Issues with swapping GenericVector for std::vector

[nocun]

Environment

• Tesseract Version: master
• Commit Number: 92b6c652..65d882f9
• Platform: Linux FC29 x86_64

Problem

Replacing GenericVector with std::vector has resulted in a small execution time performance decrease (about 4% on my simple test). I am not sure whether some other bugs have been introduced this way.

Most of the issues I could see are from changing generic_vector.init_to_size(n, x); directly into vector.resize(n, x);. It is because init_to_size() not only resizes the vector but also resets all of its values (not only the new ones).

Suggested Fix

Recheck the logic and update GenericVector's changed lines:

gv.init_to_size(n, x) and gv.resize(n, x) should become:

v.clear();
v.resize(n, x);

gv.truncate(n) (other than truncate(0) which is just v.clear()) should become:

if (n < v.size()) {
    v.resize(n);
}

gv.resize_no_init(n) (I think this one was ok):

v.resize(n);

I have added an example in comments of a commit where the first performance hit was noted .

As far as I could see the code affected is within commit range 92b6c652..65d882f9. I might be able to provide a PR but this could take some time.

[stweil]

Thank you for this valuable hint. What do you mean with the "performance decrease" of 4 %? Increased character error rates? Larger execution time?

[nocun]

Larger execution time measured between bb6dbd2c and c583ecef.

[stweil]

I am afraid that the necessary fixes will increase the execution time even further.

[nocun]

The difference is not large but measureable. After applying changes mentioned in the comments here execution time went back to normal so I have some hope.

There is yet another performance issue like that introduced in later commits but I have not found it yet.

[egorpugin]

We need some script to find such incorrect fixes in git history and update the places.

[stweil]

@nocun, can you try whether pull request #3474 changes the performance for you?

[nocun]

Sure, I'll be happy to test it but from my initial grepping it looks like there were changes in a lot of places.

I just realised I pasted a wrong commit range, the GenericVector changes started in Oct 2020 and finished in Mar 2021, commits: 92b6c652..65d882f9. Updated the issue description, sorry about that.

This is quite a lot of code to go through.

[nocun]

If you have other issues to work on I may try and change all the wrong code I can find but it will take me a few days.

[stweil]

GenericVector changes started in Oct 2020 and finished in Mar 2021, commits: 92b6c65..65d882f

My fix now covers that whole range. I looked for all replacements of init_to_size.