Weird thing in converting device depth to linear depth.

Below is the functions I used to convert device depth to linear depth. I derived this code from the projection matrix.

// Projection._34 = zn*zf/(zn-zf)

// Projection._33 = zf/(zn-zf)

//float A = Projection._34;

//float B = Projection._33;

float A = ProjectionParams.z/zf;

float B = ProjectionParams.w;

float LInearZ = A/(DeviceZ + B);

It did not work. when ZNear value is not small, resulting linear depth was smaller than the real linear depth.

But when I change the code like below, It worked.

// Projection._34 = zn*zf/(zn-zf)

// Projection._33 = zf/(zn-zf)

//float A = Projection._34/zf;

//float B = Projection._33;

float A = ProjectionParams.z;

float B = ProjectionParams.w;

float LInearZ = A/(DeviceZ + B);

The difference is deviding by 'zf' is done in CPU code. It is exactly same as the crytek's GetLinearDepth function. I still don't know what is the cause.