GPU-native AtomSpace Phase 2-3: pair counting + MI computation

[plankatron]

Summary

GPU kernels for the statistical learning stage — counting word co-occurrences and computing mutual information.

• gpu-counting.cl — Sentence pair counting with sliding-window approach. Processes tokenized sentences to extract word co-occurrence pairs using atomic increments for concurrent sentence processing.
• gpu-mi.cl — Mutual Information computation with two modes:
  • Resident mode: full recomputation over all pairs
  • Dirty mode: incremental updates for recently changed pairs only
  • Uses double precision to match AtomSpace FloatValue format

Pipeline Position

This is Phase 2-3 of the GPU-native AtomSpace pipeline:

 Phase 0-1: Hash Table + Pools (#3, dependency)
[Phase 2-3: Counting + MI] ← THIS PR
 Phase 4-5: Sections + Cosine  (#5)
 Phase 6:   Substitution       (#6)

Depends on PR #3 — these kernels reference WordPool, PairPool, and hash table structs defined in gpu-atomspace.cl and gpu-hashtable.cl. PR #3 should be merged first. No dependencies on PRs #5 or #6.

Files

File	Lines	Description
gpu-counting.cl	296	Sliding-window pair counting kernel
gpu-mi.cl	224	MI computation (resident + dirty modes)
test-counting.c	720	Pair counting unit tests
test-mi.c	961	MI computation unit tests

Performance

• MI throughput: 385K pairs/sec (1.24M pairs in ~3.2 seconds)
• Dirty mode avoids recomputing unchanged pairs, enabling incremental updates

Building & Testing

cd opencog/opencl/atomspace

# Pair counting tests
gcc -o test-counting test-counting.c -lOpenCL -lm
./test-counting

# MI computation tests
gcc -o test-mi test-mi.c -lOpenCL -lm
./test-mi

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