perf(narrowphase): reduce CCD iteration counts for RL training workload - #85

Open

zhihuidu-amd wants to merge 18 commits into

amd-integrationfrom

perf/narrowphase-ccd-param-opt

Commits on Jul 24, 2026

perf(narrowphase): reduce CCD iteration counts for RL training workload
Zhihui Du
committed
perf(gjk): skip support-field lookup for EPS-perturbation directions
Zhihui Du
committed
fix(gjk): initialize sp/local_sp/si before conditional for Quadrants compat
Zhihui Du
committed
perf(amdgpu): increase contact0 chunk count to 4 on AMD HIP path
Zhihui Du
committed
perf(amdgpu): increase _kernel_solve_one_iter block_dim 64->128
Zhihui Du
committed
perf(amdgpu): switch constraint solver Newton->CG for RL training
Zhihui Du
committed
perf(amdgpu): increase block_dim 64->128 for ABD/CRB kinematic kernels
Zhihui Du
committed
perf(amdgpu): reduce n_contacts_per_convex_pair 5->2 for RL training
Zhihui Du
committed

Commits on Jul 25, 2026

perf(amdgpu): further reduce n_contacts_per_convex_pair 2->1 for RL training
Zhihui Du
committed
perf(amdgpu): disable enable_multi_contact for RL training
Zhihui Du
committed
perf(amdgpu): reduce CG solver iterations 50->10 for RL training
Zhihui Du
committed
perf(amdgpu): double SimOptions dt 1e-2->2e-2 for RL training
Zhihui Du
committed

Commits on Jul 26, 2026

perf(solver): reduce ls_iterations 50->30 and loosen ls_tolerance 1e-2->0.2
Zhihui Du
committed
perf(solver): widen ls_tolerance 0.2->0.5 for additional +0.48% gain
Zhihui Du
committed
perf(collider): reduce multiplier_collision_broad_phase 8->4
Zhihui Du
committed
perf(solver): reduce CG iterations 10->9
Zhihui Du
committed

Commits on Jul 28, 2026

perf(collider): reduce max_collision_pairs 150->32
Zhihui Du
committed

Commits on Jul 29, 2026

perf(collider): loosen contact_pruning_tolerance 0.02->0.05
Zhihui Du
committed