#include "../ShaderLibrary/BC6H.hlsl"

Texture2D<float4> _Source;
RWTexture2D<uint4> _Target;

#pragma kernel KEncodeFast
[numthreads(4, 4, 1)]
void KEncodeFast(uint2 groupId : SV_GroupID, uint2 groupThreadId : SV_GroupThreadID, uint2 dispatchThreadId : SV_DispatchThreadID)
{
    // Load 4x4 pixel block
    float3 texels[16];

    uint2 topLeftSourceID = dispatchThreadId << 2;
    
    texels[0] = _Source.Load(uint3(topLeftSourceID              , 0)).rgb;
    texels[1] = _Source.Load(uint3(topLeftSourceID + uint2(1, 0), 0)).rgb;
    texels[2] = _Source.Load(uint3(topLeftSourceID + uint2(2, 0), 0)).rgb;
    texels[3] = _Source.Load(uint3(topLeftSourceID + uint2(3, 0), 0)).rgb;
    texels[4] = _Source.Load(uint3(topLeftSourceID + uint2(0, 1), 0)).rgb;
    texels[5] = _Source.Load(uint3(topLeftSourceID + uint2(1, 1), 0)).rgb;
    texels[6] = _Source.Load(uint3(topLeftSourceID + uint2(2, 1), 0)).rgb;
    texels[7] = _Source.Load(uint3(topLeftSourceID + uint2(3, 1), 0)).rgb;
    texels[8] = _Source.Load(uint3(topLeftSourceID + uint2(0, 2), 0)).rgb;
    texels[9] = _Source.Load(uint3(topLeftSourceID + uint2(1, 2), 0)).rgb;
    texels[10] = _Source.Load(uint3(topLeftSourceID + uint2(2, 2), 0)).rgb;
    texels[11] = _Source.Load(uint3(topLeftSourceID + uint2(3, 2), 0)).rgb;
    texels[12] = _Source.Load(uint3(topLeftSourceID + uint2(0, 3), 0)).rgb;
    texels[13] = _Source.Load(uint3(topLeftSourceID + uint2(1, 3), 0)).rgb;
    texels[14] = _Source.Load(uint3(topLeftSourceID + uint2(2, 3), 0)).rgb;
    texels[15] = _Source.Load(uint3(topLeftSourceID + uint2(3, 3), 0)).rgb;

    uint4 block = uint4(0, 0, 0, 0);
    float blockMSLE = 0;

    EncodeMode11(block, blockMSLE, texels);

    _Target[dispatchThreadId] = block;
}