Skip to content

README.md

Latest commit

 

History

History
147 lines (105 loc) · 5.79 KB

README.md

File metadata and controls

147 lines (105 loc) · 5.79 KB
Raw

GPU backends (TrueCollider / KeyCollider)

Runtime selection: -U none (default) · -U cuda · -U opencl.

Dispatcher: gpu/gpu_dispatcher.cpp.

Custom CUDA edition

keyhunt_cuda.exe is the custom CUDA edition: CLI parity with keyhunt.exe (every -m mode and every flag accepted), plus KeyHunt-class sequential device GRP (MIT clean-room). Nothing is compiled out of the CUDA binary’s mode/flag surface.

  • GRP window: TCUDA_GRP_SIZE (1024) in gpu/cuda/secp256k1_cuda.h
  • Packed targets: scripts/addr_to_hash160.py → -f file.hash160
  • Runtime: NVIDIA driver + CUDA runtime (cudart64_*.dll — copy next to the exe on Windows if needed)
  • Do not expect EC Mk/s rates from mnemonic mode

CLI parity (keyhunt.exe vs keyhunt_cuda.exe)

Help dumps match for all -m modes and flag letters (banner / exe name differ only). pub2rmd is removed on both (use -m rmd160). Research aliases (shadow160, weakrng, hybrid-dl, gaudry, wif-mask, hex-mask, kangaroo-mod, CreateAccountWithSeed) map to the same host modes on both binaries.

Mode / flag Accepted on CUDA binary Device acceleration Host-side remainder
-m address + -x sequential/rseq Yes GRP stride-1 BTC-family Filter confirm / hit I/O
-m address + -x chaos/auto/… Yes Host plans bases → CUDA sequential batches Walk planner
-m rmd160 + -x * Yes Same as address Walk planner
-m xpoint / vanity / pubkey2addr Yes GRP or per-key EC Prefix / table checks
-m bsgs Yes Baby-table + GRP giant-step Bloom / tables in RAM
-m kangaroo Yes Batch scan (≤2²⁴) / DP walkers DP table / collision
-m mnemonic / poetry / brainwallet Yes CUDA EC after derive PBKDF2 / BIP39 / paths on host
-m minikeys Yes CUDA EC batch Minikey gen on host
-c eth / etc Yes GPU EC keccak160 encode host
-c troot Yes GPU EC Taproot tweak host
-c sol Yes Device ed25519 ge when ready base58 / fallback
-e endomorphism Yes (never rejected) Skipped — CPU endo GRP Full λ-path on CPU
-x planners / --walk / research -R Yes Bases feed CUDA when secp path armed Planner logic host
-U opencl Yes if built Hash160 only EC on CPU
-U both Yes Even threads CUDA / odd CPU Hybrid

Backend matrix

Backend EC (secp256k1) Address encode Bloom / filter Vendors
CUDA On GPU (secp256k1_cuda.cu) Device hash160 (BTC-family) or host keccak (ETH) Device bloom when hash160 self-test passes; host fallback NVIDIA
OpenCL Host CPU On GPU hash160 (hash160_opencl) Host NVIDIA, AMD, Intel
CPU Host SSE / AVX2 / AVX-512 Fuse or bloom All

CUDA path (-U cuda)

  1. Init: hash160 + secp self-tests → set device vs host filter; secp_ready.
  2. Upload bloom to device when hash160 OK (else host-only copy).
  3. Batch privkeys → device EC (+ device hash160/bloom when enabled).
  4. ETH still host keccak; host confirms hits and writes keys.
  5. -e / non-accelerable bits: warn and continue on CPU (never “not available on CUDA”).

Honest limits: BIP39 PBKDF2 host; ETH keccak host; taproot tweak host; -e CPU-only; BSGS GRP correctness-first (serial cycles); prefer -t 1 with CUDA.

OpenCL path (-U opencl)

  1. Enumerate platforms (prefers NVIDIA/AMD by compute units).
  2. Self-test batch hash160 of compressed pubkeys.
  3. Main loop: CPU EC → gpu_dispatcher_hash160_33 → host filter (when AVX-512 path is not active and -e is off).

Build

# NVIDIA CUDA
cmake -B build-cuda -DENABLE_CUDA=ON
cmake --build build-cuda -j
# Windows: bats/00_build/build_cuda_vs2022.bat → keyhunt_cuda.exe

# OpenCL (AMD / NVIDIA / Intel)
cmake -B build-opencl -DENABLE_OPENCL=ON
cmake --build build-opencl -j

Examples

# BTC-family on CUDA
./keyhunt -m address -f targets.txt -U cuda -G 128 -t 1 -l compress -q -s 5

# Ethereum on CUDA
./keyhunt -m address -c eth -f eth.txt -U cuda -G 128 -t 1 -q -s 5

# Taproot on CUDA
./keyhunt -m address -c troot -f troot.txt -U cuda -G 128 -t 1 -q -s 5

# OpenCL hash offload
./keyhunt -m address -f targets.txt -U opencl -t 8 -l compress

Windows:

keyhunt_cuda.exe -m address -f tests\66.txt -b 66 -l compress -U cuda -G 128 -t 1 -q -s 5

Or double-click / run: examples/run_gpu_cuda_example.bat.

Flags

Flag Meaning
-U cuda|opencl|none Backend
-G N Batch size hint (clamped by -M / VRAM plan; large grids chunked)
-M MB|auto|2G Host/VRAM memory budget (CUDA sizes batches; BSGS scales blooms)

API surface (for developers)

Symbol Role
gpu_dispatcher_search_privkeys(..., encode) GPU EC + host hash160 or ETH keccak + bloom
gpu_dispatcher_pubkey_batch GPU EC only (Taproot / future BSGS helpers)
gpu_dispatcher_hash160_33 CUDA/OpenCL hash160 of compressed pubs
GPU_ENCODE_HASH160 / GPU_ENCODE_ETH Encode selectors

Roadmap (GPU)

  1. Trusted device hash160 → full on-device bloom path. Done.
  2. GPU BSGS giant-step GRP kernels (baby tables stay RAM-heavy). Done (serial; optimize next).
  3. GPU Kangaroo (scan + multi-walker DP). Done (not RCKangaroo/SOTA throughput yet; correct collisions).
  4. Throughput-tune device Solana ed25519 grind + base58 prefilter.
  5. OpenCL secp EC twin for AMD.

See also: docs/ROADMAP.md.

Layout

gpu/
  gpu_dispatcher.cpp|.h     # -U routing
  cuda/                     # CUDA secp + hash160 + bridge
  opencl/                   # OpenCL hash160