Developed a GPU-accelerated diagonal matrix construction feature for the opencv/opencv repository, focusing on improving performance and correctness for small matrices. Designed and implemented a dedicated OpenCL kernel for the UMat::diag operation, enabling single-invocation diagonal matrix creation while eliminating aliasing races and memory-visibility hazards. The solution incorporated explicit synchronization semantics and robust memory-model safeguards, ensuring reliable operation across diverse OpenCL backends such as Mesa. When OpenCL was unavailable, the implementation provided a seamless CPU fallback to maintain portability. This work leveraged C++, OpenCL, and GPU programming expertise to enhance reliability and efficiency in matrix operations within the OpenCV framework.