Repository navigation
This test is silly. All it does is fill up the VRAM until it's completely full, and then fails. #802
Description
Activity
You might want to try different version and/or provider of ffmpeg.
I ran into the same kind of issue where my RAM would just fill up to a certain point and the test failed.
This did not occur when I used the ffmpeg from directory of Shotcut, OBS Studio or ffmpeg from gyan.devSame here on Manjaro Linux:
$ uname -a Linux foo 6.6.30-2-MANJARO #1 SMP PREEMPT_DYNAMIC Wed May 8 17:46:43 UTC 2024 x86_64 GNU/Linux $ ffmpeg -version ffmpeg version n6.1.1 Copyright (c) 2000-2023 the FFmpeg developers built with gcc 13.2.1 (GCC) 20230801 [...] $ nvidia-smi Tue Jun 11 18:17:18 2024 +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 550.78 Driver Version: 550.78 CUDA Version: 12.4 | [...] | 0 NVIDIA GeForce GTX 1050 Off | 00000000:01:00.0 Off | N/A |Same problem for me
Ubuntu 24.04
NVIDIA-SMI 555.58.02
Driver Version: 555.58.02
CUDA Version: 12.5
ffmpeg version 6.1.1-3ubuntu5 Copyright (c) 2000-2023 the FFmpeg developers
built with gcc 13 (Ubuntu 13.2.0-23ubuntu3)CUDA_ERROR_OUT_OF_MEMORYmeans CUDA couldn't allocate. That's not a session-limit failure. So the test result tells you nothing about whether the patch works.The tester fires off multiple simultaneous sessions, but it can't tell the difference between hitting the session limit, running out of GPU memory, and some other encoding failure. Those all look the same from the output.
What it should do: use less GPU memory, shorter and lower-resolution workloads, confirm every requested session actually initializes and finishes, separate NVENC session-limit errors from CUDA allocation failures, and print PASS/FAIL/INCONCLUSIVE instead of making you read FFmpeg output.
Related: #791, another limitation in the same procedure.
Keeping this open. Contributions and reproducible results welcome.
What a stupid test. Just fills up the VRAM until there's none left. What is the point of this?