Refactor the deep tile matmul config and skip the single-iteration loop generation

[zhczhong]

Track: #288

1. The extra memref.copy is caused by write-after-write conflict because the canonicalization pass eliminates the single-iteration loop but preserves the extracted slice and insert slice. So here skip generating them when they are single-iteration loop.
2. Refactor the config. Infer dimType according to the contractionOpInterface.
3. introduce padding cost, which minimizes the cost on padding and use divisible block if possible

[yifeizh2]

So after the change here, innermost Kblock will only be one of 16/32/64 if allowIndivisibleInnerblock is true?

[zhczhong]

Yes, your understanding is correct

[yifeizh2]

Can you explicitly specify mlir::utils::IteratorType here?

[zhczhong]

add mlir namespace here

[ciyongch]

Overall LGTM, do you think we should add a single simple case to cover the case of skipping single-iteration loop generation?

[zhczhong]

Overall LGTM, do you think we should add a single simple case to cover the case of skipping single-iteration loop generation?

This test could cover this case. The file check will check two scf.forall before this PR but the single-iteration scf.forall will be skipped now.

graph-compiler/test/mlir/test/gc/Transforms/deepTileContractionNamedOp.mlir

Line 127 in 613a333

func.func @matmul_2Dx4D_bf16_with_dlti(%arg0: tensor<4096x4096xbf16>, %arg1: tensor<128x128x16x32x2xbf16>) -> tensor<4096x4096xbf16> {

[ciyongch]

Please help to rebase the code base, then we can merge it.

[zhczhong]

Please help to rebase the code base, then we can merge it.

The code has been rebased now