6 ms·

Would the video, slides (and corresponding source repos) be uploaded sometime ? Would love to see those. Reading between the lines, devectorize.jl and the vect

by srean 12y ago

Would the video, slides (and corresponding source repos) be uploaded sometime ? Would love to see those. Reading between the lines, devectorize.jl and the vectorization framework will make for a heady mix. I am glad that Julia is challenging the traditional mantra of "vectorize (as in the R/Numpy/Matlab sense) the loops". It is heart warming to see that the language designers get it that this style has inherent speed limitations (time is spent filling and copying temporary vectors). If any Julia core developer is seeing this, just a shout out, (i) to say you are doing awesome work, and (ii) @ macros available in the standard library are not documented well. It would be good to have a page that describes what is available now. EDIT: Just clarifying my point. Macros as a feature are quite well described in the docs. What I miss is an index of macros that come built in with the standard library. They would help both with pedagogy and with putting them to actual use, for example removing bounds checks. If they are a little more discoverable than they are now that would be very helpful to a newcomer. For a new'ish language the documentation is otherwise quite admirable. @ViralBShah I filed an issue.

ViralBShah 12y ago
The videos are being edited and will be available shortly. We'll post them on juliacon.org and julialang.org. Does the Metaprogramming section of the manual not adequately describe macros? It would be great if you submit a pull request or an issue describing what is missing.
ViralBShah 12y ago
An index of macros in the standard library is a good idea. Would you mind filing an issue?
- StefanKarpinski 12y ago
  Macros are documented along with everything else in the standard library: http://docs.julialang.org/en/latest/stdlib/base/ http://docs.julialang.org/en/latest/stdlib/base/ The stdlib docs are organized by subject matter. What would the benefit of having a separate index of macros be? If one wants a list of exported macros, that's just a matter of doing grep @ base/exports.jl in a Julia repo.
jarvist 12y ago
Slides are on GitHub: https://github.com/JuliaCon/presentations https://github.com/JuliaCon/presentations I believe the video is heading for their YouTube channel: https://www.youtube.com/user/JuliaLanguage/featured https://www.youtube.com/user/JuliaLanguage/featured
scott_s 12y ago
I had to look up Devectorize.jl: https://github.com/lindahua/Devectorize.jl https://github.com/lindahua/Devectorize.jl The point, I think, is not that vectorizing is slow. It's that you don't want to be myopic when you do it. Your point about time spent filling and copying vectors is independent of what Devectorize is trying to do. Yes, when you vectorize code, you need to have a prologue and an epilogue which gets the data in the correct format for the SIMD operations. This prologue and epilogue has a cost, and if it's larger than the gain from the SIMD operations, then it's not worth vectorizing. (This can happen if, say, you only have 4 elements to vectorize.) However, we have to do this cost-analysis all the time. Figuring out when it's beneficial to vectorize code has been around as long as we've vectorized code. Devectorize.jl is not about that problem. Rather, it's about when you have a chain of expressions, and the operators in those expressions have implied loops. The naive thing to do is to myopically execute each one of those loops, creating and passing around temporary vectors. The Devectorize framework is given an entire expression, and is able to analyze where the implied loops are, and figures out how to express that computation in a single loop. We can tell these two concepts are independent because the ideal situation is to first use Devectorize on an expression, and then vectorize the result! For the record, the approach taken by Devectorize.jl is similar to the problem that expression templates (http://en.wikipedia.org/wiki/Expression_templates http://en.wikipedia.org/wiki/Expression_templates) in C++ try to solve. Through template trickery and operator overloading, we can put off evaluating an expression, avoiding the temporaries and unnecessary memory traversal that a naive execution implies. The Boost library uBLAS (micro Base Linear Algebra Subprograms, http://www.boost.org/doc/libs/1_55_0/libs/numeric/ublas/doc/index.htm http://www.boost.org/doc/libs/1_55_0/libs/numeric/ublas/doc/...) uses this technique.
- srean 12y ago
  Not sure that I understand the point/tone of your comment, all that you say is correct and I am not unaware of any of those points nor do I contradict them in mine, but it seems that you are offering it as a correction. From your comment it seems that you are aware of the two different meanings of "vectorization" but that you are choosing the wrong one to setup a strawman. So I am a bit puzzled. In any case let me offer some clarifications: The canonical example of ET and the one that started it all is Blitz++ (unlike its modern descendants it actually handles full n-dimensional tensors). More recent solutions are Blaze, Eigen and Armadillo. I like these over uBLAS (more verbose and not as performant). Blaze is typically the fastest (in this class) as long as you are dealing with 2D arrays and linear algebraic operations. Since Julia is homoiconic and has hygienic macros, one does not have to resort to such syntactically complex ways of doing metaprogramming such as ETs in C++. ET's are great to use, not very pleasant to write the machinery for, although things like Boost::spirit and fusion make it more tolerable. Heaven forbid that you have to understand the error messages though. Now, to another point, looping is so pathetically slow in R, Numpy and Matlab (in that order) that the traditional solution offered in these languages is to vectorize (in the Matlab/numpy/R sense not in the SIMD sense) the loops into vector-expressions. Often this needs the help of extra prefilled vectors/matrices, (some languages do it smartly with 'broadcasting', some where late to pick it up, looking at you Matlab). Even with broadcasting tricks, the extra stride indirection in the iteration slows down the code from how fast it could have been. That is but one aspect of it, the other is chaining several binary and unary operations together. This also creates temporaries and unnecessary loops (unless you use solutions of somewhat limited scope like numexpr, which BTW is lovely when it is applicable, it also does thread level parallelization as well as SIMD if you can afford to link with intel MKL). In Julia there is no need to avoid loops because they are inherently fast. In fact they are faster than the vectorized expressions, because of inefficiencies mentioned. However, loops are a lot more verbose than expressions. So this is where devectorize comes in. It transforms those expressions into loops (much like ETs) which can then be JITed to obtain superior performance. Now with SIMD vectorization thrown in on top, one can gain even more, because now one can benefit further from the instruction level parallelism. So yes you devectorize and then vectorize (and you already know this) and the term "vectorization" refer to two different concepts named vectorization , the clash in terminology is rather unfortunate. EDIT: @scott_s no offense taken and upvoted. As I said, this post was just to offer some clarifications, the clash in terminology indeed gets very confusing. So, thanks for making me improve my comment, if it confused you I am sure it would have confused others as well.

by srean 12y ago

ViralBShah 12y ago

The videos are being edited and will be available shortly. We'll post them on juliacon.org and julialang.org. Does the Metaprogramming section of the manual not adequately describe macros? It would be great if you submit a pull request or an issue describing what is missing.

ViralBShah 12y ago

An index of macros in the standard library is a good idea. Would you mind filing an issue?

StefanKarpinski 12y ago
Macros are documented along with everything else in the standard library: http://docs.julialang.org/en/latest/stdlib/base/ http://docs.julialang.org/en/latest/stdlib/base/ The stdlib docs are organized by subject matter. What would the benefit of having a separate index of macros be? If one wants a list of exported macros, that's just a matter of doing grep @ base/exports.jl in a Julia repo.

jarvist 12y ago

Slides are on GitHub: https://github.com/JuliaCon/presentations https://github.com/JuliaCon/presentations I believe the video is heading for their YouTube channel: https://www.youtube.com/user/JuliaLanguage/featured https://www.youtube.com/user/JuliaLanguage/featured

scott_s 12y ago

I had to look up Devectorize.jl: https://github.com/lindahua/Devectorize.jl https://github.com/lindahua/Devectorize.jl The point, I think, is not that vectorizing is slow. It's that you don't want to be myopic when you do it. Your point about time spent filling and copying vectors is independent of what Devectorize is trying to do. Yes, when you vectorize code, you need to have a prologue and an epilogue which gets the data in the correct format for the SIMD operations. This prologue and epilogue has a cost, and if it's larger than the gain from the SIMD operations, then it's not worth vectorizing. (This can happen if, say, you only have 4 elements to vectorize.) However, we have to do this cost-analysis all the time. Figuring out when it's beneficial to vectorize code has been around as long as we've vectorized code. Devectorize.jl is not about that problem. Rather, it's about when you have a chain of expressions, and the operators in those expressions have implied loops. The naive thing to do is to myopically execute each one of those loops, creating and passing around temporary vectors. The Devectorize framework is given an entire expression, and is able to analyze where the implied loops are, and figures out how to express that computation in a single loop. We can tell these two concepts are independent because the ideal situation is to first use Devectorize on an expression, and then vectorize the result! For the record, the approach taken by Devectorize.jl is similar to the problem that expression templates (http://en.wikipedia.org/wiki/Expression_templates http://en.wikipedia.org/wiki/Expression_templates) in C++ try to solve. Through template trickery and operator overloading, we can put off evaluating an expression, avoiding the temporaries and unnecessary memory traversal that a naive execution implies. The Boost library uBLAS (micro Base Linear Algebra Subprograms, http://www.boost.org/doc/libs/1_55_0/libs/numeric/ublas/doc/index.htm http://www.boost.org/doc/libs/1_55_0/libs/numeric/ublas/doc/...) uses this technique.

srean 12y ago
Not sure that I understand the point/tone of your comment, all that you say is correct and I am not unaware of any of those points nor do I contradict them in mine, but it seems that you are offering it as a correction. From your comment it seems that you are aware of the two different meanings of "vectorization" but that you are choosing the wrong one to setup a strawman. So I am a bit puzzled. In any case let me offer some clarifications: The canonical example of ET and the one that started it all is Blitz++ (unlike its modern descendants it actually handles full n-dimensional tensors). More recent solutions are Blaze, Eigen and Armadillo. I like these over uBLAS (more verbose and not as performant). Blaze is typically the fastest (in this class) as long as you are dealing with 2D arrays and linear algebraic operations. Since Julia is homoiconic and has hygienic macros, one does not have to resort to such syntactically complex ways of doing metaprogramming such as ETs in C++. ET's are great to use, not very pleasant to write the machinery for, although things like Boost::spirit and fusion make it more tolerable. Heaven forbid that you have to understand the error messages though. Now, to another point, looping is so pathetically slow in R, Numpy and Matlab (in that order) that the traditional solution offered in these languages is to vectorize (in the Matlab/numpy/R sense not in the SIMD sense) the loops into vector-expressions. Often this needs the help of extra prefilled vectors/matrices, (some languages do it smartly with 'broadcasting', some where late to pick it up, looking at you Matlab). Even with broadcasting tricks, the extra stride indirection in the iteration slows down the code from how fast it could have been. That is but one aspect of it, the other is chaining several binary and unary operations together. This also creates temporaries and unnecessary loops (unless you use solutions of somewhat limited scope like numexpr, which BTW is lovely when it is applicable, it also does thread level parallelization as well as SIMD if you can afford to link with intel MKL). In Julia there is no need to avoid loops because they are inherently fast. In fact they are faster than the vectorized expressions, because of inefficiencies mentioned. However, loops are a lot more verbose than expressions. So this is where devectorize comes in. It transforms those expressions into loops (much like ETs) which can then be JITed to obtain superior performance. Now with SIMD vectorization thrown in on top, one can gain even more, because now one can benefit further from the instruction level parallelism. So yes you devectorize and then vectorize (and you already know this) and the term "vectorization" refer to two different concepts named vectorization , the clash in terminology is rather unfortunate. EDIT: @scott_s no offense taken and upvoted. As I said, this post was just to offer some clarifications, the clash in terminology indeed gets very confusing. So, thanks for making me improve my comment, if it confused you I am sure it would have confused others as well.