Commit Graph
43 Commits
Author SHA1 Message Date
rob thijssen 531eb7df99 refine mining metrics to reduce redundancy and improve cleanup
remove redundant estimated rate gauges and ensure all related job and
thread metrics are cleaned up on job cancellation and completion.

update GPU CUDA metrics to use backend-specific estimated rate.
2025-09-14 13:08:42 +03:00
rob thijssen ca9f97a265 adjust active thread count based on desired iterations in CUDA kernel
launch
2025-09-14 12:18:43 +03:00
rob thijssen d2842cf7ed refactor CUDA G2 batching to allocate inputs and outputs per batch 2025-09-14 11:58:40 +03:00
rob thijssen 8399bd6f7d early-exit detection and per-batch base nonce logic to CUDA miner
early-exit checks for found solutions within each batch by copying and
validating device results. Rebuild per-batch base nonces and initial
states for kernel launches using effective iters instead of
iters_per_thread. Log false positives with debug info for easier
diagnosis.
2025-09-14 11:22:44 +03:00
rob thijssen e57e91bec9 Refactor CUDA engine bounds calculation for clarity 2025-09-14 10:38:53 +03:00
rob thijssen 8288d21423 CUDA kernel batching and estimated attempt rate metrics
Support multiple batches per launch controlled by MINER_CUDA_BATCHES.
Report estimated device attempt rate (nonces/sec) to metrics per batch.
2025-09-14 10:35:51 +03:00
rob thijssen b58e54eacd Fix nonce coverage calculation to use actual covered count 2025-09-14 10:11:52 +03:00
rob thijssen 4c5fe8b36d Adjust CUDA iterations per thread based on remaining nonce range
Calculate effective iterations to avoid exceeding coverage and update
all relevant uses from configured iterations to this dynamic value.
2025-09-14 09:57:44 +03:00
rob thijssen 928f2fe4c2 Add detailed debug logs for CUDA G1 and G2 coverage prefixes 2025-09-14 09:42:15 +03:00
rob thijssen b0eedef3e5 Improve full-precision bound checks for CUDA G2 coverage and hashing 2025-09-14 09:11:55 +03:00
rob thijssen 05bb5d885f Increase CUDA kernel ABI version and add coverage bounds check
Limit per-thread iterations to prevent hashing beyond assigned coverage
range in the kernel and update kernel launch parameters accordingly.
2025-09-14 08:28:10 +03:00
rob thijssen 2a2bc1f60b debug and warn logs for CUDA G2 winner diagnostics
Update cuda-miner-3060-debug.env with optimized thread and iteration
settings
2025-09-14 05:46:26 +03:00
rob thijssen 060a851300 Bump CUDA ABI version to 2 and use sentinel values for winners
Use u32::MAX as invalid sentinel for winner coordinates to improve
result validation. Also update ABI version constant to reflect this
change.
2025-09-13 13:17:48 +03:00
rob thijssen 25fdbf6ed2 Add kernel ABI version constant and probe in CUDA engine
Probe device kernel ABI version if present for better compatibility and
add logging for fallback when winner coordinates readback fails.
2025-09-13 13:01:30 +03:00
rob thijssen fa9b90ef28 report winner thread and iteration via output buffers for nonce
reconstruction
2025-09-13 12:44:27 +03:00
rob thijssen 9739f7c08a record winner thread and iteration on device for accurate nonce
reconstruction
2025-09-13 12:33:49 +03:00
rob thijssen fbd143f1cc Fix nonce calculation by tracking per-thread base nonces 2025-09-13 12:17:15 +03:00
rob thijssen d8c96df8ba Add alternative nonce mappings to diagnose CUDA false positives 2025-09-13 12:02:40 +03:00
rob thijssen 8dada4cb61 Add additional host distance check for false positive logs 2025-09-13 11:45:24 +03:00
rob thijssen 734b4fbb74 richer diagnostics for cuda false positive logging 2025-09-13 07:56:47 +03:00
rob thijssen 5fe204349f Add debug output buffers to CUDA G2 kernel and log on false positive
rejection
2025-09-13 07:43:18 +03:00
rob thijssen 16057a8e20 Add optional CUDA sampler readback and parity check
Enable sampler debugging via MINER_CUDA_SAMPLER env var to verify
host/device parity of sampler outputs, logging warnings on mismatch

Add example env file for RTX 3060 debug configuration
2025-09-13 06:24:28 +03:00
rob thijssen 0aaacdf85d Add optional CUDA sampler support and big-endian SHA3 output
- Introduce constant memory variables and device buffers for sampler -
Initialize sampler in CudaEngine when MINER_CUDA_SAMPLER=1 - Change SHA3
kernel to output digest bytes in big-endian format - Add sample mismatch
metric counter
2025-09-13 04:02:06 +03:00
rob thijssen 191cc799b7 fix: cuda metrics and initialize estimated rate gauge 2025-09-13 02:42:48 +03:00
rob thijssen 300fa89927 Add optional metrics and big-endian limb support to CUDA engine
- Add `metrics` dependency and feature to engine-gpu-cuda crate -
Introduce constant memory for target and threshold values in CUDA kernel
- Implement limb-wise big-endian conversions for distance calculations -
Use constant memory when available for target/threshold within kernel -
Emit backend label metric for GPU CUDA engine - Track false positives
with metrics on host verification failure
2025-09-13 01:54:54 +03:00
rob thijssen 4bffb144cb Add host re-verification for CUDA results and new metrics
Reject false positives from device by verifying on host and add
Prometheus counters for estimated rates, found and false-positive
candidates per engine and job.
2025-09-13 01:16:23 +03:00
rob thijssen 92f58d1b76 immutable bindings for is_g2 and d_distance 2025-09-13 00:10:40 +03:00
rob thijssen 4ca906ac1b corrected host-side constant upload to pass CStrings 2025-09-12 23:55:04 +03:00
rob thijssen 398c9f9bc9 Use __constant__ memory for per-job constants in CUDA kernels
Add optional __constant__ symbols for G2 kernel constants and prefer
these over parameter pointers when available. Update host code to
attempt populating these constant symbols for faster access.
2025-09-12 18:14:26 +03:00
rob thijssen 69cbb7fa99 remove 1024 thread cap in CUDA kernel launches 2025-09-12 16:24:45 +03:00
rob thijssen 1d956c499e Allocate and copy-back CUDA output buffer for G1 path
Add optional pinned host buffer support and parallel host SHA3
validation over GPU results, updating hash count and early exit on
success.
2025-09-12 16:00:52 +03:00
rob thijssen 297a74181a Refactor CUDA kernel launch to conditionally allocate output buffer
Allocate d_y_out only for G1 kernel instead of always allocating before
launch, and track kernel type with is_g2 flag for clearer control flow.
2025-09-12 15:52:21 +03:00
rob thijssen 9167b8e48d Refactor CUDA kernel launch to separate G1 and G2 modes
Reorder code to launch G2 kernel with on-device SHA3 and early-exit
handling separately from G1 kernel that returns y values. Rename index
variables for clarity and adjust logging to measure kernel execution
time in both branches.
2025-09-12 15:09:32 +03:00
rob thijssen baf15fbb75 Add G2 kernel with device SHA3-512, early-exit, and threshold compare
- Implemented qpow_montgomery_g2_kernel for on-device SHA3-512 hashing -
Added early-exit flag and linear index tracking in G2 kernel -
Integrated G2 kernel launch and result handling in Rust CUDA engine -
Introduced MINER_CUDA_MODE=g2 config option to select G2 kernel - Added
MINER_CUDA_PINNED option for pinned host buffers and async copies -
Updated README with new runtime config knobs and G2 usage instructions
2025-09-12 14:38:13 +03:00
rob thijssen 5b44ef20df Use optional pinned host buffer and async copy for CUDA results
Enable using a page-locked pinned buffer with asynchronous copy-back
when MINER_CUDA_PINNED is set, otherwise fall back to synchronous copy.
2025-09-12 14:00:53 +03:00
rob thijssen c0fa1d15aa Limit SHA3 host hashing to remaining range and use scoped threads 2025-09-12 12:56:58 +03:00
rob thijssen 910e614cfb Add MINER_CUDA_IMAGE override and parallel host SHA3 workers
- Allow forcing PTX or CUBIN image via MINER_CUDA_IMAGE env var - Enable
selecting G2 kernel with fallback to G1 via MINER_CUDA_MODE=g2 -
Allocate device constant buffers once outside loop - Implement parallel
host SHA3 over GPU results using MINER_CUDA_HASH_THREADS - Update README
with new env vars and usage notes
2025-09-12 12:40:47 +03:00
rob thijssen 9bb6dc840d Log SHA3 host timing after GPU result consumption 2025-09-12 11:15:42 +03:00
rob thijssen 14115a12c1 Add timing logs for CUDA kernel and copy-back operations 2025-09-12 11:09:28 +03:00
rob thijssen aef8c865cb clippy cleanup 2025-09-12 08:34:54 +03:00
rob thijssen 8b02b790a4 wip: cubin embed 2025-09-12 04:22:15 +03:00
rob thijssen 222890dc92 cuda scaffold
- Add engine-gpu-cuda crate with basic CUDA kernel (qpow G1) - Implement
CUDA build.rs compiling .cu to PTX when "cuda" feature enabled -
Integrate GPU engine into miner-service with runtime CUDA availability
check - Add fallback to CPU engine in gpu-cuda until full GPU kernel
support is ready - Enable optional "cuda" feature in miner-cli and
miner-service crates
2025-09-11 18:51:52 +08:00
rob thijssen 0d5adf583b 01: soc and metrics
Convert repository into Cargo workspace with modular crates

- Split code into workspace members: miner-cli, miner-service, pow-core,
engine-cpu, engine-gpu-cuda, engine-gpu-opencl, metrics - Moved QPoW
math logic into local pow-core crate with compatibility API and
incremental helpers - Defined MinerEngine trait and baseline CPU engine
in engine-cpu crate - Rewrote service layer in miner-service: job
management, HTTP API, engine orchestration - Added CLI binary crate
miner-cli with flags: --port, --num-cores, --metrics-port, --engine -
Implemented runtime toggle for Prometheus metrics exporter enabled only
when --metrics-port provided - Deprecated root src/lib.rs and
src/main.rs with notices pointing to new CLI/workspace entrypoints
2025-09-07 20:21:10 +08:00