1 / 8

Decoupled Software Pipelining

Decoupled Software Pipelining. Fuyao Zhao Mark Hahnenberg. Problem. Automatically extract general thread -level parallelism from loop bodies. Other constraints: LLVM compiler framework No custom hardware support. Dependence Analysis.

tynice
Download Presentation

Decoupled Software Pipelining

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Decoupled Software Pipelining Fuyao Zhao Mark Hahnenberg

  2. Problem • Automatically extract general thread-level parallelism from loop bodies. • Other constraints: • LLVM compiler framework • No custom hardware support

  3. Dependence Analysis • Analyze the dependences in the program and build a graph, find SCCs, coalesce into a DAG • Dependence types: • Data • True • Anti • Output • Control • Normal • Loop iteration

  4. Thread Partitioning and Code Splitting • Partition the DAG into separate threads • Copy instructions from the partition into separate, newly created functions • Initializethreads with new loop functions • Main thread waits for auxiliary threads to finish

  5. Synchronization Insertion • At points where dependences need to be communicated between threads • Insert a produce in the producer thread • Insert a consume in the consumer thread

  6. Runtime Support • Built an external library in C • Fixed-size thread-safe queues • Block on pop if queue is empty • Block on push if queue is full • Functions callable from LLVM (for produce and consume) Original program Compiled by LLVM DSWP optimized program Linked against simple_sync lib executable

  7. Results

  8. Future Work • More testing and benchmarks • Persist threads between loop bodies • Improved thread synchronization cost model

More Related