Which operations would be ideal in fragment shader for multisampled depth textures? I mean, for RGBA textures, we could just take average of color values came from texelFetch().
What should be ideal shader code for multisampled depth texture?
Multisample depth textures are a Shader Model 4.1 (DX 10.1) feature first and foremost (multisample color textures are DX 10.0). OpenGL does not make this clear, but not all GL3 class hardware will support them. That said, since multisample textures are a GL 3.2 feature, this issue is largely moot in the OpenGL world; something that might come up once in a blue moon.
In any event, there is no difference between a multisample depth texture and color texture assuming your hardware supports the former. Even if the depth texture is an integer format, when you sample it using texelFetch (...) on a sampler2DMS you get a single-precision 4-component floating-point vector of the form: vec4 (depth.r, depth.r, depth.r, 1.0).
You can average the texels together if you want for multisample depth resolve, but the difference in depth between all of the samples can also be useful for quickly finding edges in your rendered scene to implement things like bilateral filtering.