Talked about this on the editors meeting. We'd want to reorganize the pipeline descriptors according to the following guidelines:

• a descriptor has a field per stage (be it programmable or fixed-point stage)
• each field is a "stage descriptor" (e.g. output targets go into the "output merger", index related things go first, primitive topology is between the vertex shader and rasterization, etc)
• fields are ordered according to the real order, in which they are specified on the GPUs (input assembly -> vertex -> primitive assembly -> rasterization -> fragment -> output merger)

The benefit is - consistency in what's inside a sub-dictionary versus just a field on the top-level descriptor, and clear understanding on when things happen on the GPU.