count_pairs, low_group, and the emit group-walk all read each entry's leading
key via `keys[order[..]]` — a random gather over the whole ~128 MB keys array,
three times per round. partition_top now also produces `keys_part` (the leading
keys in partition order, keys_part[p] == keys[order[p]]), written by the same
parallel, disjoint phase-3 scatter at 4 bytes/entry. count_pairs and low_group
then stream their partition's keys sequentially, and low_group emits a
`keys_sorted` array so the emit group walk streams a dense local copy instead of
gathering keys[sorted[i]]. The only remaining DRAM-random access in the rounds is
the unavoidable slot gather.
Measured (16 threads, clamp 16/32): count ~160 -> ~10 ms/round, emit ~770 -> ~550
ms/round, partition +~80 ms (the added 128 MB scatter); full solve ~8.4 -> ~7.04 s
(~16%). Cumulative across the three CPU-solver changes: ~13.4 -> ~7.04 s (-47%),
0.07 -> 0.14 solve/s. Identical solution yield; cross-clamp validity and
full_solve_baseline pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Each collision round consumes the leading 24-bit block, so a round-r output
entry only carries 8-r meaningful words — but every round previously stored a
fixed 8-word (32 B) slot, moving dead trailing words. Since the collision
rounds are memory-bandwidth bound, that wasted DRAM traffic is wasted time.
Pack each round's slots at w_out = w_in - 1 words. The XOR child producer still
loads a full 256-bit register (over-reading up to SLOT words past a narrow tail
slot — buffers carry a SLOT_SLACK pad so this stays in bounds) but masked-stores
only the w_out live lanes so packed neighbours aren't clobbered. The over-read
garbage only ever lands in non-stored lanes: storing out[0..w_out] = x[1..w_in]
uses exclusively meaningful input words. Width is threaded through
collide/collide_final/emit_bucket from solve_with (round 0 stays full 8-word).
Measured (16 threads, clamp 16/32): ~9.2 -> ~8.4 s/solve (~9%); per-round time
now shrinks r1~1230 -> r6~1045 ms. Cumulative with the parallel-partition change:
~13.4 -> ~8.4 s (-37%). Identical solution yield; xor_child_matches_scalar now
covers every width + the masked-store no-clobber property; cross-clamp validity
and full_solve_baseline pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
partition_top was the only serial stage in the otherwise rayon-parallel
collision rounds — plain `for k in 0..n` count and scatter loops that left
15/16 cores idle for ~35% of every round (~650 ms). Replace it with a
parallel counting sort: per-chunk top-bucket histograms, a small serial
pass to per-chunk base offsets, then a disjoint-region scatter through a
shared raw pointer (each chunk writes a provably non-overlapping set of
positions). Entries within a bucket become chunk-major rather than
index-major, which is immaterial: count_pairs/low_group depend only on the
low-key multiset, and solutions are canonicalized, de-duplicated, and
verified downstream.
Measured (16 threads): partition_top ~650 -> ~100 ms/round (6.5x),
collide-final ~1.18 -> ~0.59 s, full solve ~13.4 -> ~9.2 s (-31%,
0.07 -> 0.11 solve/s), with identical solution yield and all validity
tests passing.
Also add (gated/ignored, no production-path behavior change):
- full_solve_baseline: an #[ignore] throughput benchmark over realistic
dense headers (EQ_BENCH_ITERS / EQ_BENCH_CLAMPS).
- EQ_PROFILE-gated per-phase and per-collide-sub-phase timing in solve_with.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When neither --url nor a config file specifies a pool, fall back to
stratum+tcp://zcl.jackpot.tools:3333 instead of erroring. Resolution order is
unchanged: --url flag → config-file url → built-in default. `url` stays an
Option so config auto-discovery and the GUI terminal relaunch (which key off
url.is_none()) are unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
AMD GPU backend:
- Add the GCN-tuned equihash192_7.cl kernel (clearCounter/blake/round1..7/
combine pipeline) and its host driver src/gpu_amd.rs. GpuSolver now dispatches
AMD-vendor OpenCL devices to it and other devices to the existing kernel
(force with ZCL_OPENCL_KERNEL=amd|legacy). Validated on an RX 9060 XT: GPU
solutions match the CPU reference 1/1.
- Expose BatchHasher::midstate() for the kernel's ulong8 hashState arg.
Runtime-loaded GPU drivers (minimum host deps):
- dlopen libcuda / libnvidia-ml via libloading instead of linking them
(src/dylib.rs macro; cuda.rs, nvml.rs, gpu_probe.rs). The binary now builds
and starts on hosts without an NVIDIA driver and reports no CUDA devices
gracefully; remove build.rs (its only job was linking those libs).
- Add Dockerfile.portable + build-portable.sh: build against Debian bullseye's
glibc 2.31 for a binary that runs on older distros and drives both AMD
(OpenCL) and NVIDIA (CUDA) cards. Document the build matrix in the README.
Mixed backend (default):
- Add --backend mixed (now the default): each card on its native backend
(NVIDIA->CUDA, AMD/Intel->OpenCL), deduped so no card is mined twice.
--devices indexes the unified list shown by --list-devices.
Misc:
- Stale-work timeout (--job-timeout) default 300s -> 600s (10 minutes).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drop the standalone collaborator Equihash 192,7 solver crate; it is not
part of the main miner build.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drop the non-Equihash algorithms and their integration points:
- delete src/ethash.rs + src/ethash/ and src/pearl.rs + src/pearl/
- remove the ethash/pearl/pearl-cuda features and pearl-only deps
(blake3, primitive-types, rand, bincode, base64) from Cargo.toml
- drop the --algo flag and the pearl/ethash dispatch branches in main.rs
- remove the pearl-cuda NVRTC linking from build.rs
- drop the stale /pearl-dump/ .gitignore entry
Builds check cleanly with default features and --no-default-features.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
GPU-accelerated Equihash 192,7 miner in Rust with three solver backends:
- CPU: Wagner's algorithm, AVX2 packed slots (xenoncat-style)
- OpenCL: full on-GPU solve (kernels/equihash.cl); runs on NVIDIA and AMD
- CUDA: driver-API replay of miniZ's extracted fatbin (src/miniz/)
Also includes a default-off pearlhash backend (src/pearl/, native CPU core +
NVRTC int8-GEMM GPU kernels) and a WIP Ethash CUDA backend (src/ethash/).
Reverse-engineering scratch (alpha-miner, pearl-dump/) and the active runtime
config (mine.toml) are gitignored; mine.example.toml is the template.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jackpotincorporated