k-tuplet-search

GPU Filter Engines

Two GPU engines live here, built from the same Makefile:

kt_filter_v8 (production)

kt_filter_v8.cu is the production search engine. It is decomposed into multiple modules:

Module	File	Role
Main kernel	kt_filter_v8.cu	2-D wheel-offset launch, filter cascade
Lanes	kt_lanes.c	Per-GPU prefix-lane assignment
Checkpoint	kt_checkpoint.c	Active checkpoint / resume
Novel record	kt_novel_record.c	fsync’d hit persistence
Records	kt_records.c	Record corpus loader
CLI	kt_cli.c	Argument parser
Signal	kt_signal.cu	SIGINT / SIGTERM handler
Reporter	kt_reporter.cu	Periodic progress reporter
Tests	kt_tests.cu	Embedded unit test suite

Key features:

• Wheel-offset 2-D CUDA launch (one block per wheel slot × one thread per tile)
• --wheel-expr X#[/Y...] for non-plain-primorial Stage-0 wheels
• Barrett u128 modular reduction across L2 / ext-L2 / line-sieve stages
• Fermat-2 prefilter before host BPSW (CPU-side proving via GMP)
• --exhaustive mode with exact coverage tracking
• --prefix-mode {sequential|random} with per-GPU lane assignment
• Checkpoint / resume for multi-day campaigns
• Embedded --test suite and --validate-known gate

Build: make kt_filter_v8

Command-line options are documented in ../../docs/CUDA_CLI_REFERENCE.md. Use --wheel-expr when you want one process to search a specific quotient wheel such as 47#/31; run multiple GPU-pinned processes for a wheel pool. Measure unfamiliar expressions with --smoke --max-batches 1 before long runs because Stage-0 table size and downstream Fermat-2 load are pattern-dependent.

kt_filter_v5 (oracle / reference)

kt_filter_v5.cu is the initial GPU implementation. It uses a streams-based double-buffered launch (similar to the CC v15.cu ancestor) and a single-module build. It achieves similar aggregate throughput to CC v15, but lacks most v8 features:

• Random sampling is supported for oracle testing. v5 random mode samples raw-candidate batches with stride=1; it does not use v8’s wheel-offset tile geometry.
• No --exhaustive mode
• No checkpoint / resume
• No prefix-lane assignment
• No extended wheel (47#) support

v5 is kept as an oracle for cross-checking v8 filter correctness. It is not used for production campaigns.

Build: make kt_filter_v5

Difference summary

Feature	v5	v8
Launch style	streams + double-buffer	2-D wheel-offset
Throughput (RTX 5090)	benchmark locally	benchmark locally
Random sampling	raw-batch chunks	wheel-offset chunks
Exhaustive mode	no	yes
Checkpoint / resume	no	yes
Prefix lanes	no	yes
Extended / quotient wheels	no	47#, 47#/Y[/Z...]
Novel record persistence	shared JSONL	per-GPU JSONL
Embedded tests	legacy suite	current suite
Module decomposition	single file	8 modules

Building

cd src/cuda
make kt_filter_v8          # production binary
make kt_filter_v5          # oracle binary
make test                  # build v8 + run --test + external smokes
make clean

Default GPU architecture: sm_120 (RTX 5090 / Blackwell). Override with:

make kt_filter_v8 NVCC_ARCH=sm_89   # RTX 4090

sm_120 requires a CUDA toolkit new enough to know Blackwell; the release candidate was validated with CUDA 13.2 on RTX 5090. Older toolkits can still be used for syntax/build checks against older architectures by overriding NVCC_ARCH. For RTX 4090 / sm_89, use CUDA 11.8 or newer.

This site is open source. Improve this page.