GPU linear solver thread overlapping

compdyn / partmc

Particle-resolved stochastic atmospheric aerosol model

http://lagrange.mechse.illinois.edu/partmc/

GNU General Public License v2.0

27 stars 15 forks source link

GPU linear solver thread overlapping #141

Open cguzman95 opened 4 years ago

cguzman95 commented 4 years ago

Issue to fix the "bug" of thread overlapping that produces different results in the GPU linear solver when executing two instances of the same function with same output. More details of the problem defined in #139

cguzman95 commented 4 years ago

Seems this bug is produced cause the "atomicadd" operation in the multiplication of the RHS with the Jacobian values. This atomicadd can be avoided changuin the type of the Jac structure from CSC to CSR (or even a more efficient structure). Working with Guillermo to find the best structure.