[Zhenchao Ni (zhenchaoni)]

Summary

Add a --cgc-lowering-target auto|dml|d3d12 option to compile, eval, and perf, allowing users to select the CGC lowering target.

• Forward the option through command configuration to the Windows ML Runtime compiler.
• Validate that the option is used with the CGC backend.
• Unify ONNX and MLIR loading on CGC: compile to an output IR artifact, reload it, then build and run the pipeline.
• Preserve Runtime's Automatic target selection when the option is omitted.

Examples

Benchmark an ONNX model using CGC with DirectML lowering:

winml perf -m model.onnx --runtime winml-runtime --backend cgc --device gpu --cgc-lowering-target dml

Compile CGC Input IR with D3D12 lowering:

winml compile -m model.mlir --compiler cgc --device gpu --cgc-lowering-target d3d12 -o compiled.mlir

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

One input-validation issue remains before approval. All 538 existing affected tests passed locally; Ruff and git diff --check passed. A focused additional CLI probe reproduced acceptance of --runtime winml-ort --cgc-lowering-target d3d12 (exit 0 and benchmark config retained d3d12), although the non-Runtime model path discards that setting. Native behavior for recompiling existing Output IR was not exercised locally.