memory - Consistency Rules for cudaHostAllocMapped -
does know of documentation on memory consistency model guarantees memory region allocated cudahostalloc(..., cudahostallocmapped)
? instance, when writes device become visible reads host useful (could after kernel completes, @ earliest possible time during kernel execution, etc).
writes device guaranteed visible on host (or on peer devices) after performing thread has executed __threadfence_system()
call (which available on compute capability 2.0 or higher).
visible after kernel has finished, i.e. after cudadevicesynchronize()
or after 1 of other synchronization methods listed in "explicit synchronization" section of programming guide has been completed.
mapped memory should never modified host while kernel using or running, cuda not provide way of synchronization in direction.
Comments
Post a Comment