.CU File Format

Introduction

A .cu file holds source code in CUDA C++, an extension of C++ from NVIDIA for running code on the company’s graphics cards and accelerators. One file holds both the code that runs on the main processor and the code that runs on the GPU. .cuh is the usual name for CUDA headers.

NVIDIA’s compiler, nvcc, splits a file in two. It compiles the GPU part itself and hands the rest to the system’s C++ compiler. Clang can also compile CUDA.

Structure

A CUDA file is ordinary C++ text, with some additions:

• Function specifiers say where a function runs. __global__ marks a kernel, started from the host and run on the GPU, __device__ a function called on the GPU, and __host__ a function on the CPU.
• A kernel is launched with the triple angle brackets kernel<<<blocks, threads>>>(arguments), which give the grid size and the number of threads in each block.
• Inside a kernel, built-in variables such as blockIdx, threadIdx, blockDim and gridDim tell each thread which part of the work is its own.
• Memory qualifiers such as __shared__ place data in memory shared by a block of threads, and functions from the CUDA runtime, such as cudaMalloc, manage memory on the GPU.

#include <cuda_runtime.h>

__global__ void scale(float *data, int n, float factor) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n) {
        data[i] *= factor;
    }
}

int main() {
    const int n = 1024;
    float *data;
    cudaMalloc(&data, n * sizeof(float));
    scale<<<4, 256>>>(data, n, 2.0f);
    cudaDeviceSynchronize();
    cudaFree(data);
    return 0;
}

Links