Skip to content

v0.0.9m2

Choose a tag to compare

@TheAutomatic TheAutomatic released this 20 Nov 03:12
· 0 commits to main since this release

DLSSNR-AMD v0.0.9m2 (Windows Preview)

A major fork update over upstream v0.0.9 (9503f6b), featuring a Zero-Heap D3D12 Direct passthrough refactor, Standalone mode for DX11 / DX12 / Vulkan apps without OptiScaler or ReShade (including a D3D11 → D3D12 Direct bridge for apps like OBS Studio), and full read-only install directory support.

Highlights

1. Zero-Heap D3D12 Direct Passthrough (Lower VRAM & CPU Overhead)

  • Zero descriptor heaps: Refactored D3D12 Direct outer shaders (RenoDX / helper passes) onto precompiled native ACO ELF pipelines (d3d12_outer.nra) with compute-only state tracking and a read-only heap observer — eliminating descriptor-heap switches that could disturb game/host rendering state.
  • Idle RenoDX buffer elision & expanded direct formats: Removed unnecessary intermediate buffer allocations on the D3D12 Direct path, added AMD 1-level indirect SRD support, and expanded direct format passthrough.
  • RX 9070 XT Benchmarks vs v0.0.9 (CC_SILVER_AARDWOLD_OPTIMIZED Ultra Performance, AMD 26.11.1, full details in docs/BENCHMARK-D3D12-ZERO-HEAP.md):
    • D3D12 Direct Dedicated VRAM: -34.6 MiB (-6.0%) at 1080p, -60.0 to -60.6 MiB (-7.0% to -7.2%) at 1440p, and -130.1 to -130.5 MiB (-8.7% to -9.1%) at 4K vs v0.0.9 (-106 to -308 MiB / -16% to -19% vs Vulkan Bridge).
    • D3D12 Direct CPU recording time: -14% to -19% faster vs v0.0.9 (0.24–0.27 ms vs 0.29–0.33 ms, and -71% to -74% lower than Vulkan Bridge).
    • GPU throughput: Full parity with v0.0.9 on both D3D12 Direct and Vulkan Bridge (no regression), plus clean process exit on Vulkan Bridge teardown.

2. Standalone Mode (DX11 / DX12 / Vulkan) & D3D11 → D3D12 Bridge

  • Standalone route (standalone\): Run DLSSNR directly on DX11, DX12, or Vulkan apps without OptiScaler or ReShade (OBS Studio, video players, Magpie, capture viewers, older titles), plus ReShade DirectPresent support.
  • D3D11 → D3D12 Direct bridge: Standalone D3D11 apps automatically cross via shared textures/fences into the D3D12 Zero-Heap Direct ELF backend instead of falling back to the Vulkan Bridge.
  • Multi-swapchain arbiter: Handles multi-window apps (e.g., OBS main preview + fullscreen projector) cleanly by tracking the active/largest swapchain without input or hook conflicts.
  • Read-only directory support (C:\Program Files\...): Automatically falls back to %LOCALAPPDATA%\DLSSNR-AMD\... for dlssnr-amd.ini (seeded from the install folder so NR Control menu changes & model switching persist without admin rights), logs, ACO caches, screenshots, and diagnostic reports.

3. Stability, Coexistence & Build Fixes

  • Hardened crash watcher against Windows loader lock, foreign stacks, and LastError clobbering.
  • Improved D3D12 split-direct coexistence, outer SPIR-V hash checks, and PAL variant hashes.
  • Guarded null wilsjo2 guide surfaces in SessionW2Io.
  • Fixed build/CI compatibility (older MinGW headers, read-only PE signing in sign_pe.py, offline Vulkan-Headers reuse).

Notes

  • No NVIDIA model weights are included in this package; install.ps1 extracts them from your own nvngx_dlssnr.dll (310.9.16 or 310.8.0) on first install.
  • When updating an existing Standalone install (such as OBS Studio under Program Files), replace dlssnr-amd.asi, dxgi.dll, and dlssnr-amd\aco\d3d12_outer.nra (or re-run install.ps1).

CJK Language Pack (DLSSNR-AMD-lang-CJK-0.0.9m1.zip)

  • Extract lang/ from DLSSNR-AMD-lang-CJK-0.0.9m1.zip into your game/app's dlssnr-amd\ directory (dlssnr-amd\lang\{zh-CN.txt,ja.txt,ko.txt,fonts\NotoSansCJKsc-Regular.ttf}) if you need Chinese / Japanese / Korean UI and font support. (UI strings are unchanged since v0.0.9m1.)
  • SHA-256 (DLSSNR-AMD-lang-CJK-0.0.9m1.zip): 14736ABE9D8F9E763574603780993AAA285776C285264894D749E781704BD7C4

DLSSNR-AMD v0.0.9m2(Windows 预览版)

基于上游 v0.0.9(9503f6b)的增强分支版本,重点带来 D3D12 零描述符堆(Zero-Heap)直通重构、无需 OptiScaler/ReShade 的 Standalone 独立模式(含 D3D11 → D3D12 直通桥接) 以及 Program Files 只读目录完整兼容。

主要更新

1. D3D12 Zero-Heap 直通重构(更低显存占用与 CPU 开销)

  • 零描述符堆与纯计算状态追踪:将 D3D12 直通外围着色器(RenoDX / 辅助 Pass)迁移至预编译原生 ACO ELF 管线(d3d12_outer.nra),消除 SetDescriptorHeaps 切换对宿主/游戏渲染状态的干扰。
  • 消除空闲中间缓冲与扩展直通支持:省去 RenoDX 空闲中间缓冲区分配,新增 AMD 一级间接 SRD 支持并扩展直通格式覆盖。
  • RX 9070 XT 实测对比原版 v0.0.9(CC_SILVER_AARDWOLD_OPTIMIZED Ultra Performance,驱动 26.11.1,详见 docs/BENCHMARK-D3D12-ZERO-HEAP.md):
    • D3D12 直通专用显存:1080p 降低 34.6 MiB (-6.0%),1440p 降低 60.0–60.6 MiB (-7.0%~-7.2%),4K 降低 130.1–130.5 MiB (-8.7%~-9.1%)(相比 Vulkan 桥接共省 106–308 MiB / 16%~19%)。
    • CPU 录制耗时:相比 v0.0.9 降低 14%–19%(0.24–0.27 ms vs 0.29–0.33 ms),相比 Vulkan 桥接降低 71%–74%。
    • GPU 推理速度:D3D12 直通与 Vulkan 桥接均与原版持平无回退,并修复了原版 Vulkan 桥接退出时的崩溃问题。

2. Standalone 独立模式(DX11 / DX12 / Vulkan)与 D3D11 → D3D12 桥接

  • 新增独立路线(standalone\):不依赖 OptiScaler 或 ReShade,可直接挂载到 DX11 / DX12 / Vulkan 程序(如 OBS Studio、视频播放器、Magpie、采集预览等),同时新增 ReShade DirectPresent 模式。
  • D3D11 → D3D12 直通桥接:DX11 独立程序可通过共享纹理/栅栏直接走 D3D12 Zero-Heap 直通后端,不再强制回退到高显存/高 CPU 开销的 Vulkan 桥接。
  • 多交换链仲裁器:自动适配多窗口程序(例如 OBS 主预览 + 全屏投影),避免多窗口抢占 Hook 或菜单输入错乱。
  • 只读安装目录回退(如 C:\Program Files\):无管理员权限时自动将 dlssnr-amd.ini(首次写入自动从安装目录初始化拷贝)、日志、ACO 缓存、截图和诊断报告重定向至 %LOCALAPPDATA%\DLSSNR-AMD\,确保 NR Control 菜单改参数与切换模型正常生效和持久化。

3. 稳定性与兼容性修复

  • 加固崩溃监控,避免触发 Windows Loader Lock、外部栈异常及污染 LastError。
  • 改进 D3D12 Split/Direct 共存,校验外围 SPIR-V 与 PAL 变体哈希。
  • 修复 SessionW2Io 中空引导表面保护及多项编译/签名脚本问题。

注意事项

  • 安装包不包含 NVIDIA 模型权重,首次运行 install.ps1 会从你本地的 nvngx_dlssnr.dll(310.9.16 或 310.8.0)自动提取。
  • 如之前已手动装过旧版 Standalone(例如 OBS 目录),升级时请同步更新 dlssnr-amd.asi、dxgi.dll 与 dlssnr-amd\aco\d3d12_outer.nra(或重新运行 install.ps1)。

中日韩语言包(DLSSNR-AMD-lang-CJK-0.0.9m1.zip)

  • 如需中文 / 日文 / 韩文界面与字体支持,将 DLSSNR-AMD-lang-CJK-0.0.9m1.zip 内的 lang/ 解压到游戏或程序目录的 dlssnr-amd\ 下(即 dlssnr-amd\lang\{zh-CN.txt,ja.txt,ko.txt,fonts\NotoSansCJKsc-Regular.ttf})。(相对 v0.0.9m1 界面词条无变更。)
  • SHA-256(DLSSNR-AMD-lang-CJK-0.0.9m1.zip):14736ABE9D8F9E763574603780993AAA285776C285264894D749E781704BD7C4