1 /* 2 * Copyright 1993-2010 NVIDIA Corporation. All rights reserved. 3 * 4 * NVIDIA Corporation and its licensors retain all intellectual property and 5 * proprietary rights in and to this software and related documentation. 6 * Any use, reproduction, disclosure, or distribution of this software 7 * and related documentation without an express license agreement from 8 * NVIDIA Corporation is strictly prohibited. 9 * 10 * Please refer to the applicable NVIDIA end user license agreement (EULA) 11 * associated with this source code for terms and conditions that govern 12 * your use of this NVIDIA software. 13 * 14 */ 15 16 17 #include "../common/book.h" 18 #include <cuda.h> 19 #include "cuda_runtime.h" 20 #include "device_launch_parameters.h" 21 #define SIZE (64*1024*1024) 22 23 24 float cuda_malloc_test(int size, bool up) { 25 cudaEvent_t start, stop; 26 int *a, *dev_a; 27 float elapsedTime; 28 29 HANDLE_ERROR(cudaEventCreate(&start)); 30 HANDLE_ERROR(cudaEventCreate(&stop)); 31 32 a = (int*)malloc(size * sizeof(*a)); 33 HANDLE_NULL(a); 34 HANDLE_ERROR(cudaMalloc((void**)&dev_a, 35 size * sizeof(*dev_a))); 36 37 HANDLE_ERROR(cudaEventRecord(start, 0)); 38 for (int i = 0; i<100; i++) { 39 if (up) 40 HANDLE_ERROR(cudaMemcpy(dev_a, a, 41 size * sizeof(*dev_a), 42 cudaMemcpyHostToDevice)); 43 else 44 HANDLE_ERROR(cudaMemcpy(a, dev_a, 45 size * sizeof(*dev_a), 46 cudaMemcpyDeviceToHost)); 47 } 48 HANDLE_ERROR(cudaEventRecord(stop, 0)); 49 HANDLE_ERROR(cudaEventSynchronize(stop)); 50 HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime, 51 start, stop)); 52 53 free(a); 54 HANDLE_ERROR(cudaFree(dev_a)); 55 HANDLE_ERROR(cudaEventDestroy(start)); 56 HANDLE_ERROR(cudaEventDestroy(stop)); 57 58 return elapsedTime; 59 } 60 61 62 float cuda_host_alloc_test(int size, bool up) { 63 cudaEvent_t start, stop; 64 int *a, *dev_a; 65 float elapsedTime; 66 67 HANDLE_ERROR(cudaEventCreate(&start)); 68 HANDLE_ERROR(cudaEventCreate(&stop)); 69 70 HANDLE_ERROR(cudaHostAlloc((void**)&a, 71 size * sizeof(*a), 72 cudaHostAllocDefault)); 73 HANDLE_ERROR(cudaMalloc((void**)&dev_a, 74 size * sizeof(*dev_a))); 75 76 HANDLE_ERROR(cudaEventRecord(start, 0)); 77 for (int i = 0; i<100; i++) { 78 if (up) 79 HANDLE_ERROR(cudaMemcpy(dev_a, a, 80 size * sizeof(*a), 81 cudaMemcpyHostToDevice)); 82 else 83 HANDLE_ERROR(cudaMemcpy(a, dev_a, 84 size * sizeof(*a), 85 cudaMemcpyDeviceToHost)); 86 } 87 HANDLE_ERROR(cudaEventRecord(stop, 0)); 88 HANDLE_ERROR(cudaEventSynchronize(stop)); 89 HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime, 90 start, stop)); 91 92 HANDLE_ERROR(cudaFreeHost(a)); 93 HANDLE_ERROR(cudaFree(dev_a)); 94 HANDLE_ERROR(cudaEventDestroy(start)); 95 HANDLE_ERROR(cudaEventDestroy(stop)); 96 97 return elapsedTime; 98 } 99 100 101 int main(void) { 102 float elapsedTime; 103 float MB = (float)100 * SIZE*sizeof(int) / 1024 / 1024; 104 105 106 // try it with cudaMalloc 107 elapsedTime = cuda_malloc_test(SIZE, true); 108 printf("Time using cudaMalloc: %3.1f ms ", 109 elapsedTime); 110 printf(" MB/s during copy up: %3.1f ", 111 MB / (elapsedTime / 1000)); 112 113 elapsedTime = cuda_malloc_test(SIZE, false); 114 printf("Time using cudaMalloc: %3.1f ms ", 115 elapsedTime); 116 printf(" MB/s during copy down: %3.1f ", 117 MB / (elapsedTime / 1000)); 118 119 // now try it with cudaHostAlloc 120 elapsedTime = cuda_host_alloc_test(SIZE, true); 121 printf("Time using cudaHostAlloc: %3.1f ms ", 122 elapsedTime); 123 printf(" MB/s during copy up: %3.1f ", 124 MB / (elapsedTime / 1000)); 125 126 elapsedTime = cuda_host_alloc_test(SIZE, false); 127 printf("Time using cudaHostAlloc: %3.1f ms ", 128 elapsedTime); 129 printf(" MB/s during copy down: %3.1f ", 130 MB / (elapsedTime / 1000)); 131 }
项目打包下载