|Version 1 (modified by 10 years ago) (diff),|
Other work on nested data parallelism
- There is obviously all the work of Blelloch's group on Nesl including the implementation based on flattening.
- Later Blelloch worked on multi-threaded implementations of nested parallelism (without vectorisation, but by having a thread per loop iteration). This was mostly theoretical work http://www.cs.cmu.edu/~guyb/research.html#scheduling. Blelloch's PhD student Girija Narlikar did some experiments with a scheduler using Pthreads http://www.cs.cmu.edu/~scandal/papers/spaa99.html, but there was no real implementation.
- Then there is the work of Jan Prins' group on Proteus http://www.cs.unc.edu/Research/proteus/. I think they had some kind of vector library hooked up to an interpreter for experiments, but again nothing usable. Jan also implemented a couple of algorithms, which he vectorised manually, in imperative languages (eg, Fortran); in particular, the study about manual vectorisation in Fortran http://www.cs.unc.edu/~prins/Publications/SciProg99.pdf.
- Modula-2* of Tichy's group also recognised the value of nested data parallelism. They added it in the form of a FORALL construct to Modula-2: http://www2.cs.fau.de/download/Papers/CstarCritique.pdf?language=en, http://citeseer.ist.psu.edu/philippsen91modula.html and http://citeseer.ist.psu.edu/397019.html. They started on a compiler that targeted the MasPar (one of these SIMD machines), but I think didn't get any further.
- Using nested DO or FORALL constructs, nested data parallelism can be expressed in Fortran 95 and beyond. However, parallelising Fortran compilers cannot exploit such parallelism properly. For example, they may only parallelise the inner or outer loop. Loop parallelisation of Fortran programs is however a broad topic and there are all kinds of approaches aiming to broaden the classes of loop nests that can be parallelised. It's been a while that I looked at that stuff last.