Our search-based loop fusion optimization has been successful in the Tensor Contraction Engine for minimizing storage or, together with a search-based tiling optimization, for automatically generating efficient out-of-core code from tensor contraction expressions.  In this talk, we present a generalization of the loop fusion algorithm to handwritten code as a first step towards generating GPU or out-of-core code from handwritten in-core code.  The loop fusion algorithm optimizes dense array computations by fusing producer and consumer loops to minimize the storage requirements for intermediate arrays.  It undoes any fusion provided by the programmer and, using a cost model for memory minimization, searches all loop structures for the memory-minimal code.  Then the fused loops are expanded to tile size and these tiles are used as units of data transfers between different levels in the memory hierarchy.  Finally, the innermost loop nests are replaced by index permutation and BLAS library calls. We present extensive measurements that this optimization approach for handwritten tensor contraction code is superior to polyhedral model optimization for generating single-core, multi-core, and GPU code.

Optimizing handwritten tensor contraction code: Our experience

248th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.