Originally posted by difron
View Post
1. IR conversion is considered a low cost operation (doing it many times doesn't consume too much resources compared to other operations)
2. Using a common IR format at some point in mesa allows developers from different drivers to cooperate on feature development and feature optimization, for cases where features can be developed outside of the kernel driver.
Basically, if the IR is good, then you can write common implementations that work across drivers, in a way that each driver can then figure out how to consume.
* I remember also some discussion about how one of the IRs used in Mesa (I think in Gallium) is not simple to convert to/from NIR as it is not SSA, while NIR is SSA. That might have been Ilia or Marek who mentioned that.
Comment