Open ancahamuraru opened 9 years ago
Currently gpu_min_ci_balanced_factor from nbnxn_ocl_data_mgmt.cpp is set to the same value used by the equivalent CUDA implementation.
Check if for AMD this parameter should have another value which could improve performance.
If necessary, have different coefficients depending on the vendor.
Please include the following change in your experiments, this improves the list size estimation: https://gerrit.gromacs.org/#/c/4202/
Currently gpu_min_ci_balanced_factor from nbnxn_ocl_data_mgmt.cpp is set to the same value used by the equivalent CUDA implementation.
Check if for AMD this parameter should have another value which could improve performance.
If necessary, have different coefficients depending on the vendor.