CUDA: low-64 nonce specialization, compact permutation loop, tiny result record

[n13]

Summary

Per-nonce work and hot-code footprint of the native CUDA engine (--cuda-gpu) are reduced, and the host stops spinning a CPU core per GPU.

Kernel

• The host precomputes the Poseidon2 state up to the first external-round S-box for the 448 fixed nonce bits (pow_core::mining_prestate_low64). The kernel adds only the sparse linear contribution of the two byte-swapped low-64 nonce words per nonce, so the general 16-limb nonce increment and the repeated first linear layer are gone from the hot loop.
• The two hot permutations run as a non-unrolled two-pass loop instead of two inlined copies.
• The result record is a claim flag plus the 32-bit logical index. The host reconstructs the nonce, recomputes the full hash on the CPU, and enforces strict hash < target before anything is reported.
• No per-iteration global result polling inside the nonce loop.

Host

• Blocking CUDA synchronization (CU_CTX_SCHED_BLOCKING_SYNC) instead of busy-waiting.
• Batches are split at the low-64 carry so a launch never wraps the low nonce word (unit test low64_batches_stop_before_carry).
• --gpu-batch-size now defaults to 32M nonces with --cuda-gpu (Vulkan keeps 1M). With blocking sync, 1M-nonce launches lose throughput to host wake-up latency; 32M is about 50 ms per launch on a 4090.
• Each CUDA search logs the GPU duty cycle since the previous search (GPU busy N% since previous search) so idle time on a live node is measurable.

Measurements (RTX 4090, 300 W cap, driver 580.173.02, NVRTC 12.8)

Configuration	MH/s
main (v4.1.0), 67M-nonce launches, alternating samples	585-592
This branch, 67M-nonce launches	609-611
Both at 4M-nonce launches	tied (~600)

Steady-state CPU use of the GPU worker thread drops from ~100% of a core to near zero.

Testing

• The kernel and host changes were run through the engine-cuda real-GPU tests on the 4090 host above (golden vectors, target boundaries at hash-1/hash/hash+1, nonce carries at bits 32/64/128/224, Goldilocks reduction edge cases, CPU-verified solution).
• cargo fmt, cargo clippy --workspace --all-features, and cargo test pass locally; CUDA tests self-skip without a GPU.
• The CLI default change and duty-cycle log were not re-run on a GPU; they do not touch the kernel or launch path beyond a timer.

Not included

The static analysis notes under docs/research/ stay untracked; whether to publish them is a separate decision.

[n13]

Verdict: changes requested

[P2] Recompute hash_count after removing result polling

The kernel no longer checks results[0] inside the per-thread loop, so after one thread claims a solution the other threads continue through their assigned nonces. The host still calculates completed work as total_threads * (winning_iteration + 1), which assumes the old lockstep early-exit behavior. With the new 32M default and 31 nonces per thread, a winner at j=0 can report about 1.05M hashes even though nearly the full 32M dispatch executes. That incorrect count feeds gpu_hashes_total, completion-rate logs, and MiningResult.hash_count.

Please make the returned count match the new execution model (for example, report the dispatched batch under scheduled-work semantics, or add exact accounting if exact completed hashes are required) and cover a multi-iteration found batch in a regression test.

[n13]

Fixed in 857c057: found batches now report the dispatched count. The kernel-entry early exit is gone too, so the full launch rectangle always executes and the dispatched count is exact rather than an upper bound. Regression test cuda_found_batch_counts_every_dispatched_nonce covers a 4M-nonce batch with 4 nonces per thread.

[n13]

Verdict: approve

Re-reviewed current head 857c057. The prior hash_count issue is resolved under scheduled-work semantics: found and exhausted batches now report the dispatched nonce count consistently, the kernel-entry early exit no longer makes that scheduled rectangle conditional, and cuda_found_batch_counts_every_dispatched_nonce covers the multi-iteration case.

I found no remaining blocking issues. All current GitHub checks pass. Local validation also passed: git diff --check and cargo test -p pow-core -p engine-cuda -p miner-cli --locked.