Worked on the oneapi-src/oneDNN repository to improve benchmarking reliability and cross-architecture compatibility by addressing inflated memory bandwidth measurements. Developed a GPU-based random data generation path using OpenCL and C++, replacing host-side memset with an on-device gpu_fill_random() kernel and extending the Philox PRNG to efficiently generate non-compressible uint4 vectors. This approach prevented driver compression from skewing results and enabled more accurate, consistent performance comparisons across hardware. Also implemented universal NaN and Inf filtering for all floating-point types, with a safe fallback to memset for non-Intel devices, ensuring both portability and trustworthy benchmarking across diverse platforms.