CUDA Programming - asyncAPI Learning Record

2024-07-12

asyncAPI

1. Complete code
2. Analysis of important parts

1. Complete code

#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 = %dn", 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: %.2fn", gpu_time);
  printf("time spent by CPU in CUDA calls: %.2fn", sdkGetTimerValue(&timer));
  printf("CPU executed %lu iterations while waiting for GPU to finishn",
         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);
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105

2. Analysis of important parts

equipmentinitialization: The findCudaDevice function is used to select the best CUDA device and return the device ID.

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

1
2

Get device properties: The cudaGetDeviceProperties function gets the properties of the specified device, including information such as the device name.

checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));

1
2

Memory allocation: Use cudaMallocHost to allocate page-locked memory accessible on the CPU, and cudaMalloc to allocate memory on the device.

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

1
2
3

Set thread block and grid: Here the thread block size is set to 512 threads and the grid size is calculated dynamically based on the data size.

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

1
2
3

Create CUDA events and timers: CUDA events are used to record time, and timers are used to measure CPU execution time.

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

1
2
3
4

CUDA stream processing: asynchronous memory copy using cudaMemcpyAsync, <<<blocks, threads> >> The syntax starts the concurrent execution of the CUDA kernel function 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);

1
2
3
4

Timing and waiting: cudaEventRecord records events for calculating GPU execution time. Wait for GPU operations to complete via cudaEventQuery(stop).

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

1
2
3
4

Result verification: Use the correct_output function to verify the correctness of the GPU calculation results.

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

1
2

Resource release: releases allocated memory and CUDA events.

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

1
2
3
4
5

CUDA kernel function increment_kernel:

This simple CUDA kernel function is used to increase each element in an array by a specified value inc_value. blockIdx.x and threadIdx.x are used to calculate the global index idx of each thread and then perform the addition operation.

__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;
}

1
2
3
4
5

Other helper functions
checkCudaErrors：Check CUDAWhether the function call has an error.
sdkCreateTimer and sdkResetTimer: used to create and reset timers.
sdkStartTimer and sdkStopTimer: used to start and stop the timer and record the CPU execution time.

Technology Sharing