A Domain Specific Language for Computation-in-Memory

The memory-wall problem is a big challenge that classical Von Neumann-based computer systems face. This problem is caused by the increasing performance gap between the main memory and the central processing unit. The novel Computation-in-Memory (CiM) architecture can solve the memory wall problem by using emerging memristor technology to perform both data storage and computation. The CiM architecture’s distinguishing component is a large two-dimensional crossbar on which large amounts of data can be stored, and on which many operations can be executed in parallel. When used as an accelerator to execute massively parallel applications, the CiM architecture can provide significant performance improvements over classical computer systems. However, the implementation of applications on the crossbar of the CiM architecture requires new algorithm specification and compilation techniques, as spatial information about a program is required to obtain an optimal implementation of that program. In this thesis, we introduce a domain-specific language to describe computational patterns and their spatial characteristics. We also design a compiler that is able to transform these patterns into the crossbar implementations of those patterns. We verify both the functional and spatial correctness of the systems generated for the crossbar by describing several distinctly different algorithms in the language, and comparing the compilation results to our expectations. This way, we both show the ability of our language to describe a wide array of applications, and we show that the compiler of the language works correctly.Electrical Engineering, Mathematics and Computer ScienceQuantum EngineeringComputer EngineeringCE-MS-2017-0