Телеграм канал «MadSharp: Unsafe»

Just found math equivalent optimizations for L1 SH Evaluation,
If somebody is using Bakery with SH/MonoSH lightmaps or Light Probe Volumes, here's the optimized SH sampling:

Also it's possible to get rid of pow() with the ifdef, but it's the approximation and quality will differ.

Share screenshots of the approximation results :)

float shEvaluateDiffuseL1Geomerics(float4 Vec, float3 n)
{
float L0 = Vec.x;
float3 L1 = Vec.yzw;

float lenSq = dot(L1, L1);
float invLen = rsqrt(lenSq);
float len = lenSq * invLen;

float q = mad(dot(L1, n) * invLen, 0.5f, 0.5f);

float r = len * rcp(L0);
float t = saturate(r);
float ps = 1.0f + r;
float as = (2.0f - r) * rcp(2.0f + r);

float qx = q * 0.9;
float q2 = qx * qx;
float powerResult = pow(q, ps);
#ifdef ULTRA_MOBILE_SAVE
powerResult = (q) * mad(saturate(t), qx * qx - 0.75, 1.0f);
#endif
return mad(2.0f * len, powerResult, L0 * as);
}

Been playing with 4-sample PCF shadows.
Probably the best results you can get away with 4 samples. OpenGL3.3 friendly.

Code in comments.