There's a lot of resources out there on HLSL you can find with a quick web search. I would strongly recommend to check out the documentation on the MSDN website for
HLSL and the
D3D9 effect system. There's many tutorials out there on how learning to program using HLSL too. I would also recommend checking out some of the XNA tutorials, because there is one for
normal mapping. XNA is C# and you don't directly use the D3D9 interfaces and functions, so it may not help much on the CPU side, but the HLSL code is something to check out. You may also want to check out XNA's
Per-Vertex and
Per-Pixel lighting tutorials (for the shader code and explanations).
A shader can be thought of as a material - it defines how geometry is processed on the GPU and how it gets drawn. And since you write the shader, you're able to dictate to the GPU how do perform these operations.
So given a piece of geometry you want to render, you have a shader that is used to render it. Before you tell the GPU to draw anything, you need to make sure all the parameters are set on the shader (world-view-projection matrices, your geometry's texture, etc). When you command the GPU to draw, the shader is then executed. For D3D9, a complete effect is a vertex shader and a pixel shader. Vertex shaders execute for each vertex in your geometry, and the pixel shader for each pixel (more or less). So "applying a normal map" is basically setting an effect parameter, which happens to be the normal map texture. You should look at resources out there (again, google search) on how to load and use effects. There's a lot out there. You should get well acquainted with those shader topics before you tackle normal mapping, since that can be considered an advanced technique.
As far as understanding what normal mapping and specular mapping is, the concept is rather simple, but the implementation is not. As I mentioned earlier, there's per-vertex and per-pixel lighting techniques. A normal map is a texture of normal vectors that are used for per-pixel lighting. Rather than using normal vectors from the geometry (which for a low-poly model can be a small number), you use the texture. So you have more information to work with, and as a result you can produce surfaces that look rough or bumpy *without* requiring to render more geometry. E.g. in 3DS you can generate a normal map from a 2 million triangle character, create a low-poly character with 2k triangles, and render them to be almost the same. Specular maps are similar where they encode values that impact specular calculations along the surface of a model. E.g. make the surface look shinier or duller.
Since they both have to deal with lighting techniques, I would get a good grasp of how to do per-pixel lighting before advancing to normal maps.
