hypercube.cu 4.21 KB
Newer Older
jerrrrry's avatar
jerrrrry committed
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
/*************************************************************************
 * Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
 * Modifications Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

#include "hip/hip_runtime.h"
#include "common.h"

#define ALIGN 4

void HyperCubeGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
  size_t base = (count/(ALIGN*nranks))*ALIGN;
  *sendcount = base;
  *recvcount = base*nranks;
  *sendInplaceOffset = base;
  *recvInplaceOffset = 0;
  *paramcount = base;
}

testResult_t HyperCubeInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
  size_t sendcount = args->sendBytes / wordSize(type);
  size_t recvcount = args->expectedBytes / wordSize(type);
  int nranks = args->nProcs*args->nThreads*args->nGpus*args->nRanks;

  int k=0;
  for (int i=0; i<args->nGpus; i++) {
    HIPCHECK(hipSetDevice(args->gpus[i]));

    for (int l=0; l<args->nRanks; l++) {
      int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus*args->nRanks + i*args->nRanks + l);
      HIPCHECK(hipMemset(args->recvbuffs[k], 0, args->expectedBytes));
      void* data = in_place ? ((char*)args->recvbuffs[k])+rank*args->sendBytes : args->sendbuffs[k];
      TESTCHECK(InitData(data, sendcount, 0, type, ncclSum, 33*rep + rank, 1, 0));
      for (int j=0; j<nranks; j++) {
	TESTCHECK(InitData(((char*)args->expected[k])+args->sendBytes*j, sendcount, 0, type, ncclSum, 33*rep + j, 1, 0));
      }
      k++;
    }
    HIPCHECK(hipDeviceSynchronize());
  }
  return testSuccess;
}

void HyperCubeGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
  double baseBw = (double)(count * typesize * (nranks - 1)) / 1.0E9 / sec;

  *algBw = baseBw;
  double factor = 1;
  *busBw = baseBw * factor;
}

testResult_t HyperCubeRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, hipStream_t stream) {
  char* sbuff = (char*)sendbuff;
  char* rbuff = (char*)recvbuff;
  int nRanks;
  NCCLCHECK(ncclCommCount(comm, &nRanks));
  int rank;
  NCCLCHECK(ncclCommUserRank(comm, &rank));
  size_t rankSize = count * wordSize(type);
  if (rbuff+rank*rankSize != sbuff) HIPCHECK(hipMemcpyAsync(rbuff+rank*rankSize, sbuff, rankSize, hipMemcpyDeviceToDevice, stream));

  // Hypercube AllGather
  for (int mask=1; mask<nRanks; mask<<=1) {
    NCCLCHECK(ncclGroupStart());
    int s = rank & ~(mask-1);
    int r = s ^ mask;
    NCCLCHECK(ncclSend(rbuff+s*rankSize, count*mask, type, rank^mask, comm, stream));
    NCCLCHECK(ncclRecv(rbuff+r*rankSize, count*mask, type, rank^mask, comm, stream));
    NCCLCHECK(ncclGroupEnd());
  }
  return testSuccess;
}

struct testColl hyperCubeTest = {
  "HyperCube",
  HyperCubeGetCollByteCount,
  HyperCubeInitData,
  HyperCubeGetBw,
  HyperCubeRunColl
};

void HyperCubeGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
  HyperCubeGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
}

testResult_t HyperCubeRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
  args->collTest = &hyperCubeTest;
  ncclDataType_t *run_types;
  const char **run_typenames;
  int type_count;

  if ((int)type != -1) {
    type_count = 1;
    run_types = &type;
    run_typenames = &typeName;
  } else {
    type_count = test_typenum;
    run_types = test_types;
    run_typenames = test_typenames;
  }

  // Check if this is a power of 2
  int nRanks = args->nProcs*args->nThreads*args->nGpus;
  if (nRanks && !(nRanks & (nRanks - 1))) {
    for (int i=0; i<type_count; i++) {
      TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], (ncclRedOp_t)0, "", -1));
    }
  } else {
    printf("nRanks %d is not a power of 2, skipping\n", nRanks);
  }

  return testSuccess;
}

struct testEngine ncclTestEngine = {
  HyperCubeGetBuffSize,
  HyperCubeRunTest
};