FSR2 temporal upscaling fixes: unjittered reprojection, sharpen Y-flip, MSAA guard, descriptor double-buffering

- Motion vectors: single unjittered reprojection matrix (80 bytes) instead of
  two jittered matrices (160 bytes), eliminating numerical instability from
  jitter amplification through large world coordinates
- Sharpen pass: fix Y-flip for correct UV sampling, double-buffer descriptor
  sets to avoid race with in-flight command buffers
- MSAA: auto-disable when FSR2 enabled, grey out AA setting in UI
- Accumulation: variance-based neighborhood clamping in YCoCg space,
  correct history layout transitions
- Frame index: wrap at 256 for stable Halton sequence

assets/shaders/fsr2_accumulate.comp.glsl

@@ -2,25 +2,19 @@
 
 layout(local_size_x = 8, local_size_y = 8) in;
 
-// Inputs (internal resolution)
 layout(set = 0, binding = 0) uniform sampler2D sceneColor;
 layout(set = 0, binding = 1) uniform sampler2D depthBuffer;
 layout(set = 0, binding = 2) uniform sampler2D motionVectors;
-
-// History (display resolution)
 layout(set = 0, binding = 3) uniform sampler2D historyInput;
-
-// Output (display resolution)
 layout(set = 0, binding = 4, rgba16f) uniform writeonly image2D historyOutput;
 
 layout(push_constant) uniform PushConstants {
     vec4 internalSize;   // xy = internal resolution, zw = 1/internal
     vec4 displaySize;    // xy = display resolution, zw = 1/display
-    vec4 jitterOffset;   // xy = current jitter (pixel-space), zw = unused
+    vec4 jitterOffset;   // xy = current jitter (NDC-space), zw = unused
     vec4 params;         // x = resetHistory (1=reset), y = sharpness, zw = unused
 } pc;
 
-// RGB <-> YCoCg for neighborhood clamping
 vec3 rgbToYCoCg(vec3 rgb) {
     float y  = 0.25 * rgb.r + 0.5 * rgb.g + 0.25 * rgb.b;
     float co = 0.5  * rgb.r                - 0.5  * rgb.b;
@@ -40,76 +34,52 @@ void main() {
     ivec2 outSize = ivec2(pc.displaySize.xy);
     if (outPixel.x >= outSize.x || outPixel.y >= outSize.y) return;
 
-    // Output UV in display space
     vec2 outUV = (vec2(outPixel) + 0.5) * pc.displaySize.zw;
+    vec3 currentColor = texture(sceneColor, outUV).rgb;
 
-    // Map display pixel to internal resolution UV (accounting for jitter)
-    vec2 internalUV = outUV;
-
-    // Sample current frame color at internal resolution
-    vec3 currentColor = texture(sceneColor, internalUV).rgb;
-
-    // Sample motion vector at internal resolution
-    vec2 inUV = outUV;  // Approximate — display maps to internal via scale
-    vec2 motion = texture(motionVectors, inUV).rg;
-
-    // Reproject: where was this pixel in the previous frame's history?
-    vec2 historyUV = outUV - motion;
-
-    // History reset: on teleport / camera cut, just use current frame
     if (pc.params.x > 0.5) {
         imageStore(historyOutput, outPixel, vec4(currentColor, 1.0));
         return;
     }
 
-    // Sample reprojected history
+    vec2 motion = texture(motionVectors, outUV).rg;
+    vec2 historyUV = outUV + motion;
+
+    float historyValid = (historyUV.x >= 0.0 && historyUV.x <= 1.0 &&
+                          historyUV.y >= 0.0 && historyUV.y <= 1.0) ? 1.0 : 0.0;
+
     vec3 historyColor = texture(historyInput, historyUV).rgb;
 
-    // Neighborhood clamping in YCoCg space to prevent ghosting
-    // Sample 3x3 neighborhood from current frame
+    // Neighborhood clamping in YCoCg space
     vec2 texelSize = pc.internalSize.zw;
-    vec3 samples[9];
-    int idx = 0;
-    for (int dy = -1; dy <= 1; dy++) {
-        for (int dx = -1; dx <= 1; dx++) {
-            samples[idx] = rgbToYCoCg(texture(sceneColor, internalUV + vec2(dx, dy) * texelSize).rgb);
-            idx++;
-        }
-    }
+    vec3 s0 = rgbToYCoCg(currentColor);
+    vec3 s1 = rgbToYCoCg(texture(sceneColor, outUV + vec2(-texelSize.x, 0.0)).rgb);
+    vec3 s2 = rgbToYCoCg(texture(sceneColor, outUV + vec2( texelSize.x, 0.0)).rgb);
+    vec3 s3 = rgbToYCoCg(texture(sceneColor, outUV + vec2(0.0, -texelSize.y)).rgb);
+    vec3 s4 = rgbToYCoCg(texture(sceneColor, outUV + vec2(0.0,  texelSize.y)).rgb);
+    vec3 s5 = rgbToYCoCg(texture(sceneColor, outUV + vec2(-texelSize.x, -texelSize.y)).rgb);
+    vec3 s6 = rgbToYCoCg(texture(sceneColor, outUV + vec2( texelSize.x, -texelSize.y)).rgb);
+    vec3 s7 = rgbToYCoCg(texture(sceneColor, outUV + vec2(-texelSize.x,  texelSize.y)).rgb);
+    vec3 s8 = rgbToYCoCg(texture(sceneColor, outUV + vec2( texelSize.x,  texelSize.y)).rgb);
 
-    // Compute AABB in YCoCg
-    vec3 boxMin = samples[0];
-    vec3 boxMax = samples[0];
-    for (int i = 1; i < 9; i++) {
-        boxMin = min(boxMin, samples[i]);
-        boxMax = max(boxMax, samples[i]);
-    }
+    vec3 m1 = s0 + s1 + s2 + s3 + s4 + s5 + s6 + s7 + s8;
+    vec3 m2 = s0*s0 + s1*s1 + s2*s2 + s3*s3 + s4*s4 + s5*s5 + s6*s6 + s7*s7 + s8*s8;
+    vec3 mean = m1 / 9.0;
+    vec3 variance = max(m2 / 9.0 - mean * mean, vec3(0.0));
+    vec3 stddev = sqrt(variance);
 
-    // Slightly expand the box to reduce flickering on edges
-    vec3 boxCenter = (boxMin + boxMax) * 0.5;
-    vec3 boxExtent = (boxMax - boxMin) * 0.5;
-    boxMin = boxCenter - boxExtent * 1.25;
-    boxMax = boxCenter + boxExtent * 1.25;
+    float gamma = 1.5;
+    vec3 boxMin = mean - gamma * stddev;
+    vec3 boxMax = mean + gamma * stddev;
 
-    // Clamp history to the neighborhood AABB
     vec3 historyYCoCg = rgbToYCoCg(historyColor);
     vec3 clampedHistory = clamp(historyYCoCg, boxMin, boxMax);
     historyColor = yCoCgToRgb(clampedHistory);
 
-    // Check if history UV is valid (within [0,1])
-    float historyValid = (historyUV.x >= 0.0 && historyUV.x <= 1.0 &&
-                          historyUV.y >= 0.0 && historyUV.y <= 1.0) ? 1.0 : 0.0;
-
-    // Blend factor: use more current frame for disoccluded regions
-    // Luminance difference between clamped history and original → confidence
     float clampDist = length(historyYCoCg - clampedHistory);
-    float blendFactor = mix(0.05, 0.3, clamp(clampDist * 4.0, 0.0, 1.0));
-
-    // If history is off-screen, use current frame entirely
+    float blendFactor = mix(0.05, 0.30, clamp(clampDist * 2.0, 0.0, 1.0));
     blendFactor = mix(blendFactor, 1.0, 1.0 - historyValid);
 
-    // Final blend
     vec3 result = mix(historyColor, currentColor, blendFactor);
-
     imageStore(historyOutput, outPixel, vec4(result, 1.0));
 }

assets/shaders/fsr2_motion.comp.glsl

@@ -6,10 +6,8 @@ layout(set = 0, binding = 0) uniform sampler2D depthBuffer;
 layout(set = 0, binding = 1, rg16f) uniform writeonly image2D motionVectors;
 
 layout(push_constant) uniform PushConstants {
-    mat4 invViewProj;       // Inverse of current jittered VP
-    mat4 prevViewProj;      // Previous frame unjittered VP
+    mat4 reprojMatrix;      // prevUnjitteredVP * inverse(currentUnjitteredVP)
     vec4 resolution;        // xy = internal size, zw = 1/internal size
-    vec4 jitterOffset;      // xy = current jitter (NDC), zw = previous jitter
 } pc;
 
 void main() {
@@ -20,25 +18,18 @@ void main() {
     // Sample depth (Vulkan: 0 = near, 1 = far)
     float depth = texelFetch(depthBuffer, pixelCoord, 0).r;
 
-    // Pixel center in NDC [-1, 1]
+    // Pixel center in UV [0,1] and NDC [-1,1]
     vec2 uv = (vec2(pixelCoord) + 0.5) * pc.resolution.zw;
     vec2 ndc = uv * 2.0 - 1.0;
 
-    // Reconstruct world position from depth
+    // Clip-to-clip reprojection: current unjittered clip → previous unjittered clip
     vec4 clipPos = vec4(ndc, depth, 1.0);
-    vec4 worldPos = pc.invViewProj * clipPos;
-    worldPos /= worldPos.w;
-
-    // Project into previous frame's clip space (unjittered)
-    vec4 prevClip = pc.prevViewProj * worldPos;
+    vec4 prevClip = pc.reprojMatrix * clipPos;
     vec2 prevNdc = prevClip.xy / prevClip.w;
     vec2 prevUV = prevNdc * 0.5 + 0.5;
 
-    // Remove jitter from current UV to get unjittered position
-    vec2 unjitteredUV = uv - pc.jitterOffset.xy * 0.5;
-
-    // Motion = previous position - current unjittered position (in UV space)
-    vec2 motion = prevUV - unjitteredUV;
+    // Motion = previous position - current position (both unjittered, in UV space)
+    vec2 motion = prevUV - uv;
 
     imageStore(motionVectors, pixelCoord, vec4(motion, 0.0, 0.0));
 }

assets/shaders/fsr2_sharpen.frag.glsl

@@ -10,16 +10,20 @@ layout(push_constant) uniform PushConstants {
 } pc;
 
 void main() {
+    // Undo the vertex shader Y flip (postprocess.vert flips for Vulkan overlay,
+    // but we need standard UV coords for texture sampling)
+    vec2 tc = vec2(TexCoord.x, 1.0 - TexCoord.y);
+
     vec2 texelSize = pc.params.xy;
     float sharpness = pc.params.z;
 
     // RCAS: Robust Contrast-Adaptive Sharpening
     // 5-tap cross pattern
-    vec3 center = texture(inputImage, TexCoord).rgb;
-    vec3 north  = texture(inputImage, TexCoord + vec2(0.0, -texelSize.y)).rgb;
-    vec3 south  = texture(inputImage, TexCoord + vec2(0.0,  texelSize.y)).rgb;
-    vec3 west   = texture(inputImage, TexCoord + vec2(-texelSize.x, 0.0)).rgb;
-    vec3 east   = texture(inputImage, TexCoord + vec2( texelSize.x, 0.0)).rgb;
+    vec3 center = texture(inputImage, tc).rgb;
+    vec3 north  = texture(inputImage, tc + vec2(0.0, -texelSize.y)).rgb;
+    vec3 south  = texture(inputImage, tc + vec2(0.0,  texelSize.y)).rgb;
+    vec3 west   = texture(inputImage, tc + vec2(-texelSize.x, 0.0)).rgb;
+    vec3 east   = texture(inputImage, tc + vec2( texelSize.x, 0.0)).rgb;
 
     // Compute local contrast (min/max of neighborhood)
     vec3 minRGB = min(center, min(min(north, south), min(west, east)));