Make the blur filter more accurate (Close #19516)

[xrtxn]

Fixes #19516
Honestly, I got really lucky to find this. I would love to see how this affects other flash projects.

[xrtxn]

Should I rewrite the tests? They look fine visually, but obviously they fail.

[kjarosh]

Thanks for your contribution! I think we need a test that proves this behavior (which will differentiate between 2.0 and 1.99). How did you get the value 1.99?

Should I rewrite the tests? They look fine visually, but obviously they fail.

Depends on the test. Generally we require an output from Flash Player, but some outputs might be generated from Ruffle, not FP (e.g. shumway tests that are animated).

I'll try looking into it and proposing something.

[xrtxn]

I just tested it with a color picker. I tested multiple values, 1.9 was inaccurate, 1.99 looked like a good value to me. I created an example, you can check out https://github.com/xrtxn/ruffle_bug_report master or the old branch.
What was noticable to me, even with strength 1 the middle of the blurred image is #FBFEFB while with the flash player it is #FFFFFF. And the greater the strength was, the bigger the difference.
With this change, the ruffle behavior is #FFFFFF correctly.

[torokati44]

That 2 right there is in the "spatial dimension", so to say. I'd look for (or put) this slight adjustment somewhere else. Perhaps a floor call should be added somewhere later on the color values...?

[kjarosh]

Yeah, changing 2 to 1.99 does not fully fix the issue, as it still appears for different window sizes. It's most probably a rounding / numerical stability issue somewhere.