any tips for making the library run faster?

[eilat-inbal]

Hi, Thanks for this awesome library! I'm getting great results but the computation time is very long... Any tips on how to make it run faster? I'm running on windows, c++ like this:

coacd::Params params; params.preprocess_mode = "on"; params.prep_resolution = 150; params.pca = false; params.mcts_max_depth = 2; params.mcts_iteration = 60; params.merge = true; params.threshold = 0.02; coacd::Model convexMesh; convexMesh.Load(vertices, indices); convexParts = coacd::Compute(convexMesh, params);

for my mesh, with these parameters t took a few minutes! I tried making the threshold bigger and the prep_resolution smaller. and that helps but the computation is still very long, and the results are not as good... any further options for multithreading? for GPU?

thanks! Inbal

[SarahWeiii]

Hi, the code supports using OpenMP for parallel computing, have you already enabled that?

[AccessViolationEnjoyer]

I was able to make the library quite a bit faster (at least 3X) by performing some simple optimizations. The main changes, IIRC, were:

• Passing large objects by const reference, especially vectors.
• Removing the use of std::pow() for computing small integral powers.
• Reducing heap allocations by caching containers.
• Using faster hash maps.

Also, since the library allocates quite a lot of memory, it may be worth using a different memory allocator if building with multithreading. I found Mimalloc to give good results.

[Kitsune44]

I was able to make the library quite a bit faster (at least 3X) by performing some simple optimizations. The main changes, IIRC, were:

• Passing large objects by const reference, especially vectors.
• Removing the use of std::pow() for computing small integral powers.
• Reducing heap allocations by caching containers.
• Using faster hash maps.

Also, since the library allocates quite a lot of memory, it may be worth using a different memory allocator if building with multithreading. I found Mimalloc to give good results.

Hi, Can you share your repo?

[AliRezaBeigy]

I’ve been experimenting with CoACD and wanted to share some insights on tuning parameters for better performance based on my tests:

• threshold: This drives the trade-off between detail and speed. Lower values (e.g., 0.01) produce more detailed decompositions but take much longer. I found 0.05 or higher speeds things up noticeably, though you’ll trade off some precision.
• mcts_nodes: Affects how smooth the decomposition is. Low values (e.g., 10) are faster but can fragment results; I settled on 20 as a good balance for decent quality without much slowdown.
• mcts_iterations: This tweaks detail with minimal runtime impact. I used 1000 without seeing major delays—seems efficient enough up to that range.
• mcts_max_depth: Major performance lever! Keeping it at 1 is fast and works for simpler meshes, but going to 2 or higher (like your 2) slows things down a lot due to deeper searches. Try 1 if speed’s the goal.
• merge: Enabling this tanks performance because of the extra merging step. I skipped it (false) for my case and got faster results—only use it if you need combined parts.

coacd_mesh = coacd.Mesh(mesh.vertices, mesh.faces)
parts = coacd.run_coacd(
    coacd_mesh,
    threshold=0.05,
    mcts_nodes=20,
    mcts_iterations=1000,
    mcts_max_depth=1,
    merge=False,
)