一般有两种方法可以在vs2013上添加运行cuda8.0程序：
一、直接新建一个基于CUDA8.0的项目：如下图所示，
点击确定后即可生成test_cuda项目；默认会自动生成一个kernel.cu文件；默认已经配置好Debug/Release, Win32/x64环境，直接编译运行，结果如下图所示：函数执行的是两个数组的加操作。移除kernel.cu文件，加入自己需要的cuda文件即可进行实际操作了，非常方便。
二、实际情况下，多是在已有的项目中添加一些cuda文件，用于加速，下面说下具体的操作步骤：
1、新建一个CUDA_Test x64控制台空工程；
2、新建CUDA_Test.cpp文件；
3、选中CUDA_Test项目，右键单击-->生成依赖项-->生成自定义，勾选CUDA8.0，点击确定，如下图所示：
4、完成第3步后，再次打开工程的属性配置，会多出两项，CUDA C/C++和CUDA Linker，如下图所示：
5、新建或添加几个已有的文件，包括common.hpp、simple.hpp、simple.cpp、simple.cu，各个文件内容如下：
common.hpp:
#ifndef FBC_CUDA_TEST_COMMON_HPP_ #define FBC_CUDA_TEST_COMMON_HPP_ #define PRINT_ERROR_INFO(info) { \ fprintf(stderr, "Error: %s, file: %s, func: %s, line: %d\n", #info, __FILE__, __FUNCTION__, __LINE__); \ return -1; } #endif // FBC_CUDA_TEST_COMMON_HPP_ simple.hpp: #ifndef FBC_CUDA_TEST_SIMPLE_HPP_ #define FBC_CUDA_TEST_SIMPLE_HPP_ // reference: C:\ProgramData\NVIDIA Corporation\CUDA Samples\v8.0\0_Simple int test_vectorAdd(); int vectorAdd_cpu(const float *A, const float *B, float *C, int numElements); int vectorAdd_gpu(const float *A, const float *B, float *C, int numElements); #endif // FBC_CUDA_TEST_SIMPLE_HPP_ simple.cpp: #include "simple.hpp" #include <stdlib.h> #include <iostream> #include "common.hpp" // =========================== vector add ============================= int test_vectorAdd() { // Vector addition: C = A + B, implements element by element vector addition const int numElements{ 50000 }; float* A = new float[numElements]; float* B = new float[numElements]; float* C1 = new float[numElements]; float* C2 = new float[numElements]; // Initialize vector for (int i = 0; i < numElements; ++i) { A[i] = rand() / (float)RAND_MAX; B[i] = rand() / (float)RAND_MAX; } int ret = vectorAdd_cpu(A, B, C1, numElements); if (ret != 0) PRINT_ERROR_INFO(vectorAdd_cpu); ret = vectorAdd_gpu(A, B, C2, numElements); if (ret != 0) PRINT_ERROR_INFO(vectorAdd_gpu); for (int i = 0; i < numElements; ++i) { if (fabs(C1[i] - C2[i]) > 1e-5) { fprintf(stderr, "Result verification failed at element %d!\n", i); return -1; } } delete[] A; delete[] B; delete[] C1; delete[] C2; return 0; } int vectorAdd_cpu(const float *A, const float *B, float *C, int numElements) { for (int i = 0; i < numElements; ++i) { C[i] = A[i] + B[i]; } return 0; } simple.cu: #include "simple.hpp" #include <iostream> #include <cuda_runtime.h> // For the CUDA runtime routines (prefixed with "cuda_") #include <device_launch_parameters.h> // reference: C:\ProgramData\NVIDIA Corporation\CUDA Samples\v8.0\0_Simple // =========================== vector add ============================= __global__ void vectorAdd(const float *A, const float *B, float *C, int numElements) { int i = blockDim.x * blockIdx.x + threadIdx.x; if (i < numElements) { C[i] = A[i] + B[i]; } } int vectorAdd_gpu(const float *A, const float *B, float *C, int numElements) { // Error code to check return values for CUDA calls cudaError_t err{ cudaSuccess }; size_t length{ numElements * sizeof(float) }; fprintf(stderr, "Length: %d\n", length); float* d_A{ nullptr }; float* d_B{ nullptr }; float* d_C{ nullptr }; err = cudaMalloc(&d_A, length); if (err != cudaSuccess) { fprintf(stderr, "Failed to allocate device vector A (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaMalloc(&d_B, length); if (err != cudaSuccess) { fprintf(stderr, "Failed to allocate device vector B (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaMalloc(&d_C, length); if (err != cudaSuccess) { fprintf(stderr, "Failed to allocate device vector C (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaMemcpy(d_A, A, length, cudaMemcpyHostToDevice); if (err != cudaSuccess) { fprintf(stderr, "Failed to copy vector A from host to device (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaMemcpy(d_B, B, length, cudaMemcpyHostToDevice); if (err != cudaSuccess) { fprintf(stderr, "Failed to copy vector B from host to device (error code %s)!\n", cudaGetErrorString(err)); return -1; } // Launch the Vector Add CUDA kernel int threadsPerBlock = 256; int blocksPerGrid = (numElements + threadsPerBlock - 1) / threadsPerBlock; fprintf(stderr, "CUDA kernel launch with %d blocks of %d threads\n", blocksPerGrid, threadsPerBlock); vectorAdd << <blocksPerGrid, threadsPerBlock >> >(d_A, d_B, d_C, numElements); err = cudaGetLastError(); if (err != cudaSuccess) { fprintf(stderr, "Failed to launch vectorAdd kernel (error code %s)!\n", cudaGetErrorString(err)); return -1; } // Copy the device result vector in device memory to the host result vector in host memory. err = cudaMemcpy(C, d_C, length, cudaMemcpyDeviceToHost); if (err != cudaSuccess) { fprintf(stderr, "Failed to copy vector C from device to host (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaFree(d_A); if (err != cudaSuccess) { fprintf(stderr, "Failed to free device vector A (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaFree(d_B); if (err != cudaSuccess) { fprintf(stderr, "Failed to free device vector B (error code %s)!\n", cudaGetErrorString(err)); return -1; } err = cudaFree(d_C); if (err != cudaSuccess) { fprintf(stderr, "Failed to free device vector C (error code %s)!\n", cudaGetErrorString(err)); return -1; } return err; } CUDA_Test.cpp: #include <iostream> #include "simple.hpp" int main() { int ret = test_vectorAdd(); if (ret == 0) fprintf(stderr, "***** test success *****\n"); else fprintf(stderr, "===== test fail =====\n"); return 0; } 6、调整属性配置项：(1)、CUDA C/C++-->Common中Target Machine Platform中默认是32-bit(--machine32)，因为是x64，所以将其调整为64-bit(--machine 64)；
(2)、添加附加库：链接器-->输入-->附加依赖项：cudart.lib；
(3)、消除nvcc warning: The 'compute_20', 'sm_20', and'sm_21' architectures are deprecated, and may be removed in a future release：CUDA C/C++-->Device: Code Generation:由compute_20,sm_20修改为compute_30,sm_30; compute_35,sm_35; compute_37,sm_37;compute_50,sm_50; compute_52,sm_52; compute_60,sm_60
以上code是参考NVIDIA Corporation\CUDA Samples\v8.0\0_Simple中vectorAdd例子进行的改写，输出结果如下：
GitHub：https://github.com/fengbingchun/CUDA_Test
