An efficient matrix-matrix multiplication based antisymmetric tensor contraction engine for general order coupled cluster

Abstract

In this paper, we present an efficient implementation of general tensor contractions, which is part of a new coupled-cluster program. The tensor contractions, used to evaluate the residuals in each coupled-cluster iteration are particularly important for the performance of the program. We developed a generic procedure, which carries out contractions of two tensors irrespective of their explicit structure. It can handle coupled-cluster-type expressions of arbitrary excitation level. To make the contraction efficient without loosing flexibility, we use a three-step procedure. First, the data contained in the tensors are rearranged into matrices, then a matrix-matrix multiplication is performed, and finally the result is backtransformed to a tensor. The current implementation is significantly more efficient than previous ones capable of treating arbitrary high excitations.

An efficient tensor transpose algorithm for multicore CPU, Intel Xeon Phi, and NVidia Tesla GPU

“…A conceptually close, massively parallel RRR framework, based on the block-cyclic distribution of tensors, was reported very recently [21]. Besides, let us mention the C++ tensor algebra code by Hanrath [24], a shared-memory tensor algebra library by Epifanovsky et al [25] (used in Q-Chem [26]), and a shared-memory (accelerator-enabled) tensor algebra library TAL-SH developed by the present author.…”

“…In the context of tensor algebra for electronic structure theory, this problem has been addressed before by Hammond [39] and Hanrath [24]. In particular, Hammond employed an automatic code generator in order to sample the loop optimization space for tensor transposes of relatively low rank.…”

A toolchain for the automatic generation of computer codes for correlated wavefunction calculations

“…An arbitrary order CC program was developed by Kállay, which uses diagrammatic techniques to obtain and factorize the equations, while employing string‐based techniques for the representation of wavefunction parameters . The performance was later significantly improved, as the problem was reformulated using an efficient matrix multiplication scheme . Another level of complexity arises in the explicitly correlated CCSD‐F12 and CCSD‐R12 methods, which allow a much faster convergence towards the basis set limit.…”

Multireference coupled-cluster Ansatz

“…A drawback of the MRexpT method is that, due to the determinantal indexing, the resulting amplitude equations are extraordinarily complex and cannot be derived and solved using standard techniques developed during the last decades by the coupled cluster community. A new, efficient implementation of the general (arbitrary excitation order) coupled cluster code [144] can be expected to help to circumvent this difficulty.…”