CUDA编程 - asyncAPI 学习记录

一、完整代码

#include <stdio.h>

// includes CUDA Runtime
#include <cuda_runtime.h>
#include <cuda_profiler_api.h>

// includes, project
#include <helper_cuda.h>
#include <helper_functions.h>  // helper utility functions

__global__ void increment_kernel(int *g_data, int inc_value) {
  int idx = blockIdx.x * blockDim.x + threadIdx.x;
  g_data[idx] = g_data[idx] + inc_value;
}

bool correct_output(int *data, const int n, const int x) {
  for (int i = 0; i < n; i++)
    if (data[i] != x) {
      printf("Error! data[%d] = %d, ref = %d\n", i, data[i], x);
      return false;
    }

  return true;
}

int main(int argc, char *argv[]) {
  int devID;
  cudaDeviceProp deviceProps;

  printf("[%s] - Starting...\n", argv[0]);

  // This will pick the best possible CUDA capable device
  devID = findCudaDevice(argc, (const char **)argv);

  // get device name
  checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
  printf("CUDA device [%s]\n", deviceProps.name);

  int n = 16 * 1024 * 1024;
  int nbytes = n * sizeof(int);
  int value = 26;

  // allocate host memory
  int *a = 0;
  checkCudaErrors(cudaMallocHost((void **)&a, nbytes));
  memset(a, 0, nbytes);

  // allocate device memory
  int *d_a = 0;
  checkCudaErrors(cudaMalloc((void **)&d_a, nbytes));
  checkCudaErrors(cudaMemset(d_a, 255, nbytes));

  // set kernel launch configuration
  dim3 threads = dim3(512, 1);
  dim3 blocks = dim3(n / threads.x, 1);

  // create cuda event handles
  cudaEvent_t start, stop;
  checkCudaErrors(cudaEventCreate(&start));
  checkCudaErrors(cudaEventCreate(&stop));

  StopWatchInterface *timer = NULL;
  sdkCreateTimer(&timer);
  sdkResetTimer(&timer);

  checkCudaErrors(cudaDeviceSynchronize());
  float gpu_time = 0.0f;

  // asynchronously issue work to the GPU (all to stream 0)
  checkCudaErrors(cudaProfilerStart());
  sdkStartTimer(&timer);
  cudaEventRecord(start, 0);
  cudaMemcpyAsync(d_a, a, nbytes, cudaMemcpyHostToDevice, 0);
  increment_kernel<<<blocks, threads, 0, 0>>>(d_a, value);
  cudaMemcpyAsync(a, d_a, nbytes, cudaMemcpyDeviceToHost, 0);
  cudaEventRecord(stop, 0);
  sdkStopTimer(&timer);
  checkCudaErrors(cudaProfilerStop());

  // have CPU do some work while waiting for stage 1 to finish
  unsigned long int counter = 0;

  while (cudaEventQuery(stop) == cudaErrorNotReady) {
    counter++;
  }

  checkCudaErrors(cudaEventElapsedTime(&gpu_time, start, stop));

  // print the cpu and gpu times
  printf("time spent executing by the GPU: %.2f\n", gpu_time);
  printf("time spent by CPU in CUDA calls: %.2f\n", sdkGetTimerValue(&timer));
  printf("CPU executed %lu iterations while waiting for GPU to finish\n",
         counter);

  // check the output for correctness
  bool bFinalResults = correct_output(a, n, value);

  // release resources
  checkCudaErrors(cudaEventDestroy(start));
  checkCudaErrors(cudaEventDestroy(stop));
  checkCudaErrors(cudaFreeHost(a));
  checkCudaErrors(cudaFree(d_a));

  exit(bFinalResults ? EXIT_SUCCESS : EXIT_FAILURE);
}

二、重要部分解析

设备初始化：findCudaDevice函数用于选择最佳的CUDA设备，并返回设备ID。

devID = findCudaDevice(argc, (const char **)argv);

获取设备属性：cudaGetDeviceProperties函数获取指定设备的属性，这些属性包括设备名称等信息。

checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));

内存分配：使用cudaMallocHost分配CPU上可访问的页锁定内存，cudaMalloc分配设备上的内存。

int *a = 0;
checkCudaErrors(cudaMallocHost((void **)&a, nbytes));

设置线程块和网格：这里将线程块大小设置为512个线程，网格大小根据数据大小动态计算。

dim3 threads = dim3(512, 1);
dim3 blocks = dim3(n / threads.x, 1);

创建CUDA事件和计时器：CUDA事件用于记录时间，计时器用于测量CPU执行时间。

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

CUDA流处理：使用cudaMemcpyAsync进行异步内存拷贝，<<<blocks, threads>>>语法启动并发执行的CUDA内核函数increment_kernel。

cudaMemcpyAsync(d_a, a, nbytes, cudaMemcpyHostToDevice, 0);
increment_kernel<<<blocks, threads, 0, 0>>>(d_a, value);
cudaMemcpyAsync(a, d_a, nbytes, cudaMemcpyDeviceToHost, 0);

计时和等待：cudaEventRecord记录事件，用于计算GPU执行时间。通过cudaEventQuery(stop)等待GPU操作完成。

cudaEventRecord(start, 0);
// ...
cudaEventRecord(stop, 0);

结果验证：使用correct_output函数验证GPU计算结果的正确性。

bool bFinalResults = correct_output(a, n, value);

资源释放：释放分配的内存和CUDA事件。

checkCudaErrors(cudaEventDestroy(start));
checkCudaErrors(cudaEventDestroy(stop));
checkCudaErrors(cudaFreeHost(a));
checkCudaErrors(cudaFree(d_a));

CUDA内核函数 increment_kernel ：

这个简单的CUDA内核函数用于将数组中的每个元素增加一个指定的值inc_value。blockIdx.x和threadIdx.x用于计算每个线程的全局索引idx，然后执行加法操作。

__global__ void increment_kernel(int *g_data, int inc_value) {
  int idx = blockIdx.x * blockDim.x + threadIdx.x;
  g_data[idx] = g_data[idx] + inc_value;
}

其他辅助函数
checkCudaErrors：检查CUDA函数调用是否出错。
sdkCreateTimer和sdkResetTimer：用于创建和重置计时器。
sdkStartTimer和sdkStopTimer：用于启动和停止计时器并记录CPU执行时间。