Introduce DenseTileIntersections and SparseTileIntersections classes for Gaussian rasterization

[fwilliams]

Summary

This PR refactors the tile intersection data flow in the Gaussian rasterization pipeline from loose tensors and scalar parameters into well-defined C++ classes (DenseTileIntersections and SparseTileIntersections), and consolidates the four image-dimension scalars (imageWidth, imageHeight, imageOriginW, imageOriginH) into the existing RenderWindow2D struct throughout the autograd layer. The goal is to simplify and clean up the API surface ahead of an eventual migration of the autograd functions to Python.

Motivation

Previously, tile intersection results were passed through the pipeline as a collection of individual tensors (tileOffsets, tileGaussianIds, and for sparse: activeTiles, tilePixelMask, tilePixelCumsum, pixelMap) along with scalar metadata (tileSize, blockOffset, numCameras, etc.). These tensors were:

1. Unpacked from their dispatch results at the call site in GaussianSplat3d.cpp
2. Passed individually through autograd forward() function signatures (up to 7 extra parameters)
3. Re-packed or manually threaded into RasterizeCommonArgs on the other side

This created verbose, fragile function signatures and duplicated data-management logic across the dense, sparse, and from-world autograd paths. The same pattern existed for the four image-dimension scalars which were always passed together.

What Changed

New tile intersection classes (GaussianTileIntersection.h/.cu)

• DenseTileIntersections: Encapsulates tileOffsets [C, tilesH, tilesW], tileGaussianIds [totalIntersections], and tileSize. Can be constructed either from raw tensors or by computing intersections from 2D Gaussian means/radii/depths.

• SparseTileIntersections: Extends the dense concept with additional sparse-specific tensors: activeTiles, tilePixelMask, tilePixelCumsum, and pixelMap. Same two construction paths (from tensors or from computation).

• Both classes expose CUDA Accessor structs (guarded behind __CUDACC__) that provide device-callable helpers:

  • coordinates(blockIdx) — compute camera/tile-row/tile-col from a linear block index
  • tileGaussianRangeFromBlock(blockIdx) — get [start, end) range of gaussian IDs for a tile
  • activePixelIndexFromBlock(blockIdx, threadIdx) — (sparse only) map thread to active pixel
  • pixelIndexFromBlock(blockIdx, threadIdx) — compute global pixel index
  • gaussianIdAt(idx) — look up a gaussian ID from the intersection list

• A helper dispatchTileIntersectionsAccessor() function template creates the appropriate Accessor from either class, abstracting over the dense vs. sparse distinction at the kernel dispatch level.

Simplified RasterizeCommonArgs (GaussianRasterize.cuh)

• The struct is now parameterized as RasterizeCommonArgs<ScalarType, NUM_CHANNELS, IS_PACKED, TileIntersectionsT>, where TileIntersectionsT is either DenseTileIntersections::Accessor or SparseTileIntersections::Accessor.

• Removed ~15 member fields: mBlockOffset, mNumCameras, mTotalIntersections, mTileOriginW/H, mTileSize, mNumTilesW/H, mTileGaussianIds, mTileOffsets, mSparseTileOffsets, mTileOffsetsAreSparse, mIsSparse, mActiveTiles, mTilePixelMask, mTilePixelCumsum, mPixelMap — all replaced by a single mTileIntersections member of the accessor type.

• Removed redundant accessor methods (renderWidth(), renderHeight(), renderOriginX(), renderOriginY()); callers now use mRenderWindow.width, mRenderWindow.height, etc. directly.

Simplified autograd function signatures

All three autograd function classes were updated:

Function	Parameters removed	Parameters added
RasterizeGaussiansToPixels	imageWidth, imageHeight, imageOriginW, imageOriginH, tileSize, tileOffsets, tileGaussianIds (7 params)	renderWindow, tileIntersections (2 params)
RasterizeGaussiansToPixelsSparse	imageWidth, imageHeight, imageOriginW, imageOriginH, tileSize, tileOffsets, tileGaussianIds, activeTiles, tilePixelMask, tilePixelCumsum, pixelMap (11 params)	renderWindow, tileIntersections (2 params)
RasterizeGaussiansToPixelsFromWorld3DGS	imageWidth, imageHeight, imageOriginW, imageOriginH (4 params)	renderWindow (1 param)

Updated CUDA kernels

All five rasterization kernel files were updated to use the new Accessor API:

• GaussianRasterizeForward.cu
• GaussianRasterizeBackward.cu
• GaussianRasterizeContributingGaussianIds.cu
• GaussianRasterizeTopContributingGaussianIds.cu
• GaussianRasterizeNumContributingGaussians.cu

Host compilation fix

The #include <fvdb/detail/utils/AccessorHelpers.cuh> in GaussianTileIntersection.h was moved behind a #if defined(__CUDACC__) guard so that the header can be safely included from host-only .cpp translation units (the autograd .cpp files now transitively include it).

Updated tests

• GaussianTileIntersectionTest.cpp — updated to match new dispatch* function signatures
• GaussianRasterizeForwardTest.cpp — updated tile intersection dispatch call

Files changed (17)

• src/fvdb/GaussianSplat3d.cpp — construct RenderWindow2D and tile intersection objects at call sites
• src/fvdb/detail/autograd/GaussianRasterize.{h,cpp} — simplified signature
• src/fvdb/detail/autograd/GaussianRasterizeFromWorld.{h,cpp} — simplified signature
• src/fvdb/detail/autograd/GaussianRasterizeSparse.{h,cpp} — simplified signature
• src/fvdb/detail/ops/gsplat/GaussianRasterize.cuh — templatized + slimmed RasterizeCommonArgs
• src/fvdb/detail/ops/gsplat/GaussianRasterize{Forward,Backward,ContributingGaussianIds,TopContributingGaussianIds,NumContributingGaussians}.cu — use accessor API
• src/fvdb/detail/ops/gsplat/GaussianTileIntersection.{h,cu} — new classes + accessors
• src/tests/Gaussian{TileIntersection,RasterizeForward}Test.cpp — test updates

Test plan

• C++ unit tests compile and pass (GaussianTileIntersectionTest, GaussianRasterizeForwardTest)
• Python tests pass (python -m pytest tests/ -v)
• Full build succeeds with no new warnings
• Manual smoke test: run a 3DGS training loop to verify rasterization output is unchanged

Made with Cursor