CUDA Academy · 강의

커널의 구조

시그니처, 반환 자료형과 void 규칙을 알아봅니다.

커널의 구조은(는) CoddyKit의 무료 CUDA Academy 강의입니다. 이것은 4개 중 1번째 강의입니다. 아래에서 전체 강의를 무료로 읽을 수 있으며, 내장 코드 에디터와 24/7 AI 튜터와 함께 브라우저에서 직접 실습할 수 있습니다. 이 강의는 CUDA Academy 학습 경로의 일부이며, 진행 상황이 웹과 CoddyKit 앱에 동기화됩니다. CUDA Academy 강의에는 총 4개의 강의가 포함되어 있습니다.

이 강의의 일부는 아직 번역되지 않았으며 영어로 표시됩니다.

What a Kernel Is

A kernel is a single C++ function that runs on the GPU, executed at once by thousands of threads. You write it once, the hardware fans it out. 🚀

The global Marker

You mark a kernel with the __global__ qualifier. It tells nvcc this function is called from the CPU but actually runs on the device.

__global__ void myKernel() {
    // runs on the GPU
}

Kernels Return void

Every kernel must return void. There is no return value to hand back, so results travel out through pointers to device memory instead.

__global__ void k() { /* void only */ }

Why Not Return a Value?

Thousands of threads run your kernel together, so a single return value would make no sense. Each thread writes its own slot of an output array.

Parameters Are by Value

Kernel parameters are copied to every thread, so pass small things: ints, sizes, and pointers. Never pass big objects by value.

__global__ void add(float* out, float* a, int n) {}

Pass Pointers, Not Arrays

To give a kernel an array, you pass a device pointer. The kernel reads and writes the GPU memory that pointer addresses.

__global__ void scale(float* data, float k) {}

Each Thread Picks Its Work

The same code runs in every thread, so each one uses its index to decide which element to touch. That is how one function covers a whole array.

int i = threadIdx.x; // who am I?

A Tiny Complete Kernel

Here is a full kernel: it doubles one element per thread. Notice the __global__ marker, the void return, and the pointer parameter all together.

__global__ void doubleIt(float* x) {
    int i = threadIdx.x;
    x[i] = x[i] * 2.0f;
}

Host Code Stays Separate

Your normal CPU function, often main, is host code. It sets things up and then asks the GPU to run the kernel.

int main() {
    // host side: prepare and launch
}

No Recursion or I/O Surprises

A kernel runs on hardware with tight rules, so keep it simple: avoid deep recursion and heavy standard-library calls inside it.

Naming Your Kernels

Treat a kernel like any function: give it a clear, verb-based name like vectorAdd. Good names make launches far easier to read.

__global__ void vectorAdd(float* c, float* a, float* b);

Quick Check

Let us check the kernel signature rules.

Recap: Kernel Anatomy

A kernel is a __global__ void function run by many threads. Pass pointers and small values, and let each thread use its index. Nicely done! 🎉

자주 묻는 질문