[!NOTE] This page has not been translated yet. You are reading the original English version. Want to help? See the translation contribution guide.

Auto-Exposure Optimization Report (March 2026)¶

This report details the transition of the Auto-Exposure system from a hybrid Raster/Compute pipeline to a unified 100% Compute implementation.

1. Architectural Evolution¶

Baseline (Hybrid)¶

In the baseline version, the screen-space luminance was calculated using traditional rendering passes.

graph LR
    SCENE[Scene Color Tex] --> DS[FS: Downsample 64x64]
    DS --> FBO[Downsample FBO]
    FBO -- "RASTER-TO-COMPUTE SYNC" --> ADAPT[CS: Adaptation 1x1]
    ADAPT --> EXP[Exposure Tex]

Optimized (Unified Compute)¶

The new implementation eliminates the intermediate Framebuffer (FBO) and stays entirely within the GPU's compute domain.

graph LR
    SCENE[Scene Color Tex] --> DS[CS: Downsample 64x64]
    DS -- "MEMORY BARRIER (Light)" --> ADAPT[CS: Adaptation 1x1]
    ADAPT --> EXP[Exposure Tex]

2. Statistical Results¶

Measurements taken on Intel Iris Xe Graphics with a population of 60 samples per stage.

Stage	Raster Baseline (ms)	Compute Optimized (ms)	Delta	Status
AE Downsample	0.1815	0.6370	+0.45 ms	🔻 Slower
AE Adaptation	0.4517	0.0009	-0.45 ms	🚀 Faster
Total AE Effect	0.6332	0.6379	+0.00 ms	🆗 Parity
---	---	---	---	---
Total Frame Time	2.0012	1.9516	-0.05 ms	✅ Gain

3. Post-Benchmark Analysis¶

The "Sync Tax" Analysis¶

The most striking result is the 99% reduction in the cost of the Adaptation pass.

In Baseline: The 0.45ms cost was artificial. The GPU was actually mostly idle (Bubble), waiting for the Raster Downsample to finish writing to the FBO and for the caches to be flushed before the Compute Shader could safely read the texture.
In Optimized: Since both passes are Compute, they share the same execution context. The barrier between them is an order of magnitude faster than a Raster Flush.

The "Ghost Gain" (Frame Efficiency)¶

Even though the Auto-Exposure effect itself costs roughly the same time (0.63ms), the Total Frame Time decreased by 0.05ms.

Framebuffer Management: This is due to the removal of glBindFramebuffer calls.
Driver Overhead: Reducing framebuffer switches minimizes the "Driver Overhead", allowing the GPU to stay at high utilization without constant context reconfiguration.

4. Visual Parity & Stability¶

The optimization was validated using the Visual Regression Testing System:

Error Rate: 0.0% difference against Master references.
Edge Case Handling: Verified that logarithmic black blocks (Sentinel values: -100.0) are correctly handled in the parallel reduction.

5. Dual-Path Runtime Toggle (April 2026)¶

Following benchmarking on multiple GPUs (Intel Iris Xe iGPU, NVIDIA 950M), the compute downsample path showed no clear performance advantage on older/integrated hardware — and was sometimes slower due to the graphics→compute pipeline transition cost.

Rather than reverting the compute work, the system now supports both render paths at runtime:

Path	Downsample Method	Best For
Fragment (default)	Fullscreen quad + FBO	iGPU, older dGPU (Maxwell)
Compute	`glDispatchCompute` + imageStore	Modern dGPU with async compute

Switching at Runtime¶

Ctrl+J  →  Toggle between Fragment / Compute downsample
J       →  Toggle Auto-Exposure ON/OFF (unchanged)
Shift+J →  Exposure Debug histogram (unchanged)

A notification ("AE Path: Fragment" / "AE Path: Compute") confirms the switch. The GPU Profiler label also changes dynamically to reflect which path is active.

Architecture¶

graph TD
    SCENE[Scene Color Texture]
    SCENE --> SWITCH{active_path?}
    SWITCH -->|Fragment| FRAG["FS Quad → FBO 64x64"]
    SWITCH -->|Compute| COMP["CS Dispatch 4x4 WG"]
    FRAG --> DS_TEX[Downsample Tex R32F]
    COMP --> DS_TEX
    DS_TEX --> ADAPT["CS: Adaptation 1x1"]
    ADAPT --> EXP[Exposure Tex]

Both paths write to the same shared downsample texture (R32F 64x64). The adaptation compute shader is unchanged. Resource overhead of the unused path is negligible (1 FBO + 1 shader program in memory).

Default Path Rationale¶

Fragment is the default because:

The rasterizer's spatial texture prefetch is heavily optimized for this workload (4x4 box filter on a 64x64 target)
No pipeline transition barrier is needed (stays in graphics context)
Proven stable across all tested hardware

6. Conclusion¶

While the Compute Downsampler is technically slower than Raster Hardware on some architectures (like Intel integrated graphics), the Zero-Sync architecture makes the overall frame more efficient.

This refactoring is a critical foundation for Phase 2 (Bloom), as it demonstrates that moving post-processing steps to Compute is the most viable path for long-term performance scalability.