[Parallel] Primitive OpenMP support #78

wsmoses · 2020-10-19T23:13:56Z

Experiment testing what is necessary to support cpu parallelism, starting with OpenMP for ease.

Initial support for a subset of parallel fors.

Presently this includes the following limitations:

no "diffe_out" variables passed into a parallel for (since we cannot use the return value of a parallel for)
no cached values (for same reason as above preventing return of tape from fwd pass, and also need to massage tape indicies to correct offset for parallel index.

wsmoses requested a review from vchuravy October 19, 2020 23:13

[Parallel] Primitive OpenMP support

cbb121c

wsmoses force-pushed the omp branch 4 times, most recently from 82b7bfb to 374c05e Compare October 20, 2020 01:34

Add clang-format

7c57d09

wsmoses force-pushed the omp branch from 374c05e to 7c57d09 Compare October 20, 2020 03:11

Better handle of zero-size casts

d635e3c

wsmoses force-pushed the omp branch from 1292b2e to d635e3c Compare October 20, 2020 20:37

vchuravy approved these changes Oct 20, 2020

View reviewed changes

wsmoses merged commit a351805 into master Oct 21, 2020

wsmoses deleted the omp branch October 21, 2020 01:08

FROL256 mentioned this pull request Jan 17, 2024

Bugreport for complex codebase (probably some math problem) #1613

Closed

Provide feedback