CUDA实例练习：零拷贝内存

Posted 2020-10-02 Jason&Hymer

      可以在CUDA C核函数中直接访问这种类型的主机内存。由于这种内存不需要复制到GPU，因此也称为零拷贝内存。

  1 #include "book.h"
  2 #include <stdio.h>
  3 #include <cuda_runtime.h>
  4 #include <device_launch_parameters.h>
  5 #define imin(a,b) (a<b?a:b)
  6 
  7 const int N = 33 * 1024 * 1024;
  8 const int threadsPerBlock = 256;
  9 const int blocksPerGrid = imin(32, (N + threadsPerBlock - 1) / threadsPerBlock);
 10 
 11 __global__ void dot(int size, float *a, float *b, float *c) {
 12     __shared__ float cache[threadsPerBlock];
 13     int tid = threadIdx.x + blockIdx.x * blockDim.x;
 14     int cacheIndex = threadIdx.x;
 15 
 16     float   temp = 0;
 17     while (tid < size) {
 18         temp += a[tid] * b[tid];
 19         tid += blockDim.x * gridDim.x;
 20     }
 21 
 22     // set the cache values
 23     cache[cacheIndex] = temp;
 24 
 25     // synchronize threads in this block
 26     __syncthreads();
 27 
 28     // for reductions, threadsPerBlock must be a power of 2
 29     // because of the following code
 30     int i = blockDim.x / 2;
 31     while (i != 0) {
 32         if (cacheIndex < i)
 33             cache[cacheIndex] += cache[cacheIndex + i];
 34         __syncthreads();
 35         i /= 2;
 36     }
 37 
 38     if (cacheIndex == 0)
 39         c[blockIdx.x] = cache[0];
 40 }
 41 
 42 //点积运算的主机内存版本
 43 float malloc_test(int size) {
 44     cudaEvent_t start, stop;
 45     float *a, *b,c, *partial_c;
 46     float *dev_a, *dev_b, *dev_partial_c;
 47     float elapsedTime;
 48 
 49     HANDLE_ERROR(cudaEventCreate(&start));
 50     HANDLE_ERROR(cudaEventCreate(&stop));
 51 
 52     //在CPU上分配内存
 53     a = (float *)malloc(size * sizeof(float));
 54     b = (float *)malloc(size * sizeof(float));
 55     partial_c = (float *)malloc(blocksPerGrid * sizeof(float));
 56 
 57     //在GPU上分配内存
 58     HANDLE_ERROR(cudaMalloc((void **)&dev_a, size * sizeof(float)));
 59     HANDLE_ERROR(cudaMalloc((void **)&dev_b, size * sizeof(float)));
 60     HANDLE_ERROR(cudaMalloc((void **)&dev_partial_c, blocksPerGrid * sizeof(float)));
 61     
 62     //用数据填充主机内存
 63     for (int i = 0; i < size; i++){
 64         a[i] = i;
 65         b[i] = i * 2;
 66     }
 67 
 68     HANDLE_ERROR(cudaEventRecord(start, 0));
 69     //将数组\'a\'和\'b\'复制到GPU
 70     HANDLE_ERROR(cudaMemcpy(dev_a, a, size * sizeof(float),
 71         cudaMemcpyHostToDevice));
 72     HANDLE_ERROR(cudaMemcpy(dev_b, b, size * sizeof(float),
 73         cudaMemcpyHostToDevice));
 74     dot << <blocksPerGrid, threadsPerBlock >> >(size, dev_a, dev_b, dev_partial_c);
 75 
 76     //将数组\'c\'从GPU复制到CPU
 77     HANDLE_ERROR(cudaMemcpy(partial_c, dev_partial_c,
 78         blocksPerGrid * sizeof(float), cudaMemcpyDeviceToHost));
 79     HANDLE_ERROR(cudaEventRecord(stop, 0));
 80     HANDLE_ERROR(cudaEventSynchronize(stop));
 81     HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime, start, stop));
 82 
 83     //结束CPU上的计算
 84     c = 0;
 85     for (int i = 0; i < blocksPerGrid; i++){
 86         c += partial_c[i];
 87     }
 88     HANDLE_ERROR(cudaFree(dev_a));
 89     HANDLE_ERROR(cudaFree(dev_b));
 90     HANDLE_ERROR(cudaFree(dev_partial_c));
 91 
 92     //释放CPU上的内存
 93     free(a);
 94     free(b);
 95     free(partial_c);
 96 
 97     //释放事件
 98     HANDLE_ERROR(cudaEventDestroy(start));
 99     HANDLE_ERROR(cudaEventDestroy(stop));
100 
101     printf("Value calculated: %f\\n", c);
102 
103     return elapsedTime;
104 }
105 
106 //点积运算的零拷贝内存版本
107 float cuda_host_alloc_test(int size){
108     cudaEvent_t start, stop;
109     float *a, *b, c, *partial_c;
110     float *dev_a, *dev_b, *dev_partial_c;
111     float elapsedTime;
112 
113     HANDLE_ERROR(cudaEventCreate(&start));
114     HANDLE_ERROR(cudaEventCreate(&stop));
115 
116     //在CPU上分配内存
117     HANDLE_ERROR(cudaHostAlloc((void **)&a, size * sizeof(float),
118         cudaHostAllocWriteCombined | cudaHostAllocMapped));
119     HANDLE_ERROR(cudaHostAlloc((void **)&b, size * sizeof(float),
120         cudaHostAllocWriteCombined | cudaHostAllocMapped));
121     HANDLE_ERROR(cudaHostAlloc((void **)&partial_c,
122         blocksPerGrid * sizeof(float), cudaHostAllocMapped));
123 
124     //用数据填充主机内存
125     for (int i = 0; i < size; i++){
126         a[i] = i;
127         b[i] = i * 2;
128     }
129     // find out the GPU pointers
130     HANDLE_ERROR(cudaHostGetDevicePointer(&dev_a, a, 0));
131     HANDLE_ERROR(cudaHostGetDevicePointer(&dev_b, b, 0));
132     HANDLE_ERROR(cudaHostGetDevicePointer(&dev_partial_c,
133         partial_c, 0));
134 
135     // fill in the host memory with data
136     for (int i = 0; i<size; i++) {
137         a[i] = i;
138         b[i] = i * 2;
139     }
140 
141     HANDLE_ERROR(cudaEventRecord(start, 0));
142 
143     dot << <blocksPerGrid, threadsPerBlock >> >(size, dev_a, dev_b,
144         dev_partial_c);
145 
146     HANDLE_ERROR(cudaThreadSynchronize());
147     HANDLE_ERROR(cudaEventRecord(stop, 0));
148     HANDLE_ERROR(cudaEventSynchronize(stop));
149     HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime,
150         start, stop));
151 
152     // finish up on the CPU side
153     c = 0;
154     for (int i = 0; i<blocksPerGrid; i++) {
155         c += partial_c[i];
156     }
157 
158     HANDLE_ERROR(cudaFreeHost(a));
159     HANDLE_ERROR(cudaFreeHost(b));
160     HANDLE_ERROR(cudaFreeHost(partial_c));
161 
162     // free events
163     HANDLE_ERROR(cudaEventDestroy(start));
164     HANDLE_ERROR(cudaEventDestroy(stop));
165 
166     printf("Value calculated:  %f\\n", c);
167 
168     return elapsedTime;
169 }
170 
171 int main(void) {
172     cudaDeviceProp  prop;
173     int whichDevice;
174     HANDLE_ERROR(cudaGetDevice(&whichDevice));
175     HANDLE_ERROR(cudaGetDeviceProperties(&prop, whichDevice));
176     if (prop.canMapHostMemory != 1) {
177         printf("Device can not map memory.\\n");
178         return 0;
179     }
180 
181     float           elapsedTime;
182 
183     HANDLE_ERROR(cudaSetDeviceFlags(cudaDeviceMapHost));
184 
185     // try it with malloc
186     elapsedTime = malloc_test(N);
187     printf("Time using cudaMalloc:  %3.1f ms\\n",
188         elapsedTime);
189 
190     // now try it with cudaHostAlloc
191     elapsedTime = cuda_host_alloc_test(N);
192     printf("Time using cudaHostAlloc:  %3.1f ms\\n",
193         elapsedTime);
194 }