I have a lot of shaders compiling with ShaderModel vs_1_1 and ps_2_0 (using DirectX9). If I know that the hardware supports a higher shader model (probably most will support 3.0) should I compile with that higher shader model version? Is it more performant? Or is the lesser model version more effecient?
Easiest way to check is to use fxc with different profiles and check instruction count, in higher versions you can have new intrinsics that can reduce total instruction count.
I remember struggling to fit within the 64 limitation in ps_2_0, whereas compiling in ps_3_0 gave me 63 instructions.
I would (personal opinion here), use shader model 3 tho, even phones support the feature set nowadays.
As per performance indicator, you have different things to take into account of course, since you need to balance between arithmetic/texture fetch/fill rate.
Of course the lower the better in general, but 1 less arithmetic will not make you system much faster if you're draw call bound.
For profiling, if you don't already use this feature yet, I strongly recommend to implement Queries , in your case TimeStamp is the most useful one. It allows you to measure shader execution time. Occlusion is also very useful for measure overdraw in the forward pass (fill rate).