Phase 16.9-16.10: Split oversized Render() functions

[Cle2ment]

CPUBackend::Render() (205 lines, line 39-244): extract ISPC path and C++ fallback into private methods.

CUDABackend::Render() (192 lines, line 107-299): extract interop sync, Vulkan copy, and D2H download into private methods.

[github-actions]

Review: Phase 16.9-16.10 — Split Render() Functions

Both CPUBackend::Render() and CUDABackend::Render() are indeed oversized and should be decomposed. The proposed extractions are well-identified. Here's guidance:

---

CPUBackend::Render() — Extract 2 private methods

Current structure (lines 39-244, ~206 lines):

• Lines 47-63: Setup (scene ptrs, BVH rebuild, accumulation reset)
• Lines 64-197: #ifdef PN_ISPC block — ISPC SoA packing + kernel call + unpack/accumulate
• Lines 198-242: #else — C++ std::execution::par fallback (multithreaded + single-threaded)

Suggested extraction (CPUBackend.h additions):

private:
    void RenderISPC(const Camera& camera, uint32_t* outputBuffer, uint32_t frameIndex);
    void RenderCPUFallback(uint32_t* outputBuffer, uint32_t frameIndex);

• RenderISPC → move lines 64-197 into this method (the entire #ifdef PN_ISPC block body minus the outer braces)
• RenderCPUFallback → move lines 198-242 into this method (the entire #else block body minus the outer braces)

In the cleaned Render():

if constexpr (...) { RenderISPC(...); }
else                { RenderCPUFallback(...); }

---

CUDABackend::Render() — Extract 3 private methods

Current structure (lines 117-309, ~193 lines):

• Lines 124-155: Interop toggle sync (create/destroy VkCUDAInterop)
• Lines 157-222: Scene upload + BVH upload + camera pos + ray dirs + settings
• Lines 224-242: Kernel launch + OptiX denoise
• Lines 244-308: Output delivery (interop D2H-via-buffer-copy vs CUDARenderer_GetOutput)

Suggested extraction (CUDABackend.h additions):

private:
    void SyncInteropToggle(uint32_t width, uint32_t height);
    void OutputInterop(cudaStream_t stream);
    void OutputD2H(uint32_t* outputBuffer, uint32_t width, uint32_t height);

• SyncInteropToggle → lines 131-155 (interop creation/destruction on enable change)
• OutputInterop → lines 247-299 (CUDA sync, Vulkan buffer→image barriers + copy + flush). Takes cudaStream_t as param — already extracted via CUDARenderer_GetComputeStream in the caller.
• OutputD2H → lines 302-308 (legacy device-to-host copy). Takes outputBuffer + dimensions.

In the cleaned Render():

SyncInteropToggle(width, height);
// ... scene/camera uploads ...
CUDARenderer_Render(...);
// ... OptiX denoise ...
if (m_InteropEnabled && m_Interop) {
    OutputInterop(stream);
} else {
    OutputD2H(outputBuffer, width, height);
}

---

Additional observations

1. Line count accuracy: Issue says CPUBackend is 205 lines (39-244). It's actually 206 lines. Minor nit, not worth blocking.
2. SyncInteropToggle validation: The try/catch with fprintf stderr is consistent with existing patterns in OnResize (line 51). Keep it.
3. OutputInterop encapsulation: The entire Vulkan barrier + copy pipeline is self-contained — a perfect extraction candidate. No dependencies on outer scope beyond m_Interop, m_DestinationImage, and the stream.
4. OutputD2H naming: Consider OutputCPU or OutputDeviceToHost to be more descriptive. Existing codebase uses "D2H" convention, so either works.
5. Keep the comment separators (// ─── Section ───) above each new method — the codebase uses this style throughout.
6. noexcept: Neither extracted method throws (Vulkan calls don't throw, CUDARenderer_GetOutput returns via output param). Consider marking OutputD2H as noexcept.

---

Verified: matches project conventions

• ✅ Existing private method pattern: UploadSceneToGPU (line 74) already has CUDABackend precedent
• ✅ CPP/H split consistent with file structure
• ✅ No layout contract changes needed (no GPUPacked*/GPUMaterial changes)
• ✅ No public API change — Render() stays virtual override

The refactoring is safe, correctly scoped, and follows the project's established decomposition patterns (e.g., UploadSceneToGPU was already extracted the same way).

opencode session  |  github run