Last Week on DirectX Shader Compiler (2017-09-01)

Lots of SPIR-V activity, bug fixes, and more float16 support, as has been the case recently.

Some highlights are improved front-end constant evaluation (in particular in some trivial vector cases), and a good report on a performance regression with respect to the fxc compiler toolchain, which was addressed here.

I wanted to take the opportunity the fact that we are definitely interested in seeing cases where the new toolchain isn't producing code that is on par with the fxc one. Real end-to-end results are best, because some seeming codegen improvements of yesterday may not apply or be actively counterproductive today. If you're reporting one of these, it's also very helpful to indicate whether you see this across the board or on a specific configuration.

Labor Day is coming, so if you celebrate that, enjoy the long weekend!


PS: if you have a regularly-sized weekend, please go ahead and enjoy that, too.