rewrite test_select to run in a few seconds. #1665
Conversation
|
Need more time to review this. |
|
There are major concerns with this PR. I would suggest a complete re-evaluation by the author. |
|
Sorry for that; I will naturally re-evaluate and re-submit when I have a better version. |
|
Removing "focused review" until the review issues are addressed - thanks! |
# Conflicts: # test_conformance/select/test_select.cpp
…le change report on each change's cost savings:
BEFORE:
real 47m8.497s
user 48m8.860s
sys 0m14.952s
AFTER:
real 17m53.383s
user 18m53.342s
sys 0m13.297s
initSrcBuffer generates the same random noise every iteration through the loop. There is no change to the arguments, and the host data itself doesn't need to get rewritten. Profiling realizes a 2 times speed accel from simply relying upon the buffer to remain randomized at the next loop iteration.
real 17m53.383s
user 18m53.342s
sys 0m13.297s
AFTER:
real 12m26.035s
user 13m15.505s
sys 0m15.414s
rearrange a few things in the loops to allow for vectorized / interleaved loop traversal. NB: not all loops are vectorizable obviously; but this addresses the worst offenders. Also note, to enable compiler to generate vectorized and interleaved loop traversal build with -o3.
There was a problem hiding this comment.
Thanks, this doesn't quite reduce execution time to "a few seconds" but I'm definitely seeing a nice improvement and there doesn't seem to be any lost coverage.
I tested one of our devices with the command line test_select select_int_int.
For the old version, before these changes:
$ time ./test_conformance/select/test_select select_int_int
...
real 1m19.207s
user 1m9.578s
sys 0m11.316s
For the new version, with these changes:
real 0m33.572s
user 0m23.800s
sys 0m11.626sThanks!
Note, there is a merge conflict that needs to be resolved before merging.
# Conflicts: # test_conformance/select/test_select.cpp
|
I'm reusing the PR from the first attempt at this. The first attempt does complete in a few seconds but at the expense of losing the number of test cases. I merged / resolved the merge conflict. |
|
Merging as discussed in the October 3rd teleconference. |
Test select getting optimized for runtime.