A rendering engine based on gfx-hal, which mimics the Vulkan API.
Building
This library requires standard build tools for the target platforms, except in the case of windows - the spirv-compiler feature requires Ninja to be installed for compilation. https://ninja-build.org
Features
Most importantly rendy features safer API by checking important states and invariants. It checks invariants statically using marker types and dynamically with stored values.
Capability
Queue family capability defines what operation queues of the family supports. rendy provides simple mechanisms to prevent recording unsupported commands. A queue's capability level can be stored statically by marking the Family type with one of capability types: Transfer, Graphics, Compute or General (Graphics and Compute combined). Alternatively the Capability type can be used instead of the marker type, this way actual capability level can be checked dynamically.
Command buffer
rendy provides a handy wrapper named CommandBuffer. In contrast to its raw counterpart this wrapper encodes crucial information about its state directly into the type. This means users can't accidentally:
- record commands unsupported by queue family it belongs to.
- record commands when a command buffer is not in recording state.
- record render pass commands outside renderpass.
- forget to finish recording a buffer before submitting.
- resubmit a command buffer which was created for one time use.
- record execution of a primary buffer into a secondary buffer.
- etc
Memory manager
rendy's memory manager is called Heaps. Heaps provides convenient methods to sub-allocate device-visible memory based on usage and visibility requirements. It also handles mapping for specific usage types. It is possible for gfx-hal to adopt VMA. In which case rendy will use it
Rendergraph
rendy's rendergraph allows writing rendering code in simple modular style. Note that this is not a scene graph offered by high-level graphics libraries, where nodes in the graph correspond to complex objects in the world. Instead it is a graph of render passes with different properties. This makes it much easier to compose a complex frame from simple parts. A user defines nodes which declare which buffers and images it reads and writes and the rendergraph takes responsibility for transient resource allocation and execution synchronization. The user is responsible only for intra-node synchronization.
DynNode implementation - RenderPassNode can be constructed from RenderGroups collected into subpasses. RenderPassNode will do all work for render pass creating and inter-subpass synchronization. There will be more Node, DynNode and RenderGroup implementations to further simplify usage and reduce boilerplate code required for various use cases.
Cirques
This hybrid of circus and queue simplifies synchronizing host access to resources. Cirque allocates copies of the resource from resource specific allocator (e.g. CommandPool for CommandBuffers, Factory for Buffers) and gives access to the unused copy.
CPU-GPU data flow
Rendy can help to send data between device and host. The Factory type can upload data to the device local memory choosing most appropriate technique for that.
- Memory mapping will be used if device local memory happens to be cpu-visible.
- Relatively small data will be uploaded directly to buffers.
- Staging buffer will be used for bigger uploads or any image uploads.
Factorywill automatically insert synchronization commands according to user request.
GPU-CPU data flow - Not yet implemented
Data driven pipelines - WIP
We think it is possible in many common cases to feed GPU with data in semi-automatic mode. rendy::graph::node::render::RenderGroup implementation will use spirv-reflect (or similar crate) to read layout information directly from shaders and use it to automatically populate descriptors and set index/vertex buffers based on registered data encoders and provided scene instance. Current WIP implementation will use specs::World as scene to render.
Declarative pipelines - Planned
Pipelines and descriptor sets has declarative nature and it is much easier to define them declaratively. rendy provides a trait for this called DescriptorSet. Deriving it will automatically generate code necessary for set creation, writing and binding. Deriving the GraphicsPipeline trait will generate code for graphics pipeline creation and usage. A similar ComputePipeline trait exists for compute pipelines.
Example
#[derive(DescriptorSet)]
struct Example {
/// This field will be associated with binding 1 of type `VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER`.
/// Actual `Buffer` will be allocated and kept updated by `Set<Example>`.
#[descriptor(UniformBlock)]
transform: mat4,
/// This field will be associated with binding 2 of type `VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE`.
/// `ImageView` will be fetched from `Texture` which implements `Borrow<ImageView>`.
#[descriptor(SampledImage)]
texture: Texture,
/// Raw `gfx-hal` objects can be used as well.
/// But this field will make binding `Set<Example>` to a command buffer an unsafe operation
/// since it is the user's job to ensure that this raw image view is valid during command buffer execution.
#[descriptor(unsafe, SampledImage)]
foo: RawImageView,
}
Modularity
Most of the features provided by rendy can be used independently from each other This helps to keep API clean and hopefully sound. The top-level umbrella crate rendy has features for each subcrate so that they could be enabled separately (enabling a subcrate will also enable its dependencies).
Changelog
Who is using it?
Kindly open a PR or issue if you're aware of other projects using rendy.
License
Licensed under either of
- Apache License, Version 2.0, (license/APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (license/MIT or http://opensource.org/licenses/MIT)
at your option.
![[Q] How to enable a rendering backend for project?](https://avatars.githubusercontent.com/u/3727523?v=4)
1k Dec 29, 2022
5 Nov 23, 2022
7 Aug 29, 2023
103 Dec 21, 2022
10 Jan 24, 2023
3 May 4, 2023
7 Sep 4, 2023
3 Nov 3, 2022
6.5k Jan 9, 2023
2 Oct 1, 2022
6.2k Jan 2, 2023
2 Aug 25, 2022
34 Dec 15, 2022
3 Apr 16, 2022
52 Nov 14, 2022
8 Sep 16, 2024
270 Jan 5, 2023
18 Oct 19, 2022