checkgpus/data_test.cpp at main · dylanjude/checkgpus · GitHub

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
106
107
108
109
110
111
112
113
114
115
116
117
#include <cstdio>
#include <cstdlib>
#include <string>
#include "mpi.h"
#include "cuda_runtime.h"
#include "helper.h"

#define DEVICE_BUFFER 1

// some analytical function to assign values to the memory buffer
double f(int i, int rank){
  return rank*1000 + i%1000;
}

void datatest(int size0, int size1){

  int myid, numprocs, i, j;
  int size;
  int namelen;

  MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
  MPI_Comm_rank(MPI_COMM_WORLD, &myid);

  int *s_buf;
  int *r_buf;

  if(numprocs < 2){
    printf("use more than 2 ranks!\n");
    return;
  }

  int *cpu_data = new int[size1];

  // initialize data on the CPU
  if(myid == 0){
    for(int i=0; i<size1; i++){
      cpu_data[i] = f(i,myid);
    }
  }

  // allocat/copy the data to the GPU
#ifdef DEVICE_BUFFER
  HANDLE_ERROR(cudaMalloc(&s_buf, size1*sizeof(int)));
  HANDLE_ERROR(cudaMalloc(&r_buf, size1*sizeof(int)));
  HANDLE_ERROR(cudaMemcpy(s_buf, cpu_data, size1*sizeof(int), cudaMemcpyHostToDevice));
#else
  s_buf = new int[size1];
  r_buf = new int[size1];
  for(int i=0; i<size1; i++){
    s_buf[i] = cpu_data[i];
  }
#endif

  MPI_Request request;
  MPI_Status  reqstat;

  if(myid == 0){
    printf("\n ----- Checking data integrity from rank 0 to all other ranks ------- \n\n");
  }

  // loop over the proc that will get data sent to
  for(int p=1; p<numprocs; p++){
    // loop over the packet size
    for(size = size0; size<=size1; size*=4){

      if(myid == 0){ // rank 0 always sends

        MPI_Send(s_buf, size, MPI_INT, p, 100, MPI_COMM_WORLD);

      } else if(myid == p){ // rank p receives

        MPI_Recv(r_buf, size, MPI_INT, 0, 100, MPI_COMM_WORLD, &reqstat);

      }

      if(myid == p){
        // copy data back to cpu memory
#ifdef DEVICE_BUFFER
        HANDLE_ERROR(cudaMemcpy(cpu_data, r_buf, size*sizeof(int), cudaMemcpyDeviceToHost));
#else
        for(int i=0; i<size; i++){
          cpu_data[i] = r_buf[i];
        }
#endif
        // check the data is correct
        bool error=false;
        for(int i=0; i<size; i++){
          if(cpu_data[i] != f(i,0)){
            error = true;
          }
        }

        if(error){
          printf("ERROR   with N=%10d, rank=%3d\n", size, myid);
        } else {
          printf("SUCCESS with N=%10d, rank=%3d\n", size, myid);
        }
      }

    }
  }

  if(myid == numprocs-1){
    printf("------------------------------------------\n");
  }

#ifdef DEVICE_BUFFER
  HANDLE_ERROR(cudaFree(s_buf));
  HANDLE_ERROR(cudaFree(r_buf));
#else
  delete[] s_buf;
  delete[] r_buf;
#endif

  delete[] cpu_data;

}