feat(gsplat): add from-world rasterization path and UT integration

Introduce differentiable dense from-world Gaussian rasterization with UT-aligned camera/projection APIs, gradient-capable render paths, and focused C++/Python coverage for correctness and regression safety.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: fwilliams

.github/workflows/cu130-nightly.yml

@@ -1,4 +1,4 @@
-name: fVDB Nightly Build and Tests -- CUDA 13.0.1
+name: fVDB Nightly Build and Tests -- CUDA 13.0.2
 
 on:
 #   schedule:
@@ -19,7 +19,7 @@ jobs:
     runs-on:
       - self-hosted
     container:
-      image: nvidia/cuda:13.0.1-cudnn-devel-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-devel-ubuntu22.04
       env:
         PYTHONPATH: ""
         CPM_SOURCE_CACHE: "/__w/cpm_cache"
@@ -164,7 +164,7 @@ jobs:
     runs-on:
       - self-hosted
     container:
-      image: nvidia/cuda:13.0.1-cudnn-devel-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-devel-ubuntu22.04
       env:
         PYTHONPATH: ""
       options: --rm

.github/workflows/cu130.yml

@@ -1,4 +1,4 @@
-name: fVDB CUDA 13.0.1 Build and Test
+name: fVDB CUDA 13.0.2 Build and Test
 
 on:
   pull_request_target:
@@ -61,7 +61,7 @@ jobs:
     needs: start-build-runner
     runs-on: ${{ needs.start-build-runner.outputs.label }}
     container:
-      image: nvidia/cuda:13.0.1-cudnn-devel-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-devel-ubuntu22.04
       env:
         PYTHONPATH: ""
         CPM_SOURCE_CACHE: "/__w/cpm_cache"
@@ -229,7 +229,7 @@ jobs:
     name: fVDB GTests
     runs-on: ${{ needs.start-tests-gpu-runner.outputs.label }} # run the job on the newly created runner
     container:
-      image: nvidia/cuda:13.0.1-cudnn-runtime-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-runtime-ubuntu22.04
       env:
         PYTHONPATH: ""
         CPM_SOURCE_CACHE: "/__w/cpm_cache"
@@ -332,7 +332,7 @@ jobs:
     name: fVDB PyTests
     runs-on: ${{ needs.start-tests-gpu-runner.outputs.label }} # run the job on the newly created runner
     container:
-      image: nvidia/cuda:13.0.1-cudnn-devel-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-devel-ubuntu22.04
       env:
         PYTHONPATH: ""
       options: --rm
@@ -421,7 +421,7 @@ jobs:
     name: fVDB Documentation Tests
     runs-on: ${{ needs.start-tests-gpu-runner.outputs.label }} # run the job on the newly created runner
     container:
-      image: nvidia/cuda:13.0.1-cudnn-runtime-ubuntu22.04
+      image: nvidia/cuda:13.0.2-cudnn-runtime-ubuntu22.04
       env:
         PYTHONPATH: ""
       options: --rm

build.sh

@@ -29,6 +29,7 @@ usage() {
   echo "  lineinfo       Enable CUDA lineinfo (sets FVDB_LINEINFO=ON)."
   echo "  strip_symbols  Strip symbols from the build (will be ignored if debug is enabled)."
   echo "  verbose        Enable verbose build output for pip and CMake."
+  echo "  trace          Enable CMake trace output for debugging configuration."
   echo ""
   echo "  Any modifier arguments not matching above are passed through to pip."
   exit 0
@@ -145,6 +146,70 @@ PY
   fi
 }
 
+# --- Path Sanitization for Virtualized Environments ---
+# In certain virtualized or hybrid environments, the PATH may include directories
+# on slow or remote-mounted filesystems (e.g., /mnt/* host mounts, /media/sf_* for
+# VirtualBox, /mnt/hgfs/* for VMware). CMake's find_library() searches these paths,
+# and each filesystem operation on such mounts can go through slow protocols (9P, vboxsf,
+# vmhgfs-fuse), causing significant delays during configuration.
+# By filtering out these paths, we prevent CMake from crawling slow directories.
+
+has_slow_host_mounts() {
+  # Detect virtualized environments that may have slow host filesystem mounts
+  # - WSL/WSL2: /mnt/* paths use 9P protocol
+  # - VirtualBox: /media/sf_* paths use vboxsf
+  # - VMware: /mnt/hgfs/* paths use vmhgfs-fuse
+  if [ -f /proc/version ]; then
+    grep -qi "microsoft\|wsl" /proc/version && return 0
+  fi
+  # VirtualBox Guest Additions
+  if [ -d /media/sf_* ] 2>/dev/null || grep -q vboxsf /proc/filesystems 2>/dev/null; then
+    return 0
+  fi
+  # VMware Tools
+  if [ -d /mnt/hgfs ] || grep -q vmhgfs /proc/filesystems 2>/dev/null; then
+    return 0
+  fi
+  return 1
+}
+
+filter_slow_mount_paths() {
+  local input_path="$1"
+  if has_slow_host_mounts; then
+    # Filter out paths on slow host filesystem mounts:
+    # - /mnt/* (WSL host mounts, VMware hgfs)
+    # - /media/sf_* (VirtualBox shared folders)
+    echo "$input_path" | tr ':' '\n' | grep -v -E "^/mnt/|^/media/sf_" | tr '\n' ':' | sed 's/:$//'
+  else
+    echo "$input_path"
+  fi
+}
+
+run_with_sanitized_paths() {
+  # Run a command with sanitized paths in virtualized environments, or normally otherwise.
+  # This prevents CMake from searching slow host-mounted filesystems.
+  if has_slow_host_mounts; then
+    local sanitized_path
+    sanitized_path=$(filter_slow_mount_paths "$PATH")
+
+    # Build env command with sanitized variables
+    local env_args=("PATH=$sanitized_path")
+
+    # Also sanitize CMAKE_PREFIX_PATH and CMAKE_LIBRARY_PATH if set
+    if [ -n "$CMAKE_PREFIX_PATH" ]; then
+      env_args+=("CMAKE_PREFIX_PATH=$(filter_slow_mount_paths "$CMAKE_PREFIX_PATH")")
+    fi
+    if [ -n "$CMAKE_LIBRARY_PATH" ]; then
+      env_args+=("CMAKE_LIBRARY_PATH=$(filter_slow_mount_paths "$CMAKE_LIBRARY_PATH")")
+    fi
+
+    echo "Virtualized environment detected: Running with sanitized paths (excluding slow host mounts for performance)"
+    env "${env_args[@]}" "$@"
+  else
+    "$@"
+  fi
+}
+
 if [[ "$1" == "-h" || "$1" == "--help" ]]; then
   usage
 fi
@@ -169,6 +234,40 @@ CONFIG_SETTINGS=""
 PASS_THROUGH_ARGS=""
 CUDA_ARCH_LIST_ARG="default"
 
+# --- Set CUDA_HOME / Hints ---
+# This helps scikit-build-core find the CUDA toolkit directly.
+#
+# We only auto-detect CUDA_HOME for Conda environments, and only when:
+#   1. CUDA_HOME is not already set (respect user's explicit choice)
+#   2. The Conda environment actually has nvcc installed
+#
+# In venv environments, PyTorch ships with rpathed CUDA libraries, so we leave
+# CUDA_HOME unset to avoid version mismatches. Users can still set CUDA_HOME
+# explicitly if needed.
+
+if [ -n "$CONDA_PREFIX" ] && [ -z "$CUDA_HOME" ] && [ -x "$CONDA_PREFIX/bin/nvcc" ]; then
+    echo "Conda environment with CUDA detected. Setting CUDA_HOME to $CONDA_PREFIX"
+    export CUDA_HOME="$CONDA_PREFIX"
+elif [ -n "$CONDA_PREFIX" ] && [ -n "$CUDA_HOME" ]; then
+    echo "Conda environment detected, but CUDA_HOME is already set to $CUDA_HOME (keeping existing value)"
+fi
+
+# Export hints to help CMake find CUDA (only if CUDA_HOME is set)
+if [ -n "$CUDA_HOME" ]; then
+    # Always pass CUDA_TOOLKIT_ROOT_DIR as a search hint
+    CONFIG_SETTINGS+=" --config-settings=cmake.define.CUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME"
+
+    # Only set CUDACXX and CMAKE_CUDA_COMPILER if nvcc actually exists at that location.
+    # This avoids hard failures when CUDA_HOME points to a runtime-only installation.
+    if [ -x "$CUDA_HOME/bin/nvcc" ]; then
+        export CUDACXX="${CUDA_HOME}/bin/nvcc"
+        CONFIG_SETTINGS+=" --config-settings=cmake.define.CMAKE_CUDA_COMPILER=$CUDACXX"
+    else
+        echo "Warning: CUDA_HOME is set to $CUDA_HOME but nvcc not found at $CUDA_HOME/bin/nvcc"
+        echo "         CMake will attempt to find nvcc elsewhere."
+    fi
+fi
+
 # Default values for nanovdb_editor build options
 NANOVDB_EDITOR_SKIP=OFF
 NANOVDB_EDITOR_FORCE=OFF
@@ -188,6 +287,10 @@ while (( "$#" )); do
       echo "Enabling verbose build"
       CONFIG_SETTINGS+=" -v -C build.verbose=true"
       is_config_arg_handled=true
+    elif [[ "$1" == "trace" ]]; then
+      echo "Enabling CMake trace"
+      CONFIG_SETTINGS+=" --config-settings=cmake.args=--trace-expand"
+      is_config_arg_handled=true
     elif [[ "$1" == "debug" ]]; then
       echo "Enabling debug build"
       CONFIG_SETTINGS+=" --config-settings=cmake.build-type=Debug  -C cmake.define.CMAKE_BUILD_TYPE=Debug"
@@ -248,16 +351,16 @@ fi
 if [ "$BUILD_TYPE" == "wheel" ]; then
     echo "Build wheel"
     echo "pip wheel . --no-deps --wheel-dir dist/ $PIP_ARGS"
-    pip wheel . --no-deps --wheel-dir dist/ $PIP_ARGS
+    run_with_sanitized_paths pip wheel . --no-deps --wheel-dir dist/ $PIP_ARGS
+
 elif [ "$BUILD_TYPE" == "install" ]; then
     echo "Build and install package"
+    # Always use --force-reinstall to ensure the freshly built package is installed,
+    # even if pip thinks the version is already satisfied. The --no-deps flag ensures
+    # this only affects fvdb-core itself, not dependencies like torch.
     echo "pip install --no-deps --force-reinstall $PIP_ARGS ."
-    pip install --no-deps --force-reinstall $PIP_ARGS .
-# TODO: Fix editable install
-# else
-#     echo "Build and install editable package"
-#     echo "pip install $PIP_ARGS -e .  "
-#     pip install $PIP_ARGS -e .
+    run_with_sanitized_paths pip install --no-deps --force-reinstall $PIP_ARGS .
+
 elif [ "$BUILD_TYPE" == "ctest" ]; then
 
     # --- Ensure Test Data is Cached via CMake Configure Step ---
@@ -279,7 +382,7 @@ elif [ "$BUILD_TYPE" == "ctest" ]; then
 
     echo "Running CMake configure in temporary directory ($TEMP_BUILD_DIR) to trigger data download..."
     pushd "$TEMP_BUILD_DIR" > /dev/null
-    cmake "$SOURCE_DIR/src/cmake/download_test_data"
+    run_with_sanitized_paths cmake "$SOURCE_DIR/src/cmake/download_test_data"
     popd > /dev/null # Back to SOURCE_DIR
 
     # Clean up temporary directory
@@ -289,14 +392,24 @@ elif [ "$BUILD_TYPE" == "ctest" ]; then
 
     # --- Find and Run Tests ---
     echo "Searching for test build directory..."
-    # Find the directory containing the compiled tests (adjust if needed)
-    # Using -print -quit to stop after the first match for efficiency
-    BUILD_DIR=$(find build -name tests -type d -print -quit)
-
-    if [ -z "$BUILD_DIR" ]; then
-        echo "Error: Could not find build directory with tests"
-        echo "Please enable tests by building with pip argument"
-        echo "-C cmake.define.FVDB_BUILD_TESTS=ON"
+    # Find CMakeCache.txt to locate the build root
+    CMAKE_CACHE=$(find build -name CMakeCache.txt -type f -print -quit 2>/dev/null)
+
+    if [ -z "$CMAKE_CACHE" ]; then
+        echo "Error: Could not find CMakeCache.txt in build directory"
+        echo "Please build the project first with tests enabled:"
+        echo "pip install . -C cmake.define.FVDB_BUILD_TESTS=ON"
+        exit 1
+    fi
+
+    # Construct the test directory path (where CTestTestfile.cmake is generated)
+    # This discovers all tests from both src/tests/ and src/dispatch/
+    BUILD_DIR="$(dirname "$CMAKE_CACHE")/src"
+
+    if [ ! -f "$BUILD_DIR/CTestTestfile.cmake" ]; then
+        echo "Error: No CTestTestfile.cmake found in $BUILD_DIR"
+        echo "Please enable tests by building with:"
+        echo "pip install . -C cmake.define.FVDB_BUILD_TESTS=ON"
         exit 1
     fi
     echo "Found test build directory: $BUILD_DIR"
@@ -306,9 +419,10 @@ elif [ "$BUILD_TYPE" == "ctest" ]; then
     add_python_pkg_lib_to_ld_path "nanovdb_editor" "NanoVDB Editor" "libpnanovdb*.so"
 
     # Run ctest within the test build directory
+    # Note: ctest doesn't need path sanitization as it doesn't search for libraries
     pushd "$BUILD_DIR" > /dev/null
     echo "Running ctest..."
-    ctest --output-on-failure
+    ctest --output-on-failure -LE compile_fail
     CTEST_EXIT_CODE=$?
     popd > /dev/null # Back to SOURCE_DIR
 

docs/conf.py

@@ -15,7 +15,7 @@
 
 # -- Project information -----------------------------------------------------
 
-project = "fVDB"
+project = "ƒVDB"
 copyright = "Contributors to the OpenVDB Project"
 author = "Contributors to the OpenVDB Project"
 
@@ -76,6 +76,9 @@
 # relative to this directory. They are copied after the builtin static files,
 # so a file named "default.css" will overwrite the builtin "default.css".
 html_static_path = ["imgs"]
+html_css_files = [
+    "css/custom.css",
+]
 
 
 # -- Custom hooks ------------------------------------------------------------

docs/imgs/css/custom.css

@@ -0,0 +1,3 @@
+.wy-side-nav-search {
+    background-color: #76b900;
+}

docs/index.rst

@@ -1,7 +1,7 @@
 Welcome to ƒVDB!
 =================
 
-fVDB is a Python library of data structures and algorithms for building high-performance and large-domain
+ƒVDB is a Python library developed and maintained by NVIDIA, containing a collection of data structures and algorithms for building high-performance and large-domain
 spatial applications using `NanoVDB <https://dl.acm.org/doi/abs/10.1145/3450623.3464653>`_ on the GPU
 in `PyTorch <https://pytorch.org/>`_.
 Applications of fVDB include 3D deep learning, computer graphics/vision, robotics, and scientific computing.
@@ -14,18 +14,18 @@ Applications of fVDB include 3D deep learning, computer graphics/vision, robotic
 
 |
 
-fVDB aims to be production ready with a focus on robustness, usability, and extensibility.
+ƒVDB aims to be production ready with a focus on robustness, usability, and extensibility.
 It is designed to be easily integrated into existing pipelines and workflows, and to support a
-wide range of use cases and applications. To this end, fVDB has a minimal set of dependencies and
+wide range of use cases and applications. To this end, ƒVDB has a minimal set of dependencies and
 is open source under the Apache 2.0 license as part of the `The Academy Software Foundation's
 OpenVDB project <https://www.openvdb.org>`_.  Contributions and feedback from the community are welcome
-to fVDB's `GitHub repository <https://github.com/openvdb/fvdb-core>`_.
+to ƒVDB's `GitHub repository <https://github.com/openvdb/fvdb-core>`_.
 
 
 Features
 --------
 
-fVDB provides the following key features:
+ƒVDB provides the following key features:
 
 -   A sparse volumetric grid data structure optimized for GPU memory efficiency and performance.
 -   A highly optimized Gaussian splat data structure for representing radiance fields on the GPU.
@@ -35,40 +35,40 @@ fVDB provides the following key features:
 -   Modular neural network components for building 3D deep learning models that scale to large input sizes.
 -   Seamless integration with PyTorch for easy use in deep learning workflows.
 
-The videos below show fVDB being used for large-scale 3D reconstruction, simulation, and interactive visualization.
+The videos below show ƒVDB being used for large-scale 3D reconstruction, simulation, and interactive visualization.
 
 .. raw:: html
 
    <p style="text-align: center; font-weight: bold; font-style: italic; text-decoration: underline; font-size: medium; text-decoration-skip-ink: none; margin-bottom: 0.5em;">
-   fVDB being used to reconstruct radiance (25 million splats) fields and TSDF volumes (100 million voxels) from images and points</p>
+   ƒVDB being used to reconstruct radiance (25 million splats) fields and TSDF volumes (100 million voxels) from images and points</p>
   <video autoplay loop controls muted width="90%" style="display: block; margin: 0 auto;">
      <source src="https://fvdb-data.s3.us-east-2.amazonaws.com/fvdb-reality-capture/spot_airport_480p.mp4" type="video/mp4" />
   </video>
 
    <br>
    <p style="text-align: center; font-weight: bold; font-style: italic; text-decoration: underline; font-size: medium; text-decoration-skip-ink: none; margin-bottom: 0.5em;">
-   fVDB being used to process a sparse SDF on a grid with 181 million voxels. Visualized in a browser.</p>
+   ƒVDB being used to process a sparse SDF on a grid with 181 million voxels. Visualized in a browser.</p>
   <video autoplay loop controls muted width="90%" style="display: block; margin: 0 auto;">
      <source src="https://fvdb-data.s3.us-east-2.amazonaws.com/fvdb-reality-capture/crawler_480p.mp4" type="video/mp4" />
   </video>
 
 |
 
-About fVDB
+About ƒVDB
 --------------
 
-fVDB was first developed by the `NVIDIA High-Fidelity Physics Research Group <https://research.nvidia.com/labs/prl/>`_
+ƒVDB was first developed by the `NVIDIA High-Fidelity Physics Research Group <https://research.nvidia.com/labs/prl/>`_
 within the `NVIDIA Spatial Intelligence Lab <https://research.nvidia.com/labs/sil/>`_, and continues to be
 developed with the OpenVDB community to suit the growing needs for a robust framework for
 spatial intelligence research and applications.
 
 
-fVDB Reality Capture Toolbox
+ƒVDB Reality Capture
 --------------------------------
 
-In addition to the core fVDB library, we also provide the `fVDB Reality Capture <https://fvdb.ai/reality-capture>`_ toolbox,
-which is a collection of tools and utilities for 3D reconstruction and scene understanding using fVDB. Analogous to how `torchvision <https://pytorch.org/vision/stable/index.html>`_
-provides datasets, models, and transforms for computer vision tasks, `fVDB Reality Capture <https://fvdb.ai/reality-capture>`_ provides datasets, models, and
+In addition to the core ƒVDB library, we also provide the `ƒVDB Reality Capture <https://fvdb.ai/reality-capture>`_ toolbox,
+which is a collection of tools and utilities for 3D reconstruction and scene understanding using ƒVDB. Analogous to how `torchvision <https://pytorch.org/vision/stable/index.html>`_
+provides datasets, models, and transforms for computer vision tasks, `ƒVDB Reality Capture <https://fvdb.ai/reality-capture>`_ provides datasets, models, and
 algorithms for 3D reconstruction from sensor data.
 
 .. toctree::
@@ -77,6 +77,12 @@ algorithms for 3D reconstruction from sensor data.
 
    self
    installation
+   GitHub Repository <https://github.com/openvdb/fvdb-core>
+
+.. toctree::
+   :caption: Applications
+
+   ƒVDB Reality Capture <https://fvdb.ai/reality-capture>
 
 .. toctree::
    :maxdepth: 1