Archives Discussions

[lollo]

Hi,
I'm experiencing some problems with a Brook+ code (a GMRES solver for sparse matrix), which runs perfectly with BRT_RUNTIME=cpu, but which gives totally different results if BRT_RUNTIME is set to cpu.

I'm using double precision, sparse matrix-vector multiplications and dense matrix-matrix multiplications (with parts of code taken from Brook+ legacy samples "sparse_matrix_vector" and "double_precision_simple_matmult").

The matrix-matrix multiplication's gather input parameters and the output parameters are often subsets of bigger matrices. To select the sub-matrices, I'm using the "domain()" operator.

I thought that the problem could be related to the domain() operator together with a gather indexing, but I had fine results while running simple tests of matrix multiplication with input and output parameters selected with domain() operator.

I also found this post http://forums.amd.com/devforum...id=97566&enterthread=y , but in my case gather input parameters are indexed with double [ ] [ ].

Does anyone have any idea of what could be a possible source of errors?

Thanks,

Lorenzo

[gaurav_garg]

Have you checked error and errorlog on your streams?

[lollo]

Ok, I found out the cause of my problems.

To multiplicate a vector for a scalar, let us say y = a*x, I was running a simple function like this:

    kernel void stream_mult ( double instream1<>, double instream2<>, double outstream<> )
    {
         outstream = instream1*instream2;
    }

calling then from the main body :  stream_mult ( xstream, astream, ystream );

This worked fine with cpu, but not with BRT_RUNTIME=cal.

To make the function run correctly on GPU, the scalar coefficient a must be passed as a gather parameter.

Here is the correct way:

    kernel void stream_scalar_mult ( double instream<>, double coef[], double outstream<> )

    {
        outstream = instream*coef[0];
    }

calling in the main body: stream_scalar_mult ( xstream, astream, ystream ) ;

I don't understand exactly why should it be incorrect to pass the scalar coefficient as a normal input stream, as it is of dimension 1 and it should be compared with no problems with streams of all dimension (there is no problem of multiplicity of dimensions! ).

Thank you

[gaurav_garg]

When you are using regular streams, all of the streams should be of same size.

You can pass this value as a constant instead.

kernel void stream_scalar_mult ( double instream<>, double coef, double outstream<> )

{
        outstream = instream*coef;

}