2026-05 ROCm/onnxruntime — Performance-focused delivery for KleidiAi Key deliverables: - NHWC-optimized convolution path in ONNX Runtime: introduced a new FusedNhwcConv kernel implemented in MLAS, with unit tests; CPU Execution Provider updated to channels_last when data format is NHWC to avoid unnecessary transposes. Impact and value: - Eliminates layout-conversion penalties, boosting convolution throughput and reducing latency for models deployed on KleidiAi. Bug fixes and stability: - Correct data-format handling guarded to KleidiAi to prevent cross-platform impact; fixes related to NHWC path correctness and integration. Technologies demonstrated: - C++ kernel development, MLAS integration, CPU Execution Provider changes, unit-testing, and platform-specific gating. Commit reference: - a8ba94a147eea82b49409fadcf183a9be45e60f0, "[MLAS] Add an NHWC implementation of convolution to avoid transposes (#26834)"

November 2025 highlights for intel/onnxruntime: Delivered KleidiAI Kernel Logging Enhancements to improve observability and debugging of kernel operations. Implemented new logging macros for KleidiAI kernels and completed outstanding todos from prior PR to enhance functionality and maintainability. No major bugs fixed this month in this repository. The work enhances kernel traceability, reduces debugging time, and strengthens the backbone for future profiling and performance monitoring.