Automatic selection of application-specific instruction-set extensions

Abstract

In this paper, we present a general and an efficient algorithm for automatic selection of new application-specific instructions under hardware resources constraints. The instruction selection is formulated as an ILP problem and efficient solvers can be used for finding the optimal solution. An important feature of our algorithm is that it is not restricted to basic-block level nor does it impose any limitation on the number of the newly added instructions or on the number of the inputs/outputs of these instructions. The presented results show that a significant overall application speedup is achieved even for large kernels (for ADPCM decoder the speedup ranges from x1.2 to x3.7) and that our algorithm compares well with other state-of-art algorithms for automatic instruction set extensions.

The Instruction-Set Extension Problem: A Survey

“…As the name suggests, complete customization involves the whole instruction-set which is tuned towards the requirements of an application [2,3,4,5], while partial customization involves the extension of an existing instruction-set by means of a limited number of instructions [6,7,8,9,10,11,12,13,14]. In both cases the goal is to design an Instruction-Set containing the most important operations needed by the application to maximize the performance.…”

“…A collection of templates constitutes a library of templates. Many authors assume the existence of templates which are given as an input and which are identified inside the subject graph [33,6,31], however this is not always the case and many authors develop their own templates [16,17,34,7,10,35,14,43,59].…”

“…the granularity at which code is considered: fine-grained and coarse-grained. The first one works at the operation level and implements small clusters of operations in HW [7,10,12,13,14,43,59,44]; the second one operates at the loop or procedure level and identifies critical loops or procedures in the application, and displaces them from SW to HW as a whole [16,17,18,19,20,21]. The main differences are in terms of speedups and flexibility: although a coarse-grained approach could produce a large speedup, its flexibility is limited, given that this approach is often performed on a per application basis and it is difficult that other applications have the same loop or procedure as critical part.…”

“…The processor is then manually tailored to include the new capabilities. Although human ingenuity in manual creation of custom capabilities creates high quality results, performance and time-to-market requirements as well as the growing complexity of the design space, can benefit from an automatic design flow for the use of these new capabilities [28,29,12,30,31,32,13,14,60]. Moreover the selection of multiple custom instructions from a large set of candidates involves complex tradeoff and can be difficult to be performed manually.…”

“…Graph theory has became the dominant approach and seems to provide the right analytical framework. In this context the code of the application is represented with a directed graph, called the subject graph, and the intensive segments of the code to map on HW are subgraphs of the subject graph [12,14]. Depending on the level of abstraction on which we operate, nodes represent basic operations as well as entire procedures, functions or loops; edges represent data dependencies.…”