Ghostscript_Tiger.svg is not rendered properly on macOS

[DanCraft99]

On iMac with Radeon GPU, the image rendered from cli is blocky with sketchy outlines. Line and Circle drawing is fine. Seems like the bezier path rendering is not functioning with MoltenVK (latest 1.2.154 Vulkan SDK).

[raphlinus]

That's interesting. Can you post a screenshot and repro steps for how to run it with MoltenVK? At this point, it could be just about anything - something I'm doing wrong in my code that just happens to work on NV+AMD+Intel, a mistranslation of the shader code into MSL (actually one of my top guesses), a bug in MoltenVK, etc. I think it would be useful to track it down.

Also, the more we're going to be running this on actual hardware, the more we need testing infrastructure to validate that the various pipeline stages are doing what they are supposed to. I have done a bit of this by hand (see the various mostly unused "dump" functions in cli.rs) but automation is most definitely possible.

[DanCraft99]

The repro steps are straight building and running the cli test after installing latest vulkan sdk on Mac. The image.png shows

Will post any any findings as soon as I could pick up more coding details. : )

[raphlinus]

I tried it on my mac (MBP 13", running 10.15.5, Intel HD 640), and I didn't get any image, though the collatz example does work and I get meaningful validation reports when I fiddle with the code. One question - does #44 work any better? I did fix a bug in backdrops there.

Also, are you getting meaningful timestamp reports? All 0.000ms here.

[eliasnaur]

One question - does #44 work any better? I did fix a bug in backdrops there.

Is that related to issue #23? I can't seem to find any backdrop changes in your #44, except for the right_edge removal. However, right_edge tracking was already dead code.

Edit: This one?

[raphlinus]

Yes, the missing indirection on the element index. I can see how that happened, the variable naming is very confusing. As I'm working on the code after being away from it for several months, I can see a number of ways code clarity can be improved.

[DanCraft99]

The blank image on intel GPU is another rendering issue. Reproed on my Mac Mini 2018.

[DanCraft99]

The reported issue on the blocky tiger svg rendering started from 9a8854, and it was good at 8d01aba23. Reproed on my iMac using Radeon GPU.

[raphlinus]

Thanks for looking into it, but those bisection results aren't super helpful, as 9a8854 is a pretty serious rewrite based on sort-middle, and I now consider the version before that to be something of a dead end.

I think tracking this down will require some fairly serious detective work, and a good step in that direction would be to add tests of individual stages, so we can at least track down which one of those is failing.

[DanCraft99]

Close the issue with much design changes since the first report.

With ref to #138: Beginnings of new element pipeline, the native metal backend support is now working and rendering the tiger.svg nicely. The original report was based on testing results from Vulkan port through MoltenVK. Though, it may be interesting to check whether Vulkan/MoltenVK works, that will be a separate topic.

[raphlinus]

Yes, it's valid to close, though #138 hasn't been merged yet (maybe I should just do that).

I'm also curious whether MoltenVK works, but just out of idle curiosity, I have no active plans to maintain that. I do suspect that shader translation issues / atomics are the main problems, though.