Bit of shader refactoring

res/shaders/BasePass.slang

@@ -15,7 +15,7 @@ layout(set=5)
 ParameterBlock<LightCullingData> pLightCullingData;
 
 [shader("pixel")]
-float4 fragmentMain(in FragmentParameter params : PARAMETER) : SV_Target
+float4 fragmentMain(in FragmentParameter params) : SV_Target
 {
     MaterialParameter materialParams = params.getMaterialParameter();
     LightingParameter lightingParams = params.getLightingParameter();

res/shaders/LegacyBasePass.slang

@@ -3,21 +3,12 @@ import VertexData;
 import MaterialParameter;
 import Scene;
 
-struct VertexShaderOutput
-{
-	FragmentParameter parameter : PARAMETER;
-	float4 clipPos: SV_Position;
-}
-
 [shader("vertex")]
-VertexShaderOutput vertexMain(	
+FragmentParameter vertexMain(	
 	uint vertexId: SV_VertexID,
 	uint instanceId: SV_InstanceID,
 ){
 	InstanceData inst = pScene.instances[instanceId];
 	VertexAttributes attr = pVertexData.getAttributes(vertexId, inst.transformMatrix);
-	VertexShaderOutput output;
-	output.parameter = attr.getParameter(inst.transformMatrix);
-	output.clipPos = output.parameter.position_CS;
-	return output;
+	return attr.getParameter(inst.transformMatrix);
 }

res/shaders/MeshletBasePass.slang

@@ -76,8 +76,7 @@ struct PrimitiveAttributes
 
 struct MeshOutput
 {
-	FragmentParameter parameter : PARAMETER;
-	float4 position_CS : SV_Position;
+	FragmentParameter parameter;
 };
 
 [numthreads(MESH_GROUP_SIZE, 1, 1)]
@@ -94,10 +93,9 @@ void meshMain(
 	MeshData md = pScene.meshData[meshPayload.instanceId[groupID]];
 	MeshletDescription m = pScene.meshletInfos[meshPayload.meshletId[groupID]];
 	const uint vertexLoops = (MAX_VERTICES + MESH_GROUP_SIZE - 1) / MESH_GROUP_SIZE;
-	for(uint loop = 0; loop < vertexLoops; ++loop) 
+	for(uint i = threadID; i < MAX_VERTICES; i += MESH_GROUP_SIZE)
 	{
-		uint v = threadID + loop * MESH_GROUP_SIZE;
-		v = min(v, m.vertexCount - 1);
+		uint v = min(i, m.vertexCount - 1);
 		InterlockedMax(gs_numVertices, v + 1);
 		{
 			int vertexIndex = pScene.vertexIndices[m.vertexOffset + v];
@@ -106,10 +104,9 @@ void meshMain(
 	}
 	GroupMemoryBarrierWithGroupSync();
 	const uint primitiveLoops = (MAX_PRIMITIVES + MESH_GROUP_SIZE - 1) / MESH_GROUP_SIZE;
-	for(uint loop = 0; loop < primitiveLoops; ++loop)
+	for(uint i = threadID; i < MAX_PRIMITIVES; i += MESH_GROUP_SIZE)
 	{
-		uint p = threadID + loop * MESH_GROUP_SIZE;
-		p = min(p, m.primitiveCount - 1);
+		uint p = min(i, m.primitiveCount - 1);
 		InterlockedMax(gs_numPrimitives, p + 1);
 		{
 			uint32_t local_idx0 = pScene.primitiveIndices[m.primitiveOffset + (p * 3) + 0];
@@ -128,14 +125,12 @@ void meshMain(
 	v = min(v, m.vertexCount - 1);
 	FragmentParameter parameter = gs_vertices[v].getParameter(inst.transformMatrix);
 	vertices[v].parameter = parameter;
-	vertices[v].position_CS = parameter.position_CS;
 	if(vertexLoops >= 1)
 	{
 		uint v = threadID + MESH_GROUP_SIZE;
 		v = min(v, m.vertexCount - 1);
 		FragmentParameter parameter = gs_vertices[v].getParameter(inst.transformMatrix);
 		vertices[v].parameter = parameter;
-		vertices[v].position_CS = parameter.position_CS;
 	}
 
 	uint p = threadID;

res/shaders/lib/MaterialParameter.slang

@@ -2,10 +2,10 @@ import Common;
 
 struct MaterialParameter
 {
-	float3 position_TS : POSITION0;
-	float3 position_WS : POSITION1;
-	float2 texCoords : TEXCOORDS0;
-	float3 vertexColor : COLOR0;
+	float3 position_TS;
+	float3 position_WS;
+	float2 texCoords;
+	float3 vertexColor;
 };
 
 // data used by light environment

res/shaders/lib/Scene.slang

@@ -11,9 +11,11 @@ struct MeshletDescription
 
 struct MeshData
 {
-	uint numMeshlets;
-	uint meshletOffset;
-	uint indicesOffset;
+    uint numMeshlets;
+    uint meshletOffset;
+    uint firstIndex;
+    uint numIndices;
+    uint indicesOffset;
 };
 
 static const uint MAX_VERTICES = 64;

src/Engine/Graphics/Vulkan/Command.cpp

@@ -468,10 +468,7 @@ CommandPool::CommandPool(PGraphics graphics, PQueue queue)
 
 CommandPool::~CommandPool()
 {
-    for (auto& command : allocatedBuffers)
-    {
-        command->waitForCommand();
-    }
+    vkDeviceWaitIdle(graphics->getDevice());
     allocatedRenderCommands.clear();
     allocatedComputeCommands.clear();
     allocatedBuffers.clear();