milnor_gpu: fix >2^32-element master via cubecl 64-bit addressing - #17

Draft

JoeyBF wants to merge 107 commits into

milnor-gpu-u64-addressing

Commits on Jul 27, 2026

milnor_gpu: fix >2^32-element master via cubecl 64-bit addressing

JoeyBF
and
claude
committed
af74dfa
milnor_gpu: slash GPU host memory — single stream by default + drop the resident host master

JoeyBF
and
claude
committed
59bd2d9
nassau_gpu: cap the per-build GPU work so the dense readback stays bounded

JoeyBF
and
claude
committed
e4da7ed
milnor_gpu: degree-threshold eviction of the resident device master

JoeyBF
and
claude
committed
28a1dff
milnor_gpu: CPU reference for in-kernel admissible enumeration

JoeyBF
and
claude
committed
b447de5
milnor_gpu: in-kernel admissible enumeration (CUDA-validated)

JoeyBF
and
claude
committed
0ae80f3
milnor_gpu: enumerate cold masters on-device instead of uploading

JoeyBF
and
claude
committed
e41ed2f
milnor_gpu: enum-path diagnostics + CPU-vs-GPU enum bench (direction refuted)

JoeyBF
and
claude
committed
a28e214
milnor_gpu: correct the enum-vs-upload bench (parity, not 3x slower)

JoeyBF
and
claude
committed
cc6ae7b
milnor_gpu: segmented-master read primitive (chunked-growth de-risk)

JoeyBF
and
claude
committed
b647558
milnor_gpu: segmented no-copy resident master/basis (kill the realloc-doubling spike)

JoeyBF
and
claude
committed
f362f85
milnor_gpu: move all test kernels/helpers into the tests module

JoeyBF
and
claude
committed
7265f04

Commits on Jul 28, 2026

milnor_gpu: persistent seqno (g/xi) + per-stream output buffer (kill per-launch churn)

JoeyBF
and
claude
committed
c6879d9
milnor_gpu: drop the per-thread persistent out_h (keep resident g/xi)

JoeyBF
and
claude
committed
a0af1db

Commits on Jul 29, 2026

milnor_gpu: CPU-fallback stopgap for cubecl context death + GPU multiply bench

JoeyBF
and
claude
committed
72190b3
milnor_gpu: NASSAU_GPU_CLEANUP_EVERY knob — throttle per-launch memory_cleanup

JoeyBF
and
claude
committed
1c09868

Commits on Jul 30, 2026

milnor_gpu: concurrent-growth soak #[test] — fast #1401 reproducer + correctness guard

JoeyBF
and
claude
committed
aafe902

Commits on Aug 1, 2026

milnor_gpu: pivot cubecl 0.10 fork → upstream v0.11.0-pre.1

JoeyBF
and
claude
committed
9dbe8ae
milnor_gpu: guard the multiply output write against OOB atomic (out[word])

JoeyBF
and
claude
committed
3f86eb7
Bit-pack Milnor basis elements into a u64
claude
committed
bbf7c4c
Cut per-entry overhead out of the Milnor multiplier
claude
committed
769fbcc
Derive the p=2 Milnor basis instead of storing it
claude
committed
fb242b8
Add PPartRanker behind an opt-in feature, not wired in
claude
committed
6ac8173
Reject inputs the packing cannot represent, instead of panicking
claude
committed
99ae117
Apply nightly rustfmt to the packed p-part constants
claude
committed
f3aab5b
milnor_gpu: also guard the output row against OOB (word < out.len())

JoeyBF
and
claude
committed
d356e35

Commits on Aug 2, 2026

fp,fp-cuda: lock-free GPU row reduction + cross-runtime device arbitration

JoeyBF
and
claude
committed
df162b9
milnor_gpu: fail loudly on GPU context death, and split the multiply's time

JoeyBF
and
claude
committed
f9a7467
nassau: probe whether signature steps are independent (they are not)

JoeyBF
and
claude
committed
4671199
nassau_gpu: rustfmt reflow of an assert message

JoeyBF
and
claude
committed
28ed66e
milnor_gpu: dedicated GPU submission thread (kill the starvation stalls)

JoeyBF
and
claude
committed
d2cee65
milnor_gpu: bench the hard stem-200 regime instead of resolving to it

JoeyBF
and
claude
committed
f9f595f
nassau,milnor_gpu: the multi-minute stalls were a par_iter over a 5ms loop

JoeyBF
and
claude
committed
7744b4a
milnor_gpu: zero-copy batch output, bounded bench, kernel micro-optimisation

JoeyBF
and
claude
committed
3e7c821
milnor_gpu: size the per-thread working arrays per launch (+29%)

JoeyBF
and
claude
committed
62fa89c

Commits on Aug 3, 2026

Merge PR #280 (packed PPart) and port the GPU path to it

JoeyBF
and
claude
committed
683c4d8
milnor_gpu: pack the kernel's `working` accumulator into a u64

JoeyBF
and
claude
committed
22b0a67
milnor_gpu: record the occupancy finding and prepare the cubecl change

JoeyBF
and
claude
committed
76869d2
milnor_gpu: fuse the per-thread gather into the column loop (48 -> 40 registers)

JoeyBF
and
claude
committed
93520e8
milnor_gpu: NASSAU_GPU_LAUNCH_LOG — dump per-launch shape

JoeyBF
and
claude
committed
7887f24
milnor_gpu: bound the column loop per thread, not per launch (+8.5%)

JoeyBF
and
claude
committed
9e0abd9
milnor_gpu: coarse product index for the thread->product lookup (+4.2%)

JoeyBF
and
claude
committed
2aed0b0
milnor_gpu: decode matrix fastest, term slowest (+2.0%)

JoeyBF
and
claude
committed
c305f37
milnor_gpu: look up num_mats instead of dividing for it

JoeyBF
and
claude
committed
19fb7db
milnor_gpu: one thread per (matrix, term GROUP) (+51%)

JoeyBF
and
claude
committed
d54aec9
milnor_gpu: TERM_GROUP 4 -> 3 (+9%)

JoeyBF
and
claude
committed
37a298e
gitignore: exclude ext/*.log GPU run logs

JoeyBF
and
claude
committed
144727f

Commits on Aug 4, 2026

nassau: span the CPU work inside `step`

JoeyBF
and
claude
committed
d167b6f
milnor_gpu: tile over matrices as well as terms (+6.6%)

JoeyBF
and
claude
committed
d4db837
milnor_gpu: shard the resident master across GPUs (2.27x on 4 devices)

JoeyBF
and
claude
committed
9174826
milnor_gpu: assign shards to the least-loaded device (+32%)

JoeyBF
and
claude
committed
cc11687
milnor_gpu: submit shards without spawning threads; name the GPU workers

JoeyBF
and
claude
committed
1639982
milnor_gpu: marshal the shards in parallel again, wait outside

JoeyBF
and
claude
committed
a9fc729
milnor_gpu: keep the scoped-thread fan-out; it beat both rewrites

JoeyBF
and
claude
committed
0bc8956
maybe-rayon: add MaybeThreadPool; use it for the multi-GPU fan-out

JoeyBF
and
claude
committed
07d7aab
milnor_gpu: size the fan-out pool for concurrent callers, not devices

JoeyBF
and
claude
committed
b8c4470
milnor_gpu: size the fan-out pool by callers x devices

JoeyBF
and
claude
committed
8c2365c
milnor_gpu: one thread, async submissions to every device

JoeyBF
and
claude
committed
199c052
milnor_gpu: pipeline the readback; restore the memory bound

JoeyBF
and
claude
committed
a4aed87
milnor_gpu: marshal the shards in parallel again

JoeyBF
and
claude
committed
ef16f93

Commits on Aug 5, 2026

milnor_gpu: split `prep` into intern / basis / tgei

JoeyBF
and
claude
committed
3ab3792
nassau, milnor_gpu: demote the fine-grained spans below info

JoeyBF
and
claude
committed
220e8f0
milnor_gpu: flag the readback keepalive bug at the site

JoeyBF
and
claude
committed
51a8949
milnor_gpu: keep launch buffers alive until the readback completes

JoeyBF
and
claude
committed
061c667
Revert "milnor_gpu: marshal the shards in parallel again"

JoeyBF
and
claude
committed
bdbd4eb
milnor_gpu: marshal the shards on dedicated threads

JoeyBF
and
claude
committed
02dc8f0
milnor_gpu: run each shard's whole pipeline on its own thread

JoeyBF
and
claude
committed
9aa0211
milnor_gpu: persistent per-caller shard helpers, no more spawn churn

JoeyBF
and
claude
committed
63595b7
milnor_gpu: crossbeam channels for the shard helpers

JoeyBF
and
claude
committed
38bbdd7
Revert "milnor_gpu: crossbeam channels for the shard helpers"

JoeyBF
and
claude
committed
3c31f87
milnor_gpu: shard R by a mixed hash, not by accumulated num_mats

JoeyBF
and
claude
committed
8ceb3b2
milnor_gpu: cover the eviction path, and honour CUDA_VISIBLE_DEVICES

JoeyBF
and
claude
committed
5d2334c
milnor_gpu: order the enumeration by matrix count; tighten its local caps

JoeyBF
and
claude
committed
dbe1b49

Commits on Aug 6, 2026

milnor_gpu: span transient scratch across segments, not one

JoeyBF
and
claude
committed
2d765af
fp: make the multiply/reduction device-sharing test multi-GPU aware, and log it

JoeyBF
and
claude
committed
90e38c2
fp: reductions must serialize against each other, not just against the multiply

JoeyBF
and
claude
committed
32bd1f0
fp: route every fp-cuda submission through one driver thread

JoeyBF
and
claude
committed
85fa53c
fp-cuda: take a share of the device for the GEMM, not all of it

JoeyBF
and
claude
committed
527a18d
fp-cuda: name Phase 9 clusters as the co-scheduling constraint to remove

JoeyBF
and
claude
committed
9bc3012
fp-cuda: cluster-free GEMM as the default, cluster kernel opt-in

JoeyBF
and
claude
committed
d3bc656
algebra: measure the d-vs-masks enumeration shortcut, and record that it loses

JoeyBF
and
claude
committed
ec3ed45

Commits on Aug 7, 2026

algebra: measure the enum kernel's emit cost — 42.8% of it, at 1% of HBM bandwidth

JoeyBF
and
claude
committed
f3391b7
algebra: record that stack-allocating AdmissibleMatrix loses

JoeyBF
and
claude
committed
e51670c
algebra: the enum emit is store-ISSUE bound, not bandwidth bound

JoeyBF
and
claude
committed
1a7e0f9
algebra: measure R rebuild COST, not reference count — the transient head is steep

JoeyBF
and
claude
committed
ece4f45
algebra: simulate a pinned transient cache — it is linear, so theta is already optimal

JoeyBF
and
claude
committed
00f8d90
algebra: theta=125 costs 6.5x wall time to save 5.3 GB — set it as high as memory allows

JoeyBF
and
claude
committed
e7c8d2f
algebra: report which theta would have fit, instead of discovering it as an OOM

JoeyBF
and
claude
committed
a3f46eb
nassau: skip step1's kernel computation when the augmentation codomain is empty

JoeyBF
and
claude
committed
a1b832b
algebra: abort on segment-table exhaustion instead of killing one shard thread

JoeyBF
and
claude
committed
3b83d1e
algebra: profile the enum kernel — it is latency-bound, and shared state does not fix it

JoeyBF
and
claude
committed
b5d927a
algebra: the enum kernel is work-starved — registers and shared memory are not the constraint

JoeyBF
and
claude
committed
815fbdd
algebra: log enum-launch geometry — production runs the kernel at 0.2% of the grid

JoeyBF
and
claude
committed
18e79d8
algebra: block size is not the enum lever either — the launches are serial

JoeyBF
and
claude
committed
6e5b88f
milnor_gpu: GPU stream concurrency does not pay — the run is not GPU-bound

JoeyBF
and
claude
committed
78872ff
nassau_gpu: read the GPU multiply back a limb at a time, not a bit at a time

JoeyBF
and
claude
committed
72d6e1f
nassau: the missing parallelism is not inside a bidegree — two probes, both negative

JoeyBF
and
claude
committed
75faee7
nassau: span the zero-signature region — and find the run is 44% CUDA spin-wait

JoeyBF
and
claude
committed
fbf5c7e
milnor_gpu: theta=125 at stem 150 is an artificial worst case — 2.85x slower than uncapped

JoeyBF
and
claude
committed
a9d3f5e
milnor_gpu: record that FxHashMap is 8% SLOWER than SipHash here

JoeyBF
and
claude
committed
c068c48

Commits on Aug 8, 2026

nassau_gpu: one bulk memcpy per readback row instead of one per u32 limb

JoeyBF
and
claude
committed
d9f2e45
milnor_gpu: share GpuProduct term lists — 1.45x on stem 150

JoeyBF
and
claude
committed
8928356
milnor_gpu: memoize the per-product R lookup — 6-8%

JoeyBF
and
claude
committed
33c9713
milnor_gpu: scatter eviction output straight out of the device blocks

JoeyBF
and
claude
committed
3ec60d7
milnor_gpu: put the enum kernel's state in shared memory — 5.5% in the capped regime

JoeyBF
and
claude
committed
c28f431
milnor_gpu: record that de-duplicating enum work across devices buys nothing

JoeyBF
and
claude
committed
e30eceb
milnor_gpu: pin long-pole R's resident regardless of theta — 13.6% capped

JoeyBF
and
claude
committed
05bf3e0