Use new particle data layout in the P3M code - #5137
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
With the new P3M shared memory implementation, Coulomb P3M is 17% faster and Dipolar P3M is 7% faster with 2000 particles per thread. The MPI implementation is also faster now. Weak scaling benchmarks will run over the week-end. |
Use shared memory in P3M/DP3M charge/dipole/force/torque assignment and real-space force calculation in the short-range loop.
|
The benchmarks are inconclusive due to the high variance. Running the P3M benchmark with 4 threads multiple times yields two different runtimes reliably: 3.7ms/step and 4.6ms/step. Since the difference is significant, one cannot take the mean of multiple runs. Taking the minimum (or first quartile) of multiple runs requires a larger sample size than what I currently have. |
| kokkos_parallel_range_for<Kokkos::RangePolicy<execution_space>>( | ||
| "reduction", std::size_t{0}, unique_particles.size(), | ||
| [&local_force, &unique_particles, num_threads](std::size_t const i) { | ||
| Utils::Vector3d force{}; |
There was a problem hiding this comment.
I think it is possible to apply local_force directly to unique_particles.at(i)->force(), and the calculation became a bit faster due to not generating Utils::Vector3d.
There was a problem hiding this comment.
A quick benchmark reveals no major speed-up, but I'll include that in my next PR because it improves readability.
| auto pos2 = copy_aosoa_vector_elc(p2, aosoa.position); | ||
| auto pos1 = copy_aosoa_vector_elc(p1, aosoa.position); | ||
| pos1[2] *= -1.; | ||
| auto const d = box_geo.get_mi_vector(pos2, pos1); |
There was a problem hiding this comment.
When we use aosoa_pack, it would be faster to assign each value of aosoa.position directly to get_mi_vector(T const &a0, T const &a1, T const &a2, T const &b0, T const &b1, T const &b2) than to use get_mi_vector(const Utils::Vector<T, 3> &a, const Utils::Vector<T, 3> &b).
There was a problem hiding this comment.
Here I'm leveraging compiler optimizations that elide out the temporary Vector3d objects from the stack. Verified on the Compiler Explorer using GCC 15.2 and Clang 19.1.
| auto pos2 = copy_aosoa_vector_elc(p2, aosoa.position); | ||
| auto pos1 = copy_aosoa_vector_elc(p1, aosoa.position); | ||
| pos1[2] = 2. * box_h - pos1[2]; | ||
| auto const d = box_geo.get_mi_vector(pos2, pos1); |
hidekb
left a comment
hidekb
left a comment
There was a problem hiding this comment.
I think this looks good to me. Thanks a lot for the great work.
2 participants
No description provided.