Cesar, On 12/07/2015 09:55 AM, Cesar Philippidis wrote:
[snip snip] Two observations: 1. Why is deviceptr so special that gomp_map_vars can't handle it directly?
Recall that the deviceptr clause in Fortran presents as two mappings: FORCE_DEVICEPTR and POINTER. The former has the device address in hostaddr, while the latter has the host address of the pointer in hostaddr. The new code detects a properly formed pair and if found, proceeds to turn the latter into the FORCE_DEVICEPTR map, while the former is effectively discarded by setting hostaddr to NULL. This behavior is specific to OpenACC, so I decided put it where I did. Could it be put into gomp_map_vars? Probably. Would it be messy? Probably. But what may be a better solution would be to use the functionality that handles the use_device_ptr clause in OpenMP4.1 and eliminate FORCE_DEVICEPTR from gomp_map_var. Have to look into that.
2. It appears that deviceptr code in GOACC_parallel_keyed is mostly identical to GOACC_data_start. Can you put that duplicate code into a function? That would be easier to maintain in the long run.
Fixed. Thanks, Jim