Evaluation of GPU/CPU co-processing models for JPEG 2000 packetization

With the bottom-line goal of increasing the throughput of a GPU-accelerated JPEG 2000 encoder, this paper evaluates whether the post-compression rate control and packetization routines should be carried out on the CPU or on the GPU. Three co-processing models that differ in how the workload is split among the CPU and GPU are introduced. Both routines are discussed and algorithms for executing them in parallel are presented. Experimental results for compressing a detail-rich UHD sequence to 4 bits/sample indicate speed-ups of 200× for the rate control and 100× for the packetization compared to the single-threaded implementation in the commercial Kakadu library. These two routines executed on the CPU take 4× as long as all remaining coding steps on the GPU and therefore present a bottleneck. Even if the CPU bottleneck could be avoided with multi-threading, it is still beneficial to execute all coding steps on the GPU as this minimizes the required device-to-host transfer and thereby speeds up the critical path from 17.2 fps to 19.5 fps for 4 bits/sample and to 22.4 fps for 0.16 bits/sample.