Thank you for your interest and advice. How about focusing on "implementations of massively parallel dense linear algebra routines" and "implementing native Julia algorithms involving efficient, cache-conscious matrix operations on tiled matrices."? Because I interested cache-conscious parallel algorithm using tiled matrices. Can you give me advice to specify the proposal?
I'm sorry, the code I show you might be messy. It's because there are many kind of parallelism put together. and I think I can not finish all of implementation there until this Friday(Because it should make a dependency graph, create parallel queue line using it and make sub-matrices and control them to be operated in pipelining parallelism using lowered dependency and so on). How about I'll give you parallel matrix-matrix multiplication implementation in Julia using tiled matrix, I think I can do it within one or two days. P.S. I'm sorry for my late reply, I forgot checking google-groups. Best, Yonghyun. 2016년 3월 19일 토요일 오전 11시 13분 1초 UTC+9, Jiahao Chen 님의 말: > > > I tried creating my own matrix power function using Python before, I > used data parallelism(tiling), task parallelism(using topology) and > pipelining parallelism(using lowered dependency). > > As you say, attacking all levels of parallelism is very ambitious. > However, I think working on just one of these parallelism structures would > make for a good summer project. I would recommend you to pick one of these > approaches and apply it to a widely used computation such as matrix-matrix > multiplication. > > I had a look at your Python implementation > > https://github.com/usefulhyun/parallel_mmm/blob/master/prllmpow/prllmpow.py > > and it is quite hard to understand. If you can translate the essential > parts into Julia and show how you can use features like Julia types and > overloading of Julia's generic functions like * to make the code readable > yet efficient, then I think we can make a good case for your participation > in the Google Summer of Code. Without a Julia code sample to evaluate, it > is quite difficult to make a strong case for participation. >
