Optimized async GPU streams usage in Video post-treatment

edit

Thanks for your answer JasonInVegas !

So what I understood is :

1. DECODING INPUT

   • NVCUVID (ie NVIDIA VIDEO CODEC SDK) can be used to decode video files, but it depends on video formats and hardware compatibility,
   • other solution is to use ffmpeg from OpenCV, which works great, but a bit slow on CPU of course...

2. OpenCV + CUDA treatment :

   • To declare and use several streams on the GPU : cv::gpu::Stream stream[ n ] but only if "compute capability " of hardware is greater than 2.0 ? as this thread suggests : http://opencv-users.1802565.n2.nabble.com/Parallelling-CUDA-devices-td7580560.html you say : "most likely you will want to parameterize the stream counts (rather than hard-coding '12')", but on what criterion ?

What function could be used to determine if hardware is compatible with multiple streams and how much for n ?

• do you confirm the CPU could be released after each call to a new stream ?

  • is it something related with the use of : stream.enqueueHostCallback( receive_data_from_GPU, Data_to_GPU) ?
  • and after that call CPU is awaken by receive_data_from_GPU( cv::gpu::Stream&, int status, void* Data_from_GPU ) function ? The difficulty for me at this stage is I own an ATI card only : may I choose a >= 2.0 compatibility card, or 1.3 is enough ?

• Another point : is it possible to declare a memory space on GPU, that would be directly accessible by any of the "n" streams for a "second level" of the algorithm ?

ENCODING OUTPUT VIDEO

• NVIDIA VIDEO CODEC SDK with several solution to encode with GPU, compatible on Win and Linux, but not Mac actually ?

• other solution is to use ffmpeg from OpenCV, which works great, but a bit slow on CPU of course... But if all my treatment is on GPU, it could be aceptable, I'll have to test that.

Tx again !

regarding item #2: my opinion is "yes" based on the comments in the referenced opencv forum post, it appears compute level 2.0 is required in order to achieve the stream-based parallelism.

So, which nVidia cards support compute level 2.0? (or restated: What is the least expensive Fermi architecture board?) Summary page: https://developer.nvidia.com/cuda-gpus

Everything on this page will support at least compute level 2.0 (most are 2.1 and above) http://www.nvidia.com/object/quadro-desktop-gpus-specs.html

These range from $150 USD to way too much...a Quadro k600 is a $175 USD board that supports compute level 3.0

Now, about parameterizing the stream buffers based on video specs.

I'm not exactly confident which GPU memory structure(s) to use, but the GPU memory allocation is based on video frame Width x Height x Color_depth for each frame packed into an image struct like a pixel frame buffer object (FBO) in shaders....if you extract the frame(s) on CPU: you pack them, then pass them into the allocated GPU FBO memory object(s). So the frame extraction time(s) and GPU execution times ideally balance so the CPU doesn't bottleneck nor rest. The allocated amount of GPU memory must support input frames and output frame results.

My original comment was: "Why allocate 12 streams, if 6 frames 24-bit color uncompressed, HD video packed into FBOs on the GPU will consume the GPU memory?"

I believe there's a common practice of recycling the GPU algorithm input memory objects and filling them with output results so you don't have to allocate double memory....this won't work if the GPU is performing the video frame decode ... naturally...because it needs the memory for the next frame.

Last item: I believe the Cuda perf tools include a call signature (even example code) to test for GPU compute level capabilities. So that code is out-of-the-box functionality on the CUDA side.

Whew...hope this helps (and doesn't confuse things more!)

Good luck!

P.S. please reply with the use-case you're constructing!