How to render offscreen on OpenGL?
I'll assume that creating a dummy window (you don't render to it; it's just there because the API requires you to make one) that you create your main context into is an acceptable implementation strategy.
Here are your options:
Pixel buffers
A pixel buffer, or pbuffer (which isn't a pixel buffer object), is first and foremost an OpenGL context. Basically, you create a window as normal, then pick a pixel format from wglChoosePixelFormatARB (pbuffer formats must be gotten from here). Then, you call wglCreatePbufferARB, giving it your window's HDC and the pixel buffer format you want to use. Oh, and a width/height; you can query the implementation's maximum width/heights.
The default framebuffer for pbuffer is not visible on the screen, and the max width/height is whatever the hardware wants to let you use. So you can render to it and use glReadPixels to read back from it.
You'll need to share you context with the given context if you have created objects in the window context. Otherwise, you can use the pbuffer context entirely separately. Just don't destroy the window context.
The advantage here is greater implementation support (though most drivers that don't support the alternatives are also old drivers for hardware that's no longer being supported. Or is Intel hardware).
The downsides are these. Pbuffers don't work with core OpenGL contexts. They may work for compatibility, but there is no way to give wglCreatePbufferARB information about OpenGL versions and profiles.
Framebuffer Objects
Framebuffer Objects are more "proper" offscreen rendertargets than pbuffers. FBOs are within a context, while pbuffers are about creating new contexts.
FBOs are just a container for images that you render to. The maximum dimensions that the implementation allows can be queried; you can assume it to be GL_MAX_VIEWPORT_DIMS (make sure an FBO is bound before checking this, as it changes based on whether an FBO is bound).
Since you're not sampling textures from these (you're just reading values back), you should use renderbuffers instead of textures. Their maximum size may be larger than those of textures.
The upside is the ease of use. Rather than have to deal with pixel formats and such, you just pick an appropriate image format for your glRenderbufferStorage call.
The only real downside is the narrower band of hardware that supports them. In general, anything that AMD or NVIDIA makes that they still support (right now, GeForce 6xxx or better [note the number of x's], and any Radeon HD card) will have access to ARB_framebuffer_object or OpenGL 3.0+ (where it's a core feature). Older drivers may only have EXT_framebuffer_object support (which has a few differences). Intel hardware is potluck; even if they claim 3.x or 4.x support, it may still fail due to driver bugs.
It all starts with glReadPixels, which you will use to transfer the pixels stored in a specific buffer on the GPU to the main memory (RAM). As you will notice in the documentation, there is no argument to choose which buffer. As is usual with OpenGL, the current buffer to read from is a state, which you can set with glReadBuffer.
So a very basic offscreen rendering method would be something like the following. I use c++ pseudo code so it will likely contain errors, but should make the general flow clear:
//Before swapping
std::vector<std::uint8_t> data(width*height*4);
glReadBuffer(GL_BACK);
glReadPixels(0,0,width,height,GL_BGRA,GL_UNSIGNED_BYTE,&data[0]);
This will read the current back buffer (usually the buffer you're drawing to). You should call this before swapping the buffers. Note that you can also perfectly read the back buffer with the above method, clear it and draw something totally different before swapping it. Technically you can also read the front buffer, but this is often discouraged as theoretically implementations were allowed to make some optimizations that might make your front buffer contain rubbish.
There are a few drawbacks with this. First of all, we don't really do offscreen rendering do we. We render to the screen buffers and read from those. We can emulate offscreen rendering by never swapping in the back buffer, but it doesn't feel right. Next to that, the front and back buffers are optimized to display pixels, not to read them back. That's where Framebuffer Objects come into play.
Essentially, an FBO lets you create a non-default framebuffer (like the FRONT and BACK buffers) that allow you to draw to a memory buffer instead of the screen buffers. In practice, you can either draw to a texture or to a renderbuffer. The first is optimal when you want to re-use the pixels in OpenGL itself as a texture (e.g. a naive "security camera" in a game), the latter if you just want to render/read-back. With this the code above would become something like this, again pseudo-code, so don't kill me if mistyped or forgot some statements.
//Somewhere at initialization
GLuint fbo, render_buf;
glGenFramebuffers(1,&fbo);
glGenRenderbuffers(1,&render_buf);
glBindRenderbuffer(render_buf);
glRenderbufferStorage(GL_RENDERBUFFER, GL_BGRA8, width, height);
glBindFramebuffer(GL_DRAW_FRAMEBUFFER,fbo);
glFramebufferRenderbuffer(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, render_buf);
//At deinit:
glDeleteFramebuffers(1,&fbo);
glDeleteRenderbuffers(1,&render_buf);
//Before drawing
glBindFramebuffer(GL_DRAW_FRAMEBUFFER,fbo);
//after drawing
std::vector<std::uint8_t> data(width*height*4);
glReadBuffer(GL_COLOR_ATTACHMENT0);
glReadPixels(0,0,width,height,GL_BGRA,GL_UNSIGNED_BYTE,&data[0]);
// Return to onscreen rendering:
glBindFramebuffer(GL_DRAW_FRAMEBUFFER,0);
This is a simple example, in reality you likely also want storage for the depth (and stencil) buffer. You also might want to render to texture, but I'll leave that as an exercise. In any case, you will now perform real offscreen rendering and it might work faster then reading the back buffer.
Finally, you can use pixel buffer objects to make read pixels asynchronous. The problem is that glReadPixels blocks until the pixel data is completely transfered, which may stall your CPU. With PBO's the implementation may return immediately as it controls the buffer anyway. It is only when you map the buffer that the pipeline will block. However, PBO's may be optimized to buffer the data solely on RAM, so this block could take a lot less time. The read pixels code would become something like this:
//Init:
GLuint pbo;
glGenBuffers(1,&pbo);
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
glBufferData(GL_PIXEL_PACK_BUFFER, width*height*4, NULL, GL_DYNAMIC_READ);
//Deinit:
glDeleteBuffers(1,&pbo);
//Reading:
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
glReadPixels(0,0,width,height,GL_BGRA,GL_UNSIGNED_BYTE,0); // 0 instead of a pointer, it is now an offset in the buffer.
//DO SOME OTHER STUFF (otherwise this is a waste of your time)
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo); //Might not be necessary...
pixel_data = glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
The part in caps is essential. If you just issue a glReadPixels to a PBO, followed by a glMapBuffer of that PBO, you gained nothing but a lot of code. Sure the glReadPixels might return immediately, but now the glMapBuffer will stall because it has to safely map the data from the read buffer to the PBO and to a block of memory in main RAM.
Please also note that I use GL_BGRA everywhere, this is because many graphics cards internally use this as the optimal rendering format (or the GL_BGR version without alpha). It should be the fastest format for pixel transfers like this. I'll try to find the nvidia article I read about this a few monts back.
When using OpenGL ES 2.0, GL_DRAW_FRAMEBUFFER might not be available, you should just use GL_FRAMEBUFFER in that case.