Baunsgaard commented on PR #1910:
URL: https://github.com/apache/systemds/pull/1910#issuecomment-1725404341
> Hey Sebastian, I have closed the pr for this function and made a new one
aiming my two PR's would be seperate but it seems they still share the same
commit history. Other than that the translate operator in the first PR had some
serious performance issues. I have tried to change it as much as possible. I
did get rid of one of the loops. In the initial implementation we were
assigning values to every index one by one which was inefficient for large
matrices. This is not the case anymore. I also used parfor for the outer loop
which helped with the performance. I have tried to compare it with the non
linear translate operator with random matrices that are 512x512 size. It seems
aproximately when we call it with more than 10 images . the linear function
seems to catch up to the non linear one . I did this by comparing the system DS
execution time that appears after the terminal after compiling the script. Is
there a better way of comparing ? I think I could also compare with the tim
e command but it gives 3 different measurements ( real, sys and .. ) I didnt
know which one I should primarly focus on. I also added the dml script for test
and the java test and looking forward to your feedback.
Good progress.
- Please experiment with other image sizes like non square images, and
larger/smaller sizes, i suggest up to 8k by 8k and down to 64 by 64.
- I suggest adding the -stats flag to the execution and use the time of the
image function call, and please do it in a loop like:
```
res = read($1)
print(sum(res))
for(i in 1:$2) {
res2 = OPERATION(res)
}
print(sum(res2))
```
- For the time command you get 3 values yes, user system and total. User is
wall clock that is most likely the one you want to use if you want to time the
entire execution. I recommend not to do so for your use case, unless you are
measuring a sequence of image operations.
If i missed something in the text just ask again :)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
