From e1ffe253222b5bd5b9b052acf31f55010a46aa06 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tur=C3=A1nszki=20J=C3=A1nos?=
 <turanszkij@users.noreply.github.com>
Date: Wed, 27 Mar 2024 19:23:41 +0100
Subject: [PATCH] hbao changes

---
 WickedEngine/shaders/hbaoCS.hlsl | 14 ++++++--------
 WickedEngine/wiRenderer.cpp      | 30 ++++++++++++++----------------
 WickedEngine/wiVersion.cpp       |  2 +-
 3 files changed, 21 insertions(+), 25 deletions(-)

diff --git a/WickedEngine/shaders/hbaoCS.hlsl b/WickedEngine/shaders/hbaoCS.hlsl
index 68afbeddd..5c5dff827 100644
--- a/WickedEngine/shaders/hbaoCS.hlsl
+++ b/WickedEngine/shaders/hbaoCS.hlsl
@@ -10,11 +10,11 @@ RWTexture2D<unorm float> output : register(u0);
 // Step size in number of pixels
 #define STEP_SIZE 4
 // Number of shared-memory samples per direction
-#define NUM_STEPS 8
+#define NUM_STEPS 16
 // The last sample has weight = exp(-KERNEL_FALLOFF)
-#define KERNEL_FALLOFF 3.0f
-#define TAN_ANGLE_BIAS 0
-#define DEPTH_FIX 2.0f
+#define KERNEL_FALLOFF 2.0
+#define TAN_ANGLE_BIAS 0.1
+#define DEPTH_FIX 0.4
 
 static const int TILE_BORDER = NUM_STEPS * STEP_SIZE;
 static const int CACHE_SIZE = TILE_BORDER + POSTPROCESS_HBAO_THREADCOUNT + TILE_BORDER;
@@ -47,16 +47,14 @@ void IntegrateDirection(inout float ao, float2 P, float tanT, int threadId, int
 	float tanH = tanT;
 	float sinH = TanToSin(tanH);
 	float sinT = TanToSin(tanT);
-
-	[unroll]
+	
 	for (int sampleId = 0; sampleId < NUM_STEPS; ++sampleId)
 	{
 		float2 S = cache[threadId + sampleId * deltaX + deltaX];
 		float2 V = S - P;
 		float tanS = Tangent(V);
 		float d2 = dot(V, V);
-
-		[flatten]
+		
 		if ((d2 < DEPTH_FIX) && (tanS > tanH))
 		{
 			// Accumulate AO between the horizon and the sample
diff --git a/WickedEngine/wiRenderer.cpp b/WickedEngine/wiRenderer.cpp
index 08c2c8956..f088524aa 100644
--- a/WickedEngine/wiRenderer.cpp
+++ b/WickedEngine/wiRenderer.cpp
@@ -11232,11 +11232,13 @@ void Postprocess_SSAO(
 	{
 		GPUBarrier barriers[] = {
 			GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
+			GPUBarrier::Image(&res.temp, res.temp.desc.layout, ResourceState::UNORDERED_ACCESS),
 		};
 		device->Barrier(barriers, arraysize(barriers), cmd);
 	}
 
 	device->ClearUAV(&output, 0, cmd);
+	device->ClearUAV(&res.temp, 0, cmd);
 	device->Barrier(GPUBarrier::Memory(&output), cmd);
 
 	device->Dispatch(
@@ -11248,8 +11250,8 @@ void Postprocess_SSAO(
 
 	{
 		GPUBarrier barriers[] = {
-			GPUBarrier::Memory(),
 			GPUBarrier::Image(&output, ResourceState::UNORDERED_ACCESS, output.desc.layout),
+			GPUBarrier::Image(&res.temp, ResourceState::UNORDERED_ACCESS, res.temp.desc.layout),
 		};
 		device->Barrier(barriers, arraysize(barriers), cmd);
 	}
@@ -11274,7 +11276,6 @@ void Postprocess_HBAO(
 
 	device->BindComputeShader(&shaders[CSTYPE_POSTPROCESS_HBAO], cmd);
 
-
 	const TextureDesc& desc = output.GetDesc();
 
 	PostProcess postprocess;
@@ -11313,20 +11314,26 @@ void Postprocess_HBAO(
 
 		{
 			GPUBarrier barriers[] = {
+				GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
 				GPUBarrier::Image(&res.temp, res.temp.desc.layout, ResourceState::UNORDERED_ACCESS),
 			};
 			device->Barrier(barriers, arraysize(barriers), cmd);
 		}
 
+		device->ClearUAV(&output, 0, cmd);
+		device->ClearUAV(&res.temp, 0, cmd);
+		device->Barrier(GPUBarrier::Memory(&res.temp), cmd);
+
 		device->Dispatch(
 			(postprocess.resolution.x + POSTPROCESS_HBAO_THREADCOUNT - 1) / POSTPROCESS_HBAO_THREADCOUNT,
 			postprocess.resolution.y,
 			1,
 			cmd
-			);
+		);
 
 		{
 			GPUBarrier barriers[] = {
+				GPUBarrier::Memory(&output),
 				GPUBarrier::Image(&res.temp, ResourceState::UNORDERED_ACCESS, res.temp.desc.layout),
 			};
 			device->Barrier(barriers, arraysize(barriers), cmd);
@@ -11346,26 +11353,15 @@ void Postprocess_HBAO(
 		};
 		device->BindUAVs(uavs, 0, arraysize(uavs), cmd);
 
-		{
-			GPUBarrier barriers[] = {
-				GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
-			};
-			device->Barrier(barriers, arraysize(barriers), cmd);
-		}
-
-		device->ClearUAV(&output, 0, cmd);
-		device->Barrier(GPUBarrier::Memory(&output), cmd);
-
 		device->Dispatch(
 			postprocess.resolution.x,
 			(postprocess.resolution.y + POSTPROCESS_HBAO_THREADCOUNT - 1) / POSTPROCESS_HBAO_THREADCOUNT,
 			1,
 			cmd
-			);
+		);
 
 		{
 			GPUBarrier barriers[] = {
-				GPUBarrier::Memory(),
 				GPUBarrier::Image(&output, ResourceState::UNORDERED_ACCESS, output.desc.layout),
 			};
 			device->Barrier(barriers, arraysize(barriers), cmd);
@@ -11373,13 +11369,15 @@ void Postprocess_HBAO(
 
 	}
 
-	Postprocess_Blur_Bilateral(output, lineardepth, res.temp, output, cmd, 1.2f, -1, -1, true);
+	//Postprocess_Blur_Bilateral(output, lineardepth, res.temp, output, cmd, 1.2f, -1, -1, true);
 
 	wi::profiler::EndRange(prof_range);
 	device->EventEnd(cmd);
 }
 void CreateMSAOResources(MSAOResources& res, XMUINT2 resolution)
 {
+	res.cleared = false;
+
 	TextureDesc saved_desc;
 	saved_desc.format = Format::R32_FLOAT;
 	saved_desc.width = resolution.x;
diff --git a/WickedEngine/wiVersion.cpp b/WickedEngine/wiVersion.cpp
index 16b7a5560..c0eadeb9b 100644
--- a/WickedEngine/wiVersion.cpp
+++ b/WickedEngine/wiVersion.cpp
@@ -9,7 +9,7 @@ namespace wi::version
 	// minor features, major updates, breaking compatibility changes
 	const int minor = 71;
 	// minor bug fixes, alterations, refactors, updates
-	const int revision = 414;
+	const int revision = 415;
 
 	const std::string version_string = std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(revision);