April 2026: Implemented RISC-V CI test balancing and partitioning in oneDNN to optimize CI pipelines by distributing tests across partitions based on estimated runtimes. This enables faster feedback loops, more predictable build times, and better resource utilization across CI infrastructure. The change includes weekly CI test rebalance for RV64 tests, improving stability and throughput. Co-authored by Fei Zhang; commit: cpu: rv64: rebalance weekly CI test partitions (#4955).

March 2026 performance highlights: Delivered JIT-accelerated kernels and stability fixes across two oneDNN repositories to boost neural-network throughput on RV64/RVV architectures and improve numerical correctness. Key outcomes include a FP32 softmax JIT kernel with runtime ISA checks on RV64, fully JIT-based LayerNorm forward and RVV-backed GEMM acceleration, and a stability fix for softmax to treat -Inf inputs as zero. These changes enhance forward-pass throughput and matrix-multiply performance on RISC-V targets, enabling better performance-per-watt for edge and server workloads.

Month 2026-02 highlights for oneDNN: Delivered correctness improvements and performance-oriented features on RV64. Key outcomes include a bug fix for the convolution im2col width-padding path and a new JIT kernel for RV64 layer normalization with optional scaling and shifting. Re-enabled CI tests to validate changes and maintain release quality. These changes strengthen robustness for convolution workloads and accelerate normalization-critical paths on RV64, aligning with performance and portability goals.

In 2026-01, focused on hardening and expanding RISC-V support in oneDNN. Delivered stability improvements, JIT enhancements, and half-precision softmax capabilities to enable more reliable and efficient DL workloads on RISC-V-based systems. The work improves portability, performance, and memory efficiency for production deployments.

December 2025: Delivered f16 pooling support for NHWC and NCHW in oneDNN with a dedicated F16 NCHW AvgPoolingExcludePadding variant to boost performance and memory efficiency on RISC-V. Implemented governance improvements by recording ISCAS copyright notices and designating the RISC-V team as code owners for xbyak_riscv. No major bugs fixed this month; focus was on feature delivery, code quality, and governance.

November 2025 performance highlights for oneapi-src/oneDNN focused on FP16 enablement, RISC-V runtime optimizations, and GEMM performance improvements, delivering measurable business value through faster FP16 workloads, better CPU feature utilization, and improved maintainability.

Month: 2025-10 — Focused on improving reliability and quality of GEMM Convolution on RISC-V in oneDNN. Key feature delivered: Validation and code quality improvements for the RISC-V GEMM convolution path. Major bug fix: Correctness enhancements in post_ops_ok to correctly handle sum, binary, and PReLU with broadcasting. Code hygiene: formatting cleanup and ensuring newline at end of rvv_gemm_convolution.cpp. Commits are well-documented and traceable.

Month: 2025-09 — Delivered high-performance RV64 convolution path and architecture extension support in oneDNN, focused on business value, performance, and maintainability. Key outcomes include a vectorized RV64 GEMM convolution path with RVV post-ops, robust build-time support for the ZVFH extension, and code-quality improvements to validation and type handling.

August 2025: Delivered performance-focused enhancements for RISC-V vector workloads in two DNN-quality builds. Implemented RVV SIMD-accelerated f32 GEMM for RV64 with new kernels and utilities, added conditional wiring into builds when RV64 with RVV intrinsics are enabled, and stabilized the path for broader hardware coverage. Extended RISC-V vector-extension readiness by adding Zvfh support detection to the uxlfoundation/oneDNN build system. These efforts jointly boost GEMM throughput on supported hardware, improve build-time adaptability for future ISA extensions, and demonstrate strong cross-repo collaboration and engineering rigor.