[ADMB Users] A nontrivial cuda example

[dave fournier]

On 11-09-14 01:52 PM, CHRIS GRANDIN wrote:

It is not so bad to work in C.  For example for the df12variable
suppose you have

struct df12variable  a,b;

and the product  u=a*b

you can replace this with the prod fucntion below and write

    u=prod(a,b);

It would be nice if the function could be inlined.



   struct df12variable
   {
     double v[3]; // v[0] holds value,
                 //v[1] holds derivative wrt first variable
                 //v[2] holds derivative wrt second variable
   };

   struct df12variable prod( struct df12variable * a, struct 
df12variable *b)
   {
     struct df12variable  c;
     c.v[1]=a->v[0]*b->v[1] + a->v[1]*b->v[0];
     c.v[2]=a->v[0]*b->v[2] + a->v[2]*b->v[0];
     c.v[0]=a->v[0]*b->v[0];
     return c;
   }







> Thank you very much Dave, I will work to convert this to opencl, as I 
> already have the opencl library included into the admb source build 
> steps it shouldn't be too big a deal.  I knew we needed to change the 
> operator overloading to structs but figured it was too much work, i.e. 
> rewriting all the overloads in admb... I was kinda hoping opencl would 
> come out with a C++-friendly library..
> Chris
>
> ----- Original Message -----
> From: dave fournier <davef at otter-rsch.com>
> Date: Wednesday, September 14, 2011 1:13 pm
> Subject: A nontrivial cuda example
> To: CHRIS GRANDIN <cgrandin at shaw.ca>
> Cc: users at admb-project.org
>
> > As Chris anticipated (ruining my surprise) I figured out that
> > difficulty in understanding
> > derivative stuff might be holding up some of the parallel
> > processing so I
> > took a few hours to write an example of GPU processing within ADMB.
> >
> > Here is a vectorized log_negative binomial density
> > example.  It is a rather silly
> > negative binomial regression, but it does call the
> > log_negative_binomial density function
> > a lot of times.  I has to do forward AD for functions of
> > one or two variables in C
> > rather than C++
> > which means you put in the AD code rather than overloading the
> > operators.
> > It runs on the GPU or CPU depending on whether you enter 1 or 2
> > when prompted.
> >
> > The example shows how to calculate the derivatives and how to
> > make sure they are correct.
> >
> > I hope we can all build on this to clean up the GPU interface
> > and move it to opencl.
> > the current version is cuda.
> >
> > In order to build the example I used the makefile from the cuda
> > sdk examples
> > modified to include the admb stuff.  The cpp file was
> > renamed to a cu file
> > 'cause cuda seems to like that.
> >
> > It should run a lot faster on Chris's  fancy government thing.
> >
> > Have fun,
> >
> >         Dave
> >