Adding Single-Floating Point Precision #7
Description
I believe most of the cases of post-processing (e.g. QTAIM, Hirshfeld, electrostatics) doesn't require double precision (7 versus 15-16 decimal places) the way that Quantum chemistry codes require for the computation of the energy.
This would provide much greater speed-ups due to the less register usage per thread (max 255 32-bit registers), and higher FLOPS of 32- bit, e.g. TITAN V has 14.9 TFLOPS (FP32) versus 7.450 TFLOPS (FP64). I'll guess for the largest peptide (2390 basis-funcs, 9 amino-acids) this would bring atleast >2x speed-ups.
It would be tedious to implement but not hard at all, as you would use templates template<typename T> and use the correct function func<float>, func<double> according to the passed NumPy array in pymolecule.cu. It may be very difficult using Eigen to expose different functions, see https://github.com/pybind/pybind11/issues/3085 recommends using lambda functions or see https://github.com/pybind/pybind11/issues/1324.
Also see the paper
"P. Pokhilko, E. Epifanovsky, and A. I. Krylov, “Double precision is not needed
for many-body calculations: Emergent conventional wisdom,” J. Chem. Theory
Comput. 14, 4088–4096 (2018)"