// Each #kernel tells which function to compile; you can have many kernels
#pragma kernel CSMainMipmapGroup0
#pragma kernel CSMainMipmapGroup0 _INPUTE_DEPTH_TEXTURE
#pragma kernel CSMainMipmapGroup0 _INPUTE_DEPTH_TEXTURE _SKIP_3_MIP
#pragma enable_d3d11_debug_symbols
#pragma target 5.0

Texture2D<float> _InputDepth;

RWTexture2D<float> _DepthMipChain;
#include "Packages/com.unity.render-pipelines.core/ShaderLibrary/Common.hlsl"

#if UNITY_REVERSED_Z
# define MIN_DEPTH(l, r) min(l, r)
#else
# define MIN_DEPTH(l, r) max(l, r)
#endif

#define GROUP_TILE_SIZE 8
float4x4 _MipOffsetAndSizeArray;
float4 _InputMipOffsetAndSize;
float _MipCount;
groupshared float _LDSDepths[GROUP_TILE_SIZE * GROUP_TILE_SIZE];


[numthreads(GROUP_TILE_SIZE, GROUP_TILE_SIZE, 1)]
void CSMainMipmapGroup0(uint3 dispatchThreadId : SV_DispatchThreadID, uint groupThreadIndex : SV_GroupIndex, uint2 groupId : SV_GroupID,  uint3 groupThreadID : SV_GroupThreadID)
{
    uint2 curMipSize = _MipOffsetAndSizeArray[0].zw;
    //Mip0从InputeTexture贴图下采样
    float2 uv = dispatchThreadId.xy / float2(curMipSize);
    uv = clamp(uv, 0, 1);
    int2 texCrood = _InputMipOffsetAndSize.xy + uv * (_InputMipOffsetAndSize.zw);
    uint2 maxIndex = _InputMipOffsetAndSize.xy + _InputMipOffsetAndSize.zw - 1;
#ifdef _INPUTE_DEPTH_TEXTURE //第一次Dispath使用Depth Texture
    float p00 = _InputDepth[min(texCrood + uint2(0, 0), maxIndex)];
    float p01 = _InputDepth[min(texCrood + uint2(1, 0), maxIndex)];
    float p10 = _InputDepth[min(texCrood + uint2(0, 1), maxIndex)];
    float p11 = _InputDepth[min(texCrood + uint2(1, 1), maxIndex)];
#else
    float p00 = _DepthMipChain[min(texCrood + uint2(0, 0), maxIndex)];
    float p01 = _DepthMipChain[min(texCrood + uint2(1, 0), maxIndex)];
    float p10 = _DepthMipChain[min(texCrood + uint2(0, 1), maxIndex)];
    float p11 = _DepthMipChain[min(texCrood + uint2(1, 1), maxIndex)];
#endif
    float4 depths = float4(p00, p10, p01, p11);
    float minDepth = MIN_DEPTH(MIN_DEPTH(depths.x, depths.y), MIN_DEPTH(depths.z, depths.w));

    if (all(dispatchThreadId.xy < curMipSize.xy))
    {
#ifndef _SKIP_3_MIP //跳过前3级,不输出到RT
        _DepthMipChain[_MipOffsetAndSizeArray[0].xy + dispatchThreadId.xy] = minDepth;
#endif
        _LDSDepths[groupThreadIndex] = minDepth;
    }
    GroupMemoryBarrierWithGroupSync();
    //Mip1
    curMipSize = curMipSize >> 1;
    uint preMipTileSize = GROUP_TILE_SIZE;
    uint TileSize = GROUP_TILE_SIZE / 2; //线程数减半
    float4 parentFurthestDeviceZ;
    uint2 xy = min(groupThreadID.xy, TileSize - 1);
    uint2 xy2 = xy * 2;//间隔索引:[0,2,4,6]
    uint index0 = xy2.x + xy2.y * preMipTileSize;
    uint index1 = (xy2.x + 1) + xy2.y * preMipTileSize;
    uint index2 = xy2.x + (xy2.y + 1) * preMipTileSize;
    uint index3 = (xy2.x + 1) + (xy2.y + 1) * preMipTileSize;

    parentFurthestDeviceZ.x = _LDSDepths[index0];
    parentFurthestDeviceZ.y = _LDSDepths[index1];
    parentFurthestDeviceZ.z = _LDSDepths[index2];
    parentFurthestDeviceZ.w = _LDSDepths[index3];

    float furthestDeviceZ = MIN_DEPTH(MIN_DEPTH(parentFurthestDeviceZ.x, parentFurthestDeviceZ.y), MIN_DEPTH(parentFurthestDeviceZ.z, parentFurthestDeviceZ.w));
    uint2 localIndex = dispatchThreadId.xy - (groupId.xy * (GROUP_TILE_SIZE - TileSize));//因为过滤了3/4的线程组,所以线程索引需要往前补上 

    if (all(localIndex < curMipSize.xy) //索引在贴图范围内的
        && all(groupThreadID.xy < TileSize) //每次每个线程组保留1/4的线程
        )
    {
#ifndef _SKIP_3_MIP  //跳过前3级,不输出到RT
        _DepthMipChain[_MipOffsetAndSizeArray[1].xy + localIndex] = furthestDeviceZ;
#endif
        int threadIndex = xy.x + xy.y * TileSize;
        _LDSDepths[threadIndex] = furthestDeviceZ;
    }

    //Mip2
    GroupMemoryBarrierWithGroupSync();

    curMipSize = curMipSize >> 1;
    preMipTileSize = TileSize;
    TileSize = TileSize / 2; //线程数减半
    xy = min(groupThreadID.xy, TileSize - 1);
    xy2 = xy * 2;//间隔索引:[0,2,4,6]
    index0 = xy2.x + xy2.y * preMipTileSize;
    index1 = (xy2.x + 1) + xy2.y * preMipTileSize;
    index2 = xy2.x + (xy2.y + 1) * preMipTileSize;
    index3 = (xy2.x + 1) + (xy2.y + 1) * preMipTileSize;

    parentFurthestDeviceZ.x = _LDSDepths[index0];
    parentFurthestDeviceZ.y = _LDSDepths[index1];
    parentFurthestDeviceZ.z = _LDSDepths[index2];
    parentFurthestDeviceZ.w = _LDSDepths[index3];

    GroupMemoryBarrierWithGroupSync();
    furthestDeviceZ = MIN_DEPTH(MIN_DEPTH(parentFurthestDeviceZ.x, parentFurthestDeviceZ.y), MIN_DEPTH(parentFurthestDeviceZ.z, parentFurthestDeviceZ.w));
    localIndex = dispatchThreadId.xy - (groupId.xy * (GROUP_TILE_SIZE - TileSize));//因为过滤了3/4的线程组,所以线程索引需要往前补上 

    if (all(localIndex < curMipSize.xy) //索引在贴图范围内的
        && all(groupThreadID.xy < TileSize) //每次每个线程组保留1/4的线程
        )
    {
#ifndef _SKIP_3_MIP //跳过前3级,不输出到RT
        _DepthMipChain[_MipOffsetAndSizeArray[2].xy + localIndex] = furthestDeviceZ;
#endif
        int threadIndex = xy.x + xy.y * TileSize;
        _LDSDepths[threadIndex] = furthestDeviceZ;
    }

    //Mip3
    GroupMemoryBarrierWithGroupSync();

    curMipSize = curMipSize >> 1;
    preMipTileSize = TileSize;
    TileSize = TileSize / 2; //线程数减半
    xy = min(groupThreadID.xy, TileSize - 1);
    xy2 = xy * 2;//间隔索引:[0,2,4,6]
    index0 = xy2.x + xy2.y * preMipTileSize;
    index1 = (xy2.x + 1) + xy2.y * preMipTileSize;
    index2 = xy2.x + (xy2.y + 1) * preMipTileSize;
    index3 = (xy2.x + 1) + (xy2.y + 1) * preMipTileSize;

    parentFurthestDeviceZ.x = _LDSDepths[index0];
    parentFurthestDeviceZ.y = _LDSDepths[index1];
    parentFurthestDeviceZ.z = _LDSDepths[index2];
    parentFurthestDeviceZ.w = _LDSDepths[index3];

    GroupMemoryBarrierWithGroupSync();
    furthestDeviceZ = MIN_DEPTH(MIN_DEPTH(parentFurthestDeviceZ.x, parentFurthestDeviceZ.y), MIN_DEPTH(parentFurthestDeviceZ.z, parentFurthestDeviceZ.w));
    localIndex = dispatchThreadId.xy - (groupId.xy * (GROUP_TILE_SIZE - TileSize));//因为过滤了3/4的线程组,所以线程索引需要往前补上 

    if (all(localIndex < curMipSize.xy) //索引在贴图范围内的
        && all(groupThreadID.xy < TileSize) //每次每个线程组保留1/4的线程
        )
    {
        _DepthMipChain[_MipOffsetAndSizeArray[3].xy + localIndex] = furthestDeviceZ;
    }
}