About this document

CuAsmRL: GPU SASS Scheduling Optimization by askermohaned2 is a document available to read on EtoBox.

The document presents CuAsmRL, an automatic optimizer for NVIDIA GPU SASS schedules utilizing deep reinforcement learning (RL) to enhance performance beyond manual optimization methods. By formulating the optimization process as an assembly game, CuAsmRL can improve existing specialized CUDA kernels by up to 26%, with an average improvement of 9%. This approach integrates seamlessly into existing compiler frameworks, providing a transparent solution for CUDA kernel developers to optimize GPU performance.

Author

askermohaned2

Language

EN