Improve TAA

AnKi/Renderer/FinalComposite.h

@@ -14,7 +14,7 @@ namespace anki {
 /// @{
 
 inline NumericCVar<F32> g_filmGrainCVar("R", "FilmGrain", 16.0f, 0.0f, 250.0f, "Film grain strength");
-inline NumericCVar<F32> g_sharpnessCVar("R", "Sharpness", (ANKI_PLATFORM_MOBILE) ? 0.0f : 0.8f, 0.0f, 1.0f, "Sharpen the image. It's a factor");
+inline NumericCVar<F32> g_sharpnessCVar("R", "Sharpness", (ANKI_PLATFORM_MOBILE) ? 0.0f : 1.0f, 0.0f, 1.0f, "Sharpen the image. It's a factor");
 
 /// Post-processing stage.
 class FinalComposite : public RendererObject

AnKi/Shaders/TemporalAA.ankiprog

@@ -3,106 +3,58 @@
 // Code licensed under the BSD License.
 // http://www.anki3d.org/LICENSE
 
+#pragma anki 16bit
+
 #pragma anki mutator VARIANCE_CLIPPING 0 1
 #pragma anki mutator YCBCR 0 1
 
 #pragma anki technique vert pixel comp
 
 #include <AnKi/Shaders/QuadVert.hlsl>
-#if ANKI_COMPUTE_SHADER || ANKI_PIXEL_SHADER
-#	include <AnKi/Shaders/Functions.hlsl>
-#	include <AnKi/Shaders/PackFunctions.hlsl>
-
-constexpr F32 kVarianceClippingGamma = 2.7f; // Variance clipping paper proposes 1.0
-constexpr F32 kBlendFactor = 1.0f / 16.0f;
+#include <AnKi/Shaders/Functions.hlsl>
+#include <AnKi/Shaders/PackFunctions.hlsl>
+#include <AnKi/Shaders/TemporalAA.hlsl>
 
 SamplerState g_linearAnyClampSampler : register(s0);
-Texture2D<RVec4> g_inputRt : register(t0);
-Texture2D<RVec4> g_historyRt : register(t1);
+Texture2D<Vec4> g_inputRt : register(t0);
+Texture2D<Vec4> g_historyRt : register(t1);
 Texture2D g_motionVectorsTex : register(t2);
 
-#	if ANKI_COMPUTE_SHADER
-RWTexture2D<RVec4> g_storageTex : register(u0);
-#	endif
+RWTexture2D<Vec4> g_outTex : register(u0);
 
-#	if YCBCR
-#		define sample(s, uv) rgbToYCbCr(s.SampleLevel(g_linearAnyClampSampler, uv, 0.0).rgb)
-#		define sampleOffset(s, uv, x, y) rgbToYCbCr(s.SampleLevel(g_linearAnyClampSampler, uv, 0.0, IVec2(x, y)).rgb)
-#	else
-#		define sample(s, uv) s.SampleLevel(g_linearAnyClampSampler, uv, 0.0).rgb
-#		define sampleOffset(s, uv, x, y) s.SampleLevel(g_linearAnyClampSampler, uv, 0.0, IVec2(x, y)).rgb
-#	endif
-
-#	if ANKI_COMPUTE_SHADER
-[numthreads(8, 8, 1)] void main(UVec3 svDispatchThreadId : SV_DISPATCHTHREADID)
-#	else
-RVec3 main(VertOut input) : SV_Target0
-#	endif
+Vec3 computeTaa(Vec2 coord)
 {
-#	if ANKI_COMPUTE_SHADER
-	Vec2 outStorageTexSize;
-	g_storageTex.GetDimensions(outStorageTexSize.x, outStorageTexSize.y);
-
-	const Vec2 uv = (Vec2(svDispatchThreadId.xy) + 0.5) / outStorageTexSize;
-#	else
-	const Vec2 uv = input.m_uv;
-#	endif
-
-	// Get prev uv coords
-	const Vec2 oldUv = uv + g_motionVectorsTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0).rg;
-
-	// Read textures
-	Vec3 historyCol = sample(g_historyRt, oldUv);
-	const Vec3 crntCol = sample(g_inputRt, uv);
-
-	// Remove ghosting by clamping the history color to neighbour's AABB
-	const Vec3 near0 = sampleOffset(g_inputRt, uv, 1, 0);
-	const Vec3 near1 = sampleOffset(g_inputRt, uv, 0, 1);
-	const Vec3 near2 = sampleOffset(g_inputRt, uv, -1, 0);
-	const Vec3 near3 = sampleOffset(g_inputRt, uv, 0, -1);
+	Vec2 viewportSize;
+	g_historyRt.GetDimensions(viewportSize.x, viewportSize.y);
 
-#	if VARIANCE_CLIPPING
-	const Vec3 m1 = crntCol + near0 + near1 + near2 + near3;
-	const Vec3 m2 = crntCol * crntCol + near0 * near0 + near1 * near1 + near2 * near2 + near3 * near3;
+	const Vec2 uv = (Vec2(coord) + 0.5f) / viewportSize;
+	const Vec2 historyUv = uv + g_motionVectorsTex.SampleLevel(g_linearAnyClampSampler, uv, 0.0).xy;
 
-	const Vec3 mu = m1 / 5.0;
-	const Vec3 sigma = sqrt(m2 / 5.0 - mu * mu);
+	const Vec3 history = g_historyRt.SampleLevel(g_linearAnyClampSampler, historyUv, 0.0);
 
-	const Vec3 boxMin = mu - kVarianceClippingGamma * sigma;
-	const Vec3 boxMax = mu + kVarianceClippingGamma * sigma;
-#	else
-	const Vec3 boxMin = min(crntCol, min(near0, min(near1, min(near2, near3))));
-	const Vec3 boxMax = max(crntCol, max(near0, max(near1, max(near2, near3))));
-#	endif
+	const Vec3 outColor = computeTemporalAA<F32>(g_inputRt, g_linearAnyClampSampler, history, coord);
 
-	historyCol = clamp(historyCol, boxMin, boxMax);
+	return outColor;
+}
 
-	// Remove jitter (T. Lottes)
-#	if YCBCR
-	const F32 lum0 = crntCol.r;
-	const F32 lum1 = historyCol.r;
-	const F32 maxLum = boxMax.r;
-#	else
-	const F32 lum0 = computeLuminance(reinhardTonemap(crntCol));
-	const F32 lum1 = computeLuminance(reinhardTonemap(historyCol));
-	const F32 maxLum = 1.0;
-#	endif
+#if ANKI_COMPUTE_SHADER
+[numthreads(64, 1, 1)] void main(COMPUTE_ARGS)
+{
+	UVec2 viewportSize;
+	g_historyRt.GetDimensions(viewportSize.x, viewportSize.y);
 
-	F32 diff = abs(lum0 - lum1) / max(lum0, max(lum1, maxLum + kEpsilonF32));
-	diff = 1.0 - diff;
-	diff = diff * diff;
-	const F32 feedback = lerp(0.0, kBlendFactor, diff);
+	const UVec2 coord = getOptimalDispatchThreadId8x8Amd(svGroupIndex, svGroupId.xy);
 
-	// Write result
-	Vec3 outColor = lerp(historyCol, crntCol, feedback);
-#	if YCBCR
-	outColor = yCbCrToRgb(outColor);
-#	endif
+	if(any(coord >= viewportSize))
+	{
+		return;
+	}
 
-#	if ANKI_COMPUTE_SHADER
-	g_storageTex[svDispatchThreadId.xy] = RVec4(outColor, 0.0);
-#	else
-	return outColor;
-#	endif
+	TEX(g_outTex, coord) = Vec4(computeTaa(coord), 0.0);
+}
+#elif ANKI_PIXEL_SHADER
+Vec3 main(VertOut input) : SV_Target0
+{
+	return computeTaa(floor(input.m_svPosition));
 }
-#endif // ANKI_COMPUTE_SHADER || ANKI_PIXEL_SHADER
+#endif