Reftest: perspective-border-radius test failed in release build

[ClarkWang-2013]

run wrench reftest:
$ ./script/headless.py reftest reftests/transforms/perspective-border-radius.yaml

I found perspective-border-radius test failed, but when run 'OPTIMIZED=false ./script/headless.py xxx', test passed

I print related information：
Release mode:
REFTEST TEST-UNEXPECTED-FAIL | reftests/transforms/perspective-border-radius.yaml == reftests/transforms/perspective-border-radius.png | image comparison, max difference: 9, number of differing pixels: 362

Debug mode:
REFTEST TEST-PASSED | reftests/transforms/perspective-border-radius.yaml == reftests/transforms/perspective-border-radius.png | image comparison, max difference: 9, number of differing pixels: 346

I want to know how to debug this issue, and the test allow num differences is 348, What is the basis of this value（348）
Thanks！

[ClarkWang-2013]

Platform：
Linux mips64el
Rust 1.41.0 (5e1a79984 2020-01-27)

[kvark]

Those values are largely just taken from what is reported here and a bit of common sense. I.e. if you are adding a test, and you know it needs to match something else, and you see it matches but there are a few minor differences due to the way an image is produced (that aren't considered wrong), then you add the num difference as reported by the reftests.

In this case, we should just bump it to 346

[ClarkWang-2013]

@kvark Thanks for your reply.
I want to know what is the main reason for the difference between debug mode and release mode, compilation options or others？

[kvark]

I suspect some arithmetic optimization causing a slight inaccuracy. If you want to investigate, you can load this in Wrench and do a WR capture (by hitting "C"). Then comparing a capture from Debug with the one from Release and see which numbers differ.