Pyromuffin

#if __NV__
groupshared uint ldsBallot[32];
uint CoolBallot(bool value, uint waveId, uint laneId)
{
	ldsBallot[waveId] = 0;
	InterlockedOr(ldsBallot[waveId], value << laneId);

	return ldsBallot[waveId];
}
#else // AMD

Pyromuffin

[numthreads(64, 1, 1)]
void BallotExample(uint gtid : SV_GroupThreadID)
{
  // assuming GCN architecture with WAVE_SIZE == 64
  bool firstHalf = gtid < 32; 
  uint2 ballotResult = WaveActiveBallot(firstHalf);
  // ballotResult == 0x0000 0000 FFFF FFFF in ALL threads.
}

Pyromuffin

uint2 IntegerRotate(int2 right, int2 up, int2 offset, uint2 pos)
{
	static const int BIG_NUMBER_EXPONENT = 20;
	uint2 rotatedPos;

	rotatedPos.x = pos.x * right.x + pos.y * right.y;
	rotatedPos.y = pos.x * up.x + pos.y * up.y;
	rotatedPos += offset;
	rotatedPos = rotatedPos >> BIG_NUMBER_EXPONENT;
	return rotatedPos;

Pyromuffin

        Matrix4x4 mat;
        mat = Matrix4x4.Translate(new Vector3(centerPoint.x, centerPoint.y));
        mat *= Matrix4x4.Rotate(Quaternion.AngleAxis(degrees, Vector3.forward));
        mat *= Matrix4x4.Translate(new Vector3(-centerPoint.x, -centerPoint.y));
        
        IntegerBasis basis;
        basis.right =  new Vector2Int( Mathf.RoundToInt(mat.m00 * BIG_NUMBER), Mathf.RoundToInt(mat.m01 * BIG_NUMBER));
        basis.up =     new Vector2Int( Mathf.RoundToInt(mat.m10 * BIG_NUMBER), Mathf.RoundToInt(mat.m11 * BIG_NUMBER));
        basis.offset = new Vector2Int( Mathf.RoundToInt(mat.m03 * BIG_NUMBER), Mathf.RoundToInt(mat.m13 * BIG_NUMBER));

Pyromuffin

        Matrix4x4 mat;
        mat = Matrix4x4.Translate(new Vector3(centerPoint.x, centerPoint.y));
        mat *= Matrix4x4.Rotate(Quaternion.AngleAxis(degrees, Vector3.forward));
        mat *= Matrix4x4.Translate(new Vector3(-centerPoint.x, -centerPoint.y));
        
        Basis basis;
        basis.right = new Vector2(mat.m00,  mat.m01 );
        basis.up = new Vector2(mat.m10 , mat.m11);
        basis.offset = new Vector2(mat.m03, mat.m13);

Pyromuffin

Texture2D<float4> originalTex : register(t1);
RWTexture2D<float4> outHdrTex : register(u0);

float3 ApplyREC2084Curve(float3 L, float maxLuminance)
{
	float m1 = 2610.0 / 4096.0 / 4;
	float m2 = 2523.0 / 4096.0 * 128;
	float c1 = 3424.0 / 4096.0;
	float c2 = 2413.0 / 4096.0 * 32;
	float c3 = 2392.0 / 4096.0 * 32;

Pyromuffin

float3 ApplyREC2084Curve(float3 L, float maxLuminance)
{
	float m1 = 2610.0 / 4096.0 / 4;
	float m2 = 2523.0 / 4096.0 * 128;
	float c1 = 3424.0 / 4096.0;
	float c2 = 2413.0 / 4096.0 * 32;
	float c3 = 2392.0 / 4096.0 * 32;

	// L = FD / 10000, so if FD == 10000, then L = 1.
	// so to scale max luminance, we want to multiply by maxLuminance / 10000

Pyromuffin

float3 RemoveSRGBCurve(float3 x)
{
	// Approximately pow(x, 2.2)
	return x < 0.04045 ? x / 12.92 : pow((x + 0.055) / 1.055, 2.4);
}

Pyromuffin

float3 REC709toREC2020(float3 RGB709)
{
	static const float3x3 ConvMat =
	{
		0.627402, 0.329292, 0.043306,
		0.069095, 0.919544, 0.011360,
		0.016394, 0.088028, 0.895578
	};
	return mul(ConvMat, RGB709);
}