cis565-fall-2016/project4-cuda-rasterizer

CIS565: CUDA Rasterizer

project4-cuda-rasterizer's Issues

Project 4: Michael Willett

Submitting an Issue rather than Pull request since I did not fork this branch.

https://github.com/GPU-Development/Project4-Iterative-Closest-Point/tree/master

Implemented Iterative-Closest-Point for point cloud registration.
Added KD-tree search to run code near real time for 30Hz depth camera (Kinect).

Multiple hierarchy model loading

tinyglTFloader doesn't handle this within the loader. Need to add some logic in the rasterizer.cu

Bug in glTF model loading

FYI: If I remembered correctly, I copied buffer data to device memory per node when traversing the node graph (in 1.0, a tree in 2.0). Buffer data should be copied to device only once for each bufferView.

Ready for Review

Hi @pjcozzi , @likangning93
Basically what I did is bringing in the tiny gltf loader, and adapt the pipeline for it. A test implementation can be found in this branch. And below is a test rendered image.

Some Issues

• tinygltfloader currently seems to have some issues loading glTF models with hierarchy.
• glTF has a big set of features which gives benefits for GL and data transfer. Yet in our low level cuda implementation, we can only take advantage of a small part of it.
  • dynamic data type (SHORT, FLOAT, etc) can not be defined as device pointer. (AFAIK) So things like glm::vec3 dev_pos still needs to be pre-defined. Our rasterizer is still glTF model specific
  • material has a lot of shader related parameters, which we cannot make most of.
  • our base code used to have only a buffer for vertex. Now glTF brings in the whole complexity of multiple bufferView and accessor. I think some knowledge of glTF is needed to write the vertex processing and primitive assembly (partly left bank for assignments), which can greatly confuse students ( I myself spend some time figuring out how to handle it ). But it can also be treated as a learning process to glTF format.
  • The pathTracer, I think will benefit even less from using glTF. I will vote for not using it in path tracer. But for WebGL project, I do think it is very beneficial.

Some updates for the assignment requirements

• I put the mouse interaction into the base code, cuz I think this is less related.

Compiling errors when using the command

HI dear author,
It's truly a honor to write a letter to you, I'm building your project nowadays and found the error when building as following, I wonder if there is a chance that you know the reason? :)

thank you
best regards to you
William

Post-mortem

• Decrease weight of Texture (maybe required) since it is much simplified with the new base code
• Base code: use constant for vertex attribute buffers and materials
• Rasterize Algorithm: Suggest starting with AABB, then Austin's scanline https://fgiesen.wordpress.com/2013/02/10/optimizing-the-basic-rasterizer/. I used this but turns out bad so I didn't recommend this, turns out mine is dumb and lack performance analysis
• Subdivide rasterize megakernel to rasterize (depth test) and interpolation
• Camera setup array and shortcut for performance analysis
• default model to something if none is specified (simplifies grading)
• rescale the ec points to 100%