A Parallel H.264 Encoder with CUDA: Mapping and Evaluation

4kUHD H264 Wireless Live Video Streaming Using CUDA

“…Furthermore their implementation allows for a latency bottleneck for critical video applications such as live streaming and real-time encoding/decoding processes as data must be transferred from and to the CPU memory before it can be accessed by other processes. We therefore extend the work of [26] to support zero-copy memory and also implement it as a directshow filter to enable it to accept a feed from live capture. The modules in [26] are Interprediction, intra prediction, entropy encoding, and deblocking filter.…”

“…We therefore extend the work of [26] to support zero-copy memory and also implement it as a directshow filter to enable it to accept a feed from live capture. The modules in [26] are Interprediction, intra prediction, entropy encoding, and deblocking filter. We also work on the interprediction module where we implement dynamic parallelism for motion estimation and on the intraprediction module where we reduce the predictions; however, every other module within their implementation is available to perform our experiments.…”

“…This function is called when the stream is being executed. It initializes the encoder using the API already provided by the work of [26]. After the default settings are fetched, some changes are made to customize the encoding process according to user requirements (e.g., bitrate, fps, and so forth).…”

“…The authors used a raw frame as reference, though this increased data granularity, the resulting image quality was poor; for a resolution of 352 × 288 the performance achieved was 1.03 fps. In [26] the authors do a port of four major processes of H264 encoding to CUDA. They use data localization to organize data and threads to work efficiently on the GPU.…”

“…Intraframe encoding algorithm was presented in [29] which is GPUbased and does 4 × 4 block process reordering in a diagonal order using openCL (a platform dependent GPU application development environment), however, this is only one way of reordering the block for parallelism. In [26], all encoding processes are parallelized on a GPU using CUDA; however, the parallel encoder is aimed at high definition video and also for video storage as their implementation does not take into consideration real-time encoding for live video streaming due to the huge amount of memory transfer latency.…”