Software-Controlled Scratchpad Mapping Strategies for Wavelet-Based Applications

Exploiting Varying Resource Requirements in Wavelet-based Applications in Dynamic Execution Environments

“…This phenomenon can be exploited when dynamically varying Level 1 memory occurs, by switching between these localizations at run-time to achieve the lowest possible miss rate. This article is a significant extension of previous work [6] focusing on various algorithmic WT parameters, which can be used to derive mapping guidelines indicating at run-time which WT localization offers the best performance. Moreover, the mapping guidelines are extendable to general WT-like algorithms, such as hierarchical filterbanks and iterative upand downsampling as in Scalable Video Coding (SVC) standardized in MPEG [7].…”

“…3. This makes automated optimization strategies difficult to apply to wavelet-based applications, as was shown for the automated MH mapping tool MH [6]. …”

Study and implementation of wavelet-based video coding for dynamic transition environments

“…While it is feasible to remove these explicit non-linear loop and index expressions through code rewriting techniques, in this case by unrolling the WT level loop, this does not change the fact that the corresponding dependencies remain implicitly present in the code, since a level-k sample will still depend on an exponentially increasing amount of samples from each lower WT level. This makes automated optimization strategies difficult to apply to wavelet-based applications, because they cannot suitably handle these non-linear dependencies and because the code unrolling techniques increase the complexity of the internal representation used in these strategies, making it difficult to find a good solution, as was shown for the automated memory hierarchy mapping tool MH [Geelen 2006] and for the automated memory compaction tool [Ferentinos 2003. Each level of the FWT consists of a lowpass (L) and a highpass (H) filtering followed by a subsampling operation.…”

“…• Problem formulation and the proposed switching principle [Ferentinos 2003, Geelen 2006]: The difficulty of automated tools to find optimal data mappings, for applications with non-linear dependencies like the Wavelet Transform, has been demonstrated and initially solved by subdividing (localizing) the basic processing data-block (Chapter 3). By generalizing the initial approach, we have proposed a run-time switching principle between different implementation mechanisms (localizations) of wavelet-based applications.…”

“…In practice a fully manual derivation of the mapping rapidly becomes unacceptable due to the complexity of the (manifest) code, so systematic methods and partial tool support are essential. As was the case for the global loop transformation step, these tools do exist [Steinke 2002, Kandemir 2004], but unfortunately, as was also the case for the loop transformation automation, they cannot handle the loop structures and the non-affine and of high complexity indexing expressions, present in the WT [Geelen 2006]. Nevertheless, the scratchpad's energy efficiency is essential for embedded devices.…”