voxel improvements

Author: Cheaterdev

sources/RenderSystem/Effects/VoxelGI/VoxelGIGraph.cpp

@@ -504,52 +504,40 @@ VoxelGI::VoxelGI(Scene::ptr& scene) :scene(scene), VariableContext(L"VoxelGI")
 		compute.set_signature(Layouts::DefaultLayout);
 		context.graph->set_slot(SlotID::VoxelInfo, compute);
 
-		compute.set_pipeline<PSOS::VoxelZero>();
+		const uint mip_levels = tex_lighting.tex_result->get_desc().as_texture().MipLevels;
+
 		{
-			PROFILE_GPU(L"ZERO");
-			for (uint i = 1; i < tex_lighting.tex_result->get_desc().as_texture().MipLevels; i++)
-			{
+			// Only the tail mips without a tile list (packed) still need a
+			// clear — they are never downsampled and the cone trace samples
+			// them at high LODs. Tiny textures, cost is negligible.
+			PROFILE_GPU(L"clear tail mips");
+			for (uint i = 1; i < mip_levels; i++)
 				if (i >= gpu_tiles_buffer.size() || !gpu_tiles_buffer[i])
-				{
 					list.clear_uav(tex_lighting.tex_result->texture_3d().mips[i].rwTexture3D,
 						vec4(0, 0, 0, 0));
-					continue;
-				}
-
-				{
-					Slots::VoxelZero utils;
-					utils.GetTarget() = tex_lighting.tex_result->texture_3d().mips[i].rwTexture3D;
-					auto& params = utils.GetParams();
-					params.GetTiles() = gpu_tiles_buffer[i]->buffer;
-					params.GetVoxels_per_tile().xyz =
-						tex_lighting.tex_result->resource->get_tiled_manager().get_tile_shape();
-					compute.set(utils);
-				}
-
-				compute.exec_indirect(gpu_tiles_buffer[i]->dispatch_buffer, 1);
-			}
 		}
 
 		{
+			// One dispatch per mip over that mip's OWN tile list: every texel
+			// of every loaded tile is written, and unmapped source tiles read
+			// as 0 — so the old per-tile VoxelZero pre-pass is unnecessary.
+			// (The old 3-mips-per-dispatch batches covered mips 2,3,5,6 only
+			// under loaded finer tiles, which is what the zero pass patched.)
 			PROFILE_GPU(L"EXEC");
-			unsigned int mip_count = 1;
-			while (mip_count < tex_lighting.tex_result->get_desc().as_texture().MipLevels)
+			compute.set_pipeline<PSOS::VoxelDownsample>(PSOS::VoxelDownsample::Count(1));
+
+			for (uint mip = 1; mip < mip_levels; mip++)
 			{
-				if (!gpu_tiles_buffer[mip_count]) break;
-				unsigned int current_mips = std::min(
-					3u, tex_lighting.tex_result->get_desc().as_texture().MipLevels - mip_count);
-				compute.set_pipeline<PSOS::VoxelDownsample>(
-					PSOS::VoxelDownsample::Count(current_mips));
+				if (mip >= gpu_tiles_buffer.size() || !gpu_tiles_buffer[mip]) break;
 
 				{
 					Slots::VoxelMipMap mipmapping;
 					mipmapping.GetSrcMip() =
-						tex_lighting.tex_result->texture_3d().mips[mip_count - 1].texture3D;
-					for (unsigned int i = 0; i < current_mips; i++)
-						mipmapping.GetOutMips()[i] =
-							tex_lighting.tex_result->texture_3d().mips[mip_count + i].rwTexture3D;
+						tex_lighting.tex_result->texture_3d().mips[mip - 1].texture3D;
+					mipmapping.GetOutMips()[0] =
+						tex_lighting.tex_result->texture_3d().mips[mip].rwTexture3D;
 					auto& params = mipmapping.GetParams();
-					params.GetTiles() = gpu_tiles_buffer[mip_count]->buffer;
+					params.GetTiles() = gpu_tiles_buffer[mip]->buffer;
 					params.GetVoxels_per_tile().xyz =
 						tex_lighting.tex_result->resource->get_tiled_manager().get_tile_shape();
 					compute.set(mipmapping);
@@ -560,9 +548,8 @@ VoxelGI::VoxelGI(Scene::ptr& scene) :scene(scene), VariableContext(L"VoxelGI")
 					L"mip_8",  L"mip_9",  L"mip_10", L"mip_11",
 					L"mip_12", L"mip_13", L"mip_14", L"mip_15",
 				};
-				PROFILE_GPU(mip_names[mip_count < 16 ? mip_count : 0]);
-				compute.exec_indirect(gpu_tiles_buffer[mip_count]->dispatch_buffer, 1);
-				mip_count += current_mips;
+				PROFILE_GPU(mip_names[mip < 16 ? mip : 0]);
+				compute.exec_indirect(gpu_tiles_buffer[mip]->dispatch_buffer, 1);
 			}
 		}
 	};

sources/RenderSystem/TiledTextures/DynamicTileGenerator.cpp

@@ -43,7 +43,7 @@ void TileDynamicGenerator::set_scene_volume(vec3 min, vec3 max)
 	this->min = min;
 	this->max = max;
 
-	one_tile_size = (max - min) / tile_count;
+	one_tile_size = (max - min) / vec3(tile_count);
 }
 
 void TileDynamicGenerator::begin(vec3 min, vec3 max)
@@ -63,6 +63,12 @@ void TileDynamicGenerator::add(vec3 min, vec3 max)
 	ivec3 from = (min - this->min) / one_tile_size;
 	ivec3 to = (max - this->min + one_tile_size ) / one_tile_size;
 
+	// Clamp to the tile grid: an AABB at/outside the scene volume produces
+	// negative or past-end indices, and all_tiles[pos] would index the grid
+	// out of bounds. Fully-outside objects end up with from >= to (no tiles).
+	from = ivec3::max(from, ivec3(0, 0, 0));
+	to   = ivec3::min(to, ivec3(tile_count));
+
 	for (int x = from.x; x < to.x; x++)
 		for (int y = from.y; y < to.y; y++)
 			for (int z = from.z; z < to.z; z++)

sources/RenderSystem/TiledTextures/DynamicTileGenerator.ixx

@@ -8,7 +8,11 @@ export {
 		uint3 tile_count;
 		vec3 min, max;
 
-		uint3 one_tile_size;
+		// World-space tile extent. MUST stay float: scene_size/tile_count is
+		// fractional (e.g. 200/16 = 12.5) and integer truncation makes the
+		// tile ranges in add() drift with distance from the scene min —
+		// dynamic objects far from the origin mark the wrong tiles.
+		vec3 one_tile_size;
 		grid<uint3, uint64> all_tiles;
 		std::list<uint3> old_tiles;
 

workdir/shaders/voxel_mipmap.hlsl

@@ -12,17 +12,21 @@ static const Texture3D<float4> SrcMip = GetVoxelMipMap().GetSrcMip();
 static const VoxelInfo voxel_info = GetVoxelInfo();
 
 
-void calc(inout float4 color)
+// For the first stage the /8 average is done by the trilinear fetch itself;
+// the groupshared stages below still sum 8 values and need the divide.
+void calc_avg(inout float4 color)
 {
-//	color.rgb/= color.w ;//+ 1;
-	//color.w = saturate(color.w/4);
-
-    color /= 8;
 	color.w=saturate(color.w*1.3f);
 	//todo: investigate negative values
 	color = max(0, color);
 }
 
+void calc(inout float4 color)
+{
+    color /= 8;
+	calc_avg(color);
+}
+
 void add_color(inout float4 result, float4 c)
 {
 	result += float4(c.rgb,c.w);
@@ -43,27 +47,22 @@ void CS(
  uint local_index = groupThreadID.x+groupThreadID.y*4+groupThreadID.z*16;
  uint3 index = GetVoxelMipMap().GetParams().get_voxel_pos(dispatchID);
 
+	float3 dims;
+	SrcMip.GetDimensions(dims.x, dims.y, dims.z);
 
+	// One trilinear fetch at the 2x2x2 block corner IS the 8-texel average —
+	// the filtering hardware does the 8 loads + divide. Unmapped source tiles
+	// read as 0, so loaded output tiles over unloaded input regions come out
+	// zero without a separate clear pass.
+	float4 c = SrcMip.SampleLevel(linearClampSampler, (2 * index + 1) / dims, 0);
 
-//	float4 color = SrcMip.SampleLevel(linearSampler,(2*index+1)/dims,0);
-	float4 c = 0;
-
-	int4 tc = int4(2 * index , 0);
-
-	add_color(c, SrcMip.Load(tc, int3(0, 0, 0)));
-	add_color(c, SrcMip.Load(tc, int3(0, 0, 1)));
-	add_color(c, SrcMip.Load(tc, int3(0, 1, 0)));
-	add_color(c, SrcMip.Load(tc, int3(0, 1, 1)));
-
-	add_color(c, SrcMip.Load(tc, int3(1, 0, 0)));
-	add_color(c, SrcMip.Load(tc, int3(1, 0, 1)));
-	add_color(c, SrcMip.Load(tc, int3(1, 1, 0)));
-	add_color(c, SrcMip.Load(tc, int3(1, 1, 1)));
+	calc_avg(c);
 
-	calc(c);
-		
 	OutMip1[index] = c;
+
+	#if COUNT >= 2
 	data[local_index] = c;
+	#endif
 
 	#if COUNT >=2
 	GroupMemoryBarrierWithGroupSync();