[AscendNPU-IR][A5] surpport for cumsum interleave and deinterleave

[james1314520]

support for cumsum interleave and deinterleave

[github-actions]

👋 Hi! Thank you for contributing to the TileLang project.

Please remember to run bash format.sh in the root directory of the project to ensure your changes are properly linted and formatted. This will help ensure your contribution passes the format check.

We appreciate you taking this step! Our team will review your contribution, and we look forward to your awesome work!

🚀

[gemini-code-assist]

Code Review

This pull request migrates the Cumsum, Interleave, and Deinterleave operations from the hivm dialect to the hfusion dialect in the NPU codegen and adds associated unit tests. However, the implementation of VinterleaveCodegen and VdeinterleaveCodegen contains critical issues: they use memref types where tensor types are expected and incorrectly overwrite destination buffers, leading to potential data loss when dealing with subregions. Additionally, the DeinterleaveOp is missing the necessary channel_nums attribute.

[gemini-code-assist]

The current implementation of VinterleaveCodegen has several issues:

1. Type Mismatch: It uses GenSubviewFromRegion, which returns a memref, but hfusion operations and the SetVarValue mapping system in this codegen are designed for tensors. You should use GenExtractSliceFromRegion instead.
2. Partial Update Bug: By calling SetVarValue directly with the result of the interleave operation, you are overwriting the entire destination buffer with a tensor that only represents the subregion (if dst_range is not the full buffer). This will cause data loss for the rest of the buffer. You must use ReshapeCastAndInsertSlice to correctly update the destination tensor.
3. Source Types: Note that the srcs vector (populated in the context above) also contains memrefs from GenSubviewFromRegion. These should ideally be converted to tensors as well for compatibility with hfusion ops.

  mlir::Value dst_tensor = GenExtractSliceFromRegion(npuirop.dst, npuirop.dst_range);

  auto interleaveOp = builder.create<mlir::hfusion::InterleaveOp>(
    builder.getUnknownLoc(),
    dst_tensor.getType(),
    srcs_vr);
  mlir::Value result = ReshapeCastAndInsertSlice(interleaveOp->getResult(0), GetVarValue(npuirop.dst), npuirop.dst_range);
  SetVarValue(npuirop.dst, result);

[baberrehman]

These are important to address.

[james1314520]

done

[gemini-code-assist]

The VdeinterleaveCodegen implementation contains critical issues similar to VinterleaveCodegen:

1. Data Loss: SetVarValue overwrites the full destination buffer with a subregion-sized tensor. Use ReshapeCastAndInsertSlice to perform a proper tensor.insert_slice into the destination.
2. Type Mismatch: Use GenExtractSliceFromRegion to obtain a tensor instead of a memref for the destination type.
3. Missing Attribute: The DeinterleaveOp appears to be missing the total number of channels (npuirop.channel_nums). Without this, the operation cannot determine the correct stride for deinterleaving.
4. Namespace Consistency: Use mlir::hfusion::DeinterleaveOp to match the style used in other parts of the file.

    mlir::Value dst_tensor = GenExtractSliceFromRegion(npuirop.dsts[i], npuirop.dsts_range[i]);

    int64_t current_channel_idx = i;
    auto channelIdxAttr = builder.getI64IntegerAttr(current_channel_idx);
    auto channelNumsAttr = builder.getI64IntegerAttr(npuirop.channel_nums);

    auto deinterleaveOp = builder.create<mlir::hfusion::DeinterleaveOp>(
        builder.getUnknownLoc(),
        dst_tensor.getType(),
        src,
        channelIdxAttr,
        channelNumsAttr
    );
    mlir::Value result = ReshapeCastAndInsertSlice(deinterleaveOp->getResult(0), GetVarValue(npuirop.dsts[i]), npuirop.dsts_range[i]);
    SetVarValue(npuirop.dsts[i], result);

[baberrehman]

Also these