Re: [Viennacl-support] Out of resources.

[Karl R. <ru...@iu...>]

Hi Riaan,

the kernel is indeed not very efficient and can be the cause of the 
problem. I think it's better to iterate within one work group over the 
transposed matrix in a column-wise manner and synchronize after every 
column. Moreover, the sparse matrix can be split into k sets of columns 
(one set per work group) that are processed in parallel, each leading to 
a partial result vector y_k. At the end, the result vector is obtained 
by partial reduction as y = sum_k y_k.

Best regards,
Karli



On 05/24/2011 09:05 AM, Riaan van den Dool wrote:
> I do wonder if my kernel does not take too long to execute...
>
>
>
> __kernel void trans_vec_mul(
>            __global const unsigned int * row_indices,
>            __global const unsigned int * column_indices,
>            __global const float * elements,
>            __global const float * vector,
>            __global float * result,
>            unsigned int size1,
>            unsigned int size2)
> {
>    unsigned int max_index = row_indices[size1];
>    for (unsigned int col = get_global_id(0); col < size2; col +=
> get_global_size(0))
>    {
>      float dot_prod = 0.0f;
>      unsigned int row = 0;
>      for (unsigned int i = 0; i <= max_index; i++)
>      {
>        while (row_indices[row] < i)
>          row++;
>        if (column_indices[i] == col)
>        {
>          dot_prod += elements[i] * vector[row];
>        }
>      }
>      result[col] = dot_prod;
>    }
> }
>
>
> On Mon, May 23, 2011 at 2:16 PM, Karl Rupp <ru...@iu...
> <mailto:ru...@iu...>> wrote:
>
>     Hi,
>
>     if you get the error due to not providing an estimate, then it's
>     likely that you try to access some invalid piece of memory and the
>     out_of_resources exception is only a consequence.
>
>     Conservative work sizes are 64 for local and 64*64 for global.
>
>     Best regards,
>     Karli
>
>
>
>     On 05/23/2011 02:01 PM, Riaan van den Dool wrote:
>
>         I keep getting terminate called after throwing an instance of
>         'viennacl::ocl::out_of_resources' now (seems that if I don;t set
>         up the
>         solver with a prior (estimate of solution)) then it falls over.
>
>         What would the most conservative workgroup sizes look like?
>
>                    local_work_size_[0] = 1; local_work_size_[1] = 0;
>                    global_work_size_[0] = 512*512; global_work_size_[1] = 0;
>
>         These settings still give problems.
>
>         R
>
>
>