Skip to content

[3.0] Exanite's Tracking Issue #2586

Open
Open
[3.0] Exanite's Tracking Issue#2586
Assignees
Exanite

Description

@Exanite
Exanite
opened on Jun 6, 2026

Exanite's Tracking Issue

This tracks a bunch of miscellaneous tasks that I accumulated while working on Silk 3.
I'm open for anyone to work on these tasks, but please discuss with me in the Silk Discord first.

A task being here on this tracker does not mean it is guaranteed to be completed.
This tracker's primary goal is so that we do not accidentally forget about tasks and that any tasks not completed are intentionally left uncompleted.

In-progress

Tasks in progress will be moved to the corresponding pull request.

Uncategorized

This section contains informal reports, suggestions, or complaints about Silk 3.
I'm logging them here so they don't get forgotten about until I either open a formal issue or address them.

From Aqua:

From Ethereal:

  • SilkTouch generator takes an immense amount of RAM: https://discord.com/channels/521092042781229087/607634593201520651/1513592461493145601
    • I need to also check with Curin to see how much RAM their device has. If their device doesn't have enough RAM, this might explain why the Microsoft job takes so long (aside from it already processing a lot of data).
    • Run the Microsoft job locally and see how much RAM it takes on my device.

From Curin:

  • For D3D12, the Agility SDK is loaded from the System32 DLL instead of the working directory's D3D12 subdirectory. This means that Silk 2.x (report was for 2.x) and maybe Silk 3.0 (needs verification) cannot access Agility SDK features. See https://discord.com/channels/521092042781229087/587346162802229298/1549273531668373566
    • My comment: seems to be another search path issue. This seems similar to my issue in Silk 2 where system binaries are preferred over the vendored/local ones. I'm personally very interested in changing Silk so that it always prefers local binaries whenever possible.

From me (Exanite):

From me (Exanite), as part of my OpenCL example program:

Original comments:

Silk.NET/examples/CSharp/OpenCL/Tutorial001.SimpleKernel/Program.cs

Lines 75 to 80 in 021738d

// Load API
// TODO: These todos apply to all Silk bindings.
// TODO: Consider making INativeContext public. It's nice for debugging.
// TODO: Consider returning the concrete class instead of the interface. Having to cast to access Cl.CurrentPlatform is annoying.
// TODO: I liked how INativeContext only required the function name in Silk 2. Silk 3 is currently `LoadFunction(functionName, libraryNameHint)`.
var cl = Cl.Create(); // TODO: This should be disposable

  • Expose INativeContext from the concrete API object where applicable (but not the interface)
  • Discuss INativeContext API. Silk 3 is currently LoadFunction(functionName, libraryNameHint), but I liked how Silk 2 only required the function name.

From me (Exanite), as part of the code review during #2624:

  • Open a follow up PR that removes Autofac. We discussed this and while MS DI isn't the best DI solution feature-wise, I think we can avoid adding another dependency without much downside.
  • Add dedicated documentation section/folder for using the SilkTouch generator as a non-contributor.
  • Add single file SilkTouch example project
  • Add SilkTouch example project that implements a custom mod
  • If possible, change the public API docs analyzer to also require API docs on internal APIs
    • This is to ensure that we keep our internal docs up to par even after we make everything internal instead of public.

Bindings and feature requests

From Nolram (https://discord.com/channels/521092042781229087/607634593201520651/1524213012402147348):

From dfkeenan:

High priority

Bindings usage:

Documentation:

  • Add Vulkan example project
  • Add bindings usage docs
    • Explain that the native docs for each API can be used since Silk doesn't really change how you use the API
    • Explain difference between bindings and high level APIs
      • Ie, high level APIs do not represent bindings so usually the native docs don't apply for those
    • Explain the different DSL types and magical constants (eg: nullptr)

Medium priority

SilkTouch generator API:

This will probably be implemented

  • Abstract out DI container from the public API so we aren't locked into a specific framework. See discussion here: https://discord.com/channels/521092042781229087/587346162802229298/1561387300443791481
    • Alternative discussed: Just switch back to MS DI. Should be possible with some trade offs.
  • Lock down public API to only the types necessary to integrate and add custom SilkTouch mods.
  • Add example project of implementing a custom SilkTouch mod with only public APIs available.

ABI compat proposal:

Native binaries:

  • (From Aqua) libSDL3.so is humongous (should not be nearly 14 MB)
    • Exanite: I checked my system version and that one is only 3 MB.
    • Edit - Exanite: Looks like running file libSDL3.so says libSDL3.so: ELF 64-bit LSB shared object, x86-64, version 1 (SYSV), dynamically linked, with debug_info, not stripped, so definitely because the binary was not stripped

Bindings:

Maintenance:

  • ClangSharp switched from System.CommandLine to a custom command line argument parser. I noticed that we had verbatim copied code from ClangSharp for handling arguments in Silk 3, so we should update that.
    • This also lets us update System.CommandLine, which we are still using a beta version of.

Low priority

Maintenance:

  • Cross reference changelog and past issues for Silk 2 to ensure that we don't have the same bugs in Silk 3

High level utilities:

  • Reimplement the Vulkan struct chaining API
  • Add image format utils for Vulkan/etc

IDE / Metadata:

Completed tasks and canceled tasks

Completed

After maintaining this section for a bit, I'll only list tasks not obviously related to a PR. Otherwise, it means that I'm tracking the same task in two places, which can get confusing.

Completed during #2588:

  • Switch to SLNX solution format to improve readability/mergability
  • From Ethereal: They had a CSharpier TypeLoadException when running the Win32 job on develop/3.0. Might be related to the error I saw half a year ago. If so, likely can be fixed by simply updating CSharpier. CSharpier has updated their Roslyn package since then.

Completed during #2591:

  • Port back the --only option from Curin's branch so I don't need to specify --skip for all but one job

Completed or planned to be completed during #2590 (not yet merged):

  • Begin work on the rest of the Khronos bindings
    • I'd like to start this early because this lets us spot issues early and we can use these as test cases for the rest of the changes.
  • Generator usage: Explain SilkTouch CLI options
  • Generator usage: Clarify that SilkTouch does not modify non .gen.cs .cs files during the Generated Bindings Output section
  • Consider reworking data type trimming regexes.
    • This was discussed more here: [3.0] Implement name-related improvements identified in previous Vulkan PR #2503 (comment)
    • The summary is that the current set of 2 regexes doesn't elegantly handle the case where the ending of a word should not be trimmed, but the word itself can have data type suffixes.
    • For example: alSourceRewindv and alSourceRewind. The d in Rewind can be mistaken as a data type suffix, so we add it to the EndingsNotToTrim regex. However, the dv in Rewindv also has the same problem.
    • We likely need something more akin to WordsNotToTrimInto than EndingsNotToTrim.
  • Investigate why we currently use DllImport in Silk 3.
    • Also added some light documentation to AddVTables in the mod MD docs.

Completed in dotnet/ClangSharp#699

Completed outside of a PR:

  • For DrawElements/etc, check if the pointer to 0 issue present in Silk 2 still exists in Silk 3
  • Investigate using the new BitCast optimizations: https://discord.com/channels/143867839282020352/312132327348240384/1517589517727436952
    • Eg: BitCasting a user struct to a Vector struct and immediately using an SIMD operation is now properly optimized. Free performance!
    • This mainly affects Otac0n's work, but I'm logging it here so I also keep it in mind as I work on the bindings.
    • Decision: Just going to keep this in mind. Don't think I need to actively review the bindings for this since there's not many places where this will be used.

Not doing

For transparency and for reference, I'll move any tasks I don't plan on doing here.

  • Figure out how to handle CLA properly. Currently native builds will block CLA due to the use of a bot account.
    • Seems to be fine now.
  • Differentiate between smoke tests and actual tutorials/examples. I would categorize the current SDL/OpenAL "tutorials" in Silk 3 to be smoke tests since they really shouldn't be referenced as sources of best practices.

Quality (probably will add this section back later):

  • [/] Add some way of verifying our bindings are well polished.
    • These would mainly involve simple automated sanity checks that ensure our bindings our consistent. Could be done as a separate CLI project from SilkTouch.
    • [/] Check for missing namespaces.
    • [/] Check for spelling/casing issues? Identifiers like Sdl.PropGpuDeviceCreateDebugmodeBoolean are technically correct (native name is SDL_PROP_GPU_DEVICE_CREATE_DEBUGMODE_BOOLEAN, so Debugmode is "one" word), but we could have an audit tool that checks for these non-ideal cases so we can add manual overrides. Worth the effort? Probably not unless we care about perfecting the bindings.

Activity

  1. Exanite
    added this to Silk.NET Development Trackeron Mar 24, 2026
  2. Exanite
    self-assigned this
    on Jun 6, 2026
  3. Exanite
    converted this from a draft issue on Jun 6, 2026

  4. Exanite commented on Jun 17, 2026

    @Exanite
    Exanite
    on Jun 17, 2026
    MemberAuthor

    PR Plan

    Khronos vendor suffixes

    Details
    • Reconsider the usage of KhronosImpliedVendor suffixes
      • Apparently aliased enums are simply replaced when bindings are updated. This means that users will naturally replace usages of those enums to point to their new versions. Eg: EXT to KHR. Issue is, with implied vendors, this also means that users might accidentally use the wrong enum member since the KHR version of the enum member now takes the name of the old EXT enum member. It might be better to forgo prettiness in favor of correctness here.
      • Also tempted to remove KhronosNonExclusiveVendors since I find this behavior to be more incorrect than it is helpful. It also lets me delete a lot of code (yay!).
      • History note: I believe both of these behaviors were ported from Silk 2. I ported these two behaviors from the INameTrimmer-based implementation to the current name-affix-based implementation.

    Khronos bindings improvements

    Remaining bindings I plan to work on

    MoltenVK would just be a native build. VMA is not on the tracking issue, but we already discussed that we want this to be in Silk 3 (and I want it for my personal projects as well).

    • Generate initial bindings
      • Assimp
      • SPIRV Reflect
      • SPIRV Cross
      • SPIRV Tools
      • Shaderc
      • VMA
    • Do a polish pass over the bindings
      • Assimp
      • SPIRV Reflect
      • SPIRV Cross
      • SPIRV Tools
      • Shaderc
      • VMA
    • Native builds
      • Assimp
      • SPIRV Reflect
      • SPIRV Cross
      • SPIRV Tools
      • Shaderc
      • VMA
      • MoltenVK
    • Update docs
      • Any mod changes
      • Add new bindings to changelog
    • Rebase onto develop/3.0
    • Generate using Windows for consistency
      • This must include regenerating every existing binding

    Bindings polish (may be split into multiple PRs)

    • OpenXR/CL/Vulkan: Add CurrentInstance/Device/Platform properties to interface partial.
    • ExtractEnumConstants does not preserve attributes placed on extracted constants (needs verification).
    • Ensure mod order for all jobs are consistent
    • (From dfkeenan) "Trim the fat" from the generated bindings. Stuff like SDL's DUMMY_ENUM_VALUE should be removed
    • Consider making function pointer structs work with nullptr. Seems like a nice QOL feature.
    • Why don't we use the InlineArrayN<T> types provided by .NET? Should we use them?
      • Someone in the C# discord mentioned that these were added in .NET 10, which is what led to me add this task. This might be the reason we don't use them?
    • OpenAL context should be disposable
      • This is noted by Perksey in the OpenAL.Tutorial001.HelloSound project. I need to figure out what this Dispose method should do exactly.
      • Check other API contexts as well for the same issue
        channels/521092042781229087/1514735149915439274/1519099432442794150
    • The magic nullptr constant does not work. It fails to resolve to anything.
      • Apparently it works in the Silk 3 examples. Probably because it uses a ProjectReference instead of a PackageReference.
    • The Vulkan bindings are double registering the LoaderInterface hook: https://discord.com/channels/521092042781229087/607634593201520651/1518220269519638599
      • I'm going to implement a temporary hack fix, but I don't believe it to be the correct fix.
        • We need a way to initialize an assembly consistently.
        • Currently there are multiple entrypoints and they aren't guaranteed to be called.
      • In Vulkan, we call RegisterHook in both Vk (defined in a manual file) and Vk.DllImport (defined in a generated file).
        • This causes a multiple registration error, which admittedly is my fault since I didn't realize OpenAL didn't do the same pattern.
        • Removing the duplicate call in Vk is the "fix" I mentioned above, but does not solve the underlying issue.
        • Static constructors are only invoked if the static constructor's type or members are accessed.
        • Notably, accessing Vk.DllImport does not trigger the static constructor for Vk, and vice versa.
        • Because we assume/require both static constructors to be called for proper behavior (Vk.DllImport registers the hook and is not extensible, so we have to register alternative shared library names in Vk), having only one of them be called is a problem.
          Enums and other integral typedefs:
    • Add SilkConstants class to generator
      • Ideally all magic numbers and constants are consolidated here and documented.
    • Add <auto-generated/> comment to start of each generated file. This has the effect of disabling IDE/linter/etc checks, which means less noise and performance cost of working in the Silk repo.

    These have been proving to be consistently annoying to work with...
    Vulkan and OpenXR have some special handling so this problem is less problematic, but still exists.

    • I don't think the OpenGL bindings specify the method parameter types properly
      • Eg: void DrawElements(uint mode, uint count, uint type, void* indices);
      • This should use the enum type instead of uint.
    • Vulkan: Some types are incorrect. uint is used instead of MaybeBool<uint>.
    • OpenCL: Same issue. Tracked in the example project as todos.
  5. Exanite
    added a commit that references this issue on Sep 8, 2026

    Cleanup todos in OpenCL example program

    fbeaa88
  6. Perksey
    added a commit that references this issue on Oct 10, 2026

    [3.0] Add OpenCL and OpenXR bindings (#2590)

    76913ed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions