InstanceBrick--nightcap/third_party/fidelityfx

FidelityFX Super Resolution 3.1.4

This directory contains the MIT-licensed upscaler, Optical Flow component and Vulkan backend from AMD FidelityFX SDK 1.1.4, commit c6efa6bf7f2027b3ec94f28578bb5965eabb9e55.

The standalone CMake target nightcap_fsr3 builds on Linux ARM64. It compiles AMD's GLSL compute passes with glslangValidator and derives descriptor names, types and bindings from SPIR-V. spirv-val validates generated code when available. Shader binaries remain build artifacts.

Supported variants use FP32 or opt-in FP16, render-resolution motion vectors, normal or inverted depth, LDR or HDR color, optional sharpening and the SDK's Lanczos choices. Jittered and unjittered motion-vector variants are generated. Camera jitter and matching scene depth and motion vectors are required by the caller. Separate upscaler contexts retain each eye's temporal history. OpenXR retains presentation ownership.

When engine motion vectors are unavailable, the wrapper can run AMD Optical Flow independently for each eye. Its signed 8 × 8 block vectors are reconstructed as a render-resolution RG16F motion field. The conversion removes camera jitter and supplies a GPU scene-cut reactive mask. This image-motion fallback adds compute work and has less information about occlusions and object surfaces than native engine vectors.

The AMD license is retained. Linux integration includes these changes:

• Opaque context storage accommodates Linux's 32-bit wchar_t binding tables.
• Checked string helpers provide the bounds checks expected by the host code.
• Vulkan memory selection accepts unified-memory device-local heaps, while retaining a preference for invisible VRAM on discrete GPUs.
• The promoted Vulkan 1.1 memory-requirements function is used when its KHR alias is absent.
• The GLSL luma-history storage declaration matches the runtime's RGBA16F image.
• Physical-device queries use a private host Vulkan loader. Production code keeps that loader local so it does not collide with Android Vulkan symbols.

compile_fsr3_shaders.py, compile_opticalflow_shaders.py, fsr3_vulkan_support.cpp and this CMake target provide Linux build integration. OpenXR owns presentation; this target does not replace its swapchain.

The gaze transport crops a fixed-size focus region with a guard band, runs the genuine temporal upscaler on that region, and blends it with a cheap spatial reconstruction of the periphery. Native motion is converted to pixel units with the changing crop origin included. The compositor maps global pixel coordinates to the crop directly, retaining fractional output placement when the input crop moves. Large gaze jumps reset the region's temporal history.

FP16 requires shaderFloat16 and shaderInt16 to have been enabled on the logical Vulkan device. The generated shaders do not require 16-bit storage-buffer or push-constant features. The SDK also requires storage-image extended formats and storage-image writes without a declared format.

GPU checks

Configure this directory with FSR3_BUILD_VULKAN_SMOKE=ON, FSR3_BUILD_TRANSPORT_SMOKE=ON and FSR3_BUILD_GAZE_SMOKE=ON to build the native Vulkan checks:

• fsr3_vulkan_smoke dispatches the genuine upscaler with independent stereo histories and checks finite FP16 output.
• fsr3_vulkan_smoke opticalflow verifies known image translation. A texture moving right by two input pixels produces backward vectors of (-2, 0).
• fsr3_transport_smoke checks dark, mid and bright encoded colors through native RGBA/BGRA UNORM → sRGB sample → linear FSR → sRGB blit → raw UNORM copy. It checks the opaque alpha used by the VR eye images.
• fsr3_transport_smoke 1440 2160 measures GPU timestamps around one eye's transport and upscaling. Add opticalflow to include motion estimation. These synthetic timings describe the upscaler's cost rather than application FPS.
• fsr3_gaze_transport_smoke checks color transport, native motion conversion, changing crop origins and history resets for both eyes and channel orders. Its optional resolution arguments measure the gaze-region GPU cost.

Run these checks with the Khronos validation layer and synchronization validation enabled when changing resources, shaders or barriers.