CPUFJ fastpath for binary instances + latency changes to find better solutions earlier - #1750

Open

aliceb-nv wants to merge 61 commits into

cpufj-fastpath-isolated

Commits on Aug 25, 2026

Add a binary/int8 fast path to the CPU feasibility jump

aliceb-nv
and
claude
committed
0fe2abb
Take the tabu test out of the binary fast path's global argmax

aliceb-nv
and
claude
committed
ac9d47e
perf(cpufj-bin): drop the per-iteration bitmap clear and hoist apply_move's base pointers

aliceb-nv
and
claude
committed
4f4bcb1
perf(cpufj-bin): vectorize apply_move's row walk

aliceb-nv
and
claude
committed
f334465
docs: record the Highway third-party licence
aliceb-nv
committed
dd1501d
bit of cleanup
aliceb-nv
committed
57e2ef7
cleanup
aliceb-nv
committed
febac7b
perf: fold the patch width choice into the Highway target
aliceb-nv
committed
2dc1f79
perf: normalize every row to a'x <= b and drop the per-row sign
aliceb-nv
committed
eb4ce75
refactor: replace the row record with a plain weight array
aliceb-nv
committed
b0c1924
fix: derive the argmax tile from the L1 data cache size
aliceb-nv
committed
2f63ab8
debug: add CUOPT_NO_BINFJ to force the general path
aliceb-nv
committed
9bcb1b7
fix: widen the packed staged score to int64
aliceb-nv
committed
bd9c383
add row integralize
aliceb-nv
committed
c308ce4
fix solve_CPUFJ link
aliceb-nv
committed
2f7cf86
latency tweaks
aliceb-nv
committed
5805269
more logs
aliceb-nv
committed
10ddf0e
ai review
aliceb-nv
committed
69ac5a2
fix build
aliceb-nv
committed
64abba9
some optimizing
aliceb-nv
committed
0c771a3
some bug fixes regarding lift moves
aliceb-nv
committed
9d46a9f
hiverge inspired improvements
aliceb-nv
committed
996c832
ddfw improvements
aliceb-nv
committed
075ded0
changes for the hiverge harness
aliceb-nv
committed
cdc805b
replace operator overloaded arithmetic with explicit highway functions, + some fusing
aliceb-nv
committed
25e4479
save user_callbacks in run_mip
aliceb-nv
committed
71d07a6
let B&B publih solutions via the user callbacks as well to reduce latency
aliceb-nv
committed
27bae2f
publish cpufj scratch solutions early as well
aliceb-nv
committed
8b26bcd
absorb cuda driver startup in run_mip.cpp
aliceb-nv
committed
abdad3e
fix multigpu runs
aliceb-nv
committed
04a25fa
lhs refresh period stretch, lazy mtm cache invalidate, magnitude aware objective term
aliceb-nv
committed
a977cf9
objective weight tweaks
aliceb-nv
committed
b4ed104
perf optimization
aliceb-nv
committed
64b5bfd
more optimization
aliceb-nv
committed
9c94889
weight_escalation_delta
aliceb-nv
committed
08a858a
infeasible region kicks
aliceb-nv
committed
ff86b74
2opt lift move thing
aliceb-nv
committed
4ff4650
portfolio lane seeding
aliceb-nv
committed
01b44ae
early cpu portfolio
aliceb-nv
committed
af02f91
lift move handle fractionals
aliceb-nv
committed
5293291
vectorized argmax for binary path in feasible region
aliceb-nv
committed
a8bb991
kahan sum the objective; audit validity in the harness
aliceb-nv
committed
868142c
Binary perturb re-anchors to the incumbent
aliceb-nv
committed
a2ccbc9
apply_exact_k_seed
aliceb-nv
committed
e02bd00
more seeds
aliceb-nv
committed
0e4453f
bipartite seeding
aliceb-nv
committed
6c8e2ce
share incumbent in portfolio
aliceb-nv
committed
32b6b52
perturbation gated on objective stall
aliceb-nv
committed
428575a
infeasible pair repair
aliceb-nv
committed
21f2b89
dualsimplex LP seed, 2s
aliceb-nv
committed
dd70c32
some optimization work
aliceb-nv
committed
e885c11
some optimization efforts
aliceb-nv
committed
0795c61
bounds prop in CPUFJ
aliceb-nv
committed
d32d935
ddfw scheme
aliceb-nv
committed
98130fc
encode small domain integers as binary to enable the FJ binary fastpath
aliceb-nv
committed
e4caa19
feaspump in early cpufj
aliceb-nv
committed
555848a
cpufj batching
aliceb-nv
committed
b6f4f43
fix Werror=unused flags
aliceb-nv
committed
c28f2ab
match main post rebase
aliceb-nv
committed
72d7335
run cpufj portfolio before root LP solve
aliceb-nv
committed
67ccc35
many lanes during early heuristics
aliceb-nv
committed
7eda27d