srdatalog.ir.dialects.relation.d2l.cuda

D2L dialect ↔ CUDA target plugin.

How the D2L (Device2LevelIndex) dialect emits as CUDA: registers an IndexPlugin with codegen.cuda.plugin’s registry that supplies the per-version view_count (FULL=2 for HEAD+FULL, DELTA/NEW=1) and the host-side view-setup code emitted around CUDA kernels. Importing this module — typically via import srdatalog.ir.dialects.relation.d2l — is enough to register the plugin (the D2L package’s __init__ re-imports this side-effectfully).

D2L design recap:

  • HEAD + FULL sorted arrays per relation.

  • Merge: DELTA → HEAD, O(D+H) instead of O(D+F).

  • Compaction: HEAD → FULL only when HEAD exceeds ratio% of FULL, or at fixpoint exit.

  • Join (DELTA read): 1 sorted array (same as DSAI).

  • Join (FULL read): 2 sorted arrays — kernel iterates both via the D2lSegmentLoop IIR op.

Usage in schema: Path {. index: “SRDatalog::GPU::Device2LevelIndex” .}: Relation[int, int]

Module Contents

Functions

new_two_level_plugin

Build a Device2LevelIndex plugin by starting from the default DSAI plugin and overriding the view-level hooks. Expression-level hooks are unchanged — once the kernel has a view, prefix/degree/valid/get all work identically.

Data

two_level_plugin

API

srdatalog.ir.dialects.relation.d2l.cuda.new_two_level_plugin() → srdatalog.ir.codegen.cuda.plugin.IndexPlugin[source]

Build a Device2LevelIndex plugin by starting from the default DSAI plugin and overriding the view-level hooks. Expression-level hooks are unchanged — once the kernel has a view, prefix/degree/valid/get all work identically.

srdatalog.ir.dialects.relation.d2l.cuda.two_level_plugin

‘new_two_level_plugin(…)’