Support for additional Affine dialect operations in Affine loop lowering

[guosran]

Description

Currently in PR #173 , the affine-to-neura lowering pass supports the core loop-based operations (affine.for, affine.load, affine.store, affine.apply) but does not support some other Affine dialect operations.

Currently Supported

• affine.for - Loop structures with constant bounds
• affine.load / affine.store - Memory access operations
• affine.apply - Affine expression computations

Currently Unsupported

Examples include (but not limited to):

• affine.if - Conditional operations
• affine.parallel - Parallel loops
• affine.min / affine.max - Min/max expressions
• affine.for with dynamic bounds - Loops with runtime-dependent bounds
• Other advanced affine constructs

Current Behavior

When encountering unsupported operations, the pass fails with legalization errors. For example, with affine.if:

error: failed to legalize operation 'affine.if' that was explicitly marked illegal

The affine operations remain unchanged in the IR, and the pass reports a failure.

A test file test/Conversion/AffineToNeura/unsupported-affine-if.mlir demonstrates one such case.

Example of Unsupported Case

func.func @affine_if_example(%arg0: memref<10xf32>, %N: index) {
  affine.for %i = 0 to 10 {
    affine.if affine_set<(d0) : (d0 - 5 >= 0)>(%i) {
      %val = affine.load %arg0[%i] : memref<10xf32>
    }
  }
  return
}

Reason for Current Limitation

1. The Neura dataflow model focuses on loop-based computation with static control flow
2. The current implementation targets the most common loop kernels for CGRA acceleration
3. Some constructs (like conditionals or dynamic bounds) would require:
   • Predicated operations in the dataflow graph
   • Additional control flow handling in hardware
   • Runtime reconfiguration support

[tancheng]

Thanks for filing the issue @guosran. Most supports needed from HW side are already there. e.g., dynamic bounds can be represented as a computation/producer towards the loop_control operation. What we need is actually a proper way to lower this affine.for(variable loop bound) into a sequence of existing neura operations.

But we indeed lack of loop_control HW RTL module, where you might be able to help implementing it in https://github.com/tancheng/VectorCGRA.

Anyways, the advanced affine related operations' support can have lower priority for our current work.