Description how to reproduce the bug
Many of our solvers use KLU in the first iteration, which assumes CSC storage. However, our data is coming in in CSR, and the solvers themselves also use CSR. This means that unless the matrices are symmetric, we should not expect the solvers to work as is.
gpuRefactor and kluRefactor produce large residuals.
This will reproduce the result on Frontier as is, if run from the build directory of ReSolve (the matrices are already there).
nohup ./examples/gpuRefactor.exe -m /ccs/proj/eng151/jacobian_residual_1k/ScaleMicrogrid_Jacobian_N1000_ -r /ccs/proj/eng151/jacobian_residual_1k/ScaleMicrogrid_Residual_N1000_ -n 92 -b hip >> gpuRefactor_results.txt
Replace gpuRefactor with kluRefactor to see other errors.
I've started to address this already, so work with me if you want to contribute.
Re::Solve version
develop
System and environment details
All backends
Additional information
No response
Description how to reproduce the bug
Many of our solvers use KLU in the first iteration, which assumes CSC storage. However, our data is coming in in CSR, and the solvers themselves also use CSR. This means that unless the matrices are symmetric, we should not expect the solvers to work as is.
gpuRefactor and kluRefactor produce large residuals.
This will reproduce the result on Frontier as is, if run from the build directory of ReSolve (the matrices are already there).
Replace
gpuRefactorwithkluRefactorto see other errors.I've started to address this already, so work with me if you want to contribute.
Re::Solve version
develop
System and environment details
All backends
Additional information
No response