Echo-Storm/ls-addon-manager v0.9.33

Addon Manager for Lossless Scaling 0.9.33: the upscalers on a 3440x1440 screen, second try

on GitHub

The upscalers on a 3440x1440 screen, second try (issue #13). 0.9.31 and 0.9.32 were still re-reading NIS's settings many times a second there: Lossless Scaling makes a new constants buffer for its extra 48-byte pass at every frame, and that pushed the real pass's reader out. Reproduced on the test host (236 refusals and 30 re-reads before, 2 and 2 after); not yet seen in a game. The source also has a prototype Video Super Resolution addon (not in this package: it needs your own NVIDIA Video Effects SDK).

• The upscalers on a 3440x1440 screen, second try (issue #13, still broken on 0.9.32). The reporter's new log shows the refused 48-byte NIS-looking pass many times a second and the real pass read again about every second. Lossless Scaling makes a new constants buffer for that pass at every frame, and the readers of NIS's viewports (one for each pass, in 0.9.31) were keyed by the buffer's address: every frame made a new one, and after eight the real pass's was pushed out and had to be read again (the upscaler then falls back to NIS until it is). They are keyed by the buffer's size now, and a reader that is known good is never pushed out for one that was refused. Test host: nisdecoy=2 gives the other pass a new buffer at every frame, scenario scaler_crop_3440_decoy2; before the fix that run logged 236 refusals and read the real pass 30 times, after it 2 and 2 (one for each device).
• A prototype addon: Video Super Resolution (VSRUPSC, issue #8). NVIDIA's RTX Video Super Resolution, from the person's own NVIDIA Video Effects SDK (x64, with the VideoSuperRes feature; nothing of NVIDIA's is shipped or fetched), in place of Lossless Scaling's NIS pass: the frame is read through the pass's own input view into an 8-bit texture (an HDR frame as its SDR view), the SDK's Direct3D 11 transfers take it to CUDA, VSR runs, and the result is placed in the pass's output (viewport and borders as the other upscalers keep them). Measured on an RTX 4070 Ti SUPER: Low 2.3 ms at 2560x1440 to 3840x2160 and 0.65 ms of hand-over; 82 to 93 % of the game's detail against 65 to 78 % for bicubic, about a quarter more frame-to-frame shimmer; the WoW interface text and icons come out crisp (docs/vsr.md). It runs synchronously on the render thread for now. Work in progress: off by default, built only with the SDK present (external/vfx, -DVFX_DIR), left out of release packages (-IncludeWip), tools\deploy.ps1 -What vsr. Test host scenarios vsr and vsr_4_3 (need the SDK; VFX_DIR). Tools: nr_vfxprobe (time, detail, shimmer, hand-over on frames and clips).
• Video Super Resolution: the motion gate. The owner's first live test: VSR looked about like FSR or XeSS, and on what moved it blurred and shifted colour; at rest it was sharp. NIS now runs as usual and VSR's picture is blended over it per pixel only where the frame did not change from the pass before (a 3x3 neighbourhood), so the interface and a still camera get VSR and the world in motion keeps NIS's picture. Settings "Only where the picture is still" (on) and "Motion sensitivity". The test host now calls post-dispatch callbacks as the manager does; scenarios vsr_move (NIS kept on what moves) and vsr_nogate.
• The detection of the NIS pass (FindNisPass, NIS's viewports) is its own file, nis_pass.cpp, so an addon without the Direct3D 12 engine can use it. The DLSS, FSR and XeSS upscalers are unchanged (their scenarios pass).

Zip LSAddonManager-0.9.33-x64.zip, SHA-256 8c6a9fa7a95ee227cd54aa72109b30af5187fa727a23b6f53469317fb88e3f44.