Lossless Image Compression Exploiting Streaming Model for Efficient Execution on Multicores

摘要

Image and video coding play a critical role in present multimedia systems ranging from entertainment to specialized applications such as telemedicine. Usually, they are hand-customized for every intended architecture in order to meet performance requirements. This approach is neither portable nor scalable. With the advent of multicores new challenges emerged for programmers related to both efficient utilization of additional resources and scalable performance. For image and video processing applications, streaming model of computation showed to be effective in tackling these challenges. In this paper, we report the efforts to improve the execution performance of the CBPC, our compute intensive lossless image compression algorithm described in [1]. The algorithm is based on highly adaptive and predictive modeling, outperforming many other methods in compression efficiency, although with increased complexity. We employ a high-level performance optimization approach which exploits streaming model for scalability and portability. We obtain this by detecting computationally demanding parts of the algorithm and implementing them in Streamlt, an architecture-independent stream language which goal is to improve programming productivity and parallelization efficiency by exposing the parallelism and communication pattern. We developed an interface that enables the integration and hosting of streaming kernels into the host application developed in general-purpose language.%Postupci obrade slikovnih podataka su iznimno zastupljeni u postoje?im multimedijskim sustavima, po?ev od zabavnih sustava pa do specijaliziranih aplikacija u telemedicini. Vrlo ?esto, zbog svojih ra?unskih zahtjeva, ovi programski odsje?ci su iznimno optimirani i to na niskoj razini, ?to predstavlja pote?ko?e u prenosivosti i skal-abilnosti kona?nog rje?enja. Nadolaskom vi?ejezgrenih ra?unala pojavljuju se novi izazovi kao ?to su u?inkovito iskori?tavanje ra?unskih jezgri i postizanje skalabilnosti rje?enja obzirom na pove?anje broja jezgri. U ovom radu prikazan je novi pristup pobolj?anja izvedbenih performansi metode za kompresiju slika bez gubitaka CBPC koja se odlikuje adaptivnim modelom predvi?anja koji omogu?uje postizanje boljih stupnjeva kompresije uz pove?anje ra?unske slo?enosti [1]. Pristup koji je primjenjen sastoji se u implementaciji ra?unski zahtjevnog predikcijskog modela u tokovnom programskom jeziku koji omogu?uje paralelizaciju izvornog programa. Ovako projektiran predikcijski model mo?e se iskoristiti kroz su?elje koje smo razvili a koje omogu?uje pozivanje tokovnih ra?unskih modula i njihovo paralelno izvo?enje uz iskori?tavanje vi?e jezgri.