rtferraz/cuda-kernel-engineering
收藏资源简介:
这是一个关于CUDA内核工程的实践项目集,基于NVIDIA L4 GPU(GCP)构建。涵盖了从第一个内核到优化平铺GEMM的完整路径,并在每一步都提供了Nsight Compute的性能分析证据。使用的硬件是NVIDIA L4(sm_89,300 GB/s,23 GB GDDR6),软件栈包括CUDA 11.8 / driver 12.4、PyTorch和Nsight Compute 2022.3。项目包括向量加法、朴素GEMM、归约和平铺GEMM,每个项目都专注于不同的关键技能,如内核解剖、构建管道、共享内存、平铺和双缓冲等。
A hands-on CUDA kernel engineering portfolio built on an NVIDIA L4 GPU (GCP). Covers the complete path from first kernel to optimized tiled GEMM, with Nsight Compute profiling evidence at every step. Hardware: NVIDIA L4 (sm_89, 300 GB/s, 23 GB GDDR6). Stack: CUDA 11.8 / driver 12.4, PyTorch, Nsight Compute 2022.3. Projects include vector addition, naive GEMM, reduction, and tiled GEMM, each focusing on different key skills such as kernel anatomy, build pipeline, shared memory, tiling, and double buffering.




