As long as your input vertices' normals are unit vectors, and you limit yourself to certain types of transformation operations in your vertex shader, then you won't notice any difference if you don't normalize in the VS.

Assuming all thevertices on your object are transformed with the exact same matrix, then:

- translation won't have any effect on normals, so you're good there.

- rotation won't change the length of normals, so you're good there.

- (uniform) scaling will change the length of your normals. That should still be ok if you're normalizing in the PS.

Non-uniform scaling (different scale factors in x, y and z), however, will mess things up, as suddenly equal length normals will have different lengths. But non-uniform scaling is going to produce messed up normals anyway, even if you re-normalize in the VS.

think of it this way:

a normal is (by definition) a unit vector.

an operation that changes the length causes a normal to become "just a vector in the desired direction with some magnitude that is not 1" (as i recall).

therefore, any operation that changes the length of the normal will require you to re-normalize before you use the normal again.

so when you perform an operation on a normal, ask yourself, "does this change the length?" if so, remember that its now an "un-normalized" normal, and you must "re-normalize" it before using it again.

keeping track of "un-normalized" normals is all part of crossing your i's and dotting your t's when doing 3d graphics.