Improving Free-slip Hydraulic Erosion

Release 0.3.0-Preview.2

CUDA compiler version has been updated to 11.3, future release of SuperTerrain+ will have CUDA 11.3 as minimum requirement due to compatibility restriction

Improvement to free-slip hydraulic erosion algorithm (#1)

• Erosion brush indices are now correctly computed with free-slip range instead of heightmap size.
• Re-program the structure of generateHeightmap() function (now is operator()) to make it more structured and execution-efficient.
• Normal map generation and formation are now done to all neighbours after erosion such that the rendering buffer is correct.
• Seam between chunks is now fixed with post-erosion edge interpolation with interpolation lookup table which is generated during class initialisation.
• A major overhaul to the thread model in host side.

Plan

• Cutting down kernel launch by using device subrountine, i.e., group edge interpolation, normal map generation and formation into one kernel launch, and use flag to identify which operation to perform.
• Change heightfield generator operator() so it can be more flexible and adaptive. For example in current build function forces user to compute normal map and formation in a free-slip manner if free-slip dimension has been defined when initialising the class. A potential solution is to let user input free-slip dimension at function call instead of storing the parameters in the class object.
• Adding support for async device memory management (cudaMallocAsync and cudaFreeAsync) and built-in memory pool introduced in CUDA 11.3
• Cache repeated global memory access on device into shared memory for further performance enhancement.
• Simpify multithread model.

Please refer to the documentation in source code and issue link (#1 (comment)) for more developer discussions regarding this lookup table.

Other improvements

• Fully addressed issue identified as in #14, mititgated by updating nvcc to 11.3.
• Reducing the number of memory pool, memory usage now is more efficient.
• Fixing over-allocted device and page-locked memory.
• Refactoring memory copy between device and host.
• Optimise CUDA memory R/W access.
• Constant numbers are computed inside kernel and stored in shared memory to improve performance.
• Erosion brush indices and weights are now cached into shared memory.
• CUDA kernel streams are pooled and reused to reduce stream creation/destroy overhead.
• Fixing data incoherence in multithreaded environment on host by force reloading updated chunks.

There is noticeable performance gain by applying those improvements, which is good.