October 2025 – ROCm/rocm-libraries: Delivered a targeted performance optimization for the triangular packed matrix-vector multiply (tpmv) kernel in rocBLAS. By offsetting the address calculation for the current index instead of recomputing it inside kernel loops, the change reduces redundant arithmetic across upper/lower triangular configurations and both conjugate and non-conjugate operations. Commit implemented: 7966015f91fa701d54639c477379023f607aa858 ([rocblas] Optimize tpmv AP address calculation (#1810)). Overall, this work lowers per-iteration arithmetic, improves kernel throughput, and benefits downstream linear algebra workloads that rely on triangular solves. No major bugs fixed this month. Repositories touched: ROCm/rocm-libraries.