Summary: Rounding has been successfully adopted by multiple products and frameworks. Time to move it out of experimental mode. Rounding can still be turned of by setting the point scale factor to 0 on the config.
Reviewed By: gkassabli
Differential Revision: D4953838
fbshipit-source-id: 3ee5f27d92f95b3ed4a01c98bc35e9157f2e91c5
Summary:
This diff introduces new, little bit sophisticated round-to-pixel-grid algorithm.
**Motivation:**
Previous simple and straightforward solution works in most cases but sometimes produce the not-so-great result. A while ago Nick Lockwood described this problem and proposed the solution in RN's RCTShadowView class:
For example, say you have the following structure:
// +--------+---------+--------+
// | |+-------+| |
// | || || |
// | |+-------+| |
// +--------+---------+--------+
Say the screen width is 320 pts so the three big views will get the following x bounds from our layout system:
{0, 106.667}, {106.667, 213.333}, {213.333, 320}
Assuming screen scale is 2, these numbers must be rounded to the nearest 0.5 to fit the pixel grid:
{0, 106.5}, {106.5, 213.5}, {213.5, 320}
You'll notice that the three widths are 106.5, 107, 106.5.
This is great for the parent views but it gets trickier when we consider rounding for the subview. When we go to round the bounds for the subview in the middle, it's relative bounds are {0, 106.667} which gets rounded to {0, 106.5}. This will cause the subview to be one pixel smaller than it should be. This is why we need to pass in the absolute position in order to do the rounding relative to the screen's grid rather than the view's grid. After passing in the absolutePosition of {106.667, y}, we do the following calculations:
absoluteLeft = round(absolutePosition.x + viewPosition.left) = round(106.667 + 0) = 106.5
absoluteRight = round(absolutePosition.x + viewPosition.left + viewSize.width) + round(106.667 + 0 + 106.667) = 213.5
width = 213.5 - 106.5 = 107
You'll notice that this is the same width we calculated for the parent view because we've taken its position into account.
I believe this is awesome. I also believe that we have to decouple this logic from RN and put it into awesome Yoga. So I did it in this diff.
**Fun fact:**
The original implementation of this algorithm in RN had (and still have) a bug, which was found by Dustin dshahidehpour and fixed in D4133643. Therefore that diff was unlanded because it broke something unrelated inside RN text engine. I will fix that problem in RN later.
**Why do we need to change test methodology?**
Because the way we receive layout metrics from Chrome browser actually directly related to rounding problem. Previously we used `offsetHeight` and `offsetWidth` properties of the DOM node, which contain naively rounded values from `computedStyle` or `getBoundingClientRect`. (Which is we are trying to fix!) So, I added the new function that computes node size using two-step-rounding approach, conceptually similar to one that implemented in Yoga. Note: Chrome browser performs rounding layout as part of rendering process and actual values that can ve computed by counting actual pixel are different from these natively rounded ones.
**Why do some tests now have different desired values?**
These changes actually prove that my approach is correct and more useful for actual view rendering goals. So, let's take a look at test with changed values `rounding_fractial_input_3`:
Previously: 64+25+24=114 (Incorrect!)
Now: 65+24+25=114 (Correct!)
Previously: 64+25+24=114 (Incorrect!)
Now: 65+24+25=114 (Correct!)
Reviewed By: emilsjolander
Differential Revision: D4941266
fbshipit-source-id: 07500f5cc93c628219500e9e07291438e9d5d36c
Summary: Rename test files to use new name
Reviewed By: gkassabli
Differential Revision: D4265235
fbshipit-source-id: 0090d3949828058baf7435f33d4068de92756bad
