Handle memref.copy/subview ops & Add compiler pass pipeline

[ShangkunLi]

In this pr:

1. Add two passes to handle the memref.copy and memref.subview ops
   For memref.subview ops, we transform such ops into affine.load/store directly, like:

// With memref.subview op
func.func @example(%arg0: memref<10x20xf32>) {
  %subview = memref.subview %arg0[2, 3][4, 5][1, 1] 
      : memref<10x20xf32> to memref<4x5xf32, strided<[20, 1], offset: 43>>
  
  affine.for %i = 0 to 4 {
    affine.for %j = 0 to 5 {
      %val = affine.load %subview[%i, %j] : memref<4x5xf32, strided<[20, 1], offset: 43>>
      // use %val
    }
  }
  return
}

// memref.subview folded
func.func @example(%arg0: memref<10x20xf32>) {
  affine.for %i = 0 to 4 {
    affine.for %j = 0 to 5 {
      // Index adjustment: [%i, %j] + offset [2, 3] = [%i+2, %j+3]
      %val = affine.load %arg0[%i + 2, %j + 3] : memref<10x20xf32>
      // use %val
    }
  }
  return
}

For memref.copy ops, we transform it to affine.for with affine.load/store, like:

// With memref.copy
func.func @copy_example(%src: memref<8x8xf32>, %dst: memref<8x8xf32>) {
  memref.copy %src, %dst : memref<8x8xf32> to memref<8x8xf32>
  return
}

// After transformation
func.func @copy_example(%src: memref<8x8xf32>, %dst: memref<8x8xf32>) {
  affine.for %i = 0 to 8 {
    affine.for %j = 0 to 8 {
      %val = affine.load %src[%i, %j] : memref<8x8xf32>
      affine.store %val, %dst[%i, %j] : memref<8x8xf32>
    }
  }
  return
}

2. Add two compiler pass pipeline:

• --tosa-to-affine-conversion: lowers from tosa dialect to affine dialect
• --taskflow-conversion: converts affine dialects to taskflow.task + neura.kernel

[tancheng]

Plz elaborate on "handle the memref.copy and memref.subview ops". How are they handled?

--taskflow-conversion: converts affine dialects to taskflow.task + neura.kernel

I thought at least two passes/stages needed for affine -> taskflow -> neura, why single pass can perform this?

[ShangkunLi]

Plz elaborate on "handle the memref.copy and memref.subview ops". How are they handled?

--taskflow-conversion: converts affine dialects to taskflow.task + neura.kernel

I thought at least two passes/stages needed for affine -> taskflow -> neura, why single pass can perform this?

I have updated the description of how we handle memref.copy and memref.subview in the pr summary.

As for --taskflow-conversion, it is a pass pipeline that integrates the following 4 passes:

// This pass pipeline can convert affine dialect into taskflow dialect with neura.kernel op.
void mlir::taskflow::registerTaskflowConversionPassPipeline() {
  PassPipelineRegistration<>(
      "taskflow-conversion",
      "Converts affine dialects to taskflow dialect with neura.kernel ops.",
      [](OpPassManager &pm) {
        pm.addPass(mlir::createConvertAffineToTaskflowPass());
        pm.addPass(mlir::taskflow::createConstructHyperblockFromTaskPass());
        pm.addPass(mlir::taskflow::createClassifyCountersPass());
        pm.addPass(mlir::createConvertTaskflowToNeuraPass());
      });
}

[tancheng]

affine.load/store will later be lowered to what?

[ShangkunLi]

affine.load/store will later be lowered to what?

They will be handled by the neura lowering process. They can be either lowered to neura.load_index/stroe_idnex, or lowered to llvm.gep + llvm.load/store.