NVCC

nvcc is the NVIDIA CUDA Compiler Driver, the primary tool for building programs that run on NVIDIA GPUs. It is part of the NVIDIA CUDA Toolkit. nvcc compiles CUDA C and CUDA C++ source files and also orchestrates the surrounding host code compilation. When invoked as nvcc file.cu, it produces object files or executables and manages the separation of host and device code, delegating each portion to the appropriate compiler toolchain.

Internally, nvcc delegates host code to a conventional host C/C++ compiler (such as gcc, clang, or MSVC) and device code to the CUDA device compiler. It can emit device code as PTX for Just-In-Time compilation or as native cubin binaries. Options such as -arch and -gencode determine the target compute capability and code generation, allowing multiple architectures in a single build. For large projects, -rdc enables separate device compilation units, with final device linking performed by the driver.

Common usage involves specifying architecture, include paths, and libraries. Typical commands include compiling a simple program with nvcc -arch=sm_70 -O3 -o myprog myprog.cu, or compiling with -c for a library and linking against CUDA libraries like -lcudart. nvcc forwards several host-compiler options and supports -ccbin to select the host compiler, as well as -I and -L for include and library search paths.

nvcc is designed for CUDA-enabled development and relies on the CUDA toolkit and compatible NVIDIA drivers. It targets NVIDIA GPUs and is not a general C/C++ compiler. The toolkit evolves with new CUDA toolchains, while host compilers and drivers must remain compatible to ensure successful builds.