Re: [Help-gsl] Re: Help with porting GSL to CUDA

[James Bergstra]

Are you sure to have thought through whether it is a good idea to port
GSL to the CUDA platform?  To me it sounds like a bad idea for two
main reasons:

1) GSL's data types are not opaque, so you can't very well "hide" the
fact that data lives on a GPU.  Anyone who tries to access his vector
or matrix data in the conventional ways will cause a segfault.

2) None of the blas, linear algebra, or minimization routines (the
ones that use the matrix and vector) will work without heavy
modification.  Good use of the GPU will require rewriting (and
rethinking) every function.

Maybe you should think of your port as a new library, inspired by the
GSL.  Are you sure that library would come out looking sufficiently
different from cuBLAS, cuFFT, and many others that are being developed
for the GPU?

James

On Sat, Sep 5, 2009 at 10:14 AM, Rohit Garg<address@hidden> wrote:
> There is too little information in your mail for anyone to offer you
> any kind of help.
>
> On Thu, Sep 3, 2009 at 4:02 PM, Mahesh
> Narayanamurthi<address@hidden> wrote:
>> Hello,
>>
>> My friends and I are trying to make a port of GSL onto NVIDIA CUDA platform.
>> We first started with segregating files and folders from the source tree so
>> that we got a closed set - only to support Matrix and Vectors for the time
>> being. Following that we have made suitable changes to configure.ac and
>> Makefile.am such that, we are now able to compile the sources using NVCC and
>> we patched the libtool script file a bit to generate the shared library
>> consisting of only those files we have extracted from the original archive.
>> To test the generated library file, we wrote a fairly trivial program that
>> worked with vectors. Now, when I am trying to link the new program with the
>> library we have generated, we are getting the following error:
>>
>> test1.o: In function `main':
>> test1.c:(.text+0x20): undefined reference to `gsl_vector_alloc'
>> test1.c:(.text+0x2f): undefined reference to `gsl_vector_alloc'
>> test1.c:(.text+0x43): undefined reference to `gsl_vector_set_all'
>> test1.c:(.text+0x58): undefined reference to `gsl_vector_set_all'
>> test1.c:(.text+0x73): undefined reference to `gsl_vector_get'
>> test1.c:(.text+0x88): undefined reference to `gsl_vector_get'
>> test1.c:(.text+0xc9): undefined reference to `gsl_vector_add'
>> .
>> .
>> .
>>
>>
>> On Fri, Aug 28, 2009 at 7:20 PM, Brian Gough wrote:
>>
>>> Thanks for your email.  I'm just going on hoiday for a week so I am
>>> unable to reply to it properly right now.  I suggest you post your
>>> question to address@hidden and see if someone there can help you.  I
>>> will look at it when I get back if needed.
>>>
>>> --
>>> best regards
>>>
>>> Brian Gough
>>>
>>> Network Theory Ltd,
>>> Publishing Free Software Manuals --- http://www.network-theory.co.uk/
>>>
>>
>> On my communications with Dr. Brian Gough, I was advised to write to the
>> address@hidden
>>
>> What could be going wrong here?
>>
>> Thanks & Cheers!
>> Mahesh Narayanamurthi
>> --
>> Research Intern
>> IIT Madras
>> Computer Science Engineering Department
>> _______________________________________________
>> Help-gsl mailing list
>> address@hidden
>> http://lists.gnu.org/mailman/listinfo/help-gsl
>>
>
>
>
> --
> Rohit Garg
>
> http://rpg-314.blogspot.com/
>
> Senior Undergraduate
> Department of Physics
> Indian Institute of Technology
> Bombay
>
>
> _______________________________________________
> Help-gsl mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/help-gsl
>



-- 
http://www-etud.iro.umontreal.ca/~bergstrj