101 lines
3.4 KiB
HLSL
101 lines
3.4 KiB
HLSL
// Copyright Epic Games, Inc. All Rights Reserved.
|
|
|
|
// Scene writer - used to inject final instance data into the GPU Scene. Dispatched in the scene update timeline.
|
|
|
|
#include "/Engine/Private/Common.ush"
|
|
#include "/Engine/Private/ComputeShaderUtils.ush"
|
|
#include "/Engine/Private/GPUScene/GPUSceneWriter.ush"
|
|
#include "/Engine/Public/Platform.ush"
|
|
|
|
uint InPrimitiveIndex;
|
|
uint InNumInstancesAllocatedInGPUScene;
|
|
uint InInstanceOffset;
|
|
StructuredBuffer<uint4> InInstanceData;
|
|
StructuredBuffer<uint> InInstanceCustomFloatData;
|
|
StructuredBuffer<uint> InWriteCounters;
|
|
uint InPrimitiveId;
|
|
uint InCustomDataCount;
|
|
uint InPayloadDataFlags;
|
|
uint InSeed;
|
|
|
|
// This follows the implementation in PCGHelpers.
|
|
uint ComputeSeed(uint A, uint B)
|
|
{
|
|
return ((A * 196314165U) + 907633515U) ^ ((B * 73148459U) + 453816763U);
|
|
}
|
|
|
|
// This follows the implementation in FRandomStream.
|
|
float FRand(inout uint InOutSeed)
|
|
{
|
|
InOutSeed = (InOutSeed * 196314165U) + 907633515U;
|
|
|
|
const float Result = asfloat(0x3F800000U | (InOutSeed >> 9));
|
|
|
|
return Result - 1.0f;
|
|
}
|
|
|
|
void CullInstance(uint InstanceId)
|
|
{
|
|
// This will minimally initialize the instance data such that it will be ignored and culled
|
|
WriteInstancePrimitiveIdAndFlags(InstanceId, InPrimitiveId, INSTANCE_SCENE_DATA_FLAG_HIDDEN);
|
|
}
|
|
|
|
[numthreads(NUM_THREADS_PER_GROUP, 1, 1)]
|
|
void SceneWriter_CS(const uint3 GroupId : SV_GroupID, const uint GroupIndex : SV_GroupIndex)
|
|
{
|
|
const uint LocalInstanceId = GetUnWrappedDispatchThreadId(GroupId, GroupIndex, NUM_THREADS_PER_GROUP);
|
|
|
|
if (LocalInstanceId >= InNumInstancesAllocatedInGPUScene)
|
|
{
|
|
return;
|
|
}
|
|
|
|
// Retrieve the primitive data
|
|
FPrimitiveSceneData PrimitiveData = GetPrimitiveData(InPrimitiveId);
|
|
|
|
const uint InstanceId = PrimitiveData.InstanceSceneDataOffset + LocalInstanceId;
|
|
|
|
// Cull any instances at tail end (assume all valid instances are written).
|
|
if (LocalInstanceId >= InWriteCounters[InPrimitiveIndex])
|
|
{
|
|
CullInstance(InstanceId);
|
|
return;
|
|
}
|
|
|
|
const uint StrideFloat4s = 3;
|
|
float4x4 LocalToWorld;
|
|
LocalToWorld[0] = asfloat(InInstanceData[(InInstanceOffset + LocalInstanceId) * StrideFloat4s + 0]);
|
|
LocalToWorld[1] = asfloat(InInstanceData[(InInstanceOffset + LocalInstanceId) * StrideFloat4s + 1]);
|
|
LocalToWorld[2] = asfloat(InInstanceData[(InInstanceOffset + LocalInstanceId) * StrideFloat4s + 2]);
|
|
LocalToWorld[3] = float4(0.0f, 0.0f, 0.0f, 1.0f);
|
|
LocalToWorld = transpose(LocalToWorld);
|
|
|
|
uint Seed = ComputeSeed(InSeed, InstanceId);
|
|
|
|
// todo_pcg: We should write local space transforms to the instance data buffer from the SM Spawner and use InitializeInstanceSceneData (non-WS) here instead.
|
|
InitializeInstanceSceneDataWS(
|
|
InstanceId,
|
|
InPrimitiveId,
|
|
LocalInstanceId,
|
|
InPayloadDataFlags,
|
|
InCustomDataCount,
|
|
FRand(Seed),
|
|
DFFromTileOffset(LWCPromote(LocalToWorld))
|
|
);
|
|
|
|
FInstancePayloadDataOffsets PayloadOffsets = GetInstancePayloadDataOffsets(InPrimitiveId, InPayloadDataFlags, LocalInstanceId);
|
|
|
|
const uint NumCustomFloat4s = (InCustomDataCount + 3) / 4;
|
|
for (uint Float4Index = 0; Float4Index < NumCustomFloat4s; ++Float4Index)
|
|
{
|
|
const uint BaseIndex = (InInstanceOffset + LocalInstanceId) * InCustomDataCount + Float4Index * 4;
|
|
|
|
uint4 CustomData;
|
|
CustomData[0] = InInstanceCustomFloatData[BaseIndex];
|
|
CustomData[1] = InInstanceCustomFloatData[BaseIndex + 1];
|
|
CustomData[2] = InInstanceCustomFloatData[BaseIndex + 2];
|
|
CustomData[3] = InInstanceCustomFloatData[BaseIndex + 3];
|
|
WriteInstanceCustomData(PayloadOffsets, Float4Index, asfloat(CustomData));
|
|
}
|
|
}
|