/*************************************************************************
 * Copyright (c) 2015-2019, NVIDIA CORPORATION. All rights reserved.
 * Modifications Copyright (c) 2019-2020 Advanced Micro Devices, Inc. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

#include "group.h"
#include "debug.h"
#include "enqueue.h"

#define MAX_ASYNC_OPS 128
thread_local pthread_t ncclGroupThreads[MAX_ASYNC_OPS];
thread_local int ncclGroupIndex = 0;
thread_local int ncclGroupMode = 0;
thread_local ncclResult_t ncclGroupError = ncclSuccess;

bool ncclAsyncMode() {
  return ncclGroupMode > 0;
}

ncclResult_t ncclAsyncErrCheck(ncclResult_t ret) {
  if (ncclGroupError == ncclSuccess || ret != ncclSuccess) ncclGroupError = ret;
  return ret;
}

struct ncclInitArgs {
  ncclInitFunc_t func;
  int cudaDev;
  ncclComm_t* newcomm;
  int ndev;
  ncclUniqueId commId;
  int myrank;
};
struct ncclCollArgs {
  ncclComm_t comm;
};

enum ncclAsyncFuncType {
  ASYNC_FUNC_INVALID = 0,
  ASYNC_FUNC_INIT = 1,
  ASYNC_FUNC_COLL = 2,
};
struct ncclAsyncArgs {
  ncclResult_t ret;
  enum ncclAsyncFuncType funcType;
  union {
    ncclCollArgs coll;
    ncclInitArgs init;
  };
};

thread_local struct ncclAsyncArgs ncclGroupArgs[MAX_ASYNC_OPS];

#define CHECK(a) do { \
  if ((args->ret = (a)) != ncclSuccess) { \
    INFO(NCCL_INIT,"%s:%d -> %d [Async thread]", __FILE__, __LINE__, args->ret); \
    return args; \
  } \
} while(0)

void* ncclAsyncThreadMain(void* args_) {
  struct ncclAsyncArgs* args = (struct ncclAsyncArgs*)args_;
  CHECK(args->init.func(args->init.newcomm, args->init.ndev, args->init.commId, args->init.myrank, args->init.cudaDev));
  return args;
}

ncclResult_t ncclAsyncInit(ncclInitFunc_t func, ncclComm_t* newcomm, int ndev, ncclUniqueId commId, int myrank, int cudaDev) {
  if (ncclGroupIndex >= MAX_ASYNC_OPS) {
    WARN("Too many async operations in progress, max is %d", MAX_ASYNC_OPS);
    return ncclAsyncErrCheck(ncclInvalidUsage);
  }
  int index = ncclGroupIndex++;
  struct ncclAsyncArgs* args = ncclGroupArgs+index;
  args->funcType = ASYNC_FUNC_INIT;
  args->init.func = func;
  args->init.cudaDev = cudaDev;
  args->init.newcomm = newcomm;
  args->init.ndev = ndev;
  memcpy(&args->init.commId, &commId, sizeof(commId));
  args->init.myrank = myrank;
  return ncclSuccess;
}

ncclResult_t ncclAsyncColl(ncclComm_t comm) {
  struct ncclAsyncArgs* args = ncclGroupArgs;
  for (int i=0; i<ncclGroupIndex; i++) {
    if (args->coll.comm == comm) return ncclSuccess;
    args++;
  }
  if (ncclGroupIndex >= MAX_ASYNC_OPS) {
    WARN("Too many async operations in progress, max is %d", MAX_ASYNC_OPS);
    return ncclAsyncErrCheck(ncclInvalidUsage);
  }
  ncclGroupIndex++;
  args->funcType = ASYNC_FUNC_COLL;
  args->coll.comm = comm;
  return ncclSuccess;
}

NCCL_API(ncclResult_t, ncclGroupStart);
ncclResult_t ncclGroupStart() {
  ncclGroupMode++;
  return ncclSuccess;
}

NCCL_API(ncclResult_t, ncclGroupEnd);
ncclResult_t ncclGroupEnd() {
  ncclGroupMode--;
  if (ncclGroupMode > 0) return ncclSuccess;
  int savedDev;
  CUDACHECK(hipGetDevice(&savedDev));
  int done = ncclGroupIndex;
  int doneArray[MAX_ASYNC_OPS];
  for (int i=0; i<ncclGroupIndex; i++) doneArray[i] = 0;

  ncclResult_t ret = ncclGroupError;
  if (ret != ncclSuccess) goto group_cleanup;

  /* Launch async ncclCommInitRank */
  for (int i=0; i<ncclGroupIndex; i++) {
    struct ncclAsyncArgs* args = ncclGroupArgs+i;
    if (args->funcType == ASYNC_FUNC_INIT) {
      pthread_create(ncclGroupThreads+i, NULL, ncclAsyncThreadMain, args);
    }
  }

  /* Collectives are done in three steps :
   * 1. Barrier Check In. Only the last call may call cudaLaunchKernel[cooperative]
   * 2. Barrier Wait. No CUDA call is permitted
   * 3. Enqueue Events. CUDA event wait/enqueue.
   * This is needed because step 2 cannot call any CUDA primitive, otherwise if
   * hipFree happens between 1 and 3, it could block that CUDA call and
   * prevent some ranks from launching their network threads, which would
   * prevent the NCCL call from completing, blocking the hipFree call.
   */
  for (int i=0; i<ncclGroupIndex; i++) {
    struct ncclAsyncArgs* args = ncclGroupArgs+i;
    if (args->funcType == ASYNC_FUNC_COLL) {
      if (args->coll.comm->userStream == NULL)
        CUDACHECKGOTO(hipSetDevice(args->coll.comm->cudaDev), ret, end);
      NCCLCHECKGOTO(ncclBarrierEnqueue(args->coll.comm), ret, end);
    }
  }
  for (int i=0; i<ncclGroupIndex; i++) {
    struct ncclAsyncArgs* args = ncclGroupArgs+i;
    if (args->funcType == ASYNC_FUNC_COLL) {
      CUDACHECKGOTO(hipSetDevice(args->coll.comm->cudaDev), ret, end);
      NCCLCHECKGOTO(ncclBarrierEnqueueWait(args->coll.comm), ret, end);
    }
  }
  for (int i=0; i<ncclGroupIndex; i++) {
    struct ncclAsyncArgs* args = ncclGroupArgs+i;
    if (args->funcType == ASYNC_FUNC_COLL) {
      if (args->coll.comm->userStream == NULL)
        CUDACHECKGOTO(hipSetDevice(args->coll.comm->cudaDev), ret, end);
      NCCLCHECKGOTO(ncclEnqueueEvents(args->coll.comm), ret, end);
      doneArray[i] = 1;
      done--;
    }
  }

  /* For init, since we use threads, we just wait for threads to complete */
  while (done) {
    for (int i=0; i<ncclGroupIndex; i++) {
      struct ncclAsyncArgs* args = ncclGroupArgs+i;
      if (args->funcType == ASYNC_FUNC_INIT && doneArray[i] == 0) {
        int err = pthread_tryjoin_np(ncclGroupThreads[i], NULL);
        if (err == EBUSY) continue;
        if (err != 0) ret = ncclSystemError;
        if (args->ret != ncclSuccess) ret = args->ret;
        doneArray[i] = 1;
        done--;
      }
    }
  }
  goto end;
group_cleanup:
  if (ret != ncclSuccess) {
    // At least one call in the group failed. Since we want to make that group
    // an atomic operation, we need to cancel all operations.
    for (int i=0; i<ncclGroupIndex; i++) {
      struct ncclAsyncArgs* args = ncclGroupArgs+i;
      if (args->funcType == ASYNC_FUNC_INIT && doneArray[i] == 0) {
        if (args->init.newcomm) NCCLCHECK(ncclCommDestroy(*args->init.newcomm));
        *args->init.newcomm = NULL;
      } else {
        struct ncclComm* comm = args->coll.comm;
        for (int c=0; c<comm->nChannels; c++) {
          struct ncclChannel* channel = comm->channels+c;
          for (int i=0; i<channel->collCount; i++) {
            channel->collectives[(channel->collStart + i)%NCCL_MAX_OPS].active = 0;
          }
          channel->collFifoTail = channel->collStart;
          channel->collCount = 0;
        }
        /* Cancel all proxy ops : mark them as ncclProxyOpNone and they should be freed later on */
        struct ncclProxyState* state = &comm->proxyState;
        struct ncclProxyArgs *op, *start;
        pthread_mutex_lock(&state->mutex);
        op = start = state->ops;
        while (op) {
          if (op->opCount >= comm->lastOpCount) op->state = ncclProxyOpNone;
          struct ncclProxyArgs* peerOp = op->nextPeer;
          while (peerOp) {
            if (peerOp->opCount >= comm->lastOpCount) peerOp->state = ncclProxyOpNone;
            peerOp = peerOp->nextPeer;
          }
          op = op->next;
          if (op == start) break;
        }
        comm->opCount = comm->lastOpCount;
        pthread_cond_signal(&state->cond);
        pthread_mutex_unlock(&state->mutex);

        comm->myParams->gridDim.x = comm->myParams->blockDim.x = 0;
        comm->userStreamSet = false;
      }
    }
  }
end:
  ncclGroupError = ncclSuccess;
  ncclGroupIndex = 0;
  CUDACHECK(hipSetDevice(savedDev)); // do other clean-ups first before calling hipSetDevice, because this call can fail too
  return ret;
}