CUDA 粗大的笔记

一、基础数据流：
分为Host和Device如下图所示，数据流需要从内存传入显存中，然后Host(CPU) 启动Device(GPU)来计算，GPU计算完成数据后再从显存后传回内存中。
贴段代码感受一下：（CUDA 8.0 sample with VS2013）
  1 #include "cuda_runtime.h"
  2 #include "device_launch_parameters.h"
  3 
  4 #include <stdio.h>
  5 
  6 cudaError_t addWithCuda(int *c, const int *a, const int *b, unsigned int size);
  7 
  8 __global__ void addKernel(int *c, const int *a, const int *b)
  9 {
 10     int i = threadIdx.x;
 11     c[i] = a[i] + b[i];
 12 }
 13 
 14 int main()
 15 {
 16     const int arraySize = 5;
 17     const int a[arraySize] = { 1, 2, 3, 4, 5 };
 18     const int b[arraySize] = { 10, 20, 30, 40, 50 };
 19     int c[arraySize] = { 0 };
 20 
 21     // Add vectors in parallel.
 22     cudaError_t cudaStatus = addWithCuda(c, a, b, arraySize);
 23     if (cudaStatus != cudaSuccess) {
 24         fprintf(stderr, "addWithCuda failed!");
 25         return 1;
 26     }
 27 
 28     printf("{1,2,3,4,5} + {10,20,30,40,50} = {%d,%d,%d,%d,%d}\n",
 29         c[0], c[1], c[2], c[3], c[4]);
 30 
 31     // cudaDeviceReset must be called before exiting in order for profiling and
 32     // tracing tools such as Nsight and Visual Profiler to show complete traces.
 33     cudaStatus = cudaDeviceReset();
 34     if (cudaStatus != cudaSuccess) {
 35         fprintf(stderr, "cudaDeviceReset failed!");
 36         return 1;
 37     }
 38 
 39     return 0;
 40 }
 41 
 42 // Helper function for using CUDA to add vectors in parallel.
 43 cudaError_t addWithCuda(int *c, const int *a, const int *b, unsigned int size)
 44 {
 45     int *dev_a = 0;
 46     int *dev_b = 0;
 47     int *dev_c = 0;
 48     cudaError_t cudaStatus;
 49 
 50     // Choose which GPU to run on, change this on a multi-GPU system.
 51     cudaStatus = cudaSetDevice(0);
 52     if (cudaStatus != cudaSuccess) {
 53         fprintf(stderr, "cudaSetDevice failed!  Do you have a CUDA-capable GPU installed?");
 54         goto Error;
 55     }
 56 
 57     // Allocate GPU buffers for three vectors (two input, one output)    .
 58     cudaStatus = cudaMalloc((void**)&dev_c, size * sizeof(int));
 59     if (cudaStatus != cudaSuccess) {
 60         fprintf(stderr, "cudaMalloc failed!");
 61         goto Error;
 62     }
 63 
 64     cudaStatus = cudaMalloc((void**)&dev_a, size * sizeof(int));
 65     if (cudaStatus != cudaSuccess) {
 66         fprintf(stderr, "cudaMalloc failed!");
 67         goto Error;
 68     }
 69 
 70     cudaStatus = cudaMalloc((void**)&dev_b, size * sizeof(int));
 71     if (cudaStatus != cudaSuccess) {
 72         fprintf(stderr, "cudaMalloc failed!");
 73         goto Error;
 74     }
 75 
 76     // Copy input vectors from host memory to GPU buffers.
 77     cudaStatus = cudaMemcpy(dev_a, a, size * sizeof(int), cudaMemcpyHostToDevice);
 78     if (cudaStatus != cudaSuccess) {
 79         fprintf(stderr, "cudaMemcpy failed!");
 80         goto Error;
 81     }
 82 
 83     cudaStatus = cudaMemcpy(dev_b, b, size * sizeof(int), cudaMemcpyHostToDevice);
 84     if (cudaStatus != cudaSuccess) {
 85         fprintf(stderr, "cudaMemcpy failed!");
 86         goto Error;
 87     }
 88 
 89     // Launch a kernel on the GPU with one thread for each element.
 90     addKernel<<<1, size>>>(dev_c, dev_a, dev_b);
 91 
 92     // Check for any errors launching the kernel
 93     cudaStatus = cudaGetLastError();
 94     if (cudaStatus != cudaSuccess) {
 95         fprintf(stderr, "addKernel launch failed: %s\n", cudaGetErrorString(cudaStatus));
 96         goto Error;
 97     }
 98     
 99     // cudaDeviceSynchronize waits for the kernel to finish, and returns
100     // any errors encountered during the launch.
101     cudaStatus = cudaDeviceSynchronize();
102     if (cudaStatus != cudaSuccess) {
103         fprintf(stderr, "cudaDeviceSynchronize returned error code %d after launching addKernel!\n", cudaStatus);
104         goto Error;
105     }
106 
107     // Copy output vector from GPU buffer to host memory.
108     cudaStatus = cudaMemcpy(c, dev_c, size * sizeof(int), cudaMemcpyDeviceToHost);
109     if (cudaStatus != cudaSuccess) {
110         fprintf(stderr, "cudaMemcpy failed!");
111         goto Error;
112     }
113 
114 Error:
115     cudaFree(dev_c);
116     cudaFree(dev_a);
117     cudaFree(dev_b);
118     
119     return cudaStatus;
120 }
View CUDA Sample Code
posted @ 2016-09-28 13:17 Pink-Floyd 阅读(269) 评论(0) 收藏举报
刷新页面返回顶部
Pink Floyd

CUDA 粗大的笔记
