question about storage of coordinate matrix

[ghost]

Hello, I need to write a kernel to subtract a sparse coordinate matrix from a dense matrix.
Is there some convention for sorting in the storage of the rows and columns in the mat.handle12() data structure? A small kernel snippet would be great.

[karlrupp]

Hey,
yes, the (necessary) requirement is that the triplets in the coordinate_matrix are sorted by ascending row index. The best piece of code to look at is the copy() implementation from host to device here:
https://github.com/viennacl/viennacl-dev/blob/master/viennacl/coordinate_matrix.hpp#L47

As you can see, there is also an additional group_boundaries array, which aligns the work for each OpenCL work group such that the different work groups write into distinct parts of the result vector for a sparse matrix-vector multiplication. The implementation for the CPU backend is pretty simple (cf. https://github.com/viennacl/viennacl-dev/blob/master/viennacl/linalg/host_based/sparse_matrix_operations.hpp#L981), but a parallel implementation (CUDA, OpenCL) gets quite involved when one wants to deal with corner cases correctly:
https://github.com/viennacl/viennacl-dev/blob/master/viennacl/linalg/opencl/kernels/coordinate_matrix.hpp#L25
https://github.com/viennacl/viennacl-dev/blob/master/viennacl/linalg/cuda/sparse_matrix_operations.hpp#L1010

[ghost]

Thanks. How do I pass the vcl_coordinate_matrix.handle12() to my kernel,
if I use viennacl::ocl::enqueue(subtract_sparse_coordinate(vcl_coordinate_matrix)). It automatically takes passes vcl_coordinate_matrix .handle() so I an access the values but not the coords from vcl_coordinate_matrix.handle12().

[karlrupp]

vcl_coordinate_matrix.handle12().opencl_handle():
https://github.com/viennacl/viennacl-dev/blob/master/viennacl/linalg/opencl/sparse_matrix_operations.hpp#L559

[ghost]

Thank you. I thought I looked but not hard enough clearly.