CodegenDialectPCF

-iree-pcf-convert-forall-to-loops

Converts scf.forall ops to pcf.loop

Test pass for converting scf.forall ops without mapping attributes to pcf.loop ops with sequential scope.

The input is IR containing scf.forall ops with tensor results and tensor.parallel_insert_slice terminators. Only forall ops without mapping attributes are converted.

The output replaces each matching scf.forall with a pcf.loop: - Iteration bounds come from the forall's upper/lower bounds and steps - tensor.parallel_insert_slice ops become pcf.write_slice ops - Shared output tensors become tied pcf.sref region arguments - The scope is set to #pcf.sequential for sequential execution

The underlying conversion pattern is exposed separately via convertForallToPCF() with callbacks for mapping processor IDs to custom execution scopes.

-iree-pcf-convert-sref-to-memref

Converts pcf.sref types to concrete memref types

Converts pcf.sref (shaped reference) types to memref types with appropriate layouts and memory spaces derived from the scope attribute. After this pass, there are no remaining results tied to PCF parallel ops. All pcf.sref types are converted to concrete memref types, preparing the IR for structural lowering of PCF ops to SCF/CF.

The input is expected to be bufferized IR where tied inputs to parallel ops are already memref types. Layouts are derived from these parallel op tied inputs and propagated via dataflow analysis.

Op conversions: - pcf.alloc -> memref.alloc - pcf.write_slice -> memref.subview + memref.copy or vector.transfer_write - pcf.read_slice -> memref.subview + vector.transfer_read - pcf.generic/pcf.loop region arguments become the tied memref inputs

-iree-pcf-fuse-consumers

Fuses all consumers of pcf.generic/loop ops

Test pass for fusing consumer operations into pcf.generic and pcf.loop ops.

The input is IR containing PCF parallel ops with external consumers that implement TilingInterface or are tensor.extract_slice ops.

The pass greedily fuses each tilable consumer into the producer PCF op by cloning the consumer into the PCF region, tiling it to match the iteration space, and replacing the external use with a new pcf.write_slice.

Supported fusion scenarios: - Multiple pcf.write_slice producers for a single consumer value - Fusion along multiple operands with a single pcf.write_slice per operand - tensor.extract_slice by adding a condition based on the slice bounds.

The underlying fusion patterns are exposed via matchTilableConsumer() and fuseTilableConsumer() for use in custom pipelines.

-iree-pcf-fuse-pcf-writes

Consolidates pcf.write_slice ops in loop bodies

Test pass for composing pcf.write_slice operations with nested scf.forall producers.

The input is IR containing pcf.write_slice ops where the source value is produced by an scf.forall with tensor.parallel_insert_slice in its terminator.

The pass moves each matching pcf.write_slice inside the scf.forall body, composing the slice parameters: - New offsets: write_offset + insert_offset * write_stride - New sizes: from the parallel_insert_slice - New strides: write_stride * insert_stride

This enables further lowering by ensuring writes happen at the granularity of the inner parallel loop rather than after the entire forall completes.

The underlying pattern is exposed via composeWriteSliceWithParallelInsert() for use in custom pipelines.

-iree-pcf-fuse-producers

Fuses DPS producers into pcf.generic/loop ops through tied init args.

Pass for fusing producer operations into pcf.generic and pcf.loop ops.

The input is IR containing PCF parallel ops whose tied init values are produced by operations implementing TilingInterface and DestinationStyleOpInterface (e.g. linalg.fill).

The pass matches each tied init that is the single result of a DPS producer. For each pcf.read_slice on the corresponding sref argument, it generates a tiled version of the producer via generateResultTileValue and replaces the read with the tiled result. The scoped op's init is updated to the producer's DPS init, and the original producer is erased if unused.

The underlying fusion patterns are exposed via matchTilableProducer() and fuseTilableProducer() for use in custom pipelines.

-iree-pcf-lower-structural-pcf

Lowers pcf.generic/loop to scf.execute_region/forall

Lowers structured PCF parallel ops to SCF and CF ops. This is the final step of PCF lowering, converting the abstract parallel constructs to concrete control flow that can be further lowered to target-specific code.

The input is expected to be IR where PCF ops no longer have tied results (after iree-pcf-convert-sref-to-memref). The scope attribute on each op determines how worker IDs and counts are materialized.

Op conversions: - pcf.generic -> scf.execute_region with worker IDs from scope - pcf.loop -> serialized scf.forall with iteration bounds from count operands - pcf.return -> scf.yield (in generic) or scf.forall.in_parallel (in loop) - pcf.branch_cond_return -> cf.cond_br to a return block

If sync_on_return is set on a parallel op, a barrier (determined by the scope) is inserted after the lowered op. This attribute is introduced when converting sref to memref to retain the required synchronization semantics of any tied results.

-iree-pcf-resolve-tokens

Resolves synchronization scopes on pcf.sref types.

Resolves synchronization scopes attached to pcf.sref types by expanding them to their concrete representations. This pass should run before iree-pcf-convert-sref-to-memref.

The input is IR containing pcf.sref types with sync scope attributes. The pass expands pcf.sref<..., sync_scope> types into a pcf.sref without sync scope plus any concrete types required by the sync scope attribute. For shaped refs with sync_on_return scope, the parent pcf.generic or pcf.loop op has its sync_on_return flag set to true, ensuring a barrier is inserted when the op is lowered.

Write operations (pcf.write_slice) are updated to enqueue writes through the sync scope's interface methods.