[NVPTX] Plans to support s16x2/u16x2 instructions for sm_90

ThomasRaoux · September 1, 2023, 11:10pm

On hopper (sm_90) some instructions (add, min, max) support s16x2/u16x2 type. Is this something someone is planning to add in the near future?

@Artem-B do you have a plan for those?

ThomasRaoux · September 6, 2023, 2:15am

I sent a PR to add support:

github.com/llvm/llvm-project

[NVPTX] Make i16x2 a native type and add supported vec instructions

llvm:main ← ThomasRaoux:nvptx_i16x2_2

opened 02:13AM - 06 Sep 23 UTC

ThomasRaoux

+985 -410

On sm_90 some instructions now support i16x2 which allows hardware to execute mo…re efficiently add, min and max instructions. In order to support that we need to make i16x2 a native type in the backend. This does the necessary changes to make i16x2 a native type and adds support for the instructions natively supporting i16x2. This caused a negative test in nvptx slp to start passing. Changed the test to a positive one as the IR is correctly vectorized.

Any help reviewing it would be great.