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

${VFXGlobalInclude}
${VFXGlobalDeclaration}

#define IndirectOutputType uint

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

#ifdef VFX_IS_RAYTRACED
#pragma multi_compile _ VFX_COMPUTE_AABBS
#endif

#define USE_DEAD_LIST (VFX_USE_ALIVE_CURRENT && !HAS_STRIPS)

RWByteAddressBuffer attributeBuffer;

#if USE_DEAD_LIST
RWStructuredBuffer<uint> deadList;
#endif

#if VFX_HAS_INDIRECT_DRAW
RWStructuredBuffer<uint> indirectBuffer;
#endif

#if HAS_STRIPS_DATA
RWStructuredBuffer<uint> stripDataBuffer;
#endif

#if VFX_USE_STRIPALIVE_CURRENT
StructuredBuffer<uint> attachedStripDataBuffer;
#endif

CBUFFER_START(updateParamsConst)
    uint dispatchWidth;
    ${VFXInstancingConstants}
    float3 cameraXRSettings;
CBUFFER_END

${VFXPerPassInclude}

#if VFX_HAS_INDIRECT_DRAW
${VFXDeclareAppendOutputIndirectBuffer}
#endif

${VFXGeneratedBlockFunction}



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

	${VFXInitInstancingCompute}

	${VFXLoadContextData}
	uint systemSeed = contextData.systemSeed;
	uint nbMax = contextData.maxParticleCount;

#if VFX_COMPUTE_BOUNDS
	uint tid = groupThreadId.x;
#endif

	if (index < nbMax)
	{
#if HAS_STRIPS
		if (index == 0)
		{
			STRIP_PARTICLE_COUNTER(instanceIndex) = 0;
		}
#endif

		${VFXLoadGraphValues}

		VFXAttributes attributes = (VFXAttributes)0;
		VFXSourceAttributes sourceAttributes = (VFXSourceAttributes)0;

#if VFX_USE_ALIVE_CURRENT
		${VFXLoadAttributes}
		if (attributes.alive)
		{
#if VFX_USE_OLDVELOCITY_CURRENT && VFX_USE_VELOCITY_CURRENT
			attributes.oldVelocity = attributes.velocity;
#endif
#if HAS_STRIPS
			const StripData stripData = GetStripDataFromParticleIndex(index, instanceIndex);
			InitStripAttributes(index, attributes, stripData);
#endif

#if VFX_UPDATE_SKIP_ZERO_DELTA_TIME
			${VFXLoadParameter:{deltaTime}}
			if (deltaTime != 0.0f)
#endif
			{
				${VFXProcessBlocks}
			}

			${VFXStoreAttributes:{(?!(alive))(\b\w)}}
			if (attributes.alive)
			{

#if VFX_HAS_INDIRECT_DRAW
                AppendOutputBuffer(indirectBuffer, index, instanceActiveIndex);
#endif

#if HAS_STRIPS
				uint bufferIndex = STRIP_DATA_INDEX(instanceIndex, stripData.stripIndex);
				uint relativeIndexInStrip = GetRelativeIndex(index, stripData);
				InterlockedMin(STRIP_DATA(STRIP_MIN_ALIVE, bufferIndex), relativeIndexInStrip);
				InterlockedMax(STRIP_DATA(STRIP_MAX_ALIVE, bufferIndex), relativeIndexInStrip);
#endif
			}
			else
			{
				${VFXStoreAttributes:{alive}}
#if USE_DEAD_LIST && !VFX_USE_STRIPALIVE_CURRENT
	            uint deadIndex;
		        InterlockedAdd(deadList[instanceIndex], 1, deadIndex);
				deadIndex += DEAD_LIST_OFFSET + instanceIndex * RAW_CAPACITY;
				deadList[deadIndex] = index;
#endif
			}
		}
#if USE_DEAD_LIST && VFX_USE_STRIPALIVE_CURRENT
        else if (attributes.stripAlive)
        {
            uint bufferIndex = ((instanceIndex * ATTACHED_STRIP_COUNT) + index);
            if (index >= ATTACHED_STRIP_COUNT || STRIP_DATA_X(attachedStripDataBuffer, STRIP_MIN_ALIVE, bufferIndex) == ~1) // Attached strip is no longer alive, recycle the particle
            {
	            uint deadIndex;
		        InterlockedAdd(deadList[instanceIndex], 1, deadIndex);
				deadIndex += DEAD_LIST_OFFSET + instanceIndex * RAW_CAPACITY;
				deadList[deadIndex] = index;
                attributes.stripAlive = false;
                ${VFXStoreAttributes:{stripAlive}}
            }
        }
#endif

// For strips, we still need to render some dead particles if they are in the middle of the strip
#if HAS_STRIPS && VFX_HAS_INDIRECT_DRAW
		if (!attributes.alive)
		{
			const StripData stripData = GetStripDataFromParticleIndex(index, instanceIndex);
			uint relativeIndexInStrip = GetRelativeIndex(index, stripData);
			if (relativeIndexInStrip > 0 && relativeIndexInStrip + 1 < stripData.nextIndex)
			{
				AppendOutputBuffer(indirectBuffer, index, instanceActiveIndex);
			}
		}
#endif

#else
		${VFXLoadAttributes}
#if VFX_USE_OLDVELOCITY_CURRENT && VFX_USE_VELOCITY_CURRENT
		attributes.oldVelocity = attributes.velocity;
#endif
#if VFX_USE_CONTINUOUSCOLLISIONCOUNT_CURRENT
		//attributes.continuousCollisionCount = 0;
#endif
#if HAS_STRIPS
        const StripData stripData = GetStripDataFromParticleIndex(index, instanceIndex);
        InitStripAttributes(index, attributes, stripData);
#endif

#if VFX_UPDATE_SKIP_ZERO_DELTA_TIME
		${VFXLoadParameter:{deltaTime}}
		if (deltaTime != 0.0f)
#endif
		{
			${VFXProcessBlocks}
		}
		${VFXStoreAttributes}
#if VFX_HAS_INDIRECT_DRAW
        AppendOutputBuffer(indirectBuffer, index, instanceActiveIndex);
#endif
#endif
#if VFX_COMPUTE_BOUNDS || VFX_COMPUTE_AABBS
    {
    #if VFX_COMPUTE_BOUNDS
        ${VFXLoadSize}
        #if VFX_WORLD_SPACE
            ${VFXLoadParameter:{worldToLocal}}
        #else
            float4x4 worldToLocal = (float4x4)0;
        #endif
		InitReduction(attributes, size3, tid, worldToLocal);
    #endif
    }
    #if VFX_COMPUTE_AABBS
            ${VFXLoadSizeRT}
            int rayTracingDecimationFactor = VFX_RT_DECIMATION_FACTOR;
            FillAabbBuffer(attributes, size3, index, instanceIndex, rayTracingDecimationFactor);
    #endif
#endif
	}
	#if VFX_COMPUTE_BOUNDS
	PerformBoundsReduction(index, tid, instanceIndex, nbMax);
	#endif
}