This paper presents compiler and run-time support that simplifies the programming of adaptive parallel matrix computing. Matrices are declared with special keywords and can be referred to in high-level matrix operations specifying only their names, e.g., A=B*C, or in statements specifying individual matrix elements. Both types of references are translated into calls to procedures in a library. Procedures that carry out matrix operations are adaptive, currently in two ways: i) in selecting a parallel algorithm based on a cost model that considers various run-time conditions, and ii) in adapting to load imbalance.