GpuStorageNode Phase 2: type support, incoming sets, bulk ops

[plankatron]

Summary

• Type storage: Added uint16_t type column to word pool, repurposed pair_flags for link type. Golden-ratio type hashing prevents same-name-different-type collisions (ConceptNode "foo" vs SchemaNode "foo" now get separate GPU slots).
• Incoming set scan: New gpu-incoming.cl (OpenCL) and cuda_incoming_scan (CUDA) kernels — parallel scan of pair pool with atomic match counter. fetchIncomingByType filters by link type.
• Bulk operations: loadAtomSpace, loadType, storeAtomSpace iterate GPU pool slots directly.
• 3 new test files: BasicSaveUTest (round-trip), GpuIncomingUTest (incoming + loadType), GpuBulkUTest (multi-type, collision safety, bulk, throughput showcase).

Known limitations

• Pair pool key: (min(a,b), max(a,b)) — one slot per node pair. Multiple link types between the same two nodes share a slot (last write wins). Tests use distinct node pairs per link type. See Design exploration: Full GPU AtomSpace — unified atom table, N-ary links, performance envelope #2 for design exploration on fixing this.
• Store throughput: storeAtom does per-atom host→GPU transfers (~35K nodes/sec). GpuBulkUTest uses reduced scale to pass within CI timeout. Load path is fast (~700K atoms/sec) since it reads GPU pools in bulk.

Test results

16/16 pass (12 kernel tests + 4 StorageNode tests). The test-compartment-kernel benchmark is excluded from default timeout — it runs a full GPU learning simulation (~30 min).

Files changed (13)

File	Change
opencog/gpu/CMakeLists.txt	Register gpu-incoming.cl
opencog/gpu/gpu-incoming.cl	New — OpenCL incoming scan kernel
opencog/persist/gpu/GpuBackend.h	6 new type read/write methods
opencog/persist/gpu/CudaBackend.h + .cu	CUDA type column + incoming scan
opencog/persist/gpu/OpenCLBackend.h + .cc	OpenCL type column + incoming scan
opencog/persist/gpu/GpuStorageNode.h	TypedName maps, type-aware signatures
opencog/persist/gpu/GpuStorageNode.cc	Full rewrite: type-aware store/fetch/load
tests/persist/gpu/CMakeLists.txt	Register 3 new tests
tests/persist/gpu/BasicSaveUTest.cxxtest	New — basic round-trip test
tests/persist/gpu/GpuIncomingUTest.cxxtest	New — incoming set + loadType
tests/persist/gpu/GpuBulkUTest.cxxtest	New — 6 subtests: multi-type, collision, bulk, throughput

Test plan

• 16/16 tests pass locally (RTX 2070, CUDA backend)
• Verify OpenCL-only path on non-NVIDIA hardware
• Stress test at larger scale (10K+ atoms) outside CI

🤖 Generated with Claude Code

[plankatron]

Closing — builds on wrong foundation. The WordPool/PairPool/SectionPool design is domain-specific language learning, not a general GPU AtomSpace. Need to start from the actual atomspace/ and atoms/base/ source code per Linas' guidance.