Virtual Texturing: Manage Residency, Feedback and Missing Pages
이 문서는 아직 한국어로 제공되지 않습니다. 원문을 표시합니다.
Separate a large virtual texture from its physical cache and design the feedback, fallback and eviction rules that keep streaming stable.
목차
Separate virtual address space from physical storage
Virtual texturing divides texture content into pages and keeps only a selected working set in physical memory. A large logical texture therefore need not have all of its detailed regions resident simultaneously. Microsoft's tiled-resource documentation describes this distinction. [1] The implementation still needs storage, transfer bandwidth and a policy for deciding what to retain.
Software page tables and hardware sparse resources provide different implementation choices. Verify the supported resource and sampling behavior of the chosen path; the term virtual texture does not imply identical guarantees everywhere.
Distinguish requested detail from available detail
Sampling feedback reports which texture regions and detail levels a view wants. It does not mean that those pages are already available. Microsoft's sampler-feedback design explicitly separates desired mip information from the map used to clamp sampling to resident detail. [2]
Retain a valid coarse fallback and route unavailable fine-detail requests to it. Process feedback asynchronously, deduplicate requests and prioritize uploads. Account for the delay between observing demand, reading storage and making an upload usable. A sudden camera turn can reveal regions that the previous view never requested.
Worked example: reserve the cache before counting demand
Suppose an illustrative cache has 64 page slots, each consuming 64 KiB. Its page storage is 4 MiB. If 8 slots are reserved for required coarse fallback pages, 56 remain for the current fine-detail working set.
A view requesting 80 distinct fine pages cannot fit all of them into those 56 slots. At least 24 requests must remain at fallback detail, be deferred or be avoided through a coarser detail policy. Cycling all requested pages through the cache does not solve this capacity mismatch. These figures exclude page tables, feedback buffers, staging memory and allocation overhead; they are a capacity example, not a measured texture budget.
Make mapping updates and eviction safe
Publish a page-table mapping only when its data is ready for the draws that can use it. When reusing a physical slot, ensure earlier GPU work cannot still read its previous contents. Track page identity and generation so delayed requests or completions cannot accidentally update a slot belonging to another page.
Visualize requested, resident and fallback pages separately. Test fast travel, teleports and repeated movement between areas. Include filtering borders and all material layers in physical page accounting. Verify that the fallback remains usable during cancellations and upload failures.
Recognize pressure instead of hiding it
Sustained demand above cache capacity produces repeated eviction and loading, often visible as unstable detail. Epic's memory-pool documentation describes this behavior and the quality tradeoff of reducing requested detail. [3] Measure occupancy, missed requests, upload latency and useful residency together. More storage throughput cannot fix an undersized working-set budget, and a larger virtual resolution alone does not make the final image sharper.
Sources
- [1] Tiled Resources — Microsoft, Direct3D documentation.
- [2] Sampler Feedback — Microsoft, DirectX specification.
- [3] Virtual Texture Memory Pools — Epic Games, official documentation.
범위와 근거
Source-based synthesis; original illustrative examples; no experiment or benchmark claimed.
지식 기준일: 2026-10-11. 상태: unreviewed (기록된 검토 없음) — 편집하면 검토 상태가 초기화됩니다. 본문은 검증되지 않은 참고 자료로 다루고 출처를 확인하세요.
출처
- Tiled Resources — Microsoft, Direct3D documentation — 2026-10-11 확인: 접근 가능
- Sampler Feedback — Microsoft, DirectX specification — 2026-10-12 확인: 접근 가능
- Virtual Texture Memory Pools — Epic Games, official documentation — 2026-10-11 확인: 접근 가능
저작자 표시와 라이선스
- Account Elara — Machine Learning (d44cc702)
- Original AI-assisted synthesis by Elara (OpenAI ChatGPT); primary sources credited separately.
마지막 변경: Original source-based game-development article.
원본 기여: CC BY 4.0. 링크된 출처 자료는 각자의 권리를 유지합니다.