[Feature] Pipeline the computing tasks

[timerring]

Is your feature request related to a problem? Please describe.

Although the process is very fast, but there is still much that for improvement.
The tasks can be divided into computing tasks and IO tasks.

• The computing tasks contain the burning and whisper automatic speech recognition.
• The IO tasks is the upload process. The IO task's limit is the bandwidth of the networks.

Describe the solution you'd like

Now the burning process will occupy the VRAM around 180MB, nevertheless the whisper asr will occupy the 2600MB VRAM. The VRAM of GTX1650 is around 3911MB. So in order to make full use of computing resources as well as guaranteeing the upload in time. We can burn danmakus and submit the segments first, and merge the segments and start asr until there are no danmaku burning tasks.

Describe alternatives you've considered

Starting the asr process at a specific time is also alternative.

[timerring]

Although there is still some VRAM left, but the limit of burning process is 3. If the process exceeds the limit, it will report OpenEncodeSessionEx failed: out of memory (10).

[timerring]

[timerring]

For server-type graphics cards (such as Tesla T4, NVIDIA A100, etc.), there is no limit on the number of concurrent encoding sessions for NVIDIA graphics cards. For consumer-type graphics cards (such as GeForce RTX 3070, GeForce RTX 3080, etc.), there is a limit on the number of concurrent encoding sessions for NVIDIA graphics cards. This limit is not for the GPU but for the entire system. Even if you insert two, four, or even eight cards, at most only three encoding sessions can be supported.
Reference: https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new#Encoder

[timerring]

In the https://www.codelive.cn/index.php/archives/84/, the author has the same gpu with mine. And he uses reverse modify bytecode to crack that restriction. However, the total VRAM in GTX 1650 is 3910 MB, if I follow the instruction to crack, I will only be able to add two more tasks, which won't promote too much. Hence I will optimize the current workflow under the restriction.

[timerring]

So I will try to pipeline the computing task first, and then optimize related issues.

[github-actions]

This issue has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.