OneDNN Deep Learning Performance Engineering: Practical Techniques for Faster Workloads
Johnson, Robert U.
Synopsis "OneDNN Deep Learning Performance Engineering: Practical Techniques for Faster Workloads"
OneDNN Deep Learning Performance Engineering: Practical Techniques for Faster Workloads is a practical, technically rigorous guide for engineers, researchers, and performance-minded practitioners who want to maximize deep learning efficiency on CPUs and heterogeneous systems. It begins with the core principles of deep learning performance, explaining how arithmetic intensity, memory bandwidth, cache behavior, threading, and microarchitectural features shape real-world throughput and latency. By connecting algorithmic complexity to hardware execution, the book provides a clear foundation for understanding why some workloads scale efficiently while others become bottlenecked by data movement and execution overhead. The book then takes a deep dive into OneDNN’s design and optimization model, showing how its primitives, abstraction layers, and backend interfaces deliver portability without sacrificing speed. Readers learn how to optimize convolutions, tensor transformations, fully connected layers, and normalization operations using techniques such as primitive fusion, layout selection, tiling, low-precision computation, and effective multithreading. Throughout, the book emphasizes practical integration with major frameworks including TensorFlow, PyTorch, and ONNX, while also presenting profiling workflows and systematic bottleneck analysis to support iterative, evidence-based performance tuning. Looking beyond current deployment patterns, the final chapters explore emerging directions in deep learning performance engineering, including CPU+GPU+accelerator collaboration, dynamic and adaptive neural networks, hardware-aware AutoML, and secure inference. With a focus on actionable techniques, real-world case studies, and production-oriented guidance, OneDNN Deep Learning Performance Engineering: Practical Techniques for Faster Workloads equips readers to build, optimize, and maintain high-throughput deep learning systems that are efficient, scalable, and ready for future hardware and software demands.
