[TLERaw] Remove redundant copy && Support buffered_tensor with tle_raw.call

[lizhangyu258]

[TLERaw]

• Eliminate redundant copies caused when passing tensors as parameters in tle_raw.call().
• Support passing buffered_tensors as parameters with tle_raw.call_smem().
• Update mlir/05-topk.py by giving alloc initialization value.

[Bug]

• When passing parameters using tle_raw.call_smem(), you need to comment out this tle.passes.add_assign_local_pointers_encoding(pm) statement; otherwise, occasional compilation errors may occur.

[Fixed]

• There are no errors when testing based on the tle_topk_3.6 branch.

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

This PR updates the TLE Raw pipeline to (1) support passing ttg::MemDescType/buffered shared-memory tensors through tle_raw calls and (2) eliminate redundant shared-memory copies by rewriting specific scf.for loop argument patterns.

Changes:

• Extend TLERaw protocol flatten/pack logic to handle ttg::MemDescType and add a MemDescPattern to signatures.
• Add and wire a new tle-remove-redundant-copy MLIR pass (plus small SCF utility helpers) into the NVIDIA backend compilation pipeline.
• Add Python tle_raw.call_smem() and extend tle_gpu.alloc() with an init_value to support initialized shared-memory buffers; update tutorials accordingly.

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
third_party/tle/utils/lib/Protocol.cpp	Add MemDesc signature handling and allow packing memdesc returns in LLVMStructurePattern.
third_party/tle/utils/include/Protocol.h	Expose MemDescPattern and include TritonGPU dialect types for memdesc support.
third_party/tle/triton_tle.cc	Register the new raw pass wrapper add_tle_remove_redundant_copy.
third_party/tle/dialect/lib/Transforms/TleUtility.cpp	Add SCF helper utilities used by multiple transforms.
third_party/tle/dialect/lib/Transforms/RemoveRedundantCopy.cpp	New transform pass to rewrite loop-carried args to avoid redundant local copies.
third_party/tle/dialect/lib/Transforms/ConvertArgToMemDesc.cpp	Adjust arg-to-memdesc conversion behavior in single-loop/iter-arg cases and barrier insertion conditions.
third_party/tle/dialect/lib/Transforms/CMakeLists.txt	Build integration for the new transform sources.
third_party/tle/dialect/include/Transforms/TleUtility.h	Header for SCF helper utilities.
third_party/tle/dialect/include/Transforms/RemoveRedundantCopy.h	Header for new redundant-copy removal patterns.
third_party/tle/dialect/include/Transforms/Passes.td	Add tle-remove-redundant-copy pass definition.
third_party/nvidia/backend/compiler.py	Insert the new raw pass into the CUDA TTGIR pipeline.
python/triton/experimental/tle/language/raw/core.py	Add call_smem() returning buffered_tensor outputs.
python/triton/experimental/tle/language/raw/init.py	Export call_smem.
python/triton/experimental/tle/language/gpu/core.py	Add init_value support to tle_gpu.alloc() for initialized smem buffers.
python/tutorials/tle/raw/mlir/05-topk.py	Switch to initialized smem buffers + call_smem usage for outputs/inputs.
python/tutorials/tle/raw/mlir/03-matrix-multiplication-smem.py	New MLIR tutorial example using call_smem and buffered shared memory.
python/tutorials/tle/raw/cuda/03-matrix-multiplication-smem.py	New CUDA tutorial example using call_smem and buffered shared memory.

[sunnycase]

LGTM

[sunnycase]

LGTM

[zhzhcookie]

LG