raytraced shadow changed to full resolution

turanszkij · Jan 31, 2024 · 1bec956 · 1bec956
1 parent 2055856
commit 1bec956
Show file tree

Hide file tree

Showing 7 changed files with 30 additions and 46 deletions.
diff --git a/WickedEngine/shaders/rtshadow_denoise_filterCS.hlsl b/WickedEngine/shaders/rtshadow_denoise_filterCS.hlsl
@@ -33,11 +33,11 @@ float FFX_DNSR_Shadows_GetDepthSimilaritySigma()
 
 float FFX_DNSR_Shadows_ReadDepth(uint2 did)
 {
-	return texture_depth[did * 2];
+	return texture_depth[did];
 }
 float16_t3 FFX_DNSR_Shadows_ReadNormals(uint2 did)
 {
-	return normalize((float16_t3)normals[did] * 2 - 1);
+	return normalize((float16_t3)normals[did] - 0.5);
 }
 
 bool FFX_DNSR_Shadows_IsShadowReciever(uint2 did)

diff --git a/WickedEngine/shaders/rtshadow_denoise_temporalCS.hlsl b/WickedEngine/shaders/rtshadow_denoise_temporalCS.hlsl
@@ -68,7 +68,7 @@ inline void ResolverAABB(in uint shadow_index, float sharpness, float exposureSc
 [numthreads(POSTPROCESS_BLOCKSIZE, POSTPROCESS_BLOCKSIZE, 1)]
 void main(uint3 DTid : SV_DispatchThreadID, uint3 GTid : SV_GroupThreadID, uint3 Gid : SV_GroupID, uint groupIndex : SV_GroupIndex)
 {
-	if (texture_depth.Load(uint3(DTid.xy, 1)) == 0)
+	if (texture_depth[DTid.xy] == 0)
 		return;
 
 	// first 4 lights are denoised

diff --git a/WickedEngine/shaders/rtshadow_denoise_tileclassificationCS.hlsl b/WickedEngine/shaders/rtshadow_denoise_tileclassificationCS.hlsl
@@ -46,15 +46,15 @@ float4x4 FFX_DNSR_Shadows_GetReprojectionMatrix()
 
 float FFX_DNSR_Shadows_ReadDepth(uint2 did)
 {
-	return texture_depth[did * 2];
+	return texture_depth[did];
 }
 float FFX_DNSR_Shadows_ReadPreviousDepth(int2 idx)
 {
-	return texture_depth_history[idx * 2];
+	return texture_depth_history[idx];
 }
 float3 FFX_DNSR_Shadows_ReadNormals(uint2 did)
 {
-	return normalize(normals[did] * 2 - 1);
+	return normalize(normals[did] - 0.5);
 }
 uint FFX_DNSR_Shadows_ReadRaytracedShadowMask(uint linear_tile_index)
 {
@@ -70,7 +70,7 @@ float FFX_DNSR_Shadows_ReadHistory(float2 history_uv)
 }
 float2 FFX_DNSR_Shadows_ReadVelocity(uint2 did)
 {
-	return -texture_velocity[did * 2].xy;
+	return -texture_velocity[did].xy;
 }
 
 void FFX_DNSR_Shadows_WriteReprojectionResults(uint2 did, float2 value)

diff --git a/WickedEngine/shaders/screenspaceshadowCS.hlsl b/WickedEngine/shaders/screenspaceshadowCS.hlsl
@@ -36,7 +36,7 @@ void main(uint3 DTid : SV_DispatchThreadID, uint3 Gid : SV_GroupID, uint3 GTid :
 #endif // RTSHADOW
 
 	float3 P = reconstruct_position(uv, depth);
-	float3 N = decode_oct(texture_normal[DTid.xy * 2]);
+	float3 N = decode_oct(texture_normal[DTid.xy]);
 
 	Surface surface;
 	surface.init();
@@ -45,7 +45,7 @@ void main(uint3 DTid : SV_DispatchThreadID, uint3 Gid : SV_GroupID, uint3 GTid :
 
 	const float4 bluenoise = blue_noise(DTid.xy);
 
-	const uint2 tileIndex = uint2(floor(DTid.xy * 2 / TILED_CULLING_BLOCKSIZE));
+	const uint2 tileIndex = uint2(floor(DTid.xy / TILED_CULLING_BLOCKSIZE));
 	const uint flatTileIndex = flatten2D(tileIndex, GetCamera().entity_culling_tilecount.xy) * SHADER_ENTITY_TILE_BUCKET_COUNT;
 
 	uint shadow_mask[4] = {0,0,0,0}; // FXC issue: can't dynamically index into uint4, unless unrolling all loops

diff --git a/WickedEngine/wiRenderer.cpp b/WickedEngine/wiRenderer.cpp
@@ -13281,14 +13281,12 @@ void CreateRTShadowResources(RTShadowResources& res, XMUINT2 resolution)
 	res.frame = 0;
 
 	TextureDesc desc;
-	desc.width = resolution.x / 2;
-	desc.height = resolution.y / 2;
+	desc.width = resolution.x;
+	desc.height = resolution.y;
 	desc.bind_flags = BindFlag::SHADER_RESOURCE | BindFlag::UNORDERED_ACCESS;
 	desc.layout = ResourceState::SHADER_RESOURCE_COMPUTE;
 
 	desc.format = Format::R32G32B32A32_UINT;
-	device->CreateTexture(&desc, nullptr, &res.temp);
-	device->SetName(&res.temp, "rtshadow_temp");
 	device->CreateTexture(&desc, nullptr, &res.temporal[0]);
 	device->SetName(&res.temporal[0], "rtshadow_temporal[0]");
 	device->CreateTexture(&desc, nullptr, &res.temporal[1]);
@@ -13351,6 +13349,7 @@ void Postprocess_RTShadow(
 	{
 		// Maybe we don't need to clear them all, but it's safer this way:
 		GPUBarrier barriers[] = {
+			GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
 			GPUBarrier::Image(&res.temporal[0], res.temporal[0].desc.layout, ResourceState::UNORDERED_ACCESS),
 			GPUBarrier::Image(&res.temporal[1], res.temporal[1].desc.layout, ResourceState::UNORDERED_ACCESS),
 			GPUBarrier::Image(&res.denoised, res.denoised.desc.layout, ResourceState::UNORDERED_ACCESS),
@@ -13373,6 +13372,7 @@ void Postprocess_RTShadow(
 			GPUBarrier::Image(&res.moments[3][1], res.moments[3][1].desc.layout, ResourceState::UNORDERED_ACCESS),
 		};
 		device->Barrier(barriers, arraysize(barriers), cmd);
+		device->ClearUAV(&output, 0, cmd);
 		device->ClearUAV(&res.temporal[0], 0, cmd);
 		device->ClearUAV(&res.temporal[1], 0, cmd);
 		device->ClearUAV(&res.denoised, 0, cmd);
@@ -13400,7 +13400,7 @@ void Postprocess_RTShadow(
 		device->Barrier(barriers, arraysize(barriers), cmd);
 	}
 
-	const TextureDesc& desc = res.temp.GetDesc();
+	const TextureDesc& desc = output.GetDesc();
 
 	BindCommonResources(cmd);
 
@@ -13419,15 +13419,15 @@ void Postprocess_RTShadow(
 	device->PushConstants(&postprocess, sizeof(postprocess), cmd);
 
 	const GPUResource* uavs[] = {
-		&res.temp,
+		&output,
 		&res.normals,
 		&res.tiles
 	};
 	device->BindUAVs(uavs, 0, arraysize(uavs), cmd);
 
 	{
 		GPUBarrier barriers[] = {
-			GPUBarrier::Image(&res.temp, res.temp.desc.layout, ResourceState::UNORDERED_ACCESS),
+			GPUBarrier::Image(&output, output.desc.layout, ResourceState::UNORDERED_ACCESS),
 			GPUBarrier::Image(&res.normals, res.normals.desc.layout, ResourceState::UNORDERED_ACCESS),
 			GPUBarrier::Buffer(&res.tiles, ResourceState::SHADER_RESOURCE_COMPUTE, ResourceState::UNORDERED_ACCESS),
 		};
@@ -13443,7 +13443,7 @@ void Postprocess_RTShadow(
 
 	{
 		GPUBarrier barriers[] = {
-			GPUBarrier::Image(&res.temp, ResourceState::UNORDERED_ACCESS, res.temp.desc.layout),
+			GPUBarrier::Image(&output, ResourceState::UNORDERED_ACCESS, output.desc.layout),
 			GPUBarrier::Image(&res.normals, ResourceState::UNORDERED_ACCESS, res.normals.desc.layout),
 			GPUBarrier::Buffer(&res.tiles, ResourceState::UNORDERED_ACCESS, ResourceState::SHADER_RESOURCE_COMPUTE),
 		};
@@ -13664,7 +13664,7 @@ void Postprocess_RTShadow(
 		device->BindComputeShader(&shaders[CSTYPE_POSTPROCESS_RTSHADOW_DENOISE_TEMPORAL], cmd);
 		device->PushConstants(&postprocess, sizeof(postprocess), cmd);
 
-		device->BindResource(&res.temp, 0, cmd);
+		device->BindResource(&output, 0, cmd);
 		device->BindResource(&res.temporal[temporal_history], 1, cmd);
 		device->BindResource(&res.denoised, 3, cmd);
 
@@ -13697,26 +13697,15 @@ void Postprocess_RTShadow(
 		device->EventEnd(cmd);
 	}
 
-	Postprocess_Upsample_Bilateral(
-		res.temporal[temporal_output],
-		lineardepth,
-		output,
-		cmd
-	);
+	device->Barrier(GPUBarrier::Image(&output, output.desc.layout, ResourceState::COPY_DST), cmd);
+	device->CopyResource(&output, &res.temporal[temporal_output], cmd);
+	device->Barrier(GPUBarrier::Image(&output, ResourceState::COPY_DST, output.desc.layout), cmd);
 
 	wi::profiler::EndRange(prof_range);
 	device->EventEnd(cmd);
 }
 void CreateScreenSpaceShadowResources(ScreenSpaceShadowResources& res, XMUINT2 resolution)
 {
-	TextureDesc desc;
-	desc.width = resolution.x / 2;
-	desc.height = resolution.y / 2;
-	desc.format = Format::R32G32B32A32_UINT;
-	desc.bind_flags = BindFlag::SHADER_RESOURCE | BindFlag::UNORDERED_ACCESS;
-	desc.layout = ResourceState::SHADER_RESOURCE_COMPUTE;
-	device->CreateTexture(&desc, nullptr, &res.lowres);
-	device->SetName(&res.lowres, "screenspaceshadow.lowres");
 }
 void Postprocess_ScreenSpaceShadow(
 	const ScreenSpaceShadowResources& res,
@@ -13731,7 +13720,7 @@ void Postprocess_ScreenSpaceShadow(
 	device->EventBegin("Postprocess_ScreenSpaceShadow", cmd);
 	auto prof_range = wi::profiler::BeginRangeGPU("ScreenSpaceShadow", cmd);
 
-	const TextureDesc& desc = res.lowres.GetDesc();
+	const TextureDesc& desc = output.GetDesc();
 
 	device->BindComputeShader(&shaders[CSTYPE_POSTPROCESS_SCREENSPACESHADOW], cmd);
 
@@ -13745,17 +13734,20 @@ void Postprocess_ScreenSpaceShadow(
 	device->PushConstants(&postprocess, sizeof(postprocess), cmd);
 
 	const GPUResource* uavs[] = {
-		&res.lowres,
+		&output,
 	};
 	device->BindUAVs(uavs, 0, arraysize(uavs), cmd);
 
 	{
 		GPUBarrier barriers[] = {
-			GPUBarrier::Image(&res.lowres, res.lowres.desc.layout, ResourceState::UNORDERED_ACCESS),
+			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(
 		(desc.width + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
 		(desc.height + POSTPROCESS_BLOCKSIZE - 1) / POSTPROCESS_BLOCKSIZE,
@@ -13765,18 +13757,11 @@ void Postprocess_ScreenSpaceShadow(
 
 	{
 		GPUBarrier barriers[] = {
-			GPUBarrier::Image(&res.lowres, ResourceState::UNORDERED_ACCESS, res.lowres.desc.layout),
+			GPUBarrier::Image(&output, ResourceState::UNORDERED_ACCESS, output.desc.layout),
 		};
 		device->Barrier(barriers, arraysize(barriers), cmd);
 	}
 
-	Postprocess_Upsample_Bilateral(
-		res.lowres,
-		lineardepth,
-		output,
-		cmd
-	);
-
 	wi::profiler::EndRange(prof_range);
 	device->EventEnd(cmd);
 }

diff --git a/WickedEngine/wiRenderer.h b/WickedEngine/wiRenderer.h
@@ -595,7 +595,6 @@ namespace wi::renderer
 	);
 	struct RTShadowResources
 	{
-		wi::graphics::Texture temp;
 		wi::graphics::Texture temporal[2];
 		wi::graphics::Texture normals;
 
@@ -617,7 +616,7 @@ namespace wi::renderer
 	);
 	struct ScreenSpaceShadowResources
 	{
-		wi::graphics::Texture lowres;
+		int placeholder = 0;
 	};
 	void CreateScreenSpaceShadowResources(ScreenSpaceShadowResources& res, XMUINT2 resolution);
 	void Postprocess_ScreenSpaceShadow(

diff --git 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 = 365;
+	const int revision = 366;
 
 	const std::string version_string = std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(revision);