[julia-users] Re: obtaining pointer to function

[pevnak]

Hi Steven,
thank you very much for the clarification. It was indeed possible to pass 
the length if the vector and with pointer_to_array it works smoothly.

Tomas



On Tuesday, 2 August 2016 13:46:47 UTC+2, Steven G. Johnson wrote:
>
>
>
> On Tuesday, August 2, 2016 at 7:39:16 AM UTC-4, pev...@gmail.com wrote:
>>
>> Hi All,
>> I am trying to bind a fortran library for optimization (
>> http://napsu.karmitsa.fi/lmbm/) to Julia.
>> To do so, I would like to get a pointer to function written in Julia, 
>> which I try to do as
>>
>> function fOpt(x::Array{Float64,1},g::Array{Float64,1})
>>   fill!(g,1.0);
>>   convert(Cdouble,sum(x))::Cdouble
>> end
>> const fOptPtr = cfunction(fOpt, Cdouble, (Ptr{Cdouble}, Ptr{Cdouble}))
>>
>>
> An Array{T} is not the same thing as a Ptr{T}.
>
> You need to declare fOpt to take Ptr{Cdouble} arguments.  You can use the 
> pointer_to_array function to convert pointers to Arrays.  (However, you 
> need to know the length of the array; I'm surprised that the length is not 
> passed as an argument — are you supposed to use a global?)
>

[julia-users] obtaining pointer to function

[pevnak]

Hi All,
I am trying to bind a fortran library for optimization 
(http://napsu.karmitsa.fi/lmbm/) to Julia.
To do so, I would like to get a pointer to function written in Julia, which 
I try to do as

function fOpt(x::Array{Float64,1},g::Array{Float64,1})
  fill!(g,1.0);
  convert(Cdouble,sum(x))::Cdouble
end
const fOptPtr = cfunction(fOpt, Cdouble, (Ptr{Cdouble}, Ptr{Cdouble}))

Though the call of cfunction return error:

*ERROR: cfunction: no method exactly matched the required type signature 
(function not yet c-callable)*

* in cfunction at c.jl:9*



Can anyone suggest what I am doing wrong?

I use Julia Version 0.4.5 (2016-03-18 00:58 UTC) on OS X.


Thank you very much for help.


Tomas

Re: [julia-users] Re: Calling all users of ParallelAccelerator.

[pevnak]

Hi Todd,
I have been looking at latte and it does not seem to be useful for me, 
since I need some special constructs and they are just not available.
Nevertheless, I would like to ask you, if Latte uses parallelization? In my 
own implementation, I am struggling to exploit multi-core hw.
Thank you very much.
Best wishes,
Tomas




On Saturday, 23 July 2016 05:42:57 UTC+2, Todd Anderson wrote:
>
> You may also want to look at another IntelLabs project on GitHub called 
> Latte.  It provides a DSL for deep neural networks in Julia.
>
> Todd
>
> --
> *From: *pev...@gmail.com 
> *To: *"julia-users" 
> *Sent: *Friday, July 22, 2016 8:07:19 PM
> *Subject: *[julia-users] Re: Calling all users of ParallelAccelerator.
>
> Hi Todd,
>>
> I have tried several times to use ParallelAccelerator to speed up my toy 
> Neural Network library, but I never had any significant performance boost. 
> I like the idea of the project a lot, sadly I was never able to fully 
> utilise it.
>
> Best wishes,
> Tomas 
>
>

[julia-users] Re: Calling all users of ParallelAccelerator.

[pevnak]

>
> Hi Todd,
>
I have tried several times to use ParallelAccelerator to speed up my toy 
Neural Network library, but I never had any significant performance boost. 
I like the idea of the project a lot, sadly I was never able to fully 
utilise it.

Best wishes,
Tomas 

[julia-users] Re: asynchronous reading from file

[pevnak]

Hi James,
thank for the reply. Though in your implementation the reading is not in a 
separate process / thread, as I expect that you are bound to IO operations. 
In my problem there is computationally intensive post-processing. Should I 
modify the iotask as

iotask = @task begin
info("reading from stdin")
for i in 1:20
s @spawn loaddata()
produce(s)
end
end

Do I need to have the consumer of s wrapped as another task? Meaning my 
stochastic gradient descend loop will look like your worked and does the 
stochastic gradient descend needs to produce something? I would like to 
understand the details.

Thanks for the answer.

Tomas

[julia-users] asynchronous reading from file

[pevnak]

Hi All,
I would like to implement an asynchronous reading from file.

I am doing stochastic gradient descend and while I am doing the 
optimisation, I would like to load the data on the background. Since 
reading of the data is followed by a quite complicated parsing, it is not 
just simple IO operation that can be done without CPU cycles.

the skeleton of my current implementation looks like this

rr = RemoteChannel()
@async put!(rr, remotecall_fetch(loaddata,2)

  for ii in 1:maxiter
#do some steps of the gradient descend

   #check if the data are ready and schedule next reading
if isready(rr)
  append!(dss[1],take!(rr));
@async put!(rr, remotecall_fetch(loaddata,2)
 end
   end


nevertheless the isready(rr) always returns false, which looks like that 
the data are never loaded.

I start the julia as julia -p 2, therefore I expect there will be a 
processor.

Can anyone explain me please, what am I doing wrong?
Thank you very much.

Tomas

Re: [julia-users] poor performance of threads

[pevnak]

Dear Sam,
the output of the benchmark is following

105.290122 seconds (31.43 k allocations: 1.442 MB, 0.00% gc time)
107.445101 seconds (1.37 M allocations: 251.368 MB, 0.12% gc time)

Tomas

Re: [julia-users] poor performance of threads

[pevnak]

Thank you very much Tim.
I am using the profiler and your package ProfileView quite extensively and 
I know where is my Achille heel in the code, and it is cpu bound.  That's 
why I am so puzzled with threads.

I will try to use @code_warntype, never use it before. 

Best wishes,
Tomas

[julia-users] Bug in daxpy! ???

[pevnak]

Hello all,
I was polishing my call and I have found the following definition of daxpy! 
I was not aware of


function axpy!{Ti<:Integer,Tj<:Integer}(α, x::AbstractArray, 
rx::AbstractArray{Ti}, y::AbstractArray, ry::AbstractArray{Tj})
if length(x) != length(y)
throw(DimensionMismatch("x has length $(length(x)), but y has 
length $(length(y))"))
elseif minimum(rx) < 1 || maximum(rx) > length(x)
throw(BoundsError(x, rx))
elseif minimum(ry) < 1 || maximum(ry) > length(y)
throw(BoundsError(y, ry))
elseif length(rx) != length(ry)
throw(ArgumentError("rx has length $(length(rx)), but ry has length 
$(length(ry))"))
end
for i = 1:length(rx)
@inbounds y[ry[i]] += x[rx[i]]*α
end
y
end

Is the first check
 length(x) != length(y)
really an intended behavior? 

The multiplication goes over indexes rx and ry, should not be the check 
 length(rx) != length(ry) ?

Thanks for the clarification.
Tomas

Re: [julia-users] poor performance of threads

[pevnak]

Thanks a lot for the suggestions.
As I have mentioned, it was really a toy problem, but I am not getting a 
significant speedup on a bigger problem, where threads are nicely separated 
either and the problem is very CPU bound either. I would be very interested 
to know about tool that would point out to problems with cache and memory 
access.

Tomas

[julia-users] poor performance of threads

[pevnak]

Hi All,
I would like to ask if someone has an experience with Threads as they are 
implemented at the moment in the master branch.
After the successful compilation  (put JULIA_THREADS=1 to Make.user)
I have played with different levels of granularity, but usually the code 
was slower or more or less the same speed as single threaded version. I 
have even tried a totally stupid execution like this

using Base.Threads;
function one()
  x=randn(100);
  f=0;
  for i in x
f+=i;
  end
end

function two()
  x=randn(100);
  f=zeros(nthreads())
  @inbounds @threads for i in 1:length(x)
f[threadid()]+=x[i];
  end
  sum(f)
end

one()
@time one()

two()
@time two()

and the times on my 2013 Macbook air were

  0.068617 seconds (2.00 M allocations: 38.157 MB, 9.72% gc time)
  0.394164 seconds (5.72 M allocations: 99.015 MB, 5.00% gc time)

Wov, that is quite poor. I would expect an overhead, but not big like this.

Can anyone suggest, what is going wrong? I have been trying a profiler, but 
it does not help. It seems that it does not work with Threads at the 
moment. Or, is it because Threads are still not really supported.

I would like to get speed-up showing in this video
https://www.youtube.com/watch?v=GvLhseZ4D8M

Any suggestions welcomed.
Tomas

[julia-users] cannot compile Julia with threading support

[pevnak]

Hi All,
I wanted to try julia with threads, therefore I have cloned GIT repository, 
checkout version 0.4.3

git checkout release-0.4

I have put JULIA_THREADS=1, here is my Make.user

CC=/opt/rh/devtoolset-2/root/usr/bin/gcc
CXX=/opt/rh/devtoolset-2/root/usr/bin/g++
JULIA_THREADS=1

and compiled julia. Nevertheless I do not see the support threads, since I 
do not see module Base.Threads.

Does anyone has experience with this? Can anyone recommend, what should I 
do or what I did wrong?
Thank you very much for any help.

Best wishes,
Tomas

[julia-users] Tutorials and examples on threading

[pevnak]

Hi All,
I know that julia does not have at the moment native support for threads, but 
it exist in an experimental branch. Is there any tutorial on this, such that 
average user can try it? I would like to try it speed my code. I do separate 
processes at the moment and I think the overhead is enormous.
Thanks a lot for the answers.
Tomas

[julia-users] Type of composite type

[pevnak]

Hello,
I have a problem to which I have found a dirty solution and I am keen to 
know, if there is a principal one.

I have a composite type defined as

type Outer{T}
 A::T
 B::T
end

where A and and B are composite types

Then I want to create constructor
function Outer(k::Int)
  return(Outer(A{T}(k),B{T}(k))
end

But I have not find a way to put there the type information.
The only dirty hack I have come with to define the outer constructor as 

function Outer(k::Int;T::DataType=Float32)
  return(Outer(A{T}(k),B{T}(k))
end

But I do not like this solution too much. It is little awkward.

Thanks for suggesting a cleaner solution.

Best wishes,
Tomas

[julia-users] samples of different sizes in mocha

[pevnak]

Hi,
I have a question related to use of Mocha, particularly if I can tweak it 
to my problem. I want to use the library for multi-instance learning, which 
means that each sample is composed of multiple instances, but the number of 
instances in each sample differs from sample to sample. You can imagine 
this as a 2D image, where each image has different height, but the width of 
all images is the same.
When the pooling operation is performed over the height, then we get sample 
of fixed size and one can follow the usual Neural Nets, where each sample 
has the same size. 
What I feel is that the usual data provider cannot be used here, as it 
assumes that all samples are of fixed size. Am I right?
Is there a way, how this can be fixed? 

Thank you very much for help and I apologise if this is not the right place 
to ask my question, though I do not know about any other.

[julia-users] alignement in memory when allocating arrays

[pevnak]

Hi all,
I would like to ask if it is possible to enforce memory alignment for large 
arrays, such that that arrays are aligned to 64 bytes and can be 
efficiently with SIMD instructions. I assume to call library functions 
wirtten in c/c++.
Thanks for the answer.
Tomas