There are significant differences between ps/vs 1.x versions, and also the step to ps/vs 2.0 meant a whole new way of shader programming. But can anyone confirm that ps/vs 2.0 is a true subset of ps/vs 3.0 or are there small differences? Also, do you expect future shader versions to be based on ps/vs 2.0 as well?
ps/vs 2.0 a subset of ps/vs 3.0 and future?
- Thread starter Nick
- Start date
In any pipelined microarchitecture adding dynamic branches increases the difficulty of the ALU's design considerably. In a single core superscalar CPU it's difficult enough, but in a GPU which runs four threads (the pixel quad) concurrently it would seem to be a nightmare to implement.
Wasn't it the ARM that takes care of branching in an interesting way? It doesn't change the IP but instead has a "condition code" that simply masks the result of instructions not to be executed due to the branch. This is only for short branches of course.
Wasn't it the ARM that takes care of branching in an interesting way? It doesn't change the IP but instead has a "condition code" that simply masks the result of instructions not to be executed due to the branch. This is only for short branches of course.
akira888 said:
Actually more threads can be better, at least you don't need to worry about branch prediction. Multi-threading hides the latency for you.
D3D pixel shaders use a "cheaper" kind of predication: compute two results and choose from them using a condition. Of course, real predication is still better, but it's almost useless if you have no branch misprediction penalty.
