Writing GPU Array Type: Which Array Class to Inherit From?

[David Bateman]

cjbattagl wrote:
> David Bateman-2 wrote:
>   
>> Wouldn't it be better to inherit from octave_matrix and store a copy of 
>> the matrix in the original octave_matrix::matrix value and the version 
>> in GPU memory in the new class and synchronize the two values as 
>> needed.. You won't inherit the operators at the octave prompt if you 
>> inherit directly for a liboctave class.
>>
>>     
>
> I'm now inheriting from octave_float_matrix, which was a good start:.. the
> class allocates CPU and GPU memory, prints it out on request by reading the
> GPU vectors, and can assign the data to other types of Matrices:
>
> GPUArray (FloatMatrix A) : octave_float_matrix(A) { ... }
>
> It looks like I need to overload every single operator, though... I get the
> following behavior for every operator:
>
> A=single(rand(4)); G=NewGPUArray(A);
> G = G + 1;
> gives:  error: T& Array<T>::checkelem (5, -1): range error
>
> I haven't changed modified any important functions of FloatArray in
> GPUArray, so I suspect the fix should be easy for now.
> In examples I see use of the macros INSTALL_BINOP and CAST_BINOP_ARGS...
> would these provide a quick and dirty way for me to make GPUArray act
> exactly like FloatMatrix? (I would like GPUArray to act like a normal
> FloatArray at first, while I go through and add GPU functionality to each
> function)..
>
> If not, I will probably end up using the new @-classes in Octave to
> implement all of the functions/operators; it just seems like it wouldn't be
> as fast, though.
>
> Thanks for your help! 
> Casey
>   

As long as you don't might the operators that are overloaded returning 
float matrices, then adding the numerical conversion function from your 
GPU type to the float matrix like

octave_base_value::type_conv_info
GPUArray::numeric_conversion_function (void) const
  {
    return octave_base_value::type_conv_info 
(default_numeric_conversion_function, 
octave_float_matrix::static_type_id ());
  }

will be enough. Look at the triangular matrix type on octave-forge for 
an example of overloading in this manner

http://octave.svn.sourceforge.net/viewvc/octave/trunk/octave-forge/extra/triangular/

D.






-- 
David Bateman                                dbateman at dbateman.org
35 rue Gambetta                              +33 1 46 04 02 18 (Home)
92100 Boulogne-Billancourt FRANCE            +33 6 72 01 06 33 (Mob)