srdatalog.ir.dialects.parallel

Parallelism-strategy dialects.

Each strategy lowers a ParallelFor body to a target-specific work- distribution shape:

par.data.warp_strided GPU baseline grid-stride-by-warp loop par.data.block_group GPU work-balanced — per-block cumulative-work binary search, intra-block warp redistribution par.data.atomic_ws GPU work-stealing — WCOJTask queue + range board par.data.tbb_for CPU TBB parallel_for with blocked_range par.scalar No parallelism (debug / reference)

See docs/ir_lowering_semantics.md §16 for the full taxonomy. The strategies are independent dialects; each contributes one $\pi_W$ to iir.cf.ParallelFor via lowering registration.

Subpackages

• srdatalog.ir.dialects.parallel.data
  • srdatalog.ir.dialects.parallel.data.block_group
  • srdatalog.ir.dialects.parallel.data.print