srdatalog.ir.codegen.cuda.batchfile

JIT batch-file generation.

Ported from generate_batch_files_no_template.py on python-api-notemplate. Drives a mir_types.Program through the runner-struct skeleton: per ExecutePipeline, emit a JitRunner_<rule_name> struct wrapping the kernel_count / kernel_materialize / phase-decomposed setup functions.

Much of the original is a scaffold — _generate_body returns the literal string "unimplemented ", and generate_pipeline threads codegen state but dispatches to commented-out handlers for the first op. This port preserves that posture verbatim; filling in the actual C++ body for ColumnJoin / CartesianJoin / Scan / Negation / Aggregate / Filter / ConstantBind / BalancedScan roots is follow-up work to match what jit_kernel.nim emits today.

Also fixes one bug in the original: mhk’s generate_runner duplicated the setup() forward declaration twice in the phase-method block; the port drops the duplicate.

Module Contents

Functions

generate_batchfile	Top-level batchfile generator. Consumes our mir_types.Program (produced by compile_to_mir) plus the schema + program name. Emits the prelude followed by one JitRunner_... per pipeline.
generate_pipeline	Emit the kernel body for one pipeline. Dispatches on the first op’s shape to set up the outer iteration, then delegates the rest to _generate_body.
generate_prelude	Batch-file prelude: static C++ includes plus the schema-driven using aliases and device-DB typedefs.
generate_runner	Emit struct JitRunner_<rule_name> { ... } for one pipeline.

Data

PRELUDE

API

srdatalog.ir.codegen.cuda.batchfile.PRELUDE = <Multiline-String>

srdatalog.ir.codegen.cuda.batchfile.generate_batchfile(program: srdatalog.ir.mir.types.Program, schema: srdatalog.ir.codegen.cuda.schema.SchemaDefinition, name: str) → str

Top-level batchfile generator. Consumes our mir_types.Program (produced by compile_to_mir) plus the schema + program name. Emits the prelude followed by one JitRunner_... per pipeline.

srdatalog.ir.codegen.cuda.batchfile.generate_pipeline(pipeline: srdatalog.ir.mir.types.ExecutePipeline, ctx: srdatalog.ir.codegen.cuda.helpers.CodeGenContext) → str

Emit the kernel body for one pipeline. Dispatches on the first op’s shape to set up the outer iteration, then delegates the rest to _generate_body.

Matches mhk’s original: this is mostly state mutation + a few comments; the actual per-op handlers are still commented-out TODOs.

srdatalog.ir.codegen.cuda.batchfile.generate_prelude(schema: srdatalog.ir.codegen.cuda.schema.SchemaDefinition, name: str) → str

Batch-file prelude: static C++ includes plus the schema-driven using aliases and device-DB typedefs.

srdatalog.ir.codegen.cuda.batchfile.generate_runner(pipeline: srdatalog.ir.mir.types.ExecutePipeline, program_name: str) → tuple[str, str]

Emit struct JitRunner_<rule_name> { ... } for one pipeline.

Returns (full, skeleton). full is the complete struct; skeleton is reserved for a future header-only forward-decl variant (empty for now, matching mhk’s scaffolding).