Overlap communication with computation in multiply_module

OrderN / CONQUEST-release

Full public release of large scale and linear scaling DFT code CONQUEST

http://www.order-n.org/

MIT License

96 stars 25 forks source link

Overlap communication with computation in multiply_module #265

Open tkoskela opened 8 months ago

tkoskela commented 8 months ago

          Things to look out for

Can we overlap communication and calculation in https://github.com/OrderN/CONQUEST-release/blob/6bf8f4a8c20fd4fa8f1c7baeb8a6b1f23a6d2408/src/multiply_module.f90#L251

Originally posted by @tkoskela in https://github.com/OrderN/CONQUEST-release/issues/248#issuecomment-1697085794

Can we use OpenMP tasks to both receive data and call multiply kernels, with appropriate dependencies

tkoskela commented 7 months ago

Investigate non-blocking receives in prefetch. Fetch first iteration of data with a blocking receive (or wait) before the main loop. Then receive next iteration of data while computing the current interation.

tkoskela commented 7 months ago

Receive data for first iteration of main loop before the loop starts
Post nonblocking receives and sends for next iteration on master thread
Perform threaded computation in the compute kernels
After computation is finished, wait for nonblocking communications have finished.