CUDA-VMAF Memory Leak（preallocation method） in libvmaf

[Yangel1993]

I am calculating VMAF scores every 30 frames, using the CUDA-accelerated version of VMAF. Currently, the computed VMAF values are accurate each time; however, there is a significant and rapid increase in GPU memory usage. I'm wondering if there might be an issue with proper resource deallocation.

My code is based on the example found in libvmaf/test directory, specifically the test_cuda_picture_preallocation_method_device test case, where I ensure that I initialize (init) and close (close) the VMAF context appropriately for every 30 frames. Despite this, the GPU memory keeps rising steadily.

Could there be a problem with how resources are being released, or are there any suggestions to address this potential memory leak issue?

[kylophone]

The VmafPictures are reference-counted and should be automatically freed when the ref count drops to zero. The framework should take care of when using the preallocation method. Curious if you were able to solve this issue or are still dealing with it?

[Yangel1993]

The VmafPictures are reference-counted and should be automatically freed when the ref count drops to zero. The framework should take care of when using the preallocation method. Curious if you were able to solve this issue or are still dealing with it?

The issue has not been resolved; I encountered this problem with other methods involving preallocation as well. Ultimately, I switched to a non-preallocating approach using vmaf_picture_alloc, and the functionality without preallocation works as expected.

[gedoensmax]

Allocating on demand is not ideal as an allocation will force a sync on the CUDA device. @MorkTheOrk you are currently working with VMAF could you see if memory is leaked when using FFmpeg ? https://github.com/FFmpeg/FFmpeg/blob/e05d3c1a16c96fee59bf00e23f9d67a79b47220f/libavfilter/vf_libvmaf.c#L149-L166