《GPU高性能编程CUDA实战》第九章 原子性

▶ 本章介绍了原子操作，给出了基于原子操作的直方图计算的例子。

● 章节代码

 #include <stdio.h>
 #include "cuda_runtime.h"
 #include "device_launch_parameters.h"
 #include "D:\Code\CUDA\book\common\book.h"

 #define SIZE    (100*1024*1024)
 #define USE_SHARE_MEMORY true

 __global__ void histo_kernel(unsigned char *buffer, long size, unsigned int *histo)
 {
     int i = threadIdx.x + blockIdx.x * blockDim.x;

 #if USE_SHARE_MEMORY
     __shared__  unsigned int temp[];
     temp[threadIdx.x] = ;
     __syncthreads();

     while (i < size)
     {
         atomicAdd(&temp[buffer[i]], );
         i += blockDim.x * gridDim.x;
     }
     __syncthreads();
     atomicAdd(&(histo[threadIdx.x]), temp[threadIdx.x]);
 #else
     while (i < size)
     {
         atomicAdd(&histo[buffer[i]], );
         i += blockDim.x * gridDim.x;;
     }
 #endif
     return;
 }

 int main(void)
 {
     int i;
     unsigned char *buffer = (unsigned char*)big_random_block(SIZE);// 内置的生成随机字符数组的函数

     cudaEvent_t     start, stop;
     cudaEventCreate(&start);
     cudaEventCreate(&stop);
     cudaEventRecord(start, );

     unsigned char *dev_buffer;
     unsigned int *dev_histo;
     cudaMalloc((void**)&dev_buffer, SIZE);
     cudaMemcpy(dev_buffer, buffer, SIZE, cudaMemcpyHostToDevice);

     cudaMalloc((void**)&dev_histo,  * sizeof(int));
     cudaMemset(dev_histo, ,  * sizeof(int));

     cudaDeviceProp  prop;
     cudaGetDeviceProperties(&prop, );
     int blocks = prop.multiProcessorCount;// 书：实验表明使用MPS的两倍计算效率最高
     histo_kernel << <blocks * ,  >> >(dev_buffer, SIZE, dev_histo);

     unsigned int    histo[];
     cudaMemcpy(histo, dev_histo,  * sizeof(int), cudaMemcpyDeviceToHost);

     cudaEventRecord(stop, );
     cudaEventSynchronize(stop);
     float   elapsedTime;
     cudaEventElapsedTime(&elapsedTime, start, stop);
     printf("Time to generate:  %3.1f ms\n", elapsedTime);

     long histoCount = ;
     for (i = ; i < ; i++)
         histoCount += histo[i];
     printf("Histogram Sum:  %ld\n", histoCount);

     for (i = ; i < SIZE; i++)// 验证结果
         histo[buffer[i]]--;
     for (i = ; i < ; i++)
     {
         if (histo[i] != )
             printf("Failure at hist[%d] == %d\n", i,histo[i]);
     }
     if (i == )
         printf("\n\tSucceeded!\n");

     cudaFree(dev_histo);
     cudaFree(dev_buffer);
     free(buffer);
     cudaEventDestroy(start);
     cudaEventDestroy(stop);

     getchar();
     return ;
 }

● 使用全局内存时，只要在每次线程尝试 +1 时使用原子加法即可；使用共享内存时算法分两步，线程先用原子加法往各线程块的共享内存中写入，同步以后，再用原子加法把各共享内存的结果往全局内存中写入。减缓了全局内存的写入冲突。

● 代码使用了两倍数量的MPS作为线程块数量，认为这样计算效率最高。在 GTX1070 上 prop.multiProcesser 为 16，程序默认使用 32 个线程块进行计算，我另用 1 到 256 个线程块依次测试，结果如下。

● big_random_block()定义于book.h中

 void* big_random_block(int size)
 {
     unsigned char *data = (unsigned char*)malloc(size);
     for (int i = ; i < size; i++)
         data[i] = rand();
     return data;
 }