md-kernel, EPI Co-design on RISC-V platforms

In this experiment, we summarize the conclusions derived from the study conducted on the target RISC-V platforms. For additional details, the reader is referred to the following document:

Leonel Sousa, et al. : “D4.4 - Second report on codesign”, Version 1.0, Deliverable of the Performance Optimization and Productivity 3 (POP3) Centre of Excellence. December, 2025.

The initial version of the code used an OpenMP critical section in the array building loop of the code. This caused bad parallel efficiency, especially with the data structures used for the molecules. A new implementation using a second pass over the data structures, and a removal of the OpenMP critical section resulted in good scaling.

It was found that compiler auto vectorization was not sufficient to generate vector instructions during the main force computationally loop on any of our hardware test benches. Using the EVE SIMD library, this was achieveable with a significant speedup in computation. It was also possible to achieve vectorization using certain pragmas and by rewritting the force write loop to an explicit reduction.