PyGB: GraphBLAS DSL in Python with Dynamic Compilation Into Efficient C++

We present PyGB, a high-level Python domainspecific language (DSL) for GraphBLAS building blocks for graph algorithms. GraphBLAS is based on a small number of linear algebra operations, and it is described using mathematical notation and concepts. However, the concrete realizations of GraphBLAS concentrate on efficiency and widely usable programming interfaces. The GraphBLAS C language API standardized by the GraphBLAS Forum is the major example of such a realization, and our GraphBLAS Template Library (GBTL) implemented in the C++ language is another. PyGB exposes the \cpp interface of GBTL through a DSL that leverages Python's mature scientific computing ecosystem and high-level syntax. The syntax of PyGB more closely resembles the GraphBLAS mathematical notation than GBTL. The DSL dispatches to dynamically compiled templated classes to achieve comparable performance to the native GBTL code. We highlight the features of PyGB through code examples and discuss how Python's syntax and dynamic execution enable us to provide the high-level abstraction with minimal performance penalty. We demonstrate the stages of execution, from type inference to template compilation, to dynamic linking and invocation from the Python interpreter. Our experimental evaluation shows that for sufficiently large inputs, the overhead of PyGB is negligible and compilation times are not worse than for native GBTL implementation. Last, we outline concrete features we plan to implement in the future to extend PyGB.