Perf/fix4b lds linesearch jv jaref cache#82

Open

zhihuidu-amd wants to merge 13 commits into

amd-integrationfrom

perf/fix4b-lds-linesearch-jv-jaref-cache

Commits on Jul 2, 2026

perf(amdgpu): LDS-cache efc_force in tiled_wc Phase 4b
zhihuidu-amd
committed
perf(amdgpu): LDS-cache Jaref/jv/efc_D in linesearch pt_opt for tiled_wc
zhihuidu-amd
committed
perf(amdgpu): stacked LDS + compiler + occupancy opts for tiled_wc (path to 200%)
zhihuidu-amd
committed

Commits on Jul 5, 2026

fix: use module-level constants for LDS SharedArray dims in @qd.func
zhihuidu-amd
committed
fix: revert pt_3a to HBM reads; fix while compound condition in pt_opt
zhihuidu-amd
committed
fix: replace compound while conditions with simple while + if guard
zhihuidu-amd
committed
fix: use literal 64 instead of module-level _LS_MAX_CON in @qd.func
zhihuidu-amd
committed
fix: remove fn_attrs from tiled_wc kernel (incompatible with local quadrants whl)
zhihuidu-amd
committed
fix: pre-init LDS vars from HBM to satisfy quadrants scope rules in _func_ls_pt_opt_twc
zhihuidu-amd
committed
perf: split LDS linesearch loops at boundary-64 to avoid double HBM+LDS reads
zhihuidu-amd
committed
perf: linesearch LDS-only reads (n_con<=64 guaranteed, no HBM fallback needed)
zhihuidu-amd
committed
revert: linesearch LDS caching (regresses -50%); keep only Phase 4b efc_force LDS from PR#81
zhihuidu-amd
committed

Commits on Jul 12, 2026

fix: Phase 4b LDS loop guard + revert collider CU tuning that regresses Go2
zhihuidu-amd
committed