#define GROUP_SIZE 64

cbuffer FrameConstants : register(b0)
{
    float4x4 gViewProjection;
    float4x4 gView;
    float4x4 gProjection;
    float4 gCameraWorldPos;
    float4 gScreenParams;
    float4 gSettings;
};

ByteAddressBuffer gPositions : register(t0);
StructuredBuffer<uint> gOrderBuffer : register(t1);
RWStructuredBuffer<uint> gSortKeys : register(u0);

float3 LoadFloat3(ByteAddressBuffer buffer, uint byteOffset)
{
    return asfloat(buffer.Load3(byteOffset));
}

uint FloatToSortableUint(float value)
{
    uint bits = asuint(value);
    uint mask = (0u - (bits >> 31)) | 0x80000000u;
    return bits ^ mask;
}

[numthreads(GROUP_SIZE, 1, 1)]
void MainCS(uint3 dispatchThreadId : SV_DispatchThreadID)
{
    uint index = dispatchThreadId.x;
    uint splatCount = (uint)gSettings.x;
    if (index >= splatCount)
    {
        return;
    }

    uint splatIndex = gOrderBuffer[index];
    float3 position = LoadFloat3(gPositions, splatIndex * 12);
    float3 viewPosition = mul(float4(position, 1.0), gView).xyz;
    gSortKeys[index] = FloatToSortableUint(viewPosition.z);
}