/*************************************************************************
 * 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 "common.h"
#include "cuda_runtime.h"

void
ReduceScatterGetCollByteCount(size_t* sendcount, size_t* recvcount, size_t* paramcount,
                              size_t* sendInplaceOffset, size_t* recvInplaceOffset,
                              size_t count, size_t eltSize, int nranks)
{
    size_t base        = (count / nranks) & -(16 / eltSize);
    *sendcount         = base * nranks;
    *recvcount         = base;
    *sendInplaceOffset = 0;
    *recvInplaceOffset = base;
    *paramcount        = base;
}

testResult_t
ReduceScatterInitData(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;

    for(int i = 0; i < args->nGpus; i++)
    {
        CUDACHECK(cudaSetDevice(args->gpus[i]));
        int rank = ((args->proc * args->nThreads + args->thread) * args->nGpus + i);
        CUDACHECK(cudaMemset(args->recvbuffs[i], 0, args->expectedBytes));
        void* data = in_place ? args->recvbuffs[i] : args->sendbuffs[i];
        TESTCHECK(InitData(data, sendcount, 0, type, op, rep, nranks, rank));
        CUDACHECK(cudaMemcpy(args->expected[i], args->recvbuffs[i], args->expectedBytes,
                             cudaMemcpyDefault));
        TESTCHECK(InitDataReduce(args->expected[i], recvcount, rank * recvcount, type, op,
                                 rep, nranks));
        CUDACHECK(cudaDeviceSynchronize());
    }
    return testSuccess;
}

void
ReduceScatterGetBw(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
ReduceScatterRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type,
                     ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream)
{
    NCCLCHECK(ncclReduceScatter(sendbuff, recvbuff, count, type, op, comm, stream));
    return testSuccess;
}

struct testColl reduceScatterTest = { "ReduceScatter", ReduceScatterGetCollByteCount,
                                      ReduceScatterInitData, ReduceScatterGetBw,
                                      ReduceScatterRunColl };

void
ReduceScatterGetBuffSize(size_t* sendcount, size_t* recvcount, size_t count, int nranks)
{
    size_t paramcount, sendInplaceOffset, recvInplaceOffset;
    ReduceScatterGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset,
                                  &recvInplaceOffset, count, /*eltSize=*/1, nranks);
}

testResult_t
ReduceScatterRunTest(struct threadArgs* args, int root, ncclDataType_t type,
                     const char* typeName, ncclRedOp_t op, const char* opName)
{
    args->collTest = &reduceScatterTest;
    ncclDataType_t* run_types;
    ncclRedOp_t*    run_ops;
    const char **   run_typenames, **run_opnames;
    int             type_count, op_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;
    }

    if((int) op != -1)
    {
        run_ops     = &op;
        run_opnames = &opName;
        op_count    = 1;
    }
    else
    {
        op_count    = test_opnum;
        run_ops     = test_ops;
        run_opnames = test_opnames;
    }

    for(int i = 0; i < type_count; i++)
    {
        for(int j = 0; j < op_count; j++)
        {
#if defined(RCCL_FLOAT8)
            if((run_types[i] == ncclFp8E4M3 || run_types[i] == ncclFp8E5M2) &&
               run_ops[j] == ncclProd)
                continue;
#endif
            TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j],
                               run_opnames[j], -1));
        }
    }
    return testSuccess;
}

struct testEngine ncclTestEngine = { ReduceScatterGetBuffSize, ReduceScatterRunTest };