HIPERFITTalks on Streaming in Data Parallelism
HIPERFIT Seminar Talks: Streaming in Data Parallelism
Presenters: Andrzej Filinski (Associate Professor, DIKU) and Frederik Meisner Madsen (PhD Student, DIKU)
Time: Tuesday, April 24, 15:00-16:00
Place: Aud 07 (HCØ, Universitetsparken 5)
Abstracts:
The presentation will consist of two shorter talks:
Streaming NESL (speaker: Andrzej Filinski, associate professor, DIKU)
Streaming NESL, under development in the HIPERFIT Center, is a refinement of the nested data-parallel functional language NESL. It extends NESL’s flattening-based implementation strategy and language-integrated cost model for time (in the form of work and step measures) to also achieve predictable (and optimistic) bounds for space usage. This is done by avoiding complete materialization of intermediate vector-typed values that do not actually require random access. Instead, such values are transparently materialized and processed in “chunks”, of size proportional to the available parallel computing resources - from SIMD instructions on single-core CPUs to large GPGPUs.
We give an overview of the status of the project, including a summary of the source language and cost model, a sketch of the implementation strategy in terms of a chunked-dataflow abstract machine (which may also be relevant to other language-technology projects within HIPERFIT), and an outline of some current challenges, both theoretical and practical.
Streaming Accelerate (speaker: Frederik Meisner Madsen, PhD student, DIKU)
The relatively mature embedded language Accelerate brings the power of hardware-accelerated, shape-polymorphic programming with regular arrays to Haskell. Extending Accelerate with streaming functionality, along the lines of our current research, will hypothetically improve its usability and performance for many important use-cases. This talk presents an extension to Accelerate that enables streaming by lifting a non-trivial subset of the language from working on arrays to working on sequences of arrays. The programmer obtains a sequence by slicing an array into a sequence of sub-arrays, and may return to ordinary Accelerate through reduction. Sequence expressions are evaluated efficiently in fixed-size chunks by reusing memory and compiled kernels, allowing sequences to scale in length without requiring additional computing resources except computation time.