Automated design of domain-specific custom instructions: Geautomatiseerd ontwerp van domeinspecifieke gespecialiseerde instructies
- Autores: Cecilia Noemí González Álvarez
- Directores de la Tesis: Lieven Eeckhout (dir. tes.), Daniel Jiménez González (dir. tes.), Carlos Álvarez Martínez (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2015
- Idioma: inglés
- Tribunal Calificador de la Tesis: Luc Taerwe (presid.), Lieven Eeckhout (secret.), Xavier Martorell Bofill (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
In the last years, hardware specialization has received renewed attention as chips approach a utilization wall. Specialized accelerators can take advantage of underutilized transistors implementing custom hardware that complements the main processor. However, specialization adds complexity to the design process and limits reutilization. Application-Specific Instruction Processors (ASIPs) balance performance and reusability, extending a general-purpose processor with custom instructions (CIs) specific for an application, implemented in Specialized Functional Units (SFUs). Still, time-to-market is a major issue with application-specific designs because, if CIs are not frequently executed, the acceleration benefits will not compensate for the overall design cost. Domain-specific acceleration increases the applicability of ASIPs, as it targets several applications that run on the same hardware. Also, reconfigurable SFUs and the automation of the CIs design can solve the aforementioned problems. In this dissertation, we explore different automated approaches to the design of CIs that extend a baseline processor for domain-specific acceleration to improve both performance and energy efficiency. First, we develop automated techniques to identify code sequences within a domain to create CI candidates. Due to the disparity among coding styles of different programs, it is difficult to find patterns that are represented by a unique CI across applications. Therefore, we propose an analysis at the basic block level that identifies equivalent CIs within and across different programs. We use the Taylor Expansion Diagram (TED) canonical representation to find not only structurally equivalent CIs, but also functionally similar ones, as opposed to the commonly applied directed acyclic graph (DAG) isomorphism detection. We combine both methods into a new Hybrid DAG/TED technique to identify more patterns across applications that map to the same CI. Then, we select a subset of the CI candidates that fits in the available SFU area. Because of the complexity of the problem, we propose four scoring heuristics to reduce the design space and smooth the potential performance speedup across applications. We include these methods in the FuSInG framework, and we estimate performance with hardware models on a set of media benchmarks. Results show that, when limiting core area devoted to specialization, the SFU customization with the largest speedups includes a mix of application and domain-specific custom instructions. If we target larger CIs to obtain higher speedups, reusability across applications becomes critical; without enough equivalences, CIs cannot be generalized for a domain. We aim to share partially common operations among CIs to accelerate more code, especially across basic blocks, and to reduce the hardware area needed for specialization. Hence, we create a new canonical representation across basic blocks, the Merging Diagram, to facilitate similarity detection and improve code coverage. We also introduce clustering-based partial matching to identify partially-similar domain-specific CIs, which generally leads to better performance than application-specific ones. Yet, at small areas, merging two CIs induces a high additional overhead that might penalize energy-efficiency. Thus, we also detect fragments of CIs and we join them with the existing merged clusters resulting in minimal extra overhead. Also, using speedup as the deciding factor for CI selection may not be optimal for devices with limited power budgets. For that reason, we propose a linear programming-based selection that balances performance and energy consumption. We implement these techniques in the MInGLE framework and evaluate them with media benchmarks. The selected CIs significantly improve the energy-delay product and performance, demonstrating that we can accelerate a domain covering more code while reducing the needed area for the CI implementation.
