Issues in converting tt.get_program_id` and `tt.get_num_programs` to linalg

[fhossein-quic]

I'm trying to use triton-linalg tool to lower ttir to linalg, but realized that after conversion, tt.get_program_id and tt.get_num_programs still remain in the tt dialect and are not lowered to linalg. The add_kernel example shared in the repo's README shows this issue as well:

module {
  func.func @add_kernel(%arg0: i64, %arg1: i64, %arg2: i64, %arg3: i32) {
    %c1024 = arith.constant 1024 : index
    %c1024_i32 = arith.constant 1024 : i32
    %cst = arith.constant 0.000000e+00 : f32
    %0 = tt.get_program_id x : i32 

I'd also like to know how ops in Auxilary and LinalgExt Dialect can lowered to LLVM IR? What dialects they can be transformed to?

CC- @sethbrin and @zhangqirui

[sethbrin]

• tt.get_program_id x is target dependent, so we didn't handle these ops in triton-linalg. It can be convert to gpu.block_id x under the GPU or omp dialect for CPU.
• auxilary can be directly lower to llvm dialect as it just bind llvm.ptr to memref desc, while linalgext can be lower to vector or other target dependent dialect. you may refer to the codegen dialect overview of llvm community
• We are hoping to jointly build the triton + linalg ecosystem with the industry. If you are interested, welcome to make pull requests and discussions.