File: EnvVars.cpp

package info (click to toggle)
rccl 5.4.3-3
  • links: PTS, VCS
  • area: main
  • in suites: sid, trixie
  • size: 4,332 kB
  • sloc: cpp: 33,357; ansic: 6,717; xml: 5,265; makefile: 508; sh: 365; awk: 243; python: 85
file content (162 lines) | stat: -rw-r--r-- 5,283 bytes parent folder | download
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
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
/*************************************************************************
 * Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

#include "EnvVars.hpp"
#include "CollectiveArgs.hpp"
#include <cstdlib>

namespace RcclUnitTesting
{
  int const UT_SINGLE_PROCESS = (1<<0);
  int const UT_MULTI_PROCESS  = (1<<1);

  hsa_status_t CountGpus(hsa_agent_t agent, void* data)
  {
    int* currCount = (int*)data;
    hsa_device_type_t device;
    hsa_agent_get_info(agent, HSA_AGENT_INFO_DEVICE, &device);
    if (device == HSA_DEVICE_TYPE_GPU)
      *currCount = *currCount + 1;
    return HSA_STATUS_SUCCESS;
  }

  EnvVars::EnvVars()
  {
    // Collect number of GPUs available
    // NOTE: Cannot use HIP call prior to launching child processes via fork so use HSA
    int numDevicesAvailable = 0;
    hsa_init();
    hsa_iterate_agents(CountGpus, &numDevicesAvailable);
    hsa_shut_down();

    showNames      = GetEnvVar("UT_SHOW_NAMES"  , 1);
    minGpus        = GetEnvVar("UT_MIN_GPUS"    , 2);
    maxGpus        = GetEnvVar("UT_MAX_GPUS"    , numDevicesAvailable);
    processMask    = GetEnvVar("UT_PROCESS_MASK", UT_SINGLE_PROCESS | UT_MULTI_PROCESS);
    verbose        = GetEnvVar("UT_VERBOSE"     , 0);
    printValues    = GetEnvVar("UT_PRINT_VALUES", 0);
    maxRanksPerGpu = GetEnvVar("UT_MAX_RANKS_PER_GPU", 1);

    // Limit number of supported reduction operators to just ncclSum if only allReduce is built
#ifdef BUILD_ALLREDUCE_ONLY
    int numOps = 1;
#else
    int numOps = ncclNumOps;
#endif
    std::vector<std::string> redOpStrings = GetEnvVarsList("UT_REDOPS");
    for (auto s : redOpStrings)
    {
      for (int i = 0; i < numOps; ++i)
      {
        if (!strcmp(s.c_str(), ncclRedOpNames[i]))
        {
          redOps.push_back((ncclRedOp_t)i);
          break;
        }
      }
    }
    // Default back to all ops if no strings are found
    if (redOps.empty())
    {
      for (int i = 0; i < numOps; i++)
        redOps.push_back((ncclRedOp_t)i);
    }

    // Limit number of supported datatypes if only allReduce is built
    std::vector<std::string> dtStrings = GetEnvVarsList("UT_DATATYPES");
    for (auto s : dtStrings)
    {
      for (int i = 0; i < ncclNumTypes; ++i)
      {
        if (!strcmp(s.c_str(), ncclDataTypeNames[i]))
        {
#ifdef BUILD_ALLREDUCE_ONLY
          if (i == ncclFloat32)
#endif
          {
            dataTypes.push_back((ncclDataType_t)i);
          }
        }
      }
    }

    // Default option if no valid datatypes are found in env var
    if (dataTypes.empty())
    {
      dataTypes.push_back(ncclFloat32);
      // Skip all but 32-bit floats if only AllReduce is being built
#ifndef BUILD_ALLREDUCE_ONLY
      dataTypes.push_back(ncclInt8);
      dataTypes.push_back(ncclUint8);
      dataTypes.push_back(ncclInt32);
      dataTypes.push_back(ncclUint32);
      dataTypes.push_back(ncclInt64);
      dataTypes.push_back(ncclUint64);
      dataTypes.push_back(ncclFloat16);
      dataTypes.push_back(ncclFloat32);
      dataTypes.push_back(ncclFloat64);
      dataTypes.push_back(ncclBfloat16);
#endif
    }
  }

  std::vector<ncclRedOp_t> const& EnvVars::GetAllSupportedRedOps()
  {
    return redOps;
  }

  std::vector<ncclDataType_t> const& EnvVars::GetAllSupportedDataTypes()
  {
    return dataTypes;
  }

  int EnvVars::GetEnvVar(std::string const varname, int defaultValue)
  {
    if (getenv(varname.c_str()))
      return atoi(getenv(varname.c_str()));
    return defaultValue;
  };

  std::vector<std::string> EnvVars::GetEnvVarsList(std::string const varname)
  {
    std::vector<std::string> result;
    if (getenv(varname.c_str()))
    {
      char* token = strtok(getenv(varname.c_str()), ",;");
      while (token != NULL)
      {
        result.push_back(token);
        token = strtok(NULL, ",;");
      }
    }
    return result;
  }

  void EnvVars::ShowConfig()
  {
    std::vector<std::pair<std::string, std::string>> supported =
      {
        std::make_pair("UT_SHOW_NAMES"       , "Show test case names"),
        std::make_pair("UT_MIN_GPUS"         , "Minimum number of GPUs to use"),
        std::make_pair("UT_MAX_GPUS"         , "Maximum number of GPUs to use"),
        std::make_pair("UT_PROCESS_MASK"     , "Whether to run single/multi process"),
        std::make_pair("UT_VERBOSE"          , "Show verbose unit test output"),
        std::make_pair("UT_REDOPS"           , "List of reduction ops to test"),
        std::make_pair("UT_DATATYPES"        , "List of datatypes to test"),
        std::make_pair("UT_MAX_RANKS_PER_GPU", "Maximum number of ranks using the same GPU"),
        std::make_pair("UT_PRINT_VALUES"     , "Print array values (# of values to print, < 0 for all)")
      };

    printf("================================================================================\n");
    printf(" Environment variables:\n");
    for (auto p : supported)
    {
      printf(" - %-20s %-40s %s\n", p.first.c_str(), p.second.c_str(),
             getenv(p.first.c_str()) ? getenv(p.first.c_str()) : "<unset>");
    }
    printf("================================================================================\n");
  }
}