search

StreamHPC / gromacs

OpenCL porting of the GROMACS molecular simulation toolkit
http://www.gromacs.org
Other
25 stars 4 forks source link

Check if there are situations when memory polling should be enabled for OpenCL #51

Closed ancahamuraru closed 9 years ago

ancahamuraru commented 9 years ago

Low priority issue.

nbnxn_ocl_init decides on how command queue synchronization should be implemented: using memory polling or clFinish. For CUDA, the first option is sometimes much more efficient.

Check if this should also apply for OpenCL and if NVIDIA devices should be identified and handled similar to how they are in nbnxn_cuda_init.

The current nbnxn_ocl_init implementation always chooses clFinish and never memory polling.

ancahamuraru commented 9 years ago

Memory polling never enabled for OpenCL devices - see commit 16d9130e2c94da3b6b41ea62bfe264d0b1c683fe