Build and run on AMD RDNA4 / RADV (Linux, -DUSE_DLSS=OFF)

[ZJLi2013]

Summary

I tried vk_gaussian_splatting on AMD hardware (Radeon AI PRO R9700, RDNA4/gfx1201, RADV / Mesa 25.2.8, no NVIDIA runtime) and hit three small, unrelated issues that blocked a clean build and headless run on Linux without DLSS. This PR fixes all three; with them the raster, hardware ray tracing, and hybrid pipelines all build and render on RADV (vkCreateRayTracingPipelinesKHR succeeds). The changes are platform-agnostic and do not alter behavior when USE_DLSS is enabled.

The work is split into three focused commits so they can be reviewed (or cherry-picked) independently.

Tested on: base dfb3c79 (nvpro_core2 eb7c2f2), LunarG Vulkan SDK 1.4.350.1, Ubuntu, headless, -DUSE_DLSS=OFF -DDISABLE_DEFAULT_SCENE=ON.

Changes

1. Build with -DUSE_DLSS=OFF (no NVIDIA NGX SDK)

Two things break the build without the NGX SDK even with DLSS disabled:

• file(GLOB SOURCE_FILES src/*.*) unconditionally picks up src/dlss_*.cpp (which include NGX headers). These files are already added explicitly in the if(USE_DLSS) block, so the glob only ever double-adds them — excluding them from the glob is safe:

  file(GLOB SOURCE_FILES src/*.*)
  list(FILTER SOURCE_FILES EXCLUDE REGEX "/dlss_")

• m_dlss is declared only under #if defined(USE_DLSS), but gaussian_splatting_ui.cpp reads m_dlss.isEnabled() at four sites outside the guards → compile error. Added a small isDlssEnabled() helper that returns false when compiled without DLSS, and used it at those sites:

  inline bool isDlssEnabled() const
  {
  #if defined(USE_DLSS)
    return m_dlss.isEnabled();
  #else
    return false;
  #endif
  }

2. Portable asset paths in .vkgs projects

Sample projects store asset paths with Windows backslashes (e.g. data\teleportour\Winter-Garden-view2.ply). On POSIX, std::filesystem treats \ as a normal filename character, so assets fail to resolve and the scene loads empty (blank render, no error). Normalizing separators to / in the single choke point makeAbsolutePath() fixes it and stays valid on Windows too:

std::string normalized = relativePath;
std::replace(normalized.begin(), normalized.end(), '\\', '/');
return std::filesystem::absolute(base / normalized);

3. Headless --saveImage before the first frame

Passing --saveImage <file> as a bare command-line argument (outside a benchmark SEQUENCE) crashes with SIGSEGV: the save runs at parse time, before any frame is rendered, and the G-buffers are created lazily on the first onResize(), so getAllDumpableBuffers() dereferences not-yet-created color images. Returning an empty list when the G-buffer is still 0x0 makes it fail gracefully (the SEQUENCE-based path is unchanged):

const VkExtent2D gBufSize = m_gBuffers.getSize();
if(gBufSize.width == 0 || gBufSize.height == 0)
  return buffers;  // not rendered yet

If you'd prefer this to defer the save until after the first frame rather than no-op, I'm happy to adjust — that felt like a larger behavior change than this minimal guard.

Testing

• Builds cleanly (315/315 targets) with -DUSE_DLSS=OFF -DDISABLE_DEFAULT_SCENE=ON on RADV/Mesa 25.2.8 (RDNA4), LunarG SDK 1.4.350.1.
• Headless benchmark SEQUENCE: raster / RT / hybrid pipelines all produce images.
• The .vkgs sample project (15.3M-gaussian splat + OBJ meshes) loads and renders after the path fix.
• No NVIDIA hardware on hand to regression-test the DLSS-on path, but none of these changes affect it: the helper returns the same value under USE_DLSS, the glob filter is a no-op (those sources are added explicitly), and the path / saveImage fixes are platform-agnostic.

[marviej]

hello @ZJLi2013 ,
thank you very much for the work. could you signoff the three commits of the PR so i can process you PR.
See https://github.com/nvpro-samples/vk_gaussian_splatting/pull/33/checks?check_run_id=86054702117
thanks.