THE ART OF PTX AND SASS PROGRAMMING: Advanced CUDA Performance Engineering: 13 (intelligent Systems Engineering Library) - Softcover

Heller, Dion M.

Softcover

ISBN 13: 9798187898176

Publisher: Independently published, 2026

Synopsis

Have you ever optimized a CUDA kernel only to discover it still runs far slower than expected with no clear explanation why?

If you've experienced that frustration, you're not alone. The answer often isn't hidden in your C++ source code, it's buried in the PTX and SASS instructions your compiler generates.

Most CUDA programming books teach kernel syntax, memory models, and launch configurations. Those are essential skills, but they only tell part of the story. True GPU optimization begins below the source code, where register allocation, instruction scheduling, memory access, and warp execution determine whether your application fully utilizes the hardware or leaves performance on the table.

This book takes you beyond CUDA programming and into CUDA performance engineering.

Through fifteen comprehensive chapters, you'll follow the complete journey of a CUDA kernel from high-level source code to PTX intermediate representation and finally to the native machine instructions executed by NVIDIA GPUs. You'll learn how to read GPU disassembly with confidence, identify performance bottlenecks, and make optimization decisions based on evidence rather than guesswork.

Inside this book, you'll learn how to:

• Understand the complete CUDA compilation pipeline, from source code to native GPU instructions.
• Read and interpret PTX and SASS output to understand exactly what your compiler is generating.
• Diagnose register pressure, occupancy limitations, and instruction-level bottlenecks.
• Optimize memory access patterns for maximum bandwidth and cache efficiency.
• Identify, measure, and eliminate warp divergence that reduces execution efficiency.
• Understand Tensor Core and matrix acceleration instructions and verify when your kernels are using specialized hardware.
• Master professional profiling and performance analysis tools used in production GPU development.
• Build a systematic, repeatable workflow for diagnosing and optimizing CUDA applications with confidence.

Learn Through Real Performance Investigations

Every chapter is built around practical, real-world examples rather than isolated code fragments. You'll examine kernels before optimization, analyze their generated instructions, identify performance issues, implement targeted improvements, and verify measurable results using industry-standard tools and methodologies.

Rather than relying on trial and error, you'll develop the ability to explain why a kernel performs the way it does and how to improve it with precision.

Who Should Read This Book?

This guide is written for GPU programmers, systems engineers, high-performance computing professionals, machine learning infrastructure engineers, compiler enthusiasts, graduate students, and software developers who want a deeper understanding of CUDA performance. Whether you're building scientific applications, AI frameworks, graphics engines, or large-scale compute systems, the techniques in this book will help you optimize with greater confidence and accuracy.

Stop Guessing. Start Understanding.

The fastest CUDA developers aren't the ones who memorize optimization tricks they're the ones who understand what the GPU is actually executing.

If you're ready to move beyond surface-level optimization and learn how to analyze, diagnose, and maximize GPU performance from the instruction level upward, this book is your definitive guide.

Edition details

• Publisher: Independently published
• Publication date: 2026
• Language: English
• ISBN 13: 9798187898176
• Binding: Paperback
• Number of pages: 191