RedDukeDev/dlss5-nr-amd-custom 1.4.0

1.4.0 - New default mode for NR, merge upstream for optiscaler and d4r

on GitHub

Neural rendering

• New default The old "wait for network" mode has been rewritten, and it's now the default. It should handle it in the same way many other mods do, and doesn't have any ghosting or trailing effect, but the performance impact is huge, specially on RDNA3 and RDNA2.
• New option: Asynchronous Mode (off by default). Restores the previous behaviour: the network runs beside the game and its result is carried along the motion vectors, so the game keeps its frame rate. Almost no performance loss, but can show some artifacts, more or less visible depending on the gpu and game used.
• Removed: "Wait for the network" and "Apply on the same frame" are gone from the menu and from OptiScaler.ini. You can delete WaitForNetwork and WaitInsideFrame from an existing ini; they are ignored.
• Fixed a freeze with NVIDIA frame generation. The network now runs on a dedicated runtime thread that queues its work only once the capture is done. Before, the game could deadlock after minutes of play.
• Freeze diagnostics. If no frame is presented for six seconds, the runtime writes the fence values and every thread's stack to the log, once, with the prefix STALL. Turn on LogToFile=true to have them in OptiScaler.log.
• API change: the runtime exports dlss5nr_process_queue, and wait_for_network and wait_inside are gone from dlss5nr_settings. The runtime and the plugin must come from the same release.

OptiScaler fork

• The DLSS render preset menu is shown for DLSS on a GPU that is not NVIDIA's.
• The D3D12 device is hooked, and Detours transactions are started, one thread at a time.
• The game's command lists stay on the hooks, and ZLUDA has a single stream policy.
• ReShade can hook the factory before OptiScaler.
• Upstream OptiScaler master is merged in (about 590 commits).

Frame generation

• Under Wine, the library now receives the D3D12 device's LUID.

Dependencies and packaging

• Marged upstream changes to ZLUDA, it now moves to ROCm LLVM 10.1.
• The optiscaler and dxvk folders moved under third_party/, and the README paths are updated.
• Merged last changes from d4r:
  • Preset K reads the game's inputs in place by default (D4R_K_DIRECT_INPUTS=0 restores the copies).
  • Preset K's reconstruction filter is entirely in f32, which fixes the black speckles in Ghost of Tsushima.
  • RR F gains d3dhip and vkhip stages.
  • K's output stage is shorter, about 7% and 8% faster at 1440p and 4K.
  • M and L's Swin layers are faster.

Full Changelog: 1.3.2...1.4.0