#ifndef CAPSULE_SHADOW_INCLUDED
#define CAPSULE_SHADOW_INCLUDED
#pragma multi_compile_fragment  _ _MultiCapsule_Shadow_AO_ON
#pragma multi_compile_fragment  _ _MultiCapsule_AO_ON

#include "Packages/com.unity.render-pipelines.universal/ShaderLibrary/Core.hlsl"
struct Capsule
{
    half3 a, b;
    half radius;
};

struct Character
{
    half3 position;
    half radius;
    int capsuleStartID, capsuleEndID;
    uint capsuleAOSettingMask;
};

struct CapsuleShadowArea
{
    half AmbientIntensity;
    half ConeAngle;
    half ShadowIntensity;
    half ShadowSharpness;
    int lightStartID, lightEndID;
};

StructuredBuffer<Capsule> _CapsuleData;
uint _CapsulesCount;
StructuredBuffer<Character> _CharacterData;
uint _CharactersCount;

StructuredBuffer<CapsuleShadowArea> _CapsuleShadowData;
uint _CapsuleShadowDataCount;

half4 _CapsuleLightsDir[32];
uint _CapsuleLightsCount;

float CalcCapsuleShadowByIndex(float3 ro, float3 rd, uint s, uint e, in float k, float intensity, float intensity1)
{
    float shadow = 1.0;
    for (uint i = s; i < e; ++i)
    {
        Capsule c = _CapsuleData[i];
        float3 a = c.a;
        float3 b = c.b;
        float r = c.radius;
        float3 ba = b - a;
        float3 oa = ro - a;

        float oad = dot(oa, rd);
        float dba = dot(rd, ba);
        float baba = dot(ba, ba);
        float oaba = dot(oa, ba);
        float2 th = float2(-oad * baba + dba * oaba, oaba - oad * dba) / (baba - dba * dba);

        th.x = max(th.x, 0.0001);
        th.y = saturate(th.y);

        //线段ab上离射线最近的点p
        float3 p = a + ba * th.y;
        //射线rord 找到距离线段 ab 最近的点q
        float3 q = ro + rd * th.x;
        //沿着光的方向发射射线, 射线离胶囊体最短的距离 
        float d = length(p - q) - r;

        float s = saturate(k * d / th.x + 0.3);
        s = s * s * (3.0 - 2.0 * s);
        shadow *= s;
    }

    shadow = saturate(lerp(1, shadow, intensity));
    return saturate(lerp(1, shadow, intensity1));
}

float CalcCapsuleOcclusionByIndex(float3 p, float3 n, uint s, uint e, float intensity, float intensity1)
{
    float ao = 1.0;
    for (uint i = s; i < e; ++i)
    {
        Capsule capsule = _CapsuleData[i];
        float3 a = capsule.a;
        float3 b = capsule.b;
        float r = capsule.radius;
        float3 ba = b - a;
        float3 pa = p - a;
        float h = saturate(dot(pa, ba) / dot(ba, ba));
        // 地面点 p 到 ab 最近点 x, 向量 xp = d
        float3 d = pa - h * ba;
        float l = length(d);
        float nl = dot(-d, n);
        float o = 1.0 - max(0.0, nl) * r * r / (l * l * l);
        // multiplier o *= 1.0 + r*(l-r)/(l*l); 
        o = sqrt(o * o * o);
        ao *= o;
    }

    ao = saturate(lerp(1, ao, intensity));
    return clamp(lerp(1, ao, intensity1), 0.6, 1);
}

#define sq(x) (x * x)
float acosFast(float x)
{
    // Lagarde 2014, "Inverse trigonometric functions GPU optimization for AMD GCN architecture"
    // This is the approximation of degree 1, with a max absolute error of 9.0x10^-3
    float y = abs(x);
    float p = -0.1565827 * y + 1.570796;
    p *= sqrt(1.0 - y);
    return x >= 0.0 ? p : PI - p;
}

float acosFastPositive(float x)
{
    // Lagarde 2014, "Inverse trigonometric functions GPU optimization for AMD GCN architecture"
    float p = -0.1565827 * x + 1.570796;
    return p * sqrt(1.0 - x);
}
            
/*
    ref: https://developer.amd.com/wordpress/media/2012/10/Oat-AmbientApetureLighting.pdf
    Approximate the area of intersection of two spherical caps.

    With some modifcations proposed by the shadertoy implementation
*/
float SphericalCapsIntersectionAreaFast(float cosCap1, float cosCap2, float cap2, float cosDistance)
{
    // Precompute constants
    float radius1 = acosFastPositive(cosCap1); // First caps radius (arc length in radians)
    float radius2 = cap2; // Second caps radius (in radians)
    float dist = acosFast(cosDistance); // Distance between caps (radians between centers of caps)

    // Conditional expressions instead of if-else
    float check1 = min(radius1, radius2) <= max(radius1, radius2) - dist;
    float check2 = radius1 + radius2 <= dist;

    // Ternary operator to replace if-else
    float result = check1 ? (1.0 - max(cosCap1, cosCap2)) : (check2 ? 0.0 : 1.0 - max(cosCap1, cosCap2));

    float delta = abs(radius1 - radius2);
    float x = 1.0 - saturate((dist - delta) / max(radius1 + radius2 - delta, FLT_EPS));

    // simplified smoothstep()
    float area = sq(x) * (-2.0 * x + 3.0);

    // Multiply by (1.0 - max(cosCap1, cosCap2)) only once
    return area * result;
}

float DirectionalOcclusionSphere(float3 rayPosition, float3 spherePosition, float sphereRadius, float4 coneProperties)
{
    float3 occluderPosition = spherePosition.xyz - rayPosition;
    float occluderLength2 = dot(occluderPosition, occluderPosition);
    float3 occluderDir = occluderPosition * rsqrt(occluderLength2);

    float cosPhi = dot(occluderDir, coneProperties.xyz);
    // sq(sphere.w) should be a uniform --> capsuleRadius^2
    float cosTheta = sqrt(occluderLength2 / (sq(sphereRadius) + occluderLength2));
    float cosCone = cos(coneProperties.w);

    return 1.0 - SphericalCapsIntersectionAreaFast(cosTheta, cosCone, coneProperties.w, cosPhi) / (1.0 - cosCone);
}

float DirectionalOcclusionCapsule(float3 rayPosition, float3 capsuleA, float3 capsuleB, float capsuleRadius, float4 coneProperties)
{
    float3 Ld = capsuleB - capsuleA;
    float3 L0 = capsuleA - rayPosition;
    float a = dot(coneProperties.xyz, Ld);
    float t = saturate(dot(L0, a * coneProperties.xyz - Ld) / (dot(Ld, Ld) - a * a));
    float3 positionToRay = capsuleA + t * Ld;

    return DirectionalOcclusionSphere(rayPosition, positionToRay, capsuleRadius, coneProperties);
}

float CalcCapsuleShadowByIndexV2(float3 ro, float4 cone, uint s, uint e, float intensity, float intensity1)
{
    float shadow = 1.0;
    for (uint i = s; i < e; ++i)
    {
        Capsule c = _CapsuleData[i];
        shadow*= DirectionalOcclusionCapsule(ro, c.a, c.b, c.radius, cone);
    }
    shadow = saturate(lerp(1, shadow, intensity));
    return saturate(lerp(1, shadow, intensity1));
}
float CalcSphereOcclusion2(in float3 pos, in float3 nor, in float4 sph)
{
    float3 di = sph.xyz - pos;
    float l = length(di);
    float rad = sph.w;
    float nl = max(0.0, dot(nor, di / l));
    return (1.0 - nl * rad * rad / (l * l ));

    // float3  di = (sph.xyz - pos) / sph.w;
    //    float sqL = dot(di, di);
    //    return clamp(1.0 - dot(nor, di) * rsqrt(sqL * sqL * sqL), 0.1, 1);
}

float CalcSphereOcclusion2_( in float3 pos, in float3 nor, in float4 sph )
{
    float3  di = sph.xyz - pos;
    float l  = length(di);
    float nl = dot(nor,di/l) * _CapsuleShadowData[0].ConeAngle;
    float h  = l/sph.w;
    float h2 = h*h;
    float k2 = 1.0 - h2*nl*nl;

    // above/below horizon
    // EXACT: Quilez - https://iquilezles.org/articles/sphereao
    float res = max(0.0,nl)/h2;
    
    // intersecting horizon 
    if( k2 > 0.001 ) 
    {
        #if 0
        // EXACT : Lagarde/de Rousiers - https://seblagarde.files.wordpress.com/2015/07/course_notes_moving_frostbite_to_pbr_v32.pdf
        res = nl*acos(-nl*sqrt( (h2-1.0)/(1.0-nl*nl) )) - sqrt(k2*(h2-1.0));
        res = res/h2 + atan( sqrt(k2/(h2-1.0)));
        res /= 3.141593;
        #else
        // APPROXIMATED : Quilez - https://iquilezles.org/articles/sphereao
        res = (nl*h+1.0)/h2;
        res = 0.33*res*res;
        #endif
    }

    return 1 -res;
}

float CalcCapsuleOcclusionByIndexV2(float3 p, float3 n, uint s, uint e, float intensity, float intensity1)
{
    float ao = 1.0;
    for (uint i = s; i < e; ++i)
    {
        Capsule capsule = _CapsuleData[i];
        float3 ba = capsule.b - capsule.a;
        float3 pa = p - capsule.a;
        float l = dot(ba, ba);
        // p 在 ba 上投影长度与 ba 长度比值
        float t = /* abs(l) < 1e-8f ? 0.0 : */ saturate(dot(pa, ba) / l);
        float3 positionToRay = capsule.a + t * ba;
        ao *= CalcSphereOcclusion2_(p, n, float4(positionToRay, capsule.radius));
    }
    ao = saturate(lerp(1, ao, intensity));
    return saturate(lerp(1, ao, intensity1));
}

bool IsInBounds(float3 p, float3 c, float b, out float intensity)
{
    float3 diff = abs(p - c);
    if (diff.x < b && diff.y < b && diff.z < b)
    {
        intensity = saturate(dot(b, b) / dot(diff.xz, diff.xz));
        return true;
    }
    intensity = 0;
    return false;
}

float4 GetConeProperties(float3 lightDir, float coneAngle)
{
    float4 cone = float4(lightDir.xyz, radians(coneAngle) * 0.5);
    return cone;
}

void CalcCapsuleShadow(float3 worldPos, float3 worldNormal, out  float shadow, out float occlusion)
{
    shadow = 1.0;
    occlusion = 1.0;
    #if _MultiCapsule_Shadow_AO_ON || _MultiCapsule_AO_ON
    for (uint i = 0; i < _CharactersCount; ++i)
    {
        float intensity = 1;
        Character c = _CharacterData[i];
        if (IsInBounds(worldPos, c.position, c.radius * 3, intensity))
        {

            #ifdef _MultiCapsule_Shadow_AO_ON
                uint mask = c.capsuleAOSettingMask;
                uint areaCnt = countbits(mask);
                for (uint a = 0; a < areaCnt; ++a)
                {
                    uint index = firstbitlow(mask);
                    mask ^= 1 << index;
                    CapsuleShadowArea area = _CapsuleShadowData[index];
                    for (uint b = area.lightStartID; b < area.lightEndID; ++b)
                    {
                        float4 lightDir = _CapsuleLightsDir[b];
                        //float4 cone = GetConeProperties(lightDir.xyz, area.ConeAngle);
                        float4 cone = GetConeProperties(normalize(lightDir.xyz - worldPos), area.ConeAngle);
                        float tempIntensity = intensity / smoothstep(0.1, 2, lightDir.w);
                        float tempShadow = CalcCapsuleShadowByIndex(worldPos, cone.xyz, c.capsuleStartID, c.capsuleEndID, area.ShadowSharpness,  tempIntensity, area.ShadowIntensity);
                        // float tempShadow = CalcCapsuleShadowByIndexV2(worldPos, cone, c.capsuleStartID, c.capsuleEndID, tempIntensity, area.ShadowIntensity);
                        shadow = min(shadow, tempShadow);
                    }
                }
               // occlusion *= CalcCapsuleOcclusionByIndex(worldPos, worldNormal, c.capsuleStartID, c.capsuleEndID, intensity, _CapsuleShadowData[0].AmbientIntensity);
                occlusion *= CalcCapsuleOcclusionByIndexV2(worldPos, worldNormal, c.capsuleStartID, c.capsuleEndID, intensity, _CapsuleShadowData[0].AmbientIntensity);
            #elif defined(_MultiCapsule_AO_ON)
                occlusion *= CalcCapsuleOcclusionByIndex(worldPos, worldNormal, c.capsuleStartID, c.capsuleEndID, intensity, _CapsuleShadowData[0].AmbientIntensity);
            #endif
        }
    }
    #endif
    
}
#endif