From otter at otter-rsch.com  Mon Sep 15 12:03:48 2014
From: otter at otter-rsch.com (otter)
Date: Mon, 15 Sep 2014 12:03:48 -0700
Subject: [Developers] faster ludecomp in C++ for  creating adjoint code
Message-ID: <54173814.3030902@otter-rsch.com>

After much pain I have produced a C++ version of the LU decomposition which
is suitable for producing adjoint code for ADMB and perhaps cppad. 
(Don't know what adjoint
code for cppad is as yet!)  This code is about 25 times faster than the 
current
ADMB code for a 2,000 x 2,000 matrix and about 4 times slower than the 
Openblas
code which contains optimized assembler and Fortrash.  Any suggestion 
for improvements
would be welcome.  Per usual I will hold my breath.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: unblocked.cpp
Type: text/x-c++src
Size: 2692 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/b3e7b1f0/attachment.cpp>

From johnoel at hawaii.edu  Mon Sep 15 12:14:51 2014
From: johnoel at hawaii.edu (Johnoel Ancheta)
Date: Mon, 15 Sep 2014 09:14:51 -1000
Subject: [Developers] faster ludecomp in C++ for creating adjoint code
In-Reply-To: <54173814.3030902@otter-rsch.com>
References: <54173814.3030902@otter-rsch.com>
Message-ID: <CAJMx2XXamtGvnc5Bz-5VkzmCgrBk24RZiBg63edn+Z6wHxWfng@mail.gmail.com>

Thanks Dave!  I'll provide feedback after testing...



On Mon, Sep 15, 2014 at 9:03 AM, otter <otter at otter-rsch.com> wrote:

> After much pain I have produced a C++ version of the LU decomposition which
> is suitable for producing adjoint code for ADMB and perhaps cppad. (Don't
> know what adjoint
> code for cppad is as yet!)  This code is about 25 times faster than the
> current
> ADMB code for a 2,000 x 2,000 matrix and about 4 times slower than the
> Openblas
> code which contains optimized assembler and Fortrash.  Any suggestion for
> improvements
> would be welcome.  Per usual I will hold my breath.
>
>
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/94864821/attachment.html>

From davef at otter-rsch.com  Mon Sep 15 12:19:50 2014
From: davef at otter-rsch.com (dave fournier)
Date: Mon, 15 Sep 2014 12:19:50 -0700
Subject: [Developers] faster ludecomp in C++ for creating adjoint code
In-Reply-To: <CAJMx2XXamtGvnc5Bz-5VkzmCgrBk24RZiBg63edn+Z6wHxWfng@mail.gmail.com>
References: <54173814.3030902@otter-rsch.com>
	<CAJMx2XXamtGvnc5Bz-5VkzmCgrBk24RZiBg63edn+Z6wHxWfng@mail.gmail.com>
Message-ID: <54173BD6.4060307@otter-rsch.com>

On 09/15/2014 12:14 PM, Johnoel Ancheta wrote:

If you wan t I have code to call the Openblas version as well as a 
blocked version which stores the relevant matrices in
blocks as recommended by Dongarra et al  (but it does not seem to be 
worth the effort).


> Thanks Dave!  I'll provide feedback after testing...
>
>
>
> On Mon, Sep 15, 2014 at 9:03 AM, otter <otter at otter-rsch.com 
> <mailto:otter at otter-rsch.com>> wrote:
>
>     After much pain I have produced a C++ version of the LU
>     decomposition which
>     is suitable for producing adjoint code for ADMB and perhaps cppad.
>     (Don't know what adjoint
>     code for cppad is as yet!)  This code is about 25 times faster
>     than the current
>     ADMB code for a 2,000 x 2,000 matrix and about 4 times slower than
>     the Openblas
>     code which contains optimized assembler and Fortrash.  Any
>     suggestion for improvements
>     would be welcome.  Per usual I will hold my breath.
>
>
>
>     _______________________________________________
>     Developers mailing list
>     Developers at admb-project.org <mailto:Developers at admb-project.org>
>     http://lists.admb-project.org/mailman/listinfo/developers
>
>
>
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/4663f891/attachment.html>

From johnoel at hawaii.edu  Mon Sep 15 12:23:02 2014
From: johnoel at hawaii.edu (Johnoel Ancheta)
Date: Mon, 15 Sep 2014 09:23:02 -1000
Subject: [Developers] faster ludecomp in C++ for creating adjoint code
In-Reply-To: <54173BD6.4060307@otter-rsch.com>
References: <54173814.3030902@otter-rsch.com>
	<CAJMx2XXamtGvnc5Bz-5VkzmCgrBk24RZiBg63edn+Z6wHxWfng@mail.gmail.com>
	<54173BD6.4060307@otter-rsch.com>
Message-ID: <CAJMx2XWnSkAk+JciG1nm9=bUhrAYJ+mU_r2RLAmU1Ye_HUq2Fw@mail.gmail.com>

Sure, it would be nice to do the comparison.

Thanks in advance,

Johnoel

On Mon, Sep 15, 2014 at 9:19 AM, dave fournier <davef at otter-rsch.com> wrote:

>  On 09/15/2014 12:14 PM, Johnoel Ancheta wrote:
>
> If you wan t I have code to call the Openblas version as well as a blocked
> version which stores the relevant matrices in
> blocks as recommended by Dongarra et al  (but it does not seem to be worth
> the effort).
>
>
>  Thanks Dave!  I'll provide feedback after testing...
>
>
>
> On Mon, Sep 15, 2014 at 9:03 AM, otter <otter at otter-rsch.com> wrote:
>
>> After much pain I have produced a C++ version of the LU decomposition
>> which
>> is suitable for producing adjoint code for ADMB and perhaps cppad. (Don't
>> know what adjoint
>> code for cppad is as yet!)  This code is about 25 times faster than the
>> current
>> ADMB code for a 2,000 x 2,000 matrix and about 4 times slower than the
>> Openblas
>> code which contains optimized assembler and Fortrash.  Any suggestion for
>> improvements
>> would be welcome.  Per usual I will hold my breath.
>>
>>
>>
>> _______________________________________________
>> Developers mailing list
>> Developers at admb-project.org
>> http://lists.admb-project.org/mailman/listinfo/developers
>>
>>
>
>
> _______________________________________________
> Developers mailing listDevelopers at admb-project.orghttp://lists.admb-project.org/mailman/listinfo/developers
>
>
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/a929a7a8/attachment.html>

From davef at otter-rsch.com  Mon Sep 15 12:38:22 2014
From: davef at otter-rsch.com (dave fournier)
Date: Mon, 15 Sep 2014 12:38:22 -0700
Subject: [Developers] faster ludecomp in C++ for creating adjoint code
In-Reply-To: <CAJMx2XWnSkAk+JciG1nm9=bUhrAYJ+mU_r2RLAmU1Ye_HUq2Fw@mail.gmail.com>
References: <54173814.3030902@otter-rsch.com>	<CAJMx2XXamtGvnc5Bz-5VkzmCgrBk24RZiBg63edn+Z6wHxWfng@mail.gmail.com>	<54173BD6.4060307@otter-rsch.com>
	<CAJMx2XWnSkAk+JciG1nm9=bUhrAYJ+mU_r2RLAmU1Ye_HUq2Fw@mail.gmail.com>
Message-ID: <5417402E.9000201@otter-rsch.com>

On 09/15/2014 12:23 PM, Johnoel Ancheta wrote:

To compile it I used something like


CXX=g++
ADMB_HOME=~/admodel
echo "!!! Note using the version found in directory "
echo ${ADMB_HOME}
${CXX}  -march=native -ggdb  -Ofast -funroll-loops -DOPT_LIB 
-ffast-math  -pthread -DUSE_PTHREADS -W -fpermissive -DUSE_LAPLACE -Dlinux \
   avx_vecdot.o \
   -L/home/dave/opt/OpenBLAS/lib \
   -I/home/dave/opt/OpenBLAS/include \
   -D__GNUDOS__ -o$1 $1.cpp  \
   -I. -I${ADMB_HOME}/include \
   -I/home/dave/include \
   -L${ADMB_HOME}/lib \
   -L/home/dave/lib \
   -ladmbo \
   -lopenblas \
   -lgfortran

Near the top of main you set the matrix size n and the block size m. At 
present n must be a multiple of m.

  main()
  {
     ad_set_new_handler();
   ad_exit=&ad_boundf;

   int n=2000;   // we will work with an nxn symmetric matrix
   const int m=100;  // block size for blocked LU code

To restrict openblas to one thread for comparison you can use the 
environment string

export OPENBLAS_NUM_THREADS=1






> Sure, it would be nice to do the comparison.
>
> Thanks in advance,
>
> Johnoel
>
> On Mon, Sep 15, 2014 at 9:19 AM, dave fournier <davef at otter-rsch.com 
> <mailto:davef at otter-rsch.com>> wrote:
>
>     On 09/15/2014 12:14 PM, Johnoel Ancheta wrote:
>
>     If you wan t I have code to call the Openblas version as well as a
>     blocked version which stores the relevant matrices in
>     blocks as recommended by Dongarra et al  (but it does not seem to
>     be worth the effort).
>
>
>>     Thanks Dave!  I'll provide feedback after testing...
>>
>>
>>
>>     On Mon, Sep 15, 2014 at 9:03 AM, otter <otter at otter-rsch.com
>>     <mailto:otter at otter-rsch.com>> wrote:
>>
>>         After much pain I have produced a C++ version of the LU
>>         decomposition which
>>         is suitable for producing adjoint code for ADMB and perhaps
>>         cppad. (Don't know what adjoint
>>         code for cppad is as yet!)  This code is about 25 times
>>         faster than the current
>>         ADMB code for a 2,000 x 2,000 matrix and about 4 times slower
>>         than the Openblas
>>         code which contains optimized assembler and Fortrash.  Any
>>         suggestion for improvements
>>         would be welcome.  Per usual I will hold my breath.
>>
>>
>>
>>         _______________________________________________
>>         Developers mailing list
>>         Developers at admb-project.org <mailto:Developers at admb-project.org>
>>         http://lists.admb-project.org/mailman/listinfo/developers
>>
>>
>>
>>
>>     _______________________________________________
>>     Developers mailing list
>>     Developers at admb-project.org  <mailto:Developers at admb-project.org>
>>     http://lists.admb-project.org/mailman/listinfo/developers
>
>
>     _______________________________________________
>     Developers mailing list
>     Developers at admb-project.org <mailto:Developers at admb-project.org>
>     http://lists.admb-project.org/mailman/listinfo/developers
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/d7f25670/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: yblocklu10.cpp
Type: text/x-c++src
Size: 18081 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/d7f25670/attachment-0001.cpp>

From sibert at hawaii.edu  Mon Sep 15 14:07:39 2014
From: sibert at hawaii.edu (John Sibert)
Date: Mon, 15 Sep 2014 11:07:39 -1000
Subject: [Developers] faster ludecomp in C++ for creating adjoint code
In-Reply-To: <54173814.3030902@otter-rsch.com>
References: <54173814.3030902@otter-rsch.com>
Message-ID: <CAKKKzKjg+POHcwM92LvkLfwuCxLkVWmUpU049fMnnrR02OuYZA@mail.gmail.com>

Jolly good, old chap. (Headed to Heathrow.)
Cheers.

Sent from my phone.
On Sep 15, 2014 3:04 PM, "otter" <otter at otter-rsch.com> wrote:

> After much pain I have produced a C++ version of the LU decomposition which
> is suitable for producing adjoint code for ADMB and perhaps cppad. (Don't
> know what adjoint
> code for cppad is as yet!)  This code is about 25 times faster than the
> current
> ADMB code for a 2,000 x 2,000 matrix and about 4 times slower than the
> Openblas
> code which contains optimized assembler and Fortrash.  Any suggestion for
> improvements
> would be welcome.  Per usual I will hold my breath.
>
>
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140915/4fb40e6b/attachment.html>

From davef at otter-rsch.com  Tue Sep 16 11:35:42 2014
From: davef at otter-rsch.com (dave fournier)
Date: Tue, 16 Sep 2014 11:35:42 -0700
Subject: [Developers] c++ amp with admb
Message-ID: <541882FE.1090508@otter-rsch.com>

A while back Matthrew pointed out to me that development on C++AMP for
linux is an ongoing project which should be watched.

For those whio don't know AMP is a microsoft development which is 
supposed to greatly
simplify parallel processing on GPU's/CPUs.

Having done the LU decopmposition code I wondered it it could benefit 
from C++ AMP.

that is a bit difficult for a newbie sio I decided to start with a 
simple matrix multiplication example.

I have attached the code.  the zip file includes a script for compiling 
the example.


class dvector
{
     // ....................

  public:
    double * data (void) const { return v+index_min; }

};

class dmatrix
{
     // ............................................
  public:
    double * data (void) const { return m[index_min].data(); }

};

One technicality is that the AMP library expects the matrix or vector 
class to have a member function  data
which returns a pointer to the numbers in the vector.  I added this to 
the dvector and dmatrix classes.

Another technicality is that the AMP library expects the matrix to be 
stored in
contiguous memory.  I have accomplished this for the dmatrices in the 
example by using the make_dmatrix function.

Otherwise all seems to work except that the AMP code for a 2,000 x 2,000 
matrix runs less than half as fast as the
ADMB matrix multiply.  I have no GPU driver on my laptop so AMP is using 
the CPU.  This works as top
reports CPU usage of 792% which indicates that 8 threads are running.

Suggestions for improvements would be welcome.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: amp_mpy.zip
Type: application/zip
Size: 1984 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140916/d0fb5092/attachment.zip>