AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.13.4-1

Browse Source 
Optimize CUDA graph launch; avoid launching a CPU callback for
intra-node operations.
Simplify kernel common code to improve the latency of send/recv
operations.
Strengthen CUDA streams semantics.
Change NET API to v6, to add dmabuf support.
Add ncclGetLastError() function.
Add ncclRemoteError code and use it for remote network errors.
Support the use of a different NCCL_NET parameter per communicator.
Add support for SHM and P2P transfers using cudaMemcpy.


[ROCm/rccl commit: 19ab67d172]
This commit is contained in:
Sylvain Jeaugey
2022-05-24 02:02:31 -07:00
parent 1c5734046d
commit 91154e8df9
62 changed files with 4787 additions and 2496 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/info.h
+61
View File
@@ -10,6 +10,9 @@
#include "nccl.h"
#include "devcomm.h"
#include "collectives.h"
#include "core.h"
#include "utils.h"
#include "strongstream.h"
typedef enum : uint8_t {
ncclPatternRing,
@@ -54,4 +57,62 @@ struct ncclInfo {
int channelId;
};
inline ncclResult_t ncclInfoSetDerived(struct ncclInfo* info, int nRanks) {
info->nBytes = info->count * ncclTypeSize(info->datatype);
if (info->coll == ncclFuncAllGather || info->coll == ncclFuncBroadcast) {
info->count = info->nBytes;
info->datatype = ncclInt8;
}
if (info->coll == ncclFuncAllGather || info->coll == ncclFuncReduceScatter) info->nBytes *= nRanks; // count is per rank
return ncclSuccess;
}
struct ncclTaskColl {
struct ncclTaskColl* next;
ncclFunc_t func;
void const* sendbuff;
void* recvbuff;
size_t count;
int root;
ncclDataType_t datatype;
ncclDevRedOpFull op;
int chunkSteps, sliceSteps;
};
struct ncclTaskP2p {
ncclTaskP2p *next;
void *buff;
size_t bytes;
// Stateful chunk index. If a p2p gets "cut" over two plans this keeps track
// of where it left off.
int chunk;
};
struct ncclCudaStreamList {
struct ncclCudaStreamList *next;
cudaStream_t stream;
};
struct ncclTasks {
struct Peer {
bool sendSeen, recvSeen;
struct ncclIntruQueue<struct ncclTaskP2p, &ncclTaskP2p::next> sendQueue;
struct ncclIntruQueue<struct ncclTaskP2p, &ncclTaskP2p::next> recvQueue;
};
struct ncclIntruQueue<ncclTaskColl, &ncclTaskColl::next> collQueue;
size_t collBytesTotal;
struct Peer* peers/*[nRanks]*/;
int *p2pSendOrder/*[nRanks]*/, *p2pRecvOrder/*[nRanks]*/;
int nTasksColl, nTasksP2p;
// The list of user streams aggregated over all tasks present.
struct ncclCudaStreamList* streams;
// The most recent user stream. Ignored if streams==nullptr
cudaStream_t streamRecent;
// The graph capturing all user streams or invalid if none. Thus we restrict the
// user that all streams must be captured in the same graph or not captured
// at all. Technically we could probably relax this, but that would mean
// collecting a different `ncclTasks` per graph and one for non-graph.
struct ncclCudaGraph capturingGraph;
};
#endif