I have made a new OptiX 7 implementation of SPCBPT at here. It is much more faster, easier to compile and straight forward to read, I would recommend to check that version to get a better understanding of SPCBPT.

An unstable SPCBPT code

requirement：

• optix 5.0, incompatible with later versions
• cuda 10.0+
• Visual Studio 2015 (toolset v140)
• Cmake

How to Build:

• Start up cmake-gui from the Start Menu.

• Select the "src" directory and the source code

• Create a build directory that isn't the same as the source directory.

• Press "Configure" button and select the version of Visual Studio, our code is compatible with VS 2015 only (for its v140 toolset).

• Select "x64" as the platform

• Press "OK".

• Set OptiX_INSTALL_DIR to wherever you installed OptiX, e.g., C:/ProgramData/NVIDIA Corporation/OptiX SDK 5.0.1

• Set CUDA_SDK_ROOT_DIR to the path to cuda samples, e.g., C:/ProgramData/NVIDIA Corporation/CUDA Samples/v10.1

• Press "generate" and then "open Project"

I am struggling with Cmake. So you need to manually compile the hello_cuda.cu file currently. What you need to do is:

• Find the optixPathTracer project
  • Right click - Build dependencies - Build Customizations
  • Enable cuda
• Right click “Cuda Files” directory
  • Add-Existing item
  • find "hello_cuda.cu" in src/optixPathTracer directory and add it to the project
• Build the optixPathTracer project in release

Visual Studio may ask you to "Reload All" the project. Agree and do the above operations again to compile the "hello_cuda.cu" file.

Finally, you can Right click the optixPathTracer project and set it as Startup project and run the renderer program.

———————————————————————————————————————————————————————————————————————

The code corresponding to the key parts of our algorithms

• Determines $\Gamma$ and Q using train_api.data.get_data, measure the time for preprocessing, and transports $\Gamma$ and $Q$ to the device. The classification function $\kappa$ is determined and transferred to the device by train_api.data.get_data

  • train_api.data.get_data
    • Pre-traces the full paths used for trainning. We use LVCBPT for the pre-tracing but it may introduce some correlation problem and make a less effective $\Gamma$, we will consider replace it by a simple path tracer.
    • Determine the classification function $\kappa$ (and devide the sub-path space into subspaces thereby)
      • The labeled sub-paths used for decision tree come from get_rough_weighted_sample
        • Cut the prefix and suffix sub-paths by classification_data_get_flat
        • Sample the centroid sub-paths and label others by HS_algorithm::stra_hs, here we use a stratified sampled method to get the centroid sub-paths, but its optimization is slight compared to sampling the centroid based on the contribution directly.
      • Build the decision trees (with depth limitation of 12) and transfer ithem to the device
    • Label each each light sub-path, prefix and sufffix sub-path of full paths with its subspace ID
      • Get Q
      • Train the MIS-aware $\Gamma$
      • Trick: Starting training from the contribution-based $\Gamma$ can make a better result in the same time budget.

• Light sub-path tracing

  • device

    • Function light_vertex_launch in path_trace_camera.cu
    • Trace the light sub-path in the GPU

  • host

    • Function light_cache_process_ZGCBPT in path_trace_camera.cu
    • Make statistics on necessary parameters
    • Build the light subspace for the second stage sampling

• Eye sub-path tracing

  • Function ZGCBPT_pinhole_camera in path_trace_camera.cu
  • Trace an eye sub-path for each pixel
  • Two stage sampling for each prefix eye sub-path

• Hit program

  • hit program.cu
    • BDPT_closest_hit for eye sub-path
    • BDPT_L_closest_hit for light sub-path

• MIS computation

  • rmis.h
  • Use a RMIS style and is difficult to understand.