PDL patterns

[giuseros]

Hi all,
I am trying to use the PDL patterns to double tile a generic matmul-like operation:

def add_matmul_schedule(module):
  dimM, dimN, dimK = [0, 0], [1, 1], [0, 1]
  isa_linalg_matmul = match_op_with_dynamic_or_static_sizes(
      module,
      equivalent_op_name='linalg.matmul',
      dynamic_spec_list=['d', 'd', 'd'],
      op_dim_spec_list=[dimM, dimN, dimK])
  
  with InsertionPoint(module.body):
    sequence = transform.SequenceOp()
    with ir.InsertionPoint(sequence.body.blocks[0]):
      matched = transform.MatchOp(isa_linalg_matmul)
      tiled = transform.TileOp(matched, sizes=[128, 128, 128], interchange=[0,2,1], pad=False)
      tiled2 = transform.TileOp(tiled, sizes=[8,8,1], interchange=[0,1,2], pad=True, pack_paddings=[1,1,0], hoist_paddings=[4,3,0], transpose_paddings=[[0,1], [0,1], [0,1]])

The second TileOp hits this assertion:

python3: /home/giuseppe/gh_llvm_project/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp:220: mlir::LogicalResult padOperandToSmallestStaticBoundingBox(mlir::OpBuilder &, linalg::LinalgOp, mlir::OpOperand *, const mlir::linalg::PaddingValueComputationFunction &, const mlir::linalg::PaddingNoFoldComputationFunction &, mlir::Value &): Assertion `staticSizes.size() == shape.size() && "expect the dynamic and static ranks to match"' failed.

While investigating, I was wondering if I am doing the right thing, or something is wrong with my code. @nicolasvasilache is this the way it's supposed to be used?

Thanks,
Giuseppe

[giuseros]

Ok, so I think that the problem was here:

// Handle the sliced out types in a conservative fashion: all dimensions
  // become dynamic and a later canonicalization is expected to recover static
  // types.
  // TODO: should we relax this and use something less strict?

So basically, if I do:

tiled = transform.TileOp(matched, sizes=[128, 128, 128], interchange=[0,2,1], pad=False)
 tiled2 = transform.TileOp(tiled, sizes=[8,8,1], interchange=[0,1,2], pad=False)
padded = transform.TileOp(tiled2, [],  pack_paddings=[1,1,0], hoist_paddings=[4,3,0], transpose_paddings=[[0,1], [0,1], [0,1]]))

Everything seems to work. @nicolasvasilache , naive question: what about having a transform.PadOp operation? Any drawbacks?