When beginning this project I decided to start writing a 3D Rendering System completely from scratch as I figured that on top of learning the Deferred Rendering Algorithm, it would be a good exercise. In designing my system, I wanted to abstract the notion of a Renderer and a Render Pass in order to accommodate for future extensions to the project. The application and the scene it contains are completely agnostic to how the final image is produced. Thus making it extremely simple to exchange Renderers (Forward Lighting, Visibility Buffer,etc.) on the fly.

A sit currently stands, each frame is rendered by the Deferred Renderer in four subsequent passes, each contributing its portion to the imaged isplayed on the screen. The first pass, which I called GeometryPass, traverses the Scene Graph and fills the G-Buffer with all information necessary to perform the BRDF lighting calculation while also creating a collection of global and local lights used in subsequent passes. The second pass, which I termed Shadow Pass, takes the collection of global lights collected during the Geometry Pass and produces a shadow map for each. The Ambient Pass, uses the G-Buffer’s Kd Render Target and a constant environment value to produce the base to which other lighting passes accumulate. The fourth pass, called the Lighting Pass, is actually two distinct passes that process global lights and local lights, respectively. The last and final pass, the Tone Mapping Pass, performs the HDR to LDR conversion as well as the Gamma correction.

As previously mentioned, the Geometry Pass’ responsibility is collect all necessary surface information into the G-Buffer while simultaneously creating collections of Global Lights and Local Lights from the Scene. The Deferred Renderer maintains the G-Buffer which is comprised of four color render targets plus one depth.

1. The first color buffer holds the world position’s x, y, z, and w components in their respective r, g, b, and a channels:

   

2. The second color buffer of the G-Buffer contains the transformed normals’ x, y, and z components in the buffers r, g, and b channels leaving the alpha channel permanently set to 1:

   

3. The third color buffer contains the surface material’s Kd r, g,and b values in the buffer’s r, g, and b channels also leaving the alpha channel permanently set to 1:

   

4. The fourth and final color buffer contains the surface material’sKs r, g, and b values in the buffer’s r, g, and b channels while storing the material’s alpha value in the buffer’s alpha channel:

   

The second pass generates a shadow map for each Global Light contained in the scene by rendering the Scene from the perspective of the light sources and storing the depths of the all visible objects in the Shadow Buffer’s render target. The scene that is being used to illustrate the various stages of the Deferred Renderer’s pipeline contains two global lights.

Global Light 1	Global Light 2

The Ambient pass lays the groundwork for all subsequent lighting passes.It is the first of three passes to additively blend their results into the Light Accumulation Buffer. There are two reasons I separated the ambient term in the lighting equation from the Global and Local light passes. Since at this stage we consider the ambient light to bea constant within the scene, it would not make sense to have it be dependent on the number of lights contained in the scene and so it must be separate. Second, I plan to include Ambient Occlusion and figured it would make it easier to implement this new feature. Although at this point it is anticlimactic, below is the state of the Light Accumulation Buffer after the Ambient Pass.

As the name would indicate, this is where most of the magic happens. As stated before, this pass is split into two consecutive passes; one for global lights and one for local.

For every Global Light in the scene, this pass renders a Full-Screen-Quadrilateral and takes the world position, world normal, kd, ks, and alpha values from the G-Buffer as input and uses them to perform the BRDF lighting calculation. Additionally, the shader program takes the Shadow Map, generated by the Shadow Pass, as input to determine whether a given pixel is illuminated by its global light source. Below are the states of the Light Accumulation Buffer after the two global lights in the example scene:

Global Light 1	Global Light 2

This pass is the final of the lighting passes and must be capable of rendering thousands and up to tens-of-thousands of lights. In order to achieve this, this Deferred Renderer employs OpenGL’s built-incapability of instanced rendering. The idea behind it is that if there is a need to render a certain mesh many times during a frame, having to call a glDraw* function for each instance is wasteful due to the repeated overhead. Instead, it is encouraged to store allnecessary data associated with each instance in the GPU’s memory toallow it to optimize rendering however it sees fit. I created anobject that maintains a buffer of all local lighting information (the light’s position, intensity, and influence radius) that is uploaded onto the GPU which is then used to render all local lights in asingle API draw call. The geometric object used to render the lightvolumes is an icosahedron. I decided that it would be computationally cheaper to render a rougher approximation of a sphere rather than a traditional UV-sphere, especially when considering that 40,000 might have to transformed per frame. Additionally, the shader program that processes the pixels generated by the icosahedrons uses an attenuation factor to weaken the influence of the light as itapproaches the limit of its influence.

As is evident from the results of the last pass, the frame produced upto this point appears heavily over exposed and must go through a final processing step that accomplishes two things: converting a highdynamic range image to a low dynamic range, and calculate the gamma correction expected by the hardware’s pipeline. This pass produces the final output:

The scene which I’m using the demonstrate the rendering system is contained within a cube that is 4 x 4 x 4. The system can easily render up to 5,000 local lights with medium to large radii of influence and maintain real-time standards. Beyond this points, it can handle 10,000 to 40,000 local lights with small to medium radii of influence.

10,000 Local Lights:

40,000 Local Lights:

As was briefly mentioned in my report, I employed a couple of optimizations available through the OpenGL API. After having written a couple of shader programs for the different passes, I grew annoyed with having to resubmit the same Uniform data to different programs. After a little research online, I starting using OpenGL’s Uniform Buffer Objects that conveniently allow the sharing of uniform data between various programs while only having to send it to the GPU once. I also made use of Shader Storage Buffers in order to enable the use of instancing. I created a buffer object in the GPU’s memory that holds all local light information (world position,intensity, and radius of influence) which greatly reduces the copy operations overhead by a factor of the number of lights. In the future I plan to optimize this system further by tracking which lights have changed since the last frame in order to only have to update small chunks in the buffer object.

Aside from the Deferred Rendering aspect, the application offers various types of editors. A Scene Graph Editor, Material Editor, and TextureManager (picture below) to name a few. All GUI elements are created using dear ImGui.

Scene Editor	Material Editor	Texture Manager