// Precomputes data for IntegrateLD(). See that function for a detailed description.

#include "../../Core/ShaderLibrary/Common.hlsl"
#include "../../Core/ShaderLibrary/ImageBasedLighting.hlsl"

#ifdef SHADER_API_MOBILE
#define MAX_IBL_SAMPLE_CNT 34
#else
#define MAX_IBL_SAMPLE_CNT 89
#endif

RWTexture2D<float4> output; // [MAX_SAMPLE_CNT x UNITY_SPECCUBE_LOD_STEPS]

#pragma kernel ComputeGgxIblSampleData

[numthreads(MAX_IBL_SAMPLE_CNT, UNITY_SPECCUBE_LOD_STEPS, 1)]
void ComputeGgxIblSampleData(uint3 groupThreadId : SV_GroupThreadID)
{
    uint  sampleIndex = groupThreadId.x;
    uint  mipLevel    = groupThreadId.y + 1;
    float roughness   = PerceptualRoughnessToRoughness(MipmapLevelToPerceptualRoughness(mipLevel));
    uint  sampleCount = GetIBLRuntimeFilterSampleCount(mipLevel);

    if (sampleIndex >= sampleCount)
    {
        output[groupThreadId.xy] = float4(0, 0, 0, 0);
        return;
    }

    float3 V     = float3(0, 0, 1);
    float  NdotV = 1;

    float3 localL;
    float  NdotL, NdotH, LdotH;

    const uint validSampleCount = sampleCount;

    while (true)
    {
        uint acceptedSampleCount = 0;
        uint shiftedSampleIndex  = 0;

        for (uint i = 0; i < sampleCount; i++)
        {
            // We switch to the Golden sequence instead of the Fibonacci sequence
            // since the sample count is not guaranteed to be a Fibonacci number.
            // TODO: might be interesting to try Mitchell's Poisson disk sampling algorithm.
            // In our case, samples would not have disks associated with them, but rather solid angles.
            float2 u = Golden2dSeq(i, sampleCount);
            SampleGGXDir(u, V, k_identity3x3, roughness, localL, NdotL, NdotH, LdotH, true);

            if (NdotL > 0)
            {
                if (acceptedSampleCount == sampleIndex) shiftedSampleIndex = i;
                acceptedSampleCount++;
            }
        }

        if (acceptedSampleCount == validSampleCount)
        {
            sampleIndex = shiftedSampleIndex;
            break;
        }

        sampleCount++;
    }

    float2 u = Golden2dSeq(sampleIndex, sampleCount);

    SampleGGXDir(u, V, k_identity3x3, roughness, localL, NdotL, NdotH, LdotH, true);

    float pdf    = 0.25 * D_GGX(NdotH, roughness);
    float omegaS = rcp(sampleCount) * rcp(pdf);

    output[groupThreadId.xy] = float4(localL, omegaS);
}