// Put the following line to 0 or comment it to disable vignette weighting
#define USE_VIGNETTE_WEIGHTING 1

#include "Common.cginc"
#include "EyeAdaptation.cginc"

RWStructuredBuffer<uint> _Histogram;
Texture2D<float4> _Source;

CBUFFER_START(Params)
    float4 _ScaleOffsetRes; // x: scale, y: offset, z: width, w: height
CBUFFER_END

groupshared uint gs_histogram[HISTOGRAM_BINS];

#pragma kernel KEyeHistogram
[numthreads(HISTOGRAM_THREAD_X,HISTOGRAM_THREAD_Y,1)]
void KEyeHistogram(uint2 dispatchThreadId : SV_DispatchThreadID, uint2 groupThreadId : SV_GroupThreadID)
{
    // Pretty straightforward implementation of histogram gathering using atomic ops.
    // I tried a few methods (no atomic ops / heavy LDS leveraging) but this one turned out to be
    // the fastest on desktop (Nvidia - Kepler/Maxwell) and PS4. Still need to try it on GCN/desktop
    // but considering it runs very fast on PS4 we can expect it to run well (?).

    const uint localThreadId = groupThreadId.y * HISTOGRAM_THREAD_X + groupThreadId.x;

    // Clears the shared memory
    if (localThreadId < HISTOGRAM_BINS)
        gs_histogram[localThreadId] = 0u;

    GroupMemoryBarrierWithGroupSync();

    // Gather local group histogram
    if (dispatchThreadId.x < (uint)_ScaleOffsetRes.z && dispatchThreadId.y < (uint)_ScaleOffsetRes.w)
    {
#if USE_VIGNETTE_WEIGHTING
        // Vignette weighting to put more focus on what's in the center of the screen
        float2 uv01 = float2(dispatchThreadId) / float2(_ScaleOffsetRes.z, _ScaleOffsetRes.w);
        float2 d = abs(uv01 - (0.5).xx);
        float vfactor = Pow2(saturate(1.0 - dot(d, d)));
        uint weight = (uint)(64.0 * vfactor);
#else
        uint weight = 1u;
#endif

        float3 color = _Source[dispatchThreadId].xyz;
        float luminance = Max3(color); // Looks more natural than using a Rec.709 luminance for some reason
        float logLuminance = GetHistogramBinFromLuminance(luminance, _ScaleOffsetRes.xy);
        uint idx = (uint)(logLuminance * (HISTOGRAM_BINS - 1u));
        InterlockedAdd(gs_histogram[idx], weight);
    }

    GroupMemoryBarrierWithGroupSync();

    // Merge everything
    if (localThreadId < HISTOGRAM_BINS)
        InterlockedAdd(_Histogram[localThreadId], gs_histogram[localThreadId]);
}