wgpu-ffx

AMD FidelityFX FSR 3 Upscaler for wgpu.

A Rust implementation of AMD's FidelityFX Super Resolution 3 (FSR3) upscaler for the wgpu graphics library. Renders at lower resolution and reconstructs high-quality output at display resolution using temporal accumulation. Can also be used at a 1:1 ratio for temporal anti-aliasing (TAA) without upscaling.

Status

Early / experimental. Known limitations:

• GenerateReactive pass is unimplemented
• Lanczos-LUT, half-precision, and wave64 shader permutations are not yet selected based on device capabilities (the reference/full-precision variants are always chosen)

Requirements

• wgpu 29

Usage

use wgpu_ffx::{FsrContext, FsrContextInfo, FsrContextFlags, FsrDispatchInfo};

// Create the context once; this compiles all compute pipelines for the given
// feature flags but allocates no GPU textures.
let ctx = FsrContext::new(FsrContextInfo {
    device: device.clone(),
    flags: FsrContextFlags::HIGH_DYNAMIC_RANGE
         | FsrContextFlags::DEPTH_INVERTED
         | FsrContextFlags::DEPTH_INFINITE,
});

// Create a view for a specific maximum-resolution pair; this allocates all
// internal GPU textures and temporal-accumulation state. Multiple views may
// share a single context.
let mut view = ctx.create_view(&queue, [1920, 1080], [3840, 2160]);

// Per frame: record FSR passes into an encoder the caller submits.
let mut encoder = device.create_command_encoder(&Default::default());
ctx.dispatch(&mut view, &mut encoder, &FsrDispatchInfo { /* ... */ })?;
queue.submit([encoder.finish()]);

FsrView::resize reallocates internal textures for a new maximum-resolution pair without rebuilding the parent context's pipelines.

Sharpening (RCAS)

Set FsrDispatchInfo::enable_sharpening to run an additional Robust Contrast Adaptive Sharpening pass after accumulation. The sharpness field accepts values in [0.0, 1.0], where 0.0 is no additional sharpness and 1.0 is maximum.

Texture format requirements

Field	Type	Format	Size	Required usage flags
color	Texture	Rgba16Float	render_size	TEXTURE_BINDING
depth	Texture	Depth32Float	render_size	TEXTURE_BINDING
motion_vectors	Texture	Rg16Float	render_size (or upscale_size with DISPLAY_RESOLUTION_MOTION_VECTORS)	TEXTURE_BINDING
output	Texture	Rgba16Float	upscale_size	STORAGE_BINDING + TEXTURE_BINDING
dilated_depth	Texture	R32Float	render_size	STORAGE_BINDING + TEXTURE_BINDING
dilated_motion_vectors	Texture	Rg16Float	render_size	STORAGE_BINDING + TEXTURE_BINDING
reconstructed_previous_depth	Buffer	—	render_width × render_height × 4 bytes	STORAGE + COPY_DST
exposure	Texture (optional)	R32Float	1×1	TEXTURE_BINDING
reactive_mask	Texture (optional)	R8Unorm	render_size	TEXTURE_BINDING
transparency_and_composition	Texture (optional)	R8Unorm	render_size	TEXTURE_BINDING

Context flags

FsrContextFlags configure the upscaler behavior:

Flag	Meaning
HIGH_DYNAMIC_RANGE	Input color data uses a high-dynamic range
DISPLAY_RESOLUTION_MOTION_VECTORS	Motion vectors are at display resolution instead of render resolution
MOTION_VECTORS_JITTER_CANCELLATION	Motion vectors already have the jitter pattern applied; FSR will cancel it
DEPTH_INVERTED	Depth buffer is inverted (1 = near, 0 = far)
DEPTH_INFINITE	Depth buffer uses an infinite far plane
AUTO_EXPOSURE	FSR computes exposure automatically; do not provide an exposure texture
DYNAMIC_RESOLUTION	Application uses dynamic resolution scaling

Jitter

Use get_jitter_phase_count and get_jitter_offset to compute per-frame sub-pixel jitter offsets. Apply the jitter to your projection matrix and pass the same offsets to FsrDispatchInfo::jitter_offset.

use wgpu_ffx::{get_jitter_phase_count, get_jitter_offset};

let phase_count = get_jitter_phase_count(render_width, display_width);
let [jx, jy] = get_jitter_offset(frame_index, phase_count);

Building shaders

Requires glslc from the VulkanSDK.

cargo xtask vendor           # fetch upstream FidelityFX shader sources
cargo xtask compile-shaders  # compile to SPIR-V (requires glslc)

Minimum Supported Rust Version (MSRV)

The MSRV is 1.92. MSRV bumps are considered breaking changes and will be accompanied by a minor version bump.

License

MIT