Changes between Version 6 and Version 7 of Status/Oct12


Ignore:
Timestamp:
Oct 30, 2012 6:38:37 AM (18 months ago)
Author:
chak
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • Status/Oct12

    v6 v7  
    2323 * 7.8 plans: early 2013? 
    2424 
    25  * DPH 1.0? 
     25 * '''Data parallelism.''' We are currently completely rewriting our implementation of ''vectorisation avoidance'' [1] in GHC's vectoriser. This leads to an overall much simpler and more robust vectoriser. In particular, it will be more liberal in allowing scalar subcomputations imported from modules compiled without vectorisation (such as the standard Prelude). This should finally enable us to get rid of the specialised, mini-Prelude in the DPH libraries. 
     26 
     27 After having solved the problem of obtaining asymptotically work-efficient vectorisation [2], we are now turning to improving the constants in the DPH libraries, and in particular, to achieve more reliable fusion in the presence of segmented operations, folds, and parallelism. Ben Lippmeier has a few exciting ideas on major improvements in that direction that we will discuss in more detail once we have conducted more experiments. We plan to finish the new vectorisation-avoidance infrastructure in time for GHC 7.8, but the new fusion system will likely not be ready in time for that release. 
     28 
     29 Moreover, Trevor !McDonell has made good progress in devising a novel fusion system for the embedded Accelerate GPU language. We hope to be able to release it around the same time as GHC 7.8. 
    2630 
    2731 * dynlibs-by-default 
     
    3034  
    3135 * new codegen by default 
     36 
     37[1] http://www.cse.unsw.edu.au/~chak/papers/KCLLP12.html [[BR]] 
     38[2] http://www.cse.unsw.edu.au/~chak/papers/LCKLP12.html [[BR]]