/************************************************************************* * Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved. * Modifications Copyright (c) 2019-2022 Advanced Micro Devices, Inc. All rights reserved. * * See LICENSE.txt for license information ************************************************************************/ #include #include "common.h" #define ALIGN 4 void AllGatherGetCollByteCount(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 AllGatherInitData(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; inGpus; i++) { HIPCHECK(hipSetDevice(args->gpus[i])); for (int l=0; lnRanks; 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; jexpected[k])+args->sendBytes*j, sendcount, 0, type, ncclSum, 33*rep + j, 1, 0)); } k++; } HIPCHECK(hipDeviceSynchronize()); } return testSuccess; } void AllGatherGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) { double baseBw = (double)(count * typesize * nranks) / 1.0E9 / sec; *algBw = baseBw; double factor = ((double)(nranks - 1))/((double)nranks); *busBw = baseBw * factor; } testResult_t AllGatherRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, hipStream_t stream) { NCCLCHECK(ncclAllGather(sendbuff, recvbuff, count, type, comm, stream)); return testSuccess; } struct testColl allGatherTest = { "AllGather", AllGatherGetCollByteCount, AllGatherInitData, AllGatherGetBw, AllGatherRunColl }; void AllGatherGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) { size_t paramcount, sendInplaceOffset, recvInplaceOffset; AllGatherGetCollByteCount(sendcount, recvcount, ¶mcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks); } testResult_t AllGatherRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) { args->collTest = &allGatherTest; 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; } for (int i=0; i