Gristmill is a package based on the drudge algebra system for automatic optimization and code generation of tensor computations. In spite of being designed for applications in quantum chemistry and many-body theory, gristmill can be used for any scientific computing problem with dependency on tensor computations.
The optimizer utilizes novel advanced algorithm to efficiently parenthesize and factorize tensor computations for less arithmetic cost. For instance, a matrix chain product
\mathbf{R} = \mathbf{A} \mathbf{B} \mathbf{C}
can be parenthesized into
\mathbf{R} = \left( \mathbf{A} \mathbf{B} \right) \mathbf{C}
or
\mathbf{R} = \mathbf{A} \left( \mathbf{B} \mathbf{C} \right)
depending on which one of them incurs less arithmetic cost for the given shapes of the matrices. With just a small overhead relative to specialized dynamic programming code for matrix chain products, general tensor contractions are supported. For instance, the ladder term in the CCD theory in quantum chemistry
r_{abij} = \sum_{c,d=1}^v \sum_{k,l=1}^o v_{klcd} t_{cdij} t_{abkl}
can be automatically parenthesized into a two-step evaluation
\begin{aligned}
p_{klij} &= \sum_{c,d=1}^v v_{klcd} t_{cdij}\\
r_{abij} &= \sum_{k,l=1}^o p_{klij} t_{abkl}\\
\end{aligned}
Because of the efficiency of the algorithm, contraction of even twenty factors can be handled well.
When computing sums of multiple contractions, factorizations of some or all terms leading to savings of arithmetic cost can also be automatically found. For instance, the correlation energy of the CCSD theory in quantum chemistry,
e = \frac{1}{4} \sum_{i,j=1}^o \sum_{a,b=1}^{v} u_{ijab} t^{(2)}_{abij}
+ \frac{1}{2} \sum_{i,j=1}^o \sum_{a,b=1}^v u_{ijab} t^{(1)}_{ai} t^{(1)}_{bj}
can be automatically rewritten into
e = \frac{1}{4} \sum_{i,j=1}^o \sum_{a,v=1}^v u_{ijab} \left(
t^{(2)}_{abij} + 2 t^{(1)}_{ai} t^{(1)}_{bj}
\right)
which takes less arithmetic cost.
In addition to parenthesization and factorization, gristmill also has additional optimization heuristics like common symmetrization optimization. The same intermediates can also be guaranteed to be computed only once by the canonicalization power of drudge.
The code generator is a component orthogonal to the optimizer. Both optimized and unoptimized computation can be given for naive Fortran or C code (with optional OpenMP parallelization), or Python code using NumPy or TensorFlow libraries.
Gristmill is developed by Jinmo Zhao and Prof Gustavo E Scuseria at Rice University, and was supported as part of the Center for the Computational Design of Functional Layered Materials, an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences under Award DE-SC0012575.