User loginNavigation 
Scanbased Data Parallel LanguagesGiven the resurgence of interest in data parallel languages due to ubiquity of GPUs in laptops and now even phones, as well as the upcoming AVX extensions to CPU short fixedlength SSE instructions, I thought there'd be interest in the following papers I've stumbled upon by Guy Blelloch. First, scan primitives for vector computers, including segmented ones, can be efficiently implemented. Next, we can start to consider scans as primitive parallel operations in the sense of complexity models: each scan takes one step on an idealized vector machine. Once you have constant time permutes, scatters, and gathers, you can do some interesting tricks. For example, for invertible recursive tree computations (lenses anybody?), we can compute them in O(1)! For some fun examples, see 5.5 of vector models for dataparallel computing, showing the trick for tree sums. (It references Leiserson's work on contractions). When reading this work, it is worth thinking about hardware trends: work efficiency is important (parallelism is often important due to the power wall), the cost of indexed vector operations (contiguous sequential access may be better than indexed parallel), and the limited and fixed sizes of many vector units (e.g,, SSE). Finally, just like for memory, irregularity in the computation  e.g., instruction divergence  often makes the theoretical models hard to apply without losing to constant overheads. Indeed, aligning trends in CPUs and GPUs with these models seems tricky for the practicing parallel programmer, leading to studies showing CPU versions of GPU programs are often comparable. Finally, I've found it interesting that the work of Keller et al in Data Parallel Haskell and GPU spinoffs, while following the NESL transformation into scan primitives model, does not appear to do so in the advised manner for irregular structures like trees. I've been increasingly wondering why. By lmeyerov at 20101113 23:53  LtU Forum  previous forum topic  next forum topic  other blogs  3310 reads

Browse archivesActive forum topicsNew forum topics

Recent comments
1 hour 35 min ago
7 hours 45 min ago
8 hours 28 min ago
10 hours 11 min ago
11 hours 13 min ago
12 hours 51 min ago
13 hours 4 min ago
14 hours 56 sec ago
20 hours 12 min ago
20 hours 13 min ago