gmane.comp.lang.julia.devel

Re: RFC/Proposal: Dealing with external dependencies in the Julia Package Manager

Keno Fischer — 2013-05-25T06:48:01

Yeah, that's the idea. I'm not quite sure where to take it yet, but I'll
keep working on it.

Re: Announcing a new Web REPL

bsxfan — 2013-05-25T06:11:59

I tried the REPL in the Chrome browser, but the ENTER key does nothing. So 
I can't really do anything.


On Thursday, May 23, 2013 12:41:23 AM UTC+2, Adam Savitzky wrote:

Re: Fast (multithreaded) matrix summation

Viral Shah — 2013-05-25T04:31:16

Amit did some really cool stuff with shmem, and the trick should work on 
Windows using Mmap as well. We want to think this through, and bring this 
work into base.

https://github.com/amitmurthy/PTools.jl

While it will not give general shared memory parallelism, it will make 
these kinds of operations easy to parallelize.

-viral


On Friday, May 24, 2013 6:21:55 PM UTC+5:30, bsxfan wrote:
>

Re: Google Code downloads

Viral Shah — 2013-05-25T04:13:27

I do like sourceforge download analytics. They tend to be a bit excessive 
on ads, and I wish they had the option to provide clean hosting for a fee, 
or some alternative business model that gave the site a cleaner look. That 
was one of the main reasons we preferred having google code for downloads.

-viral

On Saturday, May 25, 2013 12:20:07 AM UTC+5:30, Alessandro Andrioni wrote:

Re: Google Code downloads

Viral Shah — 2013-05-25T04:10:06

The positive side about maintaining our own infrastructure is that we will 
have access to the webserver logs, and at least get a rough estimate of the 
julia community.

-viral

On Friday, May 24, 2013 6:04:43 PM UTC+5:30, Stefan Karpinski wrote:
>

Re: RFC/Proposal: Dealing with external dependencies in the Julia Package Manager

Harlan Harris — 2013-05-25T00:10:14

1. I read it and it makes sense to me. I like the separation of concerns,
and the clear language ("working" vs "functional").

2. You used "Julep"! I'm so happy! :)

3. BinDeps.jl is interesting. Sorta a Julian DSL for build steps?


On Fri, May 24, 2013 at 8:35 AM, Keno Fischer
<kfischer-uXcxMUcfaeudFdvTe/nMLpVzexx5G7lz< at >public.gmane.org>wrote:

Re: Fast (multithreaded) matrix summation

Diego Javier Zea — 2013-05-24T23:45:58

I reproduce your result for a 1000x1000 matrix on a i7 (8 threads).
I don't see the gain of BLAST for a 20x20 matrix, my usual contingency 
table.
Your hand coded is faster than sum(A,2) in both cases.

*Why Base.sum() is slower than your hand coded version for the second 
dimension?*

Timings for the 20x20 matrix:


julia> sum([< at >elapsed sum(A,1) for i=1:1_000])
0.0031782350000000015


julia> sum([< at >elapsed sum(A,2) for i=1:1_000])
0.005420045999999989


julia> 


julia> sum([< at >elapsed sum1(A) for i=1:1_000])
0.001123687


julia> sum([< at >elapsed sum2(A) for i=1:1_000])
0.0014608899999999957


julia> 


julia> sum([< at >elapsed sumBLAS1(A) for i=1:1_000])
0.018098350999999967


julia> sum([< at >elapsed sumBLAS2(A) for i=1:1_000])
0.008673458000000005




El viernes, 24 de mayo de 2013 17:59:22 UTC-3, Laurent S escribió:

Re: Fast (multithreaded) matrix summation

Laurent S — 2013-05-24T20:59:22

Off topic, but I like your user name :-)

Fast (multithreaded) matrix summation

bsxfan — 2013-05-24T12:51:55

This is just a comment on how to accomplish fast summation over matrices 
(especially on multicore machines).

I have been wondering how efficient Julia's sum(::Matrix,::dim) is. If you 
dig deep enough, you wil see Julia generates nested loops to do this, which 
is comparable with simple hand-coded Julia loops. But, it turns out that if 
the matrix is compatible with BLAS, you can sum much faster. On an 8-core 
machine, I get:

Create data and hand-coded sum functions:

julia> A = randn(1_000,1_000);
julia> sum1(A)=((m,n)=size(A);s = Array(Float64,n);for j=1:n t=0.0; for 
i=1:m t += A[i,j] end s[j] = t end reshape(s,1,n) );
julia> sum2(A)=((m,n)=size(A);s = zeros(m);for j=1:n, i=1:m s[i] += A[i,j] 
end;s);


Time built-in sum:

julia> sum([< at >elapsed sum(A,1) for i=1:1_000])
2.455390051000001

julia> sum([< at >elapsed sum(A,2) for i=1:1_000])
7.509482612999993


Direction 2 is about 3 times slower than direction 1. 

For the hand-coded case, we get: 

julia> sum([< at >elapsed sum1(A) for i=1:1_000])

Re: Google Code downloads

Alessandro "Jake" Andrioni — 2013-05-24T18:50:07

SourceForge even has one blog post[1] about this, wow.

[1]: http://sourceforge.net/blog/google-code-projects-downloads-are-welcome/

On 24 May 2013 12:19, Elliot Saba <staticfloat-Re5JQEeQqe8AvxtiuMwx3w< at >public.gmane.org> wrote:

Re: Re: [julia-users] .+ .- .* ./ are now broadcasting

Toivo Henningsson — 2013-05-24T18:06:31

That's a neat trick. But it can't work if S or T is abstract, right?

On Friday, 24 May 2013 19:31:42 UTC+2, Jeff Bezanson wrote:
>>>

Re: Re: [julia-users] .+ .- .* ./ are now broadcasting

Jeff Bezanson — 2013-05-24T17:31:42

This can often be done using expressions like typeof(one(T)/one(S)).
On May 23, 2013 2:28 PM, "Stefan Karpinski" <stefan-X/JvLeVQC/3srOwW+9ziJQ< at >public.gmane.org> wrote:

Re: Re: Control Systems Toolbox

Adam Savitzky — 2013-05-24T16:40:43

I'm sure we could build something on top of TK that does a lot of what
simulink does, but that would be it's own package, wouldn't it?

On Fri, May 24, 2013 at 9:31 AM, Jeremy West <jemofthewest-Re5JQEeQqe8AvxtiuMwx3w< at >public.gmane.org> wrote:

Re: Re: Control Systems Toolbox

Jeremy West — 2013-05-24T16:31:54

Steve,

I agree that there needs to be a open-source Simulink alternative, its
almost exclusively what I use to get paid.  Scilab has a hard time building
on my Archlinux box, and anytime following instructions leads to something
not building I get suspicious.  I know SimJulia is more like SimMechanics
(if i understand correctly), so it wouldn't be an overlap but it definitely
would be better suited as a separate project.  I've thought about having
some basic support in the LinearControl package for combining transfer
functions in series, parallel, feedback, etc. (because that is relatively
straightforward), but a full Simulink alternative seems to be a very large
undertaking.  Worth the discussion though, I think it would be worth it.

Jeremy


On Fri, May 24, 2013 at 12:23 PM, stevej <sjaffe-gpCf3mqfBdxWk0Htik3J/w< at >public.gmane.org> wrote:

Re: Control Systems Toolbox

stevej — 2013-05-24T16:23:29


On Thursday, May 23, 2013 5:30:30 PM UTC-4, Adam Savitzky wrote:


Perhaps scilab/xcos: 
 http://www.scilab.org/scilab/features/scilab/control 
 http://www.scilab.org/scilab/features/xcos

xcos is similar to Simulink; it would be very interesting to have a 
Simulink alternative in Julia...

Steve

Re: Automatic differentiation?

Tom Short — 2013-05-24T15:26:36

There are a couple of efforts, both just getting off the ground:

https://github.com/scidom/AutoDiff.jl

https://github.com/bsxfan/ad4julia





On Fri, May 24, 2013 at 11:22 AM, John Myles White <johnmyleswhite-Re5JQEeQqe8AvxtiuMwx3w< at >public.gmane.org

Re: Automatic differentiation?

John Myles White — 2013-05-24T15:22:15

I think there's work being done, but nothing complete. I've lost track of who's in charge of it.

 -- John

On May 24, 2013, at 11:21 AM, Douglas Bates <dmbates-Re5JQEeQqe8AvxtiuMwx3w< at >public.gmane.org> wrote:

Automatic differentiation?

Douglas Bates — 2013-05-24T15:21:00

I have seen mention of automatic differentiation on this list but I am not 
sure if there are any packages providing an implementation yet.  What is 
the state of the art?

Re: Google Code downloads

Elliot Saba — 2013-05-24T15:19:06

1. Get multiple free dropbox accounts
2. Host binaries on them, load-balanced to circumvent eventually getting
banned for excessive bandwidth use
3. ???
4. Profit!
-E


On Fri, May 24, 2013 at 5:34 AM, Stefan Karpinski <stefan-X/JvLeVQC/3srOwW+9ziJQ< at >public.gmane.org>wrote:

RFC/Proposal: Dealing with external dependencies in the Julia Package Manager

Keno Fischer — 2013-05-24T12:35:13

As promised during the IRC sprint last week, I wrote up my proposal of how
to deal with external dependencies for Julia packages:
https://gist.github.com/loladiro/5643179

Please post any comments and suggestions in this thread.

Re: Google Code downloads

Stefan Karpinski — 2013-05-24T12:34:43

Jeez, man. Nobody wants to host binary downloads. We're going to need to
switch to S3 or something.

On Friday, May 24, 2013, Alessandro "Jake" Andrioni wrote: