You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
parsec_get_best_device and parsec_gpu_data_stage_in use an approximate way to decide that a tile is NEW.
In parsec_gpu_data_stage_in, we just say it's NEW if NULL == source_repo_entry && NULL == original->dc && version == 0.
In parsec_get_best_device we say it's NEW if NULL == source_repo_entry.
All data copies in DTD, and probably in TTG, have a NULL source_repo_entry. This is not a good test.
Suggested solution: add a flag to the flow to denote it's NEW, or if it is not NEW for all flows, add a flag to the data copy? But do not overload the meaning of other fields, as this is fragile.
As a consequence, DTD multi-GPU is currently broken (a task can execute on any GPU even if the predecessor was modifying one of its input on another GPU and did not push it out).
The text was updated successfully, but these errors were encountered:
parsec_get_best_device and parsec_gpu_data_stage_in use an approximate way to decide that a tile is NEW.
In parsec_gpu_data_stage_in, we just say it's NEW if NULL == source_repo_entry && NULL == original->dc && version == 0.
In parsec_get_best_device we say it's NEW if NULL == source_repo_entry.
All data copies in DTD, and probably in TTG, have a NULL source_repo_entry. This is not a good test.
Suggested solution: add a flag to the flow to denote it's NEW, or if it is not NEW for all flows, add a flag to the data copy? But do not overload the meaning of other fields, as this is fragile.
As a consequence, DTD multi-GPU is currently broken (a task can execute on any GPU even if the predecessor was modifying one of its input on another GPU and did not push it out).
The text was updated successfully, but these errors were encountered: