Question about motivations and goals for IGL

[birbe]

Seeing as people are more familiar with the library bgfx for C++ and wgpu for Rust, what's the motivation for starting IGL? does it internally predate both of those libraries?

[rudybear]

The project was started some time ago, and we are committed to supporting a diverse range of backend APIs. Our vision for the abstraction layer is to keep it as lean as possible, with minimal logic built in. This will enable developers to build any type of rendering engine on top of it with minimal impact to performance

[rudybear]

Thanks for the great question!

[m-hugo]

soo... why use it over bgfx/wgpu?

[GunpowderGuy]

This question wasnt answered and its a very important one

[rudybear]

https://github.com/facebook/igl#readme

There are a lot of good options for abstracting GPU API's; each making different trade-offs. We designed IGL around the following priorities:

Low-level, forward-looking API. IGL embraces modern abstractions (command buffers, state containers, bindless, etc) and is designed to give more control than OpenGL's state machine API. As a result, IGL can have leaner backends for modern API's (e.g. Metal, Vulkan).
Minimal overhead for C++. IGL supports new or existing native rendering code without overhead of language interop or the need for other language runtimes.
Reach + scale in production. IGL has been globally battle-tested for broad device reliability (especially the long-tail of Android devices as well as Quest 2/3/Pro compatibility for OpenGL/Vulkan) and performance-tuned on our apps.

[GunpowderGuy]

-Webgpu can be used in c++ ( and rust among others ) without requiring interop with other languages

-It provides much of the same modern abstractions. Strictly speaking i dont think it currently supports a native bindless mode, but neither did direct x 12 and vulkan ( all 3 used a tecnique that simulated bindlessnes ) at launch so this was deferred from the mvp as it can be added later
( Btw, does igl require the vulkan and direct x 12 extensions that provide bindlessness properly speaking ?)

-Webgpu has also been battle tested. More so as its an standard, it precedes igl and it has been adopted by a variety of groups for different ends and using a variety of languages

[smumriak]

WebGPU is much younger compared to this project. The fact that it was published to opensource just now does not mean it was not used in production for a very long time. IGL is battle tested in this sense
Also, WebGPU is an API standard, not a library. This means iteration on changes in the API is much slower in WebGPU.
Compared to bgfx - IGL is a low level library, as close as possible to underlying API. Which is a great benefit for performance tuning.
Valid comparison would be to google's angle, not bgfx or WebGPU. And the difference there is even more prominent: angle provides dead opengl API, while IGL provides modern metal/vulkan/dx12 style API

To be completely honest with you, the probability of IGL being utilized in any significant way is quite low, particularly considering Facebook's reputation for lacking commitment with the Yoga layout engine. The community strongly desires and requires a high-level library with a C API, similar to BGFX but with the user-friendliness of SFML.

[smumriak]

To be completely honest with you, the probability of IGL being utilized in any significant way is quite low, particularly considering Facebook's reputation for lacking commitment with the Yoga layout engine. The community strongly desires and requires a high-level library with a C API, similar to BGFX but with the user-friendliness of SFML.

I can not give more details without breaching NDA :)

[Snowapril]

@rudybear is there any reason igl does not choose subpass-based rendering architecture in Vulkan for android tiling architecture? As image attachment for each pass in VR environment has large resolution there might be extensive GMEM load/store costs for each VkRenderPass draw. Is there any other alternative for it?

[rudybear]

@Snowapril We do have plans to support TB architecture more efficiently. Subpasses are only a part of missing features. Hopefully we will get there soon.