cl-cuda

Kernels are lazily compiled to CUDA C, compiled to PTX via NVCC, and loaded automatically via the kernel manager, reducing manual intervention for users.

Memory blocks abstract host and device memory with functions like `sync-memory-block`, simplifying data transfers and reducing common bugs in GPU programming.

Provides direct interfaces to CFFI pointers and CUDA device pointers, allowing advanced users to bypass abstractions for custom memory handling when needed.

Enables work with multiple CUDA streams for overlapping copy and computation operations, enhancing throughput in parallel tasks, as detailed in the streams section.