Add basic conversion between ttir and linalg

[vwellsTT]

Goal: The end-to-end goal is to integrate a path to compile and execute specific ops or sets of ops on the CPU.

Context:

The entire task will be split into (tentatively) 7 PRs, as follows:

1. Hoist specific ops into isolated funcs in a separate module
2. Convert TTIR ops to linalg ops within the module of hoisted funcs
3. Build a pipeline to lower linalg to llvm from existing conversion passes
4. Translate LLVM Dialect into a dynamic library for packing into flatbuffer
5. Generate helper functions so that we can call all of our hoisted funcs with a common signature
6. Insert TTNN instructions to move operands to host before executing hoisted func, then back to device afterwards
7. Update ttir-to-ttnn and ttnn-to-flatbuffer pipelines to use new passes, generate dylibs, and embed them into output flatbuffers, and update update runtime to consume dylibs from flatbuffers

This PR represents the 2nd subtask above--it converts TTIR ops in the "cpu" module into their linalg equivalents. In this case, we only enable a few basic operations for now, in an attempt to get basic end-to-end conversions working before fleshing out full conversion. Definitely open to adding some other specific ops to conversion before merging, not sure which ops would make most sense.

Example

Input:

module  {
  func.func @add(%arg0: tensor<32x32xf32>, %arg1: tensor<32x32xf32>, %arg2: tensor<32x32xf32>) -> tensor<32x32xf32> {
    %1 = "ttir.add"(%arg0, %arg1, %arg2) <{operandSegmentSizes = array<i32: 2, 1>}> : (tensor<32x32xf32>, tensor<32x32xf32>, tensor<32x32xf32>) -> tensor<32x32xf32>
    return %1 : tensor<32x32xf32>
  }
}

Output:

module {
  func.func @add(%arg0: tensor<32x32xf32>, %arg1: tensor<32x32xf32>, %arg2: tensor<32x32xf32>) -> tensor<32x32xf32> {
    %0 = linalg.add ins(%arg0, %arg1 : tensor<32x32xf32>, tensor<32x32xf32>) outs(%arg2 : tensor<32x32xf32>) -> tensor<32x32xf32>
    return %0 : tensor<32x32xf32>
  }
}

[ctodTT]

Could be wrong, but I don't see any CMakeLists.txt changed here, do these new Conversion/ files get compiled a different way?

[vwellsTT]

Could be wrong, but I don't see any CMakeLists.txt changed here, do these new Conversion/ files get compiled a different way?

Let me double-check, there were quite a few interleaved commits I tried to manually split out, so wouldn't surprise me if I missed one. Thanks!

Update: yup, was missing several commits...hopefully fixed now

[nsmithtt]

Minor comments inline, otherwise looks great!

[nsmithtt]

Stray comment

[vwellsTT]

This logic is much more complicated than I originally expected, but linalg.broadcastOp asserts something like "input rank + len(dimsToBroadcastAlong) == output rank". We want to support torch-like broadcasting, where ranks with size 1 are implicitly broadcasted => we need to do a tensor.collapseOp in such cases, which requires hairy logic.

I would appreciate careful review/re-review of this logic, especially this function 🙂

[mtopalovicTT]

Use OpTrait::util::getBroadcastedShape util function to calculate broadcasted shape. It will save you all work you did here.
Also you don't need to check if operands are broadcastable since we already check this using traits for all eltwise operations (check verifyBroadcastable).

[vwellsTT]

Makes sense, I have switched to this func. Let me know if you think I should replace this if check with some sort of assert etc instead given checks must have occurred at TTIR level