| Description |
Sparse matrix codes are found in numerous applications ranging from iterative numerical solvers to graph analytics. Achieving high performance on these codes has however been a significant challenge, mainly due to array access indirection, for example, of the form A[B[i]]. Indirect accesses make precise dependence analysis impossible at compile-time, and hence prevent many parallelizing and locality optimizing transformations from being applied. The expert user relies on manually written libraries to tailor the sparse code and data representations best suited to the target architecture from a general sparse matrix representation. However libraries have limited composability, address very specific optimization strategies, and have to be rewritten as new architectures emerge. In this dissertation, we explore the use of the inspector/executor methodology to accomplish the code and data transformations to tailor high performance sparse matrix representations. We devise and embed abstractions for such inspector/executor transformations within a compiler framework so that they can be composed with a rich set of existing polyhedral compiler transformations to derive complex transformation sequences for high performance. We demonstrate the automatic generation of inspector/executor code, which orchestrates code and data transformations to derive high performance representations for the Sparse Matrix Vector Multiply kernel in particular. We also show how the same transformations may be integrated into sparse matrix and graph applications such as Sparse Matrix Matrix Multiply and Stochastic Gradient Descent, respectively. The specific constraints of these applications, such as problem size and dependence structure, necessitate unique sparse matrix representations that can be realized using our transformations. Computations such as Gauss Seidel, with loop carried dependences at the outer most loop necessitate different strategies for high performance. Specifically, we organize the computation into level sets or wavefronts of irregular size, such that iterations of a wavefront may be scheduled in parallel but different wavefronts have to be synchronized. We demonstrate automatic code generation of high performance inspectors that do explicit dependence testing and level set construction at runtime, as well as high performance executors, which are the actual parallelized computations. For the above sparse matrix applications, we automatically generate inspector/executor code comparable in performance to manually tuned libraries. |
