Search
Monday, 19th of August 2019, 7:33:49 UTC
16:41:17
drmeister
Bike: I got my first one of these: Condition of type: OUT-OF-EXTENT-UNWIND
16:41:50
Bike
just so you know, that won't work on linux - not for any major reason, but because it's conditionalized on a USE_LIBUNWIND preprocessor flag
16:41:57
Bike
i don't know if we require libunwind on linux otherwise
16:42:15
drmeister
We could - would that help?
16:43:08
drmeister
It may have happened because of old cached fasls - I'm recompiling stuff to see.
16:44:47
Bike
i mean without this libunwind thing, out of extent unwinds will crash.
16:47:33
drmeister
Ok, but if we did have libunwind as a dependency on linux - would you be able to use it to catch this problem?
16:49:09
drmeister
I'm hitting this problem everytime I run (start-swank) in the jupyter notebook. Grrr.
16:49:26
drmeister
Is it possible that this has been a problem all this time and only now we are catching it?
16:58:16
drmeister
Forgot to clear a cache - trying again.
17:00:37
drmeister
Sorry - it was just bad old code.
17:00:57
drmeister
After clearing the caches - everything works. But seeing the error was great!
17:12:37
Bike
an out of extent returnw ithout this code would crash clasp entirely, so there are probably no lurking cases
17:13:31
drmeister
https://usercontent.irccloud-cdn.com/file/NwA3uVre/out-94492.svg
17:14:09
drmeister
Our profiling is great! This is profiling of our new algorithm. It clearly shows the bottlenecks.
17:32:32
drmeister
With a little bit of work I sped it up 3.4x
17:37:34
drmeister
And I'd left in an unneeded allocation in the velocity-verlet step that I just took out.
17:37:43
drmeister
That was visible here...
17:38:54
drmeister
https://usercontent.irccloud-cdn.com/file/mceUW2DN/out-95653.svg
17:39:34
drmeister
I mean - more apparent now that I've improved the performance.
17:41:16
drmeister
https://usercontent.irccloud-cdn.com/file/xhDBbmuW/out-96218.svg
17:41:54
drmeister
The EnergySketchNonbond_O::evaluateTerms is an N^2 algorithym
17:42:25
drmeister
I'm using double precision - I'll drop to single precision.
17:52:19
drmeister
Huh - that actually slows it down. Maybe conversions from double to float and back to double.
17:53:22
drmeister
750 ms (double) and 1000 ms (float)
17:53:41
drmeister
I'd have to retool everything to switch from double to float.
17:54:41
drmeister
I'm generating images like this from random molecule graphs
17:54:42
drmeister
https://usercontent.irccloud-cdn.com/file/6iruFV2Y/image.png
18:49:08
Bike
drmeister: do we have a way to clone a source pos info? so i can make a new one with an inlined-at, basically.
19:02:43
drmeister
(source-pos-info-copy source-pos-info)
19:03:25
drmeister
Check that it's copying everything it needs.
19:03:59
drmeister
This is a general thing I want the documentation to lay out - how do we discover this method in Cando?
19:04:18
Bike
hopefully we would organize all the source pos info related functions together.
Monday, 19th of August 2019, 19:33:49 UTC