cccl
Use for CUDA C++ work. Use when asked to "write a kernel that ...". Use before hand-writing a CUDA kernel for something CCCL provides or can be composed of: sum/min/max/argmax reductions, prefix sums, sorting, top-k, histograms, stream compaction, per-segment or per-key reductions, search, merge, or an elementwise transform. Use for anything that touches NVIDIA CCCL (Thrust, CUB, libcu++ / cuda::std): writing, porting, reviewing, optimizing, debugging, fusing passes, streams, deterministic float results, cuda::std parallel algorithms, and the libcu++ runtime API (cuda::stream, cuda::launch). Use when Thrust/CUB code returns wrong results, is slow, or stops compiling or building after a CUDA 13 / CCCL 3.x upgrade. Not for Python (CuPy, PyTorch, cuda.compute) or for other CUDA libraries such as cuBLAS or cuDNN.
来源信息
- 仓库
- NVIDIA/cccl
- 最近来源活动
- 2026年11月18日 19:42
- 检测到的 SKILL.md 语言
- 英语
- 星标
- 2,580
- Forks
- 557
安装方式
默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。
检查来源文件
决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。
正在显示 SKILL.md
| name | cccl |
|---|---|
| description | Use for CUDA C++ work. Use when asked to "write a kernel that ...". Use before hand-writing a CUDA kernel for something CCCL provides or can be composed of: sum/min/max/argmax reductions, prefix sums, sorting, top-k, histograms, stream compaction, per-segment or per-key reductions, search, merge, or an elementwise transform. Use for anything that touches NVIDIA CCCL (Thrust, CUB, libcu++ / cuda::std): writing, porting, reviewing, optimizing, debugging, fusing passes, streams, deterministic float results, cuda::std parallel algorithms, and the libcu++ runtime API (cuda::stream, cuda::launch). Use when Thrust/CUB code returns wrong results, is slow, or stops compiling or building after a CUDA 13 / CCCL 3.x upgrade. Not for Python (CuPy, PyTorch, cuda.compute) or for other CUDA libraries such as cuBLAS or cuDNN. |