- Use visibleMeshletIds to remap meshlet instance indices to support …

…scenes with more meshlet instances

- Added option to disable materials when loading a scene (use when a scene has huge materials that don't fit in VRAM)
- Fixed potential issue with VisbufferResolve.frag reading random, invalid meshlets

data/shaders/debug/DebugRect.vert.glsl

@@ -8,7 +8,7 @@
 
 layout(location = 0) out vec4 v_color;
 
-FVOG_DECLARE_ARGUMENTS(DebugAabbArguments)
+FVOG_DECLARE_ARGUMENTS(DebugRectArguments)
 {
   FVOG_UINT32 debugRectBufferIndex;
 };

data/shaders/shadows/ShadowMain.vert.glsl

@@ -10,7 +10,8 @@ layout(location = 2) out vec3 i_objectSpacePos;
 
 void main()
 {
-  const uint meshletInstanceId = (uint(gl_VertexIndex) >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  const uint visibleMeshletId = (uint(gl_VertexIndex) >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  const uint meshletInstanceId = d_visibleMeshlets.indices[visibleMeshletId];
   const uint primitiveId = uint(gl_VertexIndex) & MESHLET_PRIMITIVE_MASK;
   const MeshletInstance meshletInstance = d_meshletInstances[meshletInstanceId];
   const Meshlet meshlet = d_meshlets[meshletInstance.meshletId];

data/shaders/shadows/vsm/VsmCommon.h.glsl

@@ -29,6 +29,7 @@ FVOG_DECLARE_ARGUMENTS(VsmPushConstants)
   FVOG_INT32 currentPass;
 
   // VsmShadow + ShadowMain
+  FVOG_UINT32 visibleMeshletsIndex;
   FVOG_UINT32 meshletInstancesIndex;
   FVOG_UINT32 meshletDataIndex;
   FVOG_UINT32 meshletPrimitivesIndex;

data/shaders/visbuffer/CullMeshlets.comp.glsl

@@ -58,15 +58,6 @@ void DebugDrawMeshletAabb(in uint meshletInstanceId)
 FVOG_DECLARE_SAMPLERS;
 FVOG_DECLARE_SAMPLED_IMAGES(texture2D);
 
-//layout(std430, binding = 9) restrict buffer MeshletVisibilityBuffer
-FVOG_DECLARE_STORAGE_BUFFERS(restrict MeshletVisibilityBuffer)
-{
-  uint indices[];
-} visibleMeshletsBuffers[];
-
-#define d_visibleMeshlets visibleMeshletsBuffers[visibleMeshletsIndex]
-
-//layout(std430, binding = 10) restrict buffer CullTrianglesDispatchParams
 FVOG_DECLARE_STORAGE_BUFFERS(restrict CullTrianglesDispatchParams)
 {
   uint groupCountX;

data/shaders/visbuffer/CullTriangles.comp.glsl

@@ -9,15 +9,6 @@
 #include "../debug/DebugCommon.h.glsl"
 #include "../shadows/vsm/VsmCommon.h.glsl"
 
-//layout(std430, binding = 9) restrict buffer MeshletVisbilityBuffer
-FVOG_DECLARE_STORAGE_BUFFERS(restrict readonly MeshletVisbilityBuffer)
-{
-  uint indices[];
-} visibleMeshletsBuffers[];
-
-#define d_visibleMeshlets visibleMeshletsBuffers[visibleMeshletsIndex]
-
-//layout (std430, binding = 7) writeonly buffer MeshletPackedBuffer
 FVOG_DECLARE_STORAGE_BUFFERS(writeonly MeshletPackedBuffer)
 {
   uint data[];
@@ -240,8 +231,11 @@ void main()
   if (primitivePassed)
   {
     const uint indexOffset = sh_baseIndex + activePrimitiveId * 3;
-    d_indexBuffer.data[indexOffset + 0] = (meshletInstanceId << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 0) & MESHLET_PRIMITIVE_MASK);
-    d_indexBuffer.data[indexOffset + 1] = (meshletInstanceId << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 1) & MESHLET_PRIMITIVE_MASK);
-    d_indexBuffer.data[indexOffset + 2] = (meshletInstanceId << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 2) & MESHLET_PRIMITIVE_MASK);
+    if (indexOffset + 2 < d_indexBuffer.data.length())
+    {
+      d_indexBuffer.data[indexOffset + 0] = (gl_WorkGroupID.x << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 0) & MESHLET_PRIMITIVE_MASK);
+      d_indexBuffer.data[indexOffset + 1] = (gl_WorkGroupID.x << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 1) & MESHLET_PRIMITIVE_MASK);
+      d_indexBuffer.data[indexOffset + 2] = (gl_WorkGroupID.x << MESHLET_PRIMITIVE_BITS) | ((primitiveId + 2) & MESHLET_PRIMITIVE_MASK);
+    }
   }
 }

data/shaders/visbuffer/Visbuffer.frag.glsl

@@ -5,7 +5,7 @@
 #include "../Math.h.glsl"
 #include "../Hash.h.glsl"
 
-layout (location = 0) in flat uint i_meshletInstanceId;
+layout (location = 0) in flat uint i_visibleMeshletId;
 layout (location = 1) in flat uint i_primitiveId;
 layout (location = 2) in vec2 i_uv;
 layout (location = 3) in vec3 i_objectSpacePos;
@@ -67,5 +67,5 @@ void main()
     }
   }
 
-  o_pixel = (i_meshletInstanceId << MESHLET_PRIMITIVE_BITS) | (i_primitiveId & MESHLET_PRIMITIVE_MASK);
+  o_pixel = (i_visibleMeshletId << MESHLET_PRIMITIVE_BITS) | (i_primitiveId & MESHLET_PRIMITIVE_MASK);
 }

data/shaders/visbuffer/Visbuffer.vert.glsl

@@ -2,15 +2,16 @@
 #extension GL_GOOGLE_include_directive : enable
 #include "VisbufferCommon.h.glsl"
 
-layout (location = 0) out flat uint o_meshletInstanceId;
+layout (location = 0) out flat uint o_visibleMeshletId;
 layout (location = 1) out flat uint o_primitiveId;
 layout (location = 2) out vec2 o_uv;
 layout (location = 3) out vec3 o_objectSpacePos;
 layout (location = 4) out flat uint o_materialId;
 
 void main()
 {
-  const uint meshletInstanceId = (uint(gl_VertexIndex) >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  const uint visibleMeshletId = (uint(gl_VertexIndex) >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  const uint meshletInstanceId = d_visibleMeshlets.indices[visibleMeshletId];
   const uint primitiveId = uint(gl_VertexIndex) & MESHLET_PRIMITIVE_MASK;
   const MeshletInstance meshletInstance = d_meshletInstances[meshletInstanceId];
   const Meshlet meshlet = d_meshlets[meshletInstance.meshletId];
@@ -26,7 +27,7 @@ void main()
   const mat4 transform = d_transforms[instanceId].modelCurrent;
   const vec2 uv = PackedToVec2(vertex.uv);
 
-  o_meshletInstanceId = meshletInstanceId;
+  o_visibleMeshletId = visibleMeshletId;
   o_primitiveId = primitiveId / 3;
   o_uv = uv;
   o_objectSpacePos = position;

data/shaders/visbuffer/VisbufferCommon.h.glsl

@@ -109,7 +109,7 @@ FVOG_DECLARE_ARGUMENTS(VisbufferPushConstants)
   FVOG_UINT32 hzbSamplerIndex;
   FVOG_UINT32 cullTrianglesDispatchIndex;
 
-  // CullMeshlets.comp and CullTriangles.comp
+  // CullMeshlets.comp, CullTriangles.comp, all vertex shaders
   FVOG_UINT32 visibleMeshletsIndex;
 
   // CullTriangles.comp
@@ -206,4 +206,11 @@ FVOG_DECLARE_STORAGE_BUFFERS(restrict readonly ViewBuffer)
 
 #define d_currentView ViewBuffers[viewIndex].currentView
 
+FVOG_DECLARE_STORAGE_BUFFERS(restrict MeshletVisbilityBuffer)
+{
+  uint indices[];
+} visibleMeshletsBuffers[];
+
+#define d_visibleMeshlets visibleMeshletsBuffers[visibleMeshletsIndex]
+
 #endif // VISBUFFER_COMMON_H

data/shaders/visbuffer/VisbufferResolve.frag.glsl

@@ -9,7 +9,6 @@
 layout(early_fragment_tests) in;
 
 layout(location = 0) in vec2 i_uv;
-layout(location = 1) in flat uint i_materialId;
 
 layout(location = 0) out vec4 o_albedo;
 layout(location = 1) out vec3 o_metallicRoughnessAo;
@@ -243,7 +242,12 @@ void main()
 {
   const ivec2 position = ivec2(gl_FragCoord.xy);
   const uint payload = texelFetch(Fvog_utexture2D(visbufferIndex), position, 0).x;
-  const uint meshletInstanceId = (payload >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  if (payload == ~0u)
+  {
+    discard;
+  }
+  const uint visibleMeshletId = (payload >> MESHLET_PRIMITIVE_BITS) & MESHLET_ID_MASK;
+  const uint meshletInstanceId = d_visibleMeshlets.indices[visibleMeshletId];
   const uint primitiveId = payload & MESHLET_PRIMITIVE_MASK;
   const MeshletInstance meshletInstance = d_meshletInstances[meshletInstanceId];
   const Meshlet meshlet = d_meshlets[meshletInstance.meshletId];

data/shaders/visbuffer/VisbufferResolve.vert.glsl

@@ -3,13 +3,10 @@
 #include "VisbufferCommon.h.glsl"
 
 layout (location = 0) out vec2 v_uv;
-layout (location = 1) out flat uint v_materialId;
 
 void main()
 {
-  v_materialId = gl_BaseInstance;
-  const highp float materialId = uintBitsToFloat(0x3f7fffff - (uint(gl_BaseInstance) & MESHLET_MATERIAL_ID_MASK));
   vec2 pos = vec2(gl_VertexIndex == 0, gl_VertexIndex == 2);
   v_uv = pos.xy * 2.0;
-  gl_Position = vec4(pos * 4.0 - 1.0, materialId, 1.0);
+  gl_Position = vec4(pos * 4.0 - 1.0, 0.0, 1.0);
 }

src/FrogRenderer2.cpp

@@ -214,6 +214,7 @@ FrogRenderer2::FrogRenderer2(const Application::CreateInfo& createInfo)
   //Utility::LoadModelFromFileMeshlet(*device_, scene, "H:/Repositories/glTF-Sample-Models/downloaded schtuff/sponza_compressed_tu.glb", glm::scale(glm::vec3{1}));
   //Utility::LoadModelFromFileMeshlet(*device_, scene, "H:/Repositories/glTF-Sample-Models/downloaded schtuff/subdiv_deccer_cubes.glb", glm::scale(glm::vec3{1}));
   //Utility::LoadModelFromFileMeshlet(*device_, scene, "H:/Repositories/glTF-Sample-Models/downloaded schtuff/SM_Deccer_Cubes_Textured.glb", glm::scale(glm::vec3{1}));
+  //Utility::LoadModelFromFileMeshlet(*device_, scene, "H:/Repositories/glTF-Sample-Models/downloaded schtuff/small_city.glb", glm::scale(glm::vec3{1}));
 
   meshletIndirectCommand = Fvog::TypedBuffer<Fvog::DrawIndexedIndirectCommand>(*device_, {}, "Meshlet Indirect Command");
   cullTrianglesDispatchParams = Fvog::TypedBuffer<Fvog::DispatchIndirectCommand>(*device_, {}, "Cull Triangles Dispatch Params");
@@ -406,14 +407,18 @@ void FrogRenderer2::OnUpdate([[maybe_unused]] double dt)
     }
   }
 
+  // TODO: This is an expensive call! Seek to minimize the parts of the scene that need to be re-traversed to those that actually changed.
   sceneFlattened = scene.Flatten();
   shadingUniforms.numberOfLights = (uint32_t)sceneFlattened.lights.size();
 
   // A few of these buffers are really slow (2-3ms) to create and destroy every frame (large ones hit vkAllocateMemory), so
   // this scheme is still not ideal as e.g. adding geometry every frame will cause reallocs.
   // The current scheme works fine when the scene is mostly static.
   const auto numMeshlets = static_cast<uint32_t>(sceneFlattened.meshletInstances.size());
-  const auto maxIndices = numMeshlets * Utility::maxMeshletPrimitives * 3;
+
+  // Soft cap of 1 billion indices should prevent oversubscribing memory (on my system) when loading huge scenes.
+  // This limit should be OK as it only limits post-culling geometry.
+  const auto maxIndices  = glm::min(1'000'000'000u, numMeshlets * Utility::maxMeshletPrimitives * 3);
   if (!instancedMeshletBuffer || instancedMeshletBuffer->Size() < maxIndices)
   {
     instancedMeshletBuffer = Fvog::TypedBuffer<uint32_t>(*device_, {.count = maxIndices}, "Instanced Meshlets");
@@ -429,9 +434,14 @@ void FrogRenderer2::OnUpdate([[maybe_unused]] double dt)
     transformBuffer = Fvog::NDeviceBuffer<Utility::ObjectUniforms>(*device_, (uint32_t)sceneFlattened.transforms.size(), "Transforms");
   }
 
-  if (!visibleMeshletIds || visibleMeshletIds->Size() < numMeshlets)
+  if (!persistentVisibleMeshletIds || persistentVisibleMeshletIds->Size() < numMeshlets)
+  {
+    persistentVisibleMeshletIds = Fvog::TypedBuffer<uint32_t>(*device_, {.count = numMeshlets}, "Persistent Visible Meshlet IDs");
+  }
+
+  if (!transientVisibleMeshletIds || transientVisibleMeshletIds->Size() < numMeshlets)
   {
-    visibleMeshletIds = Fvog::TypedBuffer<uint32_t>(*device_, {.count = numMeshlets}, "Visible Meshlet IDs");
+    transientVisibleMeshletIds = Fvog::TypedBuffer<uint32_t>(*device_, {.count = numMeshlets}, "Transient Visible Meshlet IDs");
   }
 
   if (!meshletInstancesBuffer || meshletInstancesBuffer->Size() < numMeshlets)
@@ -447,7 +457,7 @@ void FrogRenderer2::OnUpdate([[maybe_unused]] double dt)
   }
 }
 
-void FrogRenderer2::CullMeshletsForView(VkCommandBuffer commandBuffer, const ViewParams& view, std::string_view name)
+void FrogRenderer2::CullMeshletsForView(VkCommandBuffer commandBuffer, const ViewParams& view, Fvog::Buffer& visibleMeshletIds, std::string_view name)
 {
   ZoneScoped;
   TracyVkZoneTransient(tracyVkContext_, tracyProfileVar, commandBuffer, name.data(), true);
@@ -493,7 +503,7 @@ void FrogRenderer2::CullMeshletsForView(VkCommandBuffer commandBuffer, const Vie
     .hzbIndex = frame.hzb->ImageView().GetSampledResourceHandle().index,
     .hzbSamplerIndex = hzbSampler.GetResourceHandle().index,
     .cullTrianglesDispatchIndex = cullTrianglesDispatchParams->GetResourceHandle().index,
-    .visibleMeshletsIndex = visibleMeshletIds->GetResourceHandle().index,
+    .visibleMeshletsIndex = visibleMeshletIds.GetResourceHandle().index,
     .debugAabbBufferIndex = debugGpuAabbsBuffer->GetResourceHandle().index,
     .debugRectBufferIndex = debugGpuRectsBuffer->GetResourceHandle().index,
   };
@@ -546,10 +556,12 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
     auto gpuMaterials = std::vector<Utility::GpuMaterial>(scene.materials.size());
     std::ranges::transform(scene.materials, gpuMaterials.begin(), [](const auto& mat) { return mat.gpuMaterial; });
 
+    // TODO: These buffers should only be updated when changed (these uploads can be quite costly in larger scenes).
     materialStorageBuffer->UpdateData(commandBuffer, gpuMaterials);
     lightBuffer->UpdateData(commandBuffer, sceneFlattened.lights);
     transformBuffer->UpdateData(commandBuffer, sceneFlattened.transforms);
     meshletInstancesBuffer->UpdateData(commandBuffer, sceneFlattened.meshletInstances);
+
     // VSM lod bias corresponds to upscaling lod bias, otherwise shadows become blocky as the upscaling ratio increases.
     auto actualVsmUniforms = vsmUniforms;
     actualVsmUniforms.lodBias += fsr2LodBias;
@@ -674,7 +686,7 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
   if (executeMeshletGeneration)
   {
     TIME_SCOPE_GPU(StatGroup::eMainGpu, eCullMeshletsMain, commandBuffer);
-    CullMeshletsForView(commandBuffer, mainView, "Cull Meshlets Main");
+    CullMeshletsForView(commandBuffer, mainView, persistentVisibleMeshletIds.value(), "Cull Meshlets Main");
   }
 
   ctx.ImageBarrierDiscard(*frame.visbuffer, VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL);
@@ -697,19 +709,20 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
     .depthAttachment = visbufferDepthAttachment,
   });
   {
-    // TIME_SCOPE_GPU(StatGroup::eMainGpu, eRenderVisbufferMain);
+    TIME_SCOPE_GPU(StatGroup::eMainGpu, eRenderVisbufferMain, commandBuffer);
     ctx.BindGraphicsPipeline(visbufferPipeline);
     auto visbufferArguments = VisbufferPushConstants{
-      .globalUniformsIndex = globalUniformsBuffer.GetDeviceBuffer().GetResourceHandle().index,
-      .meshletInstancesIndex = meshletInstancesBuffer->GetDeviceBuffer().GetResourceHandle().index,
-      .meshletDataIndex = meshletBuffer->GetResourceHandle().index,
+      .globalUniformsIndex    = globalUniformsBuffer.GetDeviceBuffer().GetResourceHandle().index,
+      .meshletInstancesIndex  = meshletInstancesBuffer->GetDeviceBuffer().GetResourceHandle().index,
+      .meshletDataIndex       = meshletBuffer->GetResourceHandle().index,
       .meshletPrimitivesIndex = primitiveBuffer->GetResourceHandle().index,
-      .meshletVerticesIndex = vertexBuffer->GetResourceHandle().index,
-      .meshletIndicesIndex = indexBuffer->GetResourceHandle().index,
-      .transformsIndex = transformBuffer->GetDeviceBuffer().GetResourceHandle().index,
-      .indirectDrawIndex = meshletIndirectCommand->GetResourceHandle().index,
-      .materialsIndex = materialStorageBuffer->GetDeviceBuffer().GetResourceHandle().index,
-      .viewIndex = viewBuffer->GetResourceHandle().index,
+      .meshletVerticesIndex   = vertexBuffer->GetResourceHandle().index,
+      .meshletIndicesIndex    = indexBuffer->GetResourceHandle().index,
+      .transformsIndex        = transformBuffer->GetDeviceBuffer().GetResourceHandle().index,
+      .indirectDrawIndex      = meshletIndirectCommand->GetResourceHandle().index,
+      .materialsIndex         = materialStorageBuffer->GetDeviceBuffer().GetResourceHandle().index,
+      .viewIndex              = viewBuffer->GetResourceHandle().index,
+      .visibleMeshletsIndex   = persistentVisibleMeshletIds->GetResourceHandle().index,
     };
     ctx.SetPushConstants(visbufferArguments);
     ctx.BindIndexBuffer(*instancedMeshletBuffer, 0, VK_INDEX_TYPE_UINT32);
@@ -766,7 +779,7 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
       };
       Math::MakeFrustumPlanes(sunCurrentClipmapView.viewProj, sunCurrentClipmapView.frustumPlanes);
 
-      CullMeshletsForView(commandBuffer, sunCurrentClipmapView, "Cull Sun VSM Meshlets, View " + std::to_string(i));
+      CullMeshletsForView(commandBuffer, sunCurrentClipmapView, transientVisibleMeshletIds.value(), "Cull Sun VSM Meshlets, View " + std::to_string(i));
 
       const auto vsmExtent = Fvog::Extent2D{Techniques::VirtualShadowMaps::maxExtent, Techniques::VirtualShadowMaps::maxExtent};
       ctx.Barrier();
@@ -802,6 +815,7 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
       pushConstants.materialSamplerIndex = materialSampler.GetResourceHandle().index;
       pushConstants.clipmapLod = vsmSun.GetClipmapTableIndices()[i];
       pushConstants.clipmapUniformsBufferIndex = vsmSun.clipmapUniformsBuffer_.GetResourceHandle().index;
+      pushConstants.visibleMeshletsIndex       = transientVisibleMeshletIds->GetResourceHandle().index,
 
       ctx.BindIndexBuffer(*instancedMeshletBuffer, 0, VK_INDEX_TYPE_UINT32);
 
@@ -921,9 +935,10 @@ void FrogRenderer2::OnRender([[maybe_unused]] double dt, VkCommandBuffer command
       .meshletIndicesIndex = indexBuffer->GetResourceHandle().index,
       .transformsIndex = transformBuffer->GetDeviceBuffer().GetResourceHandle().index,
       .materialsIndex = materialStorageBuffer->GetDeviceBuffer().GetResourceHandle().index,
+      .visibleMeshletsIndex = persistentVisibleMeshletIds->GetResourceHandle().index,
       .materialSamplerIndex = materialSampler.GetResourceHandle().index,
 
-      .visbufferIndex = frame.visbuffer->ImageView().GetSampledResourceHandle().index,
+      .visbufferIndex       = frame.visbuffer->ImageView().GetSampledResourceHandle().index,
     };
 
     ctx.SetPushConstants(pushConstants);

src/FrogRenderer2.h

@@ -164,7 +164,7 @@ class FrogRenderer2 final : public Application
   void GuiDrawSceneGraph(VkCommandBuffer commandBuffer);
   void GuiDrawSceneGraphHelper(Utility::Node* node);
 
-  void CullMeshletsForView(VkCommandBuffer commandBuffer, const ViewParams& view, std::string_view name = "Cull Meshlet Pass");
+  void CullMeshletsForView(VkCommandBuffer commandBuffer, const ViewParams& view, Fvog::Buffer& visibleMeshletIds, std::string_view name = "Cull Meshlet Pass");
   void MakeStaticSceneBuffers(VkCommandBuffer commandBuffer);
 
   enum class GlobalFlags : uint32_t
@@ -388,7 +388,14 @@ class FrogRenderer2 final : public Application
   std::optional<Fvog::TypedBuffer<Fvog::DrawIndexedIndirectCommand>> meshletIndirectCommand;
   std::optional<Fvog::TypedBuffer<uint32_t>> instancedMeshletBuffer;
   std::optional<Fvog::TypedBuffer<Fvog::DispatchIndirectCommand>> cullTrianglesDispatchParams;
-  std::optional<Fvog::TypedBuffer<uint32_t>> visibleMeshletIds;
+
+  // These buffers serve two purposes:
+  // First, they store the IDs of meshlet instances that passed meshlet culling.
+  // Second, they allow us to remap the range 0-2^24 to meshlet instances that may
+  // have an index outside that range. That can allow us to render scenes with more than 2^24
+  // meshlet instances correctly, as long as 2^24 meshlet instances or fewer are visible.
+  std::optional<Fvog::TypedBuffer<uint32_t>> persistentVisibleMeshletIds; // For when the data needs to be retrieved later (i.e. it is stored in the visbuffer)
+  std::optional<Fvog::TypedBuffer<uint32_t>> transientVisibleMeshletIds;  // For shadows or forward passes
 
   Fvog::ComputePipeline cullMeshletsPipeline;
   Fvog::ComputePipeline cullTrianglesPipeline;