nvcc로 컴파일하기

The CUDA compiler is called nvcc. It takes your .cu file with mixed CPU and GPU code and turns it into one runnable program. 🛠️

A .cu file holds two worlds: host code that runs on the CPU and device code that runs on the GPU. nvcc must handle each one differently.

nvcc splits your source automatically. It hands the host portions to your normal C++ compiler and keeps the device portions for itself.

For the GPU side, nvcc generates a special assembly called PTX, then turns it into machine code your specific GPU can execute.

nvcc does not compile CPU code itself. It quietly calls your system's host compiler, like g++ or clang, to handle the ordinary C++.

Compiling looks just like g++. Point nvcc at your .cu file and name the output with the -o flag.

The -arch flag tells nvcc which GPU generation to build for. Matching it to your card produces faster, fully optimized device code.

That sm number is a compute capability, a label for a GPU's feature set. sm_80 means an Ampere card; older cards use smaller numbers.

nvcc can pack code for several architectures into one fatbinary, so the same executable runs well on different GPUs without rebuilding.

Many g++ habits carry over. nvcc accepts -O levels and -g for debugging, passing them along to the right compiler stage.

Need a tuned library like cuBLAS? You link it just as in g++, adding the library name with an -l flag on the nvcc command.

Let us check how nvcc divides the work.

You learned that nvcc splits a .cu file, compiles device code to GPU machine code, and hands host code to g++. The -arch flag targets your card. 🎯