LLVM ERROR: mma16816 data type not supported

[mobicham]

The latest Triton build (3.1.0) throws the following error when using bitpacked data inside a loop with tl.dot:

LLVM ERROR: mma16816 data type not supported

This error happens on Ampere and Hopper, but not on older gpus like the Titan RTX/2080 Ti.

The bitpacked data is read with indices in the form offs_k[:, None] // num_elements, something like [0,0,0...1,1,1...64,64,64].

I have faced this error in the previous build and I found that replacing for k in range(0, total_blocks_k, 1): with for k in tl.range(0, total_blocks_k, 1, num_stages=1): solved the issue, but this trick no longer works with 3.1.0.

Here's a full-script to reproduce it.
https://gist.github.com/mobicham/f9eba3c07f7e497ae622194a9c5e4822