trungtle / tlvulkanrenderer Goto Github PK

We're using raw pointers everywhere. Could potentially bite us with memory leak

Instead calling the VulkanRenderer directly, instead implement a rendering interface. This would be extremely helpful to switch out for OpenGL or DirectX12 if we so desire in the future.

By looping through each leaf that has been modified, update its bounding box and its parents bounding boxes.

Possibly can speed up loading time.
Cache pipelines?

Maybe we could voxelize the space and store the light IDs that cast shadows in that voxel

Try doing this: http://www.cs.utah.edu/~awilliam/box/box.pdf

CMake has more flexibilities than VS and more light-weight than uploading VS project settings files to github

This issue layouts the big picture plan for my thesis. To render transparency, I am using ray tracing. The key implementations are:

Acceleration structure

For ray tracing, I am choosing SBVH, Spatial Splits in BVH. The advantage of SBVH includes a faster traversal similar to kd-tree while maintaining the BVH structure. The leaf nodes only split when there is an overlap between two primitives' bounding boxes. SBVH is also a good candidate for collision detection if my timeline allows.

Parallelization on CPU

Tasks

• SBVH construction. The SBVH construction is done for the next frame while the rendering threads display the frame buffer for the current from.
• Ray tracing and tree SBVH traversal. The basic algorithm is:
  • For each pixel, generate a primary ray and input into a queue. For each available thread, grabs a ray from the queue a compute its pixel value. When the queue is empty, draw the image.
  • Since each ray is computed independently, we might incur a synchronization overhead. It would likely be more beneficial to have each thread grabs more than one ray from the queue.
• Collision detection can utilize the SBVH structure. For each object in the scene, we can traverse the tree and search for bounding box intersection.

Architecture

• A thread manager manages the available thread pool. SBVH construction is done on its own dedicated thread, while the task of ray tracing can be spread across multiple threads.
• Parallelization should be able to toggle on/off for comparison.

With the advantage of Vulkan's support for multithreading, these tasks should have increasing speed by off-loading work on the CPU.

Mesh import

The engine support glTF loading via tinygltf.

Timeline

Jan - Feb: SBVH construction and ray tracing (single thread)
March: Parallelization
March - May: Start with writing the paper

Look at https://github.com/gabime/spdlog

Right now, we generate too much unnecessary memory on the stack whenever a new leaf is created. Make sure to clean this up