-drm.so.2: cannot open shared object file: No such file or directory
node: libva-drm.so.2.init.c:122: load_library: Assertion `0 && "Assertion in generated code"' failed.

I was able to circumvent this error in webcodecs-polyfill with this botched-together check:

const canUseHwAccel = () => {
	if (os.platform() !== 'linux') {
		return true; // Assume Windows/macOS work fine
	}

	// node-av downright crashes (with exit code 0, lol) if these aren't there, so we check ourselves
	const paths = [
		'/usr/lib/x86_64-linux-gnu/libva-drm.so.2',
		'/usr/lib/libva-drm.so.2',
		'/usr/lib/x86_64-linux-gnu/libcuda.so.1',
		'/usr/lib/libcuda.so.1',
	];
	return paths.some(p => fs.existsSync(p));
};

But, would be of course cool if it worked without having to do that.

To Reproduce

See above

Expected Behavior

If hardware encoders are not available, which I would assume for the default runner, then it should just not return a codec or return a software one.

Code Sample

See above

Relevant Log Output / Error Messages

See above

Node.js Version

v22

NodeAV Version

5.0.2

Operating System

Ubuntu 24.04

Installation Method

npm install

API Used

Low-Level API

Hardware Acceleration

Other

Additional Context

No response

[seydx]

Thanks for reporting this! I was already aware of this problem but hadn't analyzed the root cause yet - that's why I implemented the CI skips as a workaround in my tests.

The crash happens because my FFmpeg build (a fork of jellyfin-ffmpeg) uses Implib.so to create portable Linux binaries. These "implib stubs" try to dlopen() the actual libraries (libcuda.so, libva-drm.so, etc.) at runtime. When the library doesn't exist, the generated stub code calls assert(0) which crashes the process:

implib-gen: libva-drm.so.2: failed to load library 'libva-drm.so.2' via dlopen
node: libva-drm.so.2.init.c:122: load_library: Assertion `0 && "Assertion in generated code"' failed.

This is not a JavaScript error that can be caught - it's a hard process abort at the C level.

I'm currently testing a fix in my FFmpeg fork. The implib-gen tool supports a --dlopen-callback option that allows custom error handling. I've added a graceful callback that returns NULL instead of crashing:

void *implib_dlopen_callback(const char *lib_name) {
    void *handle = dlopen(lib_name, RTLD_LAZY | RTLD_LOCAL);
    if (!handle) {
        return NULL;
    }
    return handle;
}

This should allow FFmpeg to handle the missing library as a normal error and return null from HardwareContext.auto().

Building new Docker images with the fix now. Will report back once tested.

[seydx]

The issue was more complicated than expected... But it has now been fixed and should work without any issues.

import { HardwareContext } from '../src/index.js';

console.log('Testing hardware detection...');

using hw = HardwareContext.auto();
if (hw) {
  console.log('Hardware found:', hw.deviceTypeName);
} else {
  console.log('No hardware available');
}

Testing hardware detection...                                                                                                                                                                       
[CUDA @ 0x168378c0] Cannot load libcuda.so.1                                                                                                                                                        
[CUDA @ 0x168378c0] Could not dynamically load CUDA                                                                                                                                                 
[VAAPI @ 0x166cf3c0] Cannot load libva-drm.so.2                                                                                                                                                     
[VAAPI @ 0x166cf3c0] No VA display found for any default device.                                                                                                                                    
[VAAPI @ 0x168378c0] No VA display found for device /dev/dri/renderD128.                                                                                                                            
[Vulkan @ 0x168378c0] Unable to find device "/dev/dri/renderD128"!                                                                                                                                  
[OpenCL @ 0x168378c0] Invalid OpenCL platform/device index specification "/dev/dri/renderD128".                                                                                                     
[AMF @ 0x168378c0] DLL libamfrt64.so.1 failed to open                                                                                                                                               
No hardware available                                                                                                                                                       

https://github.com/seydx/node-av/releases/tag/v5.1.0