mavismmg/MFGAdaUnlock-RenoDx 1.1.5

Release 1.1.5 — DLSS-G Workload Optimization & Latency Fixes

on GitHub

DLSS-G workload optimization

• Added conservative fast paths for saturated confidence, interior-border and candidate-agreement cases.
• Full quality calculations remain active for difficult pixels and boundary regions.
• NVIDIA driver JIT validation reduced the synthetic kernel from 19 to 18 registers per thread.
• Common identity paths can avoid up to 28 PTX operations. The real-world gain depends on the game, resolution and scene; this is not a universal FPS or latency guarantee.

Latency measurement and guard reliability

• Fixed timing validation that could remain permanently in Timing unverified.
• Source timing now validates from Simulation + Present markers, while queue timing is validated independently.
• The UI now reports exact validation reasons, median/p95 queue and pipeline timing, active GPU work, DLSS-G workload, dominant observed stage and monitoring overhead.
• Automatic Latency Guard remains opt-in. It tests a lower addon-forced multiplier first, may then test a small source-rate trim, and rolls back when measured latency does not improve. Monitor Only remains the default.

VRAM and resource diagnostics

• Added read-only process VRAM and DLSS-G resource diagnostics for future optimization work.
• No experimental resource-release behavior is enabled in this release.

Compatibility and validation

• Existing Adaptive Quality behavior and difficult-region quality paths are preserved.
• Clean release build completed successfully.
• All 16 automated tests passed.
• Modified quality PTX was accepted by the NVIDIA driver JIT.

Thanks to everyone who tested the experimental builds and supplied pacing, latency and image-quality feedback.

SHA-256 (renodx-mfgunlock.addon64): 0D04D858A62D3D19E7E3D478C0B8C46FE3AC43EC9FD11E4ABB15617BD291D71A