Does the triton define the fast change dimension?

[chengjunlu]

I am investigating to integrate the SPIRV to triton as a new backend.

The Triton, PTX and SPIRV all define three parallel loop dimensions.

I want to know whether it is clear defined in Triton how the dimensions are mapped.

By reviewing the code it seems the mapping from Triton to PTX is:
Triton dim: 0, 1, 2
PTX dim: x, y, z

In NV/CUDA, the X dimension is the fast change dimension which means the thread are grouped along the X dim first.

Can I assume in the triton, dim 0 is the fast change dim in parallel?

[ptillet]

Exciting! Are you looking at the Triton-MLIR branch?

Yes, this is right! triton.program_id({0,1,2}) maps to blockIdx.{x,y,z}

[chengjunlu]

Thanks @ptillet

I am looking the Triton-MLIR.

I may have some basic questions about the Triton-MLIR. I will raise new discussion with the Triton-MLIR category.