Glitch experienced with SA / CUSP

[Andrea Borsic]

Hi,

First of all thanks for your efforts in developing CUSP, we are finding it extremely useful for speeding up some thermal and electrical biomedical simulations where computational performance is critical.

I have been using a development snapshot of CUSP 0.4 for about 2 years and I would like to upgrade to current release / dev. snapshots, but I am experiencing some problems, and I thought it would be good to report them.

My environment:

• Windows 10 Professional 64bit
• Visual Studio 2015 Community Edition
• CUDA 8.0
• GTX Titan Black

Problem Description:

The attached code "smoothed_aggregation_example_loading_matrix.cu" (which is basically the "smoothed_aggregation_example.cu" example modified to load a matrix from file and to initialize the preconditioner with "sa_initialize()" ) compiles and runs fine on a small test matrix (details about the matrix later) if compiled with the CUSP snapshot provided (ver 0.4, dating back 2 years approximately). The same example compiles with the current release 0.5, but crashes. This behavior does not seems to be specific to the matrix used.

Out of curiosity I have also tried to compile the example under the current dev. snapshot (0.6) but it fails to compile (the online example "smoothed_aggregation_example.cu" also does not compile).

I am providing a link to the test matrix being used (https://www.dropbox.com/s/k8lkem3eny8k0eo/FEM_22k_nodes.mtx?dl=0 - this is a small test FEM matrix  - 22,000 x 22,000 with 300K nnz). The problem shows though consistently with several matrices we have used, and seems to the connected to the initialization of the preconditioner with sa_initialize(). The same example runs fine under CUSP 0.5 if the system matrix is passed as an argument to the constructor of the preconditioner (but in our application we need to build the preconditioner at a later time).

Thanks for an advice !

Best Regards,

Andrea