April 2025 (Month: 2025-04) - Xilinx/onnx-mlir delivered compiler-level optimizations that improve compatibility, graph simplicity, and potential runtime performance across backends. Key work centers were: (1) Hardswish decomposition implemented as a compile-time option with a default kernel dialect lowering path to improve compatibility and flexibility, and (2) canonicalization and fusion of consecutive ONNXClip operations into a single Clip to reduce graph complexity and enable more efficient backends. These changes are aligned with the project’s goals of broader operator support, cleaner IR, and improved end-to-end performance. Commits include bd070eac8977c73fa3e7c3cff1ebf8d32aa9645c (Decompose Hardswish into simpler ONNX ops) and 5e96d18023a1c20abbb0e9160c5288cd47e16925 (Fuse consecutive clips pattern).