#pragma kernel CSMain
${VFXGlobalInclude}
${VFXGlobalDeclaration}
${VFXPerPassInclude}

CBUFFER_START(params)
    uint nbMax;
    uint dispatchWidth;
CBUFFER_END

CBUFFER_START(cameraParams)
    float3 cameraPosition;
CBUFFER_END

ByteAddressBuffer attributeBuffer;
StructuredBuffer<uint> inputBuffer;

#if USE_DEAD_LIST_COUNT
ByteAddressBuffer deadListCount;
#endif

struct Kvp
{
	float sortKey;
	uint index;
};

RWStructuredBuffer<Kvp> outputBuffer;

[numthreads(NB_THREADS_PER_GROUP,1,1)]
void CSMain(uint3 groupId          : SV_GroupID,
            uint3 groupThreadId    : SV_GroupThreadID)
{
	uint threshold = nbMax;
#if USE_DEAD_LIST_COUNT
	threshold -= deadListCount.Load(0);
#endif
	uint id = groupThreadId.x + groupId.x * NB_THREADS_PER_GROUP + groupId.y * dispatchWidth * NB_THREADS_PER_GROUP;
	if (id < threshold)
	{
		uint index = inputBuffer[id];
		
		Attributes attributes = (Attributes)0;
		${VFXLoadAttributes:{position}}
		
#if VFX_LOCAL_SPACE
		float3 wPos = mul(localToWorld,float4(attributes.position,1.0f)).xyz;
#else
		float3 wPos = attributes.position;
#endif
		float3 camToPos = wPos - cameraPosition;
		
		Kvp kvp;
		kvp.sortKey = dot(camToPos,camToPos); // sqr distance to the camera
		kvp.index = index;

		outputBuffer[id] = kvp;
	}
}