GPU-native AtomSpace Phase 0-1: hash table + pool layouts

[plankatron]

Summary

GPU-resident data structures for the AtomSpace pipeline — the foundation layer that all subsequent kernels build on.

• gpu-hashtable.cl — Lock-free hash table using FNV-1a hashing with CAS-based insert. Provides O(1) atom identity lookups on GPU with linear probing for collision handling and support for concurrent writers.
• gpu-atomspace.cl — SoA (Structure of Arrays) pool management for WordPool, PairPool, and SectionPool. Atomic allocation via gpu_pool_alloc(), capacity tracking, and efficient GPU-side memory layout.

Pipeline Position

This is Phase 0-1 of the GPU-native AtomSpace pipeline:

[Phase 0-1: Hash Table + Pools] ← THIS PR (foundation — merge first)
 Phase 2-3: Counting + MI      (#4, depends on this)
 Phase 4-5: Sections + Cosine  (#5, depends on this)
 Phase 6:   Substitution       (#6, depends on this)

Note: PRs #4, #5, and #6 reference the pool structs (WordPool, PairPool, SectionPool) and hash table defined here. This PR should be merged first. Each subsequent PR adds only new kernel files — no merge conflicts between them.

Files

File	Lines	Description
gpu-hashtable.cl	303	Lock-free FNV-1a hash table kernel
gpu-atomspace.cl	407	Word/Pair/Section pool management
test-hashtable.c	532	Hash table unit tests
test-atomspace.c	739	Pool management unit tests
.gitignore	8	Ignore compiled test binaries

Building & Testing

cd opencog/opencl/atomspace

# Hash table tests
gcc -o test-hashtable test-hashtable.c -lOpenCL -lm
./test-hashtable

# Pool management tests
gcc -o test-atomspace test-atomspace.c -lOpenCL -lm
./test-atomspace

Relationship to Prior Work

Builds on PR #2 (binary caching + dispatch thread build), which was merged as 2a9c544. These kernels benefit from the binary caching infrastructure for faster program load times.

🤖 Generated with Claude Code