Have you tried with different light directions? The dark areas simply look like they are dark because they are facing away from the parallel light you have hard-coded in your shader (ie. the dark sections are always on the -z axis in your image).
It may also be worth outputing the normal to the frame buffer so you can visually see any issues with the interpolation.
eg: change your color computation (in the pixel shader to) to:
color = normal * 0.5f + 0.5f;
You can also change your light vector so that it's always directly above the terrain (float3 toLight = float3( 0.0f, 1.0f, 0.0f );) which should give you a more even lighting across the terrain, again helping to see any issue with the normals. With the off-center light angle, it makes it difficult to say what is wrong really sorry But, certainly, your shader code looks fine. If it isn't the light direction confusing you, then it maybe the normals themselves.
My first guess is just that the light is at an angle really ;)
Setting the light to 0,1,0 doesn't help, the problem persists. Rendering the normals into the frame buffer shows the same problem.
It could also be that I'm misunderstanding what you're complaining about.
Sometimes people build their terrain such that the vertices look like:
Whereas you can avoid some artifacts on terrain lighting if you structure your vertices like:
I am building the mesh the way shown in the first picture, will try the other way that thanks.
From your shader code, it looks like you are doing this correctly (although you could remove the normalize call on the normal vector in the VS since you renormalize in the PS after rasterization). My guess is that your terrain is defining three vertices for each triangle face, rather than one vertex at each grid point. You can verify this by checking the number of vertices you are passing in with your draw call, or you can also check this with PIX/Graphics Debugger to see how many primitives are generated from how many input vertices.
So, I moved the rendering into indexed rendering, so there is only one normal per vertex and still get same result.
That may or may not mean that there is exactly one vertex normal being used at each grid point. How many vertices are in your vertex buffer, how many indices in your index buffer, and how many primitives are you drawing? Compare that with your grid size and make sure that you only have N+1 x N+1 vertices for a grid of size N x N.
For a 16x16 grid, there are 256 vertices, index buffer holds 1350 indices