#pragma kernel CSMain
${VFXPragmaOnlyRenderers}
${VFXPragmaRequire}

${VFXGlobalInclude}
${VFXGlobalDeclaration}
${VFXInclude("Shaders/VFXParticleCommon.template")}

CBUFFER_START(sortParamsConst)
    uint dispatchWidth;
    ${VFXInstancingConstants}
CBUFFER_END

ByteAddressBuffer attributeBuffer;
StructuredBuffer<uint> inputBuffer;

struct Kvp
{
	float sortKey;
	uint index;
};

RWStructuredBuffer<Kvp> outputBuffer;

${VFXPerPassInclude}

[numthreads(NB_THREADS_PER_GROUP,1,1)]
void CSMain(uint3 groupId          : SV_GroupID,
            uint3 groupThreadId    : SV_GroupThreadID)
{
	uint id = GetThreadId(groupId, groupThreadId, dispatchWidth);
	uint index = id;
    ${VFXInitInstancing}

	uint threshold = inputBuffer[instanceActiveIndex];
    if(index == 0)
        outputBuffer[instanceActiveIndex].index = threshold; //Store the count here

	if (index < threshold)
	{
#if HAVE_VFX_MODIFICATION
        ${VFXLoadGraphValues}

    #if VFX_LOCAL_SPACE
        ${VFXLoadParameter:{localToWorld}}
        vfxLocalToWorld = localToWorld;

        ${VFXLoadParameter:{worldToLocal}}
        vfxWorldToLocal = worldToLocal;
    #else
        vfxWorldToLocal = k_identity4x4;
        vfxLocalToWorld = k_identity4x4;
    #endif
#endif

        uint indirectId = index + instancingBatchSize + instanceActiveIndex * RAW_CAPACITY;
		uint index = inputBuffer[indirectId];

		VFXAttributes attributes = (VFXAttributes)0;
        ${VFXLoadAttributes}

        ${VFXFeedSortingKeys}

        outputBuffer[indirectId] = output;
	}
}