adjust blur radius for HTML to match CanvasKit

[flar]

The blur radius was too aggressive for HTML. The results for CanvasKit matched what we say on mobile hardware pretty well, but the HTML blurs were always too blurry. I don't know why the * 2 factor was chosen, but both the Flutter parameters and the HTML parameters are radii so they should be compatible.

I also adjusted it to be an average of the X & Y values rather than a max. I don't have a strong opinion on that, but I think it chooses a middle ground for when the dimensional radii are different.

Here are the before/after images (snapshots from Chrome, but results on Safari were pretty much the same, and Firefox doesn't seem to implement blur). Ignore the shade of the square in the middle - it was animating and the shade just depends on when I clicked "snapshot":

build	backend	resulting screenshot
before fix	HTML
before fix	CanvasKit
after fix	HTML
after fix	CanvasKit
(reference)	(Android Skia on Moto G4)

[flutter-dashboard]

It looks like this pull request may not have tests. Please make sure to add tests before merging. If you need an exemption to this rule, contact Hixie on the #hackers channel in Chat.

Reviewers: Read the Tree Hygiene page and make sure this patch meets those guidelines before LGTMing.

[flar]

In looking at the golden test failures and trying to reproduce them locally I get different test pictures. I notice in the code that it special cases MacOS and allows an extra percent of failures, so does this mean that I should generate the new goldens on a Linux machine for best matching with the CI job?

There are 18 failures:

Expected

These golden diffs were expected due to the non-doubled average radius calculation now used.
(are there canvaskit goldens to compare against?)

• 2 x backdrop filter
• compositing image filter (this test uses sigmaX: 1, sigmaY: 3 which shows my max() -> average() change above)

Unexpected

I have no idea how these tests could have been affected by my changes.

• draw_text_composite_below
• path_transform_with_arc
• smoke_test
• 8 x svg_arc
• svg_rect
• 3 x text_with_placeholders

[yjbanov]

You can download the images generated by LUCI and use them to update flutter/goldens.

Sorry for the unexpected failures. Those are warnings. Feel free to ignore them.

[flar]

Ah, now I see that the links at the bottom of the log aren't all failures, some are warnings. Most of the unrelated tests are, as you say, just warnings when I look into them.

But...

The smoke_test failure was not a warning. The entire wrong string was shown and it netted an actual test failure.

Oddly, the backdrop tests were all under the threshold and are not listed as failures, but I will update their golden images anyway. Those 2 tests probably need to be adjusted so that they better detect changes in the rendering.

At least the ImageFiltered test was accurately noted as a failure.

[flar]

Waiting on Goldens PR: flutter/goldens#107

[flar]

Forgot to mention this would fix some, but not all, of the idiosyncrasies observed in flutter/flutter#64828

[yjbanov]

The smoke test failure is intentional. It tests that our test harness can fail when necessary. Feel free to ignore it. (note to self: hide the output from the smoke test).

[yjbanov]

Oddly, the backdrop tests were all under the threshold and are not listed as failures, but I will update their golden images anyway. Those 2 tests probably need to be adjusted so that they better detect changes in the rendering.

Yeah, we're still looking for the right balance here. I was hoping we'd be on Skia Gold by now and not need to have our own logic. I tightened up some of the screenshots. You could tighten up these with a smaller threshold if you are updating the screenshots anyway.

[yjbanov]

LGTM once the tests pass.

[flar]

OK, I'll adjust the size of the canvas a bit until I get actual failures on the tests that are "passing by threshold" and then resubmit the goldens...

Hmmm, if you change the canvas size then you get a hard failure, so I'll just have to change to match the size of the tested shapes and assume that they would have been up-voted to "failure"...