Repository navigation
Tags
Tags
update: neraulgraphics 0.2.0, Python package and release pipeline One name everywhere: - import neraulgraphics, find_package(neraulgraphics), the neraulgraphics::core target, #include <neraulgraphics/renderer.h> and namespace neraulgraphics. The viewer is neraulgraphics-viewer. Public headers need no CUDA: - scene.h uses its own Vec3. The float3 versions of the scene types moved to src/device_scene.cuh with geom_gpu.cuh, converted at the boundary. Clients compile with no CUDA include path. - make_camera is computed in the library, so a -ffast-math caller gets the same camera bits as everyone else. Python package (step 3): - nanobind bindings built by scikit-build-core: pip install, then import neraulgraphics. render() returns a (height, width, 3) uint8 numpy array, a copy that later renders cannot change. - Vectors are any 3-sequence. ValueError for bad input, neraulgraphics.CudaError for CUDA failures. The GIL is released while rendering, with a per-renderer lock so threads can share one. - Stable ABI on Python 3.12+, type stubs and py.typed included. Releases: - The version lives only in project(VERSION); version.h, the find_package version file, the wheel metadata and __version__ read it. - The static CUDA runtime in wheels, so they need only the driver. - .github/workflows/release.yml: a vX.Y.Z tag checks that the tag, CMakeLists.txt and CHANGELOG.md agree, builds the sdist and manylinux wheels for sm_75 to sm_120, and publishes a GitHub Release. - CHANGELOG.md and RELEASING.md. Tests: tests/test_renderer.cpp under ctest and tests/test_python.py under pytest. Output is byte-identical to the original renderer from C++, from a -ffast-math C++ client and from Python.