srdatalog.ir.codegen.cuda.orchestrator

JIT orchestrator — step_N() method bodies.

Port of src/srdatalog/codegen/target_jit/jit_orchestrator.nim. Named orchestrator_jit to avoid colliding with the existing legacy codegen/orchestrator.py (mhk’s python-api-notemplate design, which this module replaces for byte-match purposes).

Top-level entry: gen_step_body(plan, db_type_name, is_recursive, step_num, count_only_rels) = the full template <typename DB>\n  static void step_N(...) function, including fixpoint loop for recursive steps and straight-line sequence for non-recursive.

Internal structure (mirrors Nim procs): gen_index_spec_type(rel, ver, cols) — IndexSpecT<R, integer_sequence<int,...>, V> version_string(ver) — MIR version -> C++ const extract_source_info(src_spec) — (rel, ver, idx) from any source node build_dest_stream_map(exec_ops) — dest rel -> [stream indices] all_dests_independent(dest_streams) — every dest written by ≤1 stream is_count_only_pipeline(instr, cors) — dest in count-only set gen_instruction_code(instr, indent, iter_var, dest_stream_map, count_only_rels) — BIG dispatcher collect_canonical_specs(instrs) — canonical index per rel gen_fixpoint_body(plan, db, indent, cors) — recursive path gen_non_recursive_block(plan, db, indent, cors) — non-recursive path

Scope: baseline. WS / BG / dedup / fan-out variants of ExecutePipeline emission use the same dispatcher paths Nim does (the feature flags are read off the ExecutePipeline node itself, not the context), so the generated C++ is correct for those rules too — the difference is that our kernel-side emit doesn’t yet produce the matching WS / BG kernel bodies. The orchestrator half is self-contained.

Module Contents

Functions

all_dests_independent
build_dest_stream_map	dest rel → list of stream indices (one per pipeline) that write to it.
collect_canonical_specs	First (rel, cols) per relation appearing in ComputeDelta / ComputeDeltaIndex / MergeIndex ops. Used for using R_canonical_spec_t declarations and the fixpoint convergence check.
extract_source_info	Pull (rel_name, version-as-C++-code, index) from any source node.
gen_fixpoint_body	Recursive path: fixpoint loop with convergence check + maintenance ops + post-fixpoint reconstruct for each canonical rel.
gen_index_spec_type	SRDatalog::mir::IndexSpecT<Rel, std::integer_sequence<int, ...>, VER>
gen_instruction_code	Emit imperative C++ for one fixpoint-level MIR instruction.
gen_non_recursive_block	Non-recursive path. Handles Block / FixpointPlan / ExecutePipeline / PostStratumReconstructInternCols.
gen_step_body	Matches Nim’s genStepBody. Returns the full template <typename DB> static void step_N(...) { ... } function. What the fixture-dumping tool captures for each (node, is_recursive) pair in the MIR program.
is_count_only_pipeline
version_string	Convert MIR version name to C++ constant. Nim accepts both forms ("FULL_VER" and "full"); we accept both to stay compatible with whatever shape upstream callers produce.

API

srdatalog.ir.codegen.cuda.orchestrator.all_dests_independent(dest_streams: dict[str, list[int]]) → bool

srdatalog.ir.codegen.cuda.orchestrator.build_dest_stream_map(exec_ops: list[srdatalog.ir.mir.types.ExecutePipeline]) → dict[str, list[int]]

dest rel → list of stream indices (one per pipeline) that write to it.

srdatalog.ir.codegen.cuda.orchestrator.collect_canonical_specs(instrs: list[srdatalog.ir.mir.types.MirNode]) → list[tuple[str, list[int]]]

First (rel, cols) per relation appearing in ComputeDelta / ComputeDeltaIndex / MergeIndex ops. Used for using R_canonical_spec_t declarations and the fixpoint convergence check.

srdatalog.ir.codegen.cuda.orchestrator.extract_source_info(src_spec: srdatalog.ir.mir.types.MirNode) → tuple[str, str, list[int]]

Pull (rel_name, version-as-C++-code, index) from any source node.

srdatalog.ir.codegen.cuda.orchestrator.gen_fixpoint_body(plan: srdatalog.ir.mir.types.FixpointPlan, db_type_name: str, indent: str, count_only_rels: set[str] | None = None) → str

Recursive path: fixpoint loop with convergence check + maintenance ops + post-fixpoint reconstruct for each canonical rel.

srdatalog.ir.codegen.cuda.orchestrator.gen_index_spec_type(rel_name: str, version: str, cols: list[int]) → str

SRDatalog::mir::IndexSpecT<Rel, std::integer_sequence<int, ...>, VER>

srdatalog.ir.codegen.cuda.orchestrator.gen_instruction_code(instr: srdatalog.ir.mir.types.MirNode, indent: str, iter_var: str, dest_stream_map: dict[str, list[int]], count_only_rels: set[str] | None = None) → str

Emit imperative C++ for one fixpoint-level MIR instruction.

srdatalog.ir.codegen.cuda.orchestrator.gen_non_recursive_block(plan: srdatalog.ir.mir.types.MirNode, db_type_name: str, indent: str, count_only_rels: set[str] | None = None) → str

Non-recursive path. Handles Block / FixpointPlan / ExecutePipeline / PostStratumReconstructInternCols.

srdatalog.ir.codegen.cuda.orchestrator.gen_step_body(plan: srdatalog.ir.mir.types.MirNode, db_type_name: str, is_recursive: bool, step_num: int, count_only_rels: set[str] | None = None) → str

Matches Nim’s genStepBody. Returns the full template <typename DB> static void step_N(...) { ... } function. What the fixture-dumping tool captures for each (node, is_recursive) pair in the MIR program.

srdatalog.ir.codegen.cuda.orchestrator.is_count_only_pipeline(instr: srdatalog.ir.mir.types.MirNode, count_only_rels: set[str]) → bool

srdatalog.ir.codegen.cuda.orchestrator.version_string(ver: str) → str

Convert MIR version name to C++ constant. Nim accepts both forms ("FULL_VER" and "full"); we accept both to stay compatible with whatever shape upstream callers produce.