When is Z axis flipped? (OpenGL)

Question

To improve my understanding of OpenGL and 3D I am trying to implement a simple rendering pipeline running only on the CPU. Where it makes sense, I'm trying to use the same standards as OpenGL. For example, I'm using a right handed coordinate system like OpenGL. As it often happens when learning something new, I've now run into a question I cannot find an answer to. If something in the following doesn't make sense, please forgive and correct me.

In clip space (the coordinate you give to gl_Position in your vertex shader), negative Z is away from the user. But in the depth buffer, positive Z is away (if you're using the default depth settings).

One answer to this question suggests that the flip is due to the projection transformation. But I made a little experiment (based on WebGL/OpenGL ES) that suggests the opposite: In clip space, Z is pointing away from the user even if you don't use a projection transformation.

So at some point, after you've handed it your gl_Position, OpenGL flips the Z coordinate. When and how is that?

My guess is that it is in the viewport transform but I haven't been able to find documentation that supports this.

Answer 1

Clip-Space in OpenGL is already left handed (and always was). So, in modern GL there is no space explicitely using a right-handed system, and there is no "flipping" done at all. However, the eye space in old GL was often defined right-handed, and the "flipping" was done as part of the projection matrix, glFrustum() and gluPerpective() explicitely expect positive values for near and far to refer to the repsective negative z values. However, as you could use arbitrary matrices, you never were forced to define your eye or object spaces right-handed, it was just something like a default convention. And still a lot of people follow this convention with the programmable pipeline.

Answer 2

According to WebGL Beginner's Guide, the z-axis is inverted during the perspective division which is done after the projection transform.

Here is what it says:

The perspective division transforms the viewing frustum into a cube centered in the origin with minimum coordinates [-1,-1,-1] and maximum coordinates [1,1,1]. Also, the direction of the z-axis is inverted, as shown in the following figure: