GPU Hardware Transcoding

How to use NVIDIA GPU to accelerate video transcoding

NVIDIA GPUs have one or more hardware-based decoder and encoders which provide fully-accelerated hardware-based video decoding and encoding for several popular codecs.

Nowadays, many video processings rely on deep learning methods. The deep learning models usually run on the NVIDIA GPUs and libraries developed by NVIDIA. So, transferring the decoding and encoding from CPU to GPU can obtain benefit in such cases. One obvious benefit is that we can reduce the copy overhead between CPU and GPU.

In the gpu_transcode folder, we provide various examples to show how to use GPU decoding and encoding as well as how to combine the FFmpeg CUDA filters in the BMF.

The examples are listed below. We will explain them in detail.

• Decode videos
• Decode videos using multiple threads
• Encode videos
• Encode videos using multiple threads
• Transcode
• Transcode 1 to n
• Transcode using multiple threads

Decode

In the BMF, enabling GPU decoding is really simple. What you need to do is to add "hwaccel": "cuda" into the "video_params".

You should note that if you use GPU to decode videos, the decoded frames are in the GPU memory. So if you want to manipulate them at the cpu side, don’t forget to copy these frames into cpu memory. In the BMF, you can set GPU decoding followed by a cpu_gpu_trans_module or followed by a hwdownload filter.

See more details in the test_gpu_decode().

Encode

In the BMF, you can add "codec": "h264_nvenc" or "codec": "hevc_nvenc" in the encode module’s video_params to enable GPU encoding. If the inputs of the encoder are in the GPU memory, you should add "pix_fmt": "cuda" to the video_params.

See more details in the test_gpu_encode() and test_gpu_transcode().

Transcode

For GPU transcoding, you should combine the GPU encoding and GPU encoding mentioned before. Since all the intermediate data are in the GPU memory, we don’t need to consider extra copying any more.

See more details in the test_gpu_transcode().

test_gpu_transcode_1_to_n() shows that BMF can transcode one video to several videos at the same time. Just add more encode modules with different parameters after the same decode module.

Multiple threads and multiple processes

Some GPUs may have more than one hardware-based decoders and encoders. In order to fully utilize these hardwares, we have to start as many instances as possible. BMF can launch these instances through python multi-threading and multi-processing. You can see the examples in the test_gpu_decode_multi_thread_perf, test_gpu_encode_multi_thread_perf and test_gpu_transcode_multi_thread_perf.

For multi-processing, there’s one special thing we should notice so we use a separate script test_gpu_decode_multi_processes.py to show how to do it in the BMF.

We should import bmf in the task function rather than at the beginning of the script file.