Originally Posted by bridgman
Reply With Quote
Just guessing here, but probably a bit harder the first time since it will need to include and upload software to perform vertex grouping and rasterizing / scan converting -- those tasks are done in hardware on other GPUs. I imagine the Intel open source devs will be able to pick up most of the code from other drivers so it shouldn't be too bad.
The x86-ness probably won't make much difference, since the driver will just be compiling TGSI opcodes down to SIMD instructions anyways. It's "just another instruction set" for the developers. The actual x86 opcodes presumably would not be used in shader programs (other than for flow control) because the performance would drop to 1/16 or less.
I imagine the shader translation / compilation code code would be fairly similar to the existing Nouveau code, since Nouveau only has to deal with SIMD hardware while the r6xx/7xx code needs to deal with SIMD + superscalar hardware -- perversely enough what you call a vector engine on a normal CPU counts as a scalar engine on a GPU, since everyone takes SIMD for granted in GPUs.
I think LRB has dedicated texture blocks so that part shouldn't be much different from other GPUs. I haven't really looked much at LRB though -- and probably won't until I have time to catch up on all the things I want to look at with *our* products
it's over NINE THOUSAND!
