Merge remote-tracking branch 'nccl/master' into develop
此提交包含在：
Marzieh Berenjkoub
+118
-166
@@ -23,10 +23,13 @@ const char* ncclFuncToString(ncclFunc_t fn) {
|
||||
switch (fn) {
|
||||
case ncclFuncAllGather: return "AllGather";
|
||||
case ncclFuncAllReduce: return "AllReduce";
|
||||
case ncclFuncAlltoAll: return "AlltoAll";
|
||||
case ncclFuncBroadcast: return "Broadcast";
|
||||
case ncclFuncGather: return "Gather";
|
||||
case ncclFuncRecv: return "Recv";
|
||||
case ncclFuncReduce: return "Reduce";
|
||||
case ncclFuncReduceScatter: return "ReduceScatter";
|
||||
case ncclFuncScatter: return "Scatter";
|
||||
case ncclFuncSendRecv: return "SendRecv";
|
||||
case ncclFuncSend: return "Send";
|
||||
default: return "Invalid";
|
||||
@@ -85,7 +88,6 @@ const char* ncclProtoToString(int proto) {
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAllGather, const void* sendbuff, void* recvbuff, size_t sendcount,
|
||||
ncclDataType_t datatype, ncclComm_t comm, cudaStream_t stream);
|
||||
|
||||
ncclResult_t ncclAllGather_impl(const void* sendbuff, void* recvbuff, size_t sendcount,
|
||||
ncclDataType_t datatype, ncclComm_t comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AllGather, NcclNvtxParamsAllGather,
|
||||
@@ -148,10 +150,101 @@ ncclResult_t ncclAllGather_impl(const void* sendbuff, void* recvbuff, size_t sen
|
||||
}
|
||||
}
|
||||
|
||||
RCCL_PARAM(AlltoAllPivotEnable, "ALL_TO_ALL_PIVOT_ENABLE", 0);
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAlltoAll, const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclComm* comm, cudaStream_t stream);
|
||||
ncclResult_t ncclAlltoAll_impl(const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclComm* comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AlltoAll, NcclNvtxParamsAlltoAll,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, count * ncclTypeSize(datatype), datatype));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrAllToAll, sendbuff, recvbuff, count, datatype, comm, stream));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, nullptr, nullptr, recvbuff, nullptr, nullptr,
|
||||
count, datatype, 0, 0, ncclSum, mscclFuncAllToAll, comm, stream);
|
||||
}
|
||||
|
||||
size_t rankOffset = count * ncclTypeSize(datatype);
|
||||
size_t rankAlign = rankOffset & ((~rankOffset) + 1);
|
||||
size_t msgSize = count * ncclTypeSize(datatype) * comm->nRanks;
|
||||
|
||||
struct ncclInfo info;
|
||||
if (comm->topo->pivotA2AEnabled && comm->nChannels >= comm->topo->pivotA2ANumBiRings * 2 &&
|
||||
rankOffset >= 744 * 1024 && rankAlign != 4 && rcclParamAlltoAllPivotEnable()) {
|
||||
info = { ncclFuncAlltoAllPivot, "AlltoAllPivot",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream, /* Args */
|
||||
ALLTOALL_PIVOT_CHUNKSTEPS, ALLTOALL_PIVOT_SLICESTEPS, nullptr };
|
||||
} else {
|
||||
#ifdef ENABLE_ROCSHMEM
|
||||
if (rcclUseAllToAllGda(comm) && msgSize <= comm->rocshmemThreshold) {
|
||||
struct ncclInfo info = { ncclFuncAllToAllGda, "AllToAllGda",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream,
|
||||
ALLTOALL_PIVOT_CHUNKSTEPS, ALLTOALL_PIVOT_SLICESTEPS, nullptr };
|
||||
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
#endif ENABLE_ROCSHMEM
|
||||
info = { ncclFuncAlltoAll, "AlltoAll",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream, /* Args */
|
||||
ALLTOALL_CHUNKSTEPS, ALLTOALL_SLICESTEPS };
|
||||
}
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAlltoAllv, const void *sendbuff, const size_t sendcounts[], const size_t sdispls[],
|
||||
void *recvbuff, const size_t recvcounts[], const size_t rdispls[],
|
||||
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
|
||||
ncclResult_t ncclAlltoAllv_impl(const void *sendbuff, const size_t sendcounts[], const size_t sdispls[],
|
||||
void *recvbuff, const size_t recvcounts[], const size_t rdispls[],
|
||||
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AlltoAllv, NcclNvtxParamsAlltoAllv,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, sendcounts[comm->rank] * ncclTypeSize(datatype),
|
||||
recvcounts[comm->rank] * ncclTypeSize(datatype), datatype));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrAllToAllv, sendbuff, recvbuff, 0, datatype, comm, stream, -1, sendcounts, sdispls, recvcounts, rdispls));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, sendcounts, sdispls, recvbuff, recvcounts, rdispls,
|
||||
0, datatype, 0, 0, ncclSum, mscclFuncAllToAllv, comm, stream);
|
||||
}
|
||||
|
||||
int nRanks;
|
||||
NCCLCHECK(ncclCommCount(comm, &nRanks));
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(true);
|
||||
NCCLCHECK(ncclGroupStart());
|
||||
for (int r=0; r<nRanks; r++) {
|
||||
NCCLCHECK(ncclSend(
|
||||
((char*)sendbuff) + sdispls[r]*ncclTypeSize(datatype),
|
||||
sendcounts[r],
|
||||
datatype,
|
||||
r,
|
||||
comm,
|
||||
stream));
|
||||
NCCLCHECK(ncclRecv(
|
||||
((char*)recvbuff) + rdispls[r]*ncclTypeSize(datatype),
|
||||
recvcounts[r],
|
||||
datatype,
|
||||
r,
|
||||
comm,
|
||||
stream));
|
||||
}
|
||||
NCCLCHECK(ncclGroupEnd());
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(false);
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAllReduce, const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, ncclComm* comm, cudaStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclAllReduce_impl(const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, ncclComm* comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AllReduce, NcclNvtxParamsAllReduce,
|
||||
@@ -202,116 +295,8 @@ ncclResult_t ncclAllReduceWithBias_impl(const void* sendbuff, void* recvbuff, si
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
|
||||
RCCL_PARAM(AllToAllPivotEnable, "ALL_TO_ALL_PIVOT_ENABLE", 0);
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAllToAll, const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype,
|
||||
ncclComm_t comm, hipStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclAllToAll_impl(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype,
|
||||
ncclComm_t comm, hipStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AllToAll, NcclNvtxParamsAllToAll,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, count * ncclTypeSize(datatype), datatype));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrAllToAll, sendbuff, recvbuff, count, datatype, comm, stream));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, nullptr, nullptr, recvbuff, nullptr, nullptr,
|
||||
count, datatype, 0, 0, ncclSum, mscclFuncAllToAll, comm, stream);
|
||||
}
|
||||
|
||||
size_t rankOffset = count * ncclTypeSize(datatype);
|
||||
size_t rankAlign = rankOffset & ((~rankOffset) + 1);
|
||||
size_t msgSize = count * ncclTypeSize(datatype) * comm->nRanks;
|
||||
|
||||
// Determine Pivot A2A support now that we know number of channels
|
||||
if (comm->topo->pivotA2AEnabled && comm->nChannels >= comm->topo->pivotA2ANumBiRings * 2 &&
|
||||
rankOffset >= 744 * 1024 && rankAlign != 4 && rcclParamAllToAllPivotEnable()) {
|
||||
struct ncclInfo info = { ncclFuncAllToAllPivot, "AllToAllPivot",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream, /* Args */
|
||||
ALLTOALL_PIVOT_CHUNKSTEPS, ALLTOALL_PIVOT_SLICESTEPS, nullptr };
|
||||
return ncclEnqueueCheck(&info);
|
||||
} else {
|
||||
#ifdef ENABLE_ROCSHMEM
|
||||
if (rcclUseAllToAllGda(comm) && msgSize <= comm->rocshmemThreshold) {
|
||||
struct ncclInfo info = { ncclFuncAllToAllGda, "AllToAllGda",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream,
|
||||
ALLTOALL_PIVOT_CHUNKSTEPS, ALLTOALL_PIVOT_SLICESTEPS, nullptr };
|
||||
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
#endif
|
||||
int nRanks;
|
||||
//comm->isA2a = 0;
|
||||
NCCLCHECK(ncclCommCount(comm, &nRanks));
|
||||
if (count == 0) return ncclSuccess;
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(true);
|
||||
NCCLCHECK(ncclGroupStart());
|
||||
for (int r=0; r<nRanks; r++) {
|
||||
NCCLCHECK(ncclSend(((char*)sendbuff)+r*rankOffset, count, datatype, r, comm, stream));
|
||||
NCCLCHECK(ncclRecv(((char*)recvbuff)+r*rankOffset, count, datatype, r, comm, stream));
|
||||
}
|
||||
NCCLCHECK(ncclGroupEnd());
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(false);
|
||||
return ncclSuccess;
|
||||
}
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclAllToAllv, const void *sendbuff, const size_t sendcounts[], const size_t sdispls[],
|
||||
void *recvbuff, const size_t recvcounts[], const size_t rdispls[],
|
||||
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclAllToAllv_impl(const void *sendbuff, const size_t sendcounts[], const size_t sdispls[],
|
||||
void *recvbuff, const size_t recvcounts[], const size_t rdispls[],
|
||||
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(AllToAllv, NcclNvtxParamsAllToAllv,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, sendcounts[comm->rank] * ncclTypeSize(datatype),
|
||||
recvcounts[comm->rank] * ncclTypeSize(datatype), datatype));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrAllToAllv, sendbuff, recvbuff, 0, datatype, comm, stream, -1, sendcounts, sdispls, recvcounts, rdispls));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, sendcounts, sdispls, recvbuff, recvcounts, rdispls,
|
||||
0, datatype, 0, 0, ncclSum, mscclFuncAllToAllv, comm, stream);
|
||||
}
|
||||
|
||||
int nRanks;
|
||||
NCCLCHECK(ncclCommCount(comm, &nRanks));
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(true);
|
||||
NCCLCHECK(ncclGroupStart());
|
||||
for (int r=0; r<nRanks; r++) {
|
||||
NCCLCHECK(ncclSend(
|
||||
((char*)sendbuff) + sdispls[r]*ncclTypeSize(datatype),
|
||||
sendcounts[r],
|
||||
datatype,
|
||||
r,
|
||||
comm,
|
||||
stream));
|
||||
NCCLCHECK(ncclRecv(
|
||||
((char*)recvbuff) + rdispls[r]*ncclTypeSize(datatype),
|
||||
recvcounts[r],
|
||||
datatype,
|
||||
r,
|
||||
comm,
|
||||
stream));
|
||||
}
|
||||
NCCLCHECK(ncclGroupEnd());
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(false);
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclBroadcast, const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, int root,
|
||||
ncclComm_t comm, cudaStream_t stream);
|
||||
|
||||
ncclResult_t ncclBroadcast_impl(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, int root,
|
||||
ncclComm_t comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Broadcast, NcclNvtxParamsBroadcast,
|
||||
@@ -343,46 +328,32 @@ ncclResult_t ncclBcast(void* buff, size_t count, ncclDataType_t datatype, int ro
|
||||
return ncclBroadcast(buff, buff, count, datatype, root, comm, stream);
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclGather, const void* sendbuff, void* recvbuff, size_t sendcount,
|
||||
ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream);
|
||||
|
||||
ncclResult_t ncclGather_impl(const void* sendbuff, void* recvbuff, size_t sendcount,
|
||||
ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream) {
|
||||
NCCL_API(ncclResult_t, ncclGather, const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, int root,
|
||||
ncclComm* comm, cudaStream_t stream);
|
||||
ncclResult_t ncclGather_impl(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, int root,
|
||||
ncclComm* comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Gather, NcclNvtxParamsGather,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, sendcount * ncclTypeSize(datatype), root, datatype));
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, count * ncclTypeSize(datatype), root));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrGather, sendbuff, recvbuff, sendcount, datatype, comm, stream, root));
|
||||
NCCLCHECK(Recorder::instance().record(rrGather, sendbuff, recvbuff, count, datatype, comm, stream, root));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, nullptr, nullptr, recvbuff, nullptr, nullptr,
|
||||
sendcount, datatype, root, 0, ncclSum, mscclFuncGather, comm, stream);
|
||||
count, datatype, root, 0, ncclSum, mscclFuncGather, comm, stream);
|
||||
}
|
||||
|
||||
int nRanks;
|
||||
NCCLCHECK(ncclCommCount(comm, &nRanks));
|
||||
size_t rankOffset = sendcount * ncclTypeSize(datatype);
|
||||
if (sendcount == 0) return ncclSuccess;
|
||||
int rank;
|
||||
NCCLCHECK(ncclCommUserRank(comm, &rank));
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(true);
|
||||
NCCLCHECK(ncclGroupStart());
|
||||
if (rank == root) {
|
||||
for (int r=0; r<nRanks; r++)
|
||||
NCCLCHECK(ncclRecv(((char*)recvbuff)+r*rankOffset, sendcount, datatype, r, comm, stream));
|
||||
}
|
||||
NCCLCHECK(ncclSend(sendbuff, sendcount, datatype, root, comm, stream));
|
||||
NCCLCHECK(ncclGroupEnd());
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(false);
|
||||
return ncclSuccess;
|
||||
struct ncclInfo info = { ncclFuncGather, "Gather",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, root, comm, stream, /* Args */
|
||||
GATHER_CHUNKSTEPS, GATHER_SLICESTEPS };
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclReduce, const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream);
|
||||
|
||||
ncclResult_t ncclReduce_impl(const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Reduce, NcclNvtxParamsReduce,
|
||||
@@ -408,8 +379,6 @@ ncclResult_t ncclReduce_impl(const void* sendbuff, void* recvbuff, size_t count,
|
||||
|
||||
NCCL_API(ncclResult_t, ncclReduceScatter, const void* sendbuff, void* recvbuff, size_t recvcount,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, ncclComm* comm, cudaStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclReduceScatter_impl(const void* sendbuff, void* recvbuff, size_t recvcount,
|
||||
ncclDataType_t datatype, ncclRedOp_t op, ncclComm* comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(ReduceScatter, NcclNvtxParamsReduceScatter,
|
||||
@@ -433,48 +402,32 @@ ncclResult_t ncclReduceScatter_impl(const void* sendbuff, void* recvbuff, size_t
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclScatter, const void* sendbuff, void* recvbuff, size_t recvcount, ncclDataType_t datatype, int root,
|
||||
ncclComm_t comm, hipStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclScatter_impl(const void* sendbuff, void* recvbuff, size_t recvcount, ncclDataType_t datatype, int root,
|
||||
ncclComm_t comm, hipStream_t stream) {
|
||||
NCCL_API(ncclResult_t, ncclScatter, const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, int root, ncclComm* comm, cudaStream_t stream);
|
||||
ncclResult_t ncclScatter_impl(const void* sendbuff, void* recvbuff, size_t count,
|
||||
ncclDataType_t datatype, int root, ncclComm* comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Scatter, NcclNvtxParamsScatter,
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, recvcount * ncclTypeSize(datatype), root, datatype));
|
||||
NVTX3_PAYLOAD(comm ? comm->commHash : 0, count * ncclTypeSize(datatype), root, datatype));
|
||||
|
||||
if (!mscclIsCaller()) // when msccl falls back to
|
||||
{
|
||||
NCCLCHECK(Recorder::instance().record(rrScatter, sendbuff, recvbuff, recvcount, datatype, comm, stream, root));
|
||||
NCCLCHECK(Recorder::instance().record(rrScatter, sendbuff, recvbuff, count, datatype, comm, stream, root));
|
||||
}
|
||||
|
||||
if (mscclAvailable(comm) && !mscclIsCaller()) {
|
||||
return mscclEnqueueCheck(
|
||||
sendbuff, nullptr, nullptr, recvbuff, nullptr, nullptr,
|
||||
recvcount, datatype, root, 0, ncclSum, mscclFuncScatter, comm, stream);
|
||||
count, datatype, root, 0, ncclSum, mscclFuncScatter, comm, stream);
|
||||
}
|
||||
|
||||
int nRanks;
|
||||
NCCLCHECK(ncclCommCount(comm, &nRanks));
|
||||
size_t rankOffset = recvcount * ncclTypeSize(datatype);
|
||||
if (recvcount == 0) return ncclSuccess;
|
||||
int rank;
|
||||
NCCLCHECK(ncclCommUserRank(comm, &rank));
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(true);
|
||||
NCCLCHECK(ncclGroupStart());
|
||||
if (rank == root) {
|
||||
for (int r=0; r<nRanks; r++)
|
||||
NCCLCHECK(ncclSend(((char*)sendbuff)+r*rankOffset, recvcount, datatype, r, comm, stream));
|
||||
}
|
||||
NCCLCHECK(ncclRecv(recvbuff, recvcount, datatype, root, comm, stream));
|
||||
NCCLCHECK(ncclGroupEnd());
|
||||
if (!mscclIsCaller()) Recorder::instance().skip(false);
|
||||
return ncclSuccess;
|
||||
struct ncclInfo info = { ncclFuncScatter, "Scatter",
|
||||
sendbuff, recvbuff, count, datatype, ncclSum, root, comm, stream, /* Args */
|
||||
SCATTER_CHUNKSTEPS, SCATTER_SLICESTEPS };
|
||||
return ncclEnqueueCheck(&info);
|
||||
}
|
||||
|
||||
NCCL_API(ncclResult_t, ncclSend, const void* sendbuff, size_t count, ncclDataType_t datatype, int peer,
|
||||
ncclComm_t comm, cudaStream_t stream);
|
||||
|
||||
|
||||
ncclResult_t ncclSend_impl(const void* sendbuff, size_t count, ncclDataType_t datatype, int peer,
|
||||
ncclComm_t comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Send, NcclNvtxParamsSendRecv,
|
||||
@@ -500,7 +453,6 @@ ncclResult_t ncclSend_impl(const void* sendbuff, size_t count, ncclDataType_t da
|
||||
|
||||
NCCL_API(ncclResult_t, ncclRecv, void* recvbuff, size_t count, ncclDataType_t datatype, int peer,
|
||||
ncclComm_t comm, cudaStream_t stream);
|
||||
|
||||
ncclResult_t ncclRecv_impl(void* recvbuff, size_t count, ncclDataType_t datatype, int peer,
|
||||
ncclComm_t comm, cudaStream_t stream) {
|
||||
NVTX3_FUNC_WITH_PARAMS(Recv, NcclNvtxParamsSendRecv,
|
||||
|
||||
新增問題並參考
封鎖使用者