Assessment of a semi integral-direct local multi-reference configuration interaction implementation employing shared-memory parallelization

We present different integral-direct implementations in a local Cholesky-decomposed multi-reference configuration interaction framework. We discuss their performance, parallel efficiency and scaling properties for a range of alkyne test systems. As we are able to introduce accuracy-preserving integral truncations within a direct algorithm, we observe superior performance and, through the drastically reduced I/O operations, better parallel efficiency for the truncated integral-direct kernels compared to their conventional counterparts.