API addition to expose sync/wait functionality

Question

API addition to expose sync/wait functionality

slayoo opened this issue 5 years ago · comments

Currently when profiling code that uses ThrustRTC, execution times are in a potentially misleading way associated with subsequent API calls - apparently due to asyncronous execution. Would it be possible to expose some wait/sync feature in the ThrustRTC API so that one could enforce completion of execution of launched kernels?

Sylwester Arabas commented 5 years ago

ping? :)

Sylwester Arabas commented 5 years ago

thanks!

Sylwester Arabas · Answer 1 · Mon Aug 12 2019 19:42:43 GMT+0800 (China Standard Time)

or perhaps two versions:

launch()
launch_and_wait()

or

launch(wait=False)
launch(wait=True)

Fei Yang · Answer 2 · Wed Sep 11 2019 10:08:48 GMT+0800 (China Standard Time)

Hey, sorry. Wasn't on this these days. Will take a look.
I didn't consider this as an issue, because all device->host transfers are now using the synchronized copy APIs. However, yes, profiling is a good point.

Fei Yang · Answer 3 · Wed Sep 11 2019 21:53:57 GMT+0800 (China Standard Time)

Added Wait() just now. (eb94b45)

Sylwester Arabas · Answer 4 · Wed Sep 11 2019 22:10:46 GMT+0800 (China Standard Time)

kudos! Will try soon

Sylwester Arabas · Answer 5 · Tue May 19 2020 00:11:32 GMT+0800 (China Standard Time)

It does work! Thank you!