#pragma kernel TileMerge MERGE_PASS SCATTERING
#pragma only_renderers d3d11 playstation xboxone xboxseries vulkan metal switch

#include "Packages/com.unity.render-pipelines.high-definition/Runtime/PostProcessing/Shaders/MotionBlurTileCommon.hlsl"

TEXTURE2D_X_UINT(_TileToScatterMax);
TEXTURE2D_X(_TileToScatterMin);
RW_TEXTURE2D_X(float3, _TileMaxNeighbourhood);

[numthreads(8, 8, 1)]
void TileMerge(uint3 dispatchID : SV_DispatchThreadID)
{
    UNITY_XR_ASSIGN_VIEW_INDEX(dispatchID.z);

    int2 id = dispatchID.xy;

    // If we scattered wider than a 3 tile radius, then we do an additional neighbourhood search.
    float2 maxVelData = UnpackMotionVec(_TileToScatterMax[COORD_TEXTURE2D_X(id)]);
    float maxVelLen = MotionVecLengthInPixelsFromEncoded(maxVelData);

    // TODO: In theory this could have to be even wider. In general, we need a better way to determine min velocities.
    float minVel = 0.0f;
    int2 maxCoords = int2(_TileTargetSize.xy - 1);

    if (maxVelLen > TILE_SIZE*0.5f)
    {
        int tileCount = maxVelLen / TILE_SIZE;
        if (tileCount > 4) tileCount = 4;

        minVel = 999999.0f;
        // This is sub-optimal, but the texture is tiny and should be fairly cache efficient.
        for (int x = -tileCount; x < tileCount; ++x)
        {
            for (int y = -tileCount; y < tileCount; ++y)
            {
                int2 tapCoord = clamp(id + int2(x, y), 0, _TileTargetSize.xy - 1);
                float data = _TileToScatterMin[COORD_TEXTURE2D_X(tapCoord)].x;
                minVel = min(minVel, data);
            }
        }
    }
    else
    {
        minVel = _TileToScatterMin[COORD_TEXTURE2D_X(id)].x;
    }

    _TileMaxNeighbourhood[COORD_TEXTURE2D_X(id)] = float3(UnpackMotionVec(_TileToScatterMax[COORD_TEXTURE2D_X(id)]), (minVel));
}