search

Xiangyu-Hu / SPHinXsys

SPHinXsys provides C++ APIs for engineering simulation and optimization. It aims at complex systems driven by fluid, structure, multi-body dynamics and beyond. The multi-physics library is based on a unique and unified computational framework by which strong coupling has been achieved for all involved physics.
https://www.sphinxsys.org/
Apache License 2.0
259 stars 199 forks source link

Making cache-friendly loops #124

Open FabienPean-Virtonomy opened 1 year ago

FabienPean-Virtonomy commented 1 year ago

https://github.com/Xiangyu-Hu/SPHinXsys/blob/1e0042b5040c6f59b9fa98e7634c33f63090fd1a/SPHINXsys/src/shared/particle_dynamics/solid_dynamics/solid_dynamics.cpp#L770-L778

While considering vectorization and see how it could fit, I ended up thinking specifically about this part of the code above. This kind of loop is suboptimal as it is doing one action on several different arrays. A data-oriented and cache-friendly approach would be applying the operation for each array one after the other:

for(auto index_i = 0; index_i < total_real_particles; ++index_i)
    pos_n_[index_i] += vel_n_[index_i] * dt * 0.5;
for(auto i = 0; i < total_real_particles; ++i)
    F_[index_i] += dF_dt_[index_i] * dt * 0.5;
for(auto i = 0; i < total_real_particles; ++i)
        rho_n_[index_i] = rho0_ / det(F_[index_i]);
for(auto i = 0; i < total_real_particles; ++i)
    stress_PK1_[index_i] = F_[index_i] * material_->ConstitutiveRelation(F_[index_i], index_i);

It does not seem so widespread overall in the codebase (as only one field is getting touched) but the architecture around ParticleDynamics1Level in the library favors this kind of pattern.

Xiangyu-Hu commented 1 year ago

It is good starting point. Need some test to see whether we get benefits. And how to combine with parallelization.