Coil: A gif is playing very fast

[rupinderjeet]

Plays really fast than how it is shown in Android Gallery and Chrome. Issue occurs while using ImageDecoderDecoder. If only GifDecoder is used, gif doesn't play at all.

How do I load this gif? Checkout #539

This is how it looks when I load it:

record.mp4.zip

[IljaKosynkin]

Hey @rupinderjeet, @colinrtwhite
I poked around with this issue (have a little bit of free time on my hands and decided to see if there are OS projects that might need help).
From my research, it seems that this particular GIF has 0 delay time in-between frames @rupinderjeet. The reason it plays so differently on the web is that major browsers handle this case by replacing delay with a default value if it is lesser then cut-off. ImageDecoder plays it as received (really fast) and Movie just shows the first frame.
When I replaced delay time with 15 it seems to be loading fine by Coil with both decoders.

Now the question to the @colinrtwhite: I presume that pre-processing GIF or writing custom decoder, which allows replacing delay with minimal one, is way beyond the scope of that library, so what do we do with that issue?

[colinrtwhite]

@IljaKosynkin Thanks for investigating this! Do you know how tough it would be to pre-process the GIF as it's being read? I'd be open to a solution that uses ForwardingSource to replace the GIF frame delay similar to web browsers depending the performance impact + how much code it would require.

[IljaKosynkin]

@colinrtwhite I can't tell it from the top of my head regarding complexity and performance impact, to be honest, but I'm going to poke around to tell estimation, possibly with PR, if it is not too bad in my opinion.

[IljaKosynkin]

Also, on a side note, from my investigation it seems that all files in gifs.json are actually WEBP files, maybe I should add some actual GIFs here?

[colinrtwhite]

@IljaKosynkin Are you sure? They appear to be encoded as GIFs when I download them locally. Though, feel free to add more sample gifs to that file.

[IljaKosynkin]

@colinrtwhite oh ok, seems that it might be an issue with browser. Because it defaults to WEBP for me apparently, but I can manually change link to get GIF version of the same animation.
I wonder what version does Android receive tho, ImageDecoderDecoder checks for both GIF and WEBP, so either can get fetched. I would change all examples to hardcode extension, tbh, half for GIF half for WEBP.

As for the ForwardingSource, it seems that delay time can be found by searching for 21 F9 04 sequence in the byte buffer and changing uint16_t after first byte after that sequence.
However, I'm not sure that it won't mess with the internal structure of the GIF, introducing potential artefacts in the image. On the other hand, that it is not likely and proper solution would require writing like good 30-40% of a GIF decoder, to make sure we're changing data in proper offsets.
Thoughts?

[colinrtwhite]

@IljaKosynkin I don't think we can use animated WebP in the sample since those only work on API 28+.

Interesting! Do you know what 21 F9 04 represents in the GIF spec? Is it a magic number? I'd want to avoid anything that could introduce artifacts (unless that number is safe according to the GIF spec) and instead opt for a mini GIF parser.

[IljaKosynkin]

@colinrtwhite It is not a magic number, well at least not sequence itself.
21 signals that extension will follow, F9 signals type of graphic control extension, 04 signals how many bytes are there before end of the block.
While probability of introducing artefacts is low, I believe it is still possible since image data is LZW compressed and there is no guarantee this sequence of bytes will not appear here, AFAIK.
We can extend the sequence to regex 00 21 F9 04 XX XX XX XX 00 which would lower chances of corrupting data further, but won't really make it impossible.
As for the decoder, I'm not sure writing it in Kotlin won't hit performance significantly. And writing native code might be not a best idea from the point of view of maintainability.
Or, alternatively, we can submit feature request to Android team to add minimal delay time for GIFs in ImageDecoder itself (since apparently it is not part of AOSP) and mark this issue as "won't fix".
So it's "pick your poison" kind of situation.

[colinrtwhite]

@IljaKosynkin Yep, we definitely don't want to add native code and I don't think requesting a feature for the next platform release would help much. In terms of performance I'm not too worried about adding some pre-processing of the input data using Kotlin code. Glide's GIF parser is 100% Java and it's pretty quick. As long as we're not allocating class instances I think we should be good.

Correct me if I'm wrong (not super familiar with the GIF technical specs), but maybe we could do something like this:

1. Skip the Header (6 bytes)
2. Skip the Logical Screen Descriptor (7 bytes)
3. Skip the Global Color Table (3*2^(N+1) bytes where N is the number given by the last 3 bytes of the Logical Screen Descriptor)
4. Modify the delay time in the Graphics Control Extension (if present).
5. Skip the Image Descriptor (10 bytes)
6. Skip the Local Color Table (same way we skipped the Global Color Table)
7. Skip through the image data. It gives us repeating sizes to skip through for each image block.
8. Handle skipping the Application and Plain Text and extensions (both list their size at the beginning of the block)
9. Go back to step 4 if we aren't at 00 (the termination flag).

Is that something you're interested in working on? Else I'm happy to take it on.

[IljaKosynkin]

@colinrtwhite tbh I'm not sure how much overhead processing of binary files in pure Java/Kotlin would add compared to native languages, but I feel like it might be substantial. But we can try and profile code to see how bad it actually is.
Anyway, regarding issue itself:

• Yes, that exactly what I meant by 40% of decoder. Because you're still parsing pretty much everything, just skipping most of the data.
• Who picks the issue largely depends on how much priority it has IMHO. I don't have great familiarity with either Coil internal structure or GIF decoder implementation, so if this issue is urgent you're probably better person to pick it up. Otherwise I can implement it, it just will take some while.

[colinrtwhite]

@IljaKosynkin Sounds good. I'll probably have time to try to prototype this this weekend.

[IljaKosynkin]

Ok, so I guess I better pickup something else, right? @colinrtwhite

[colinrtwhite]

@IljaKosynkin Sure, if you could investigate + fix https://github.com/coil-kt/coil/issues/539 that would be awesome.

[IljaKosynkin]

Sure, gonna take a look @colinrtwhite