Initial support for clique-based kernels (#276)

* Initial support for clique-based kernels
This commit is contained in:
gilbertlee-amd
2020-10-13 11:22:04 -06:00
committato da GitHub
parent ae008fd2db
commit 2b8184808d
27 ha cambiato i file con 1969 aggiunte e 7 eliminazioni
+1 -1
Vedi File
@@ -12,7 +12,7 @@
ncclResult_t bootstrapNetInit();
ncclResult_t bootstrapCreateRoot(ncclUniqueId* commId, bool idFromEnv);
ncclResult_t bootstrapGetUniqueId(ncclUniqueId* out);
ncclResult_t bootstrapInit(ncclUniqueId* id, int rank, int nranks, void** commState);
ncclResult_t bootstrapInit(ncclUniqueId* id, int rank, int nranks, void** commState, int* rootPid); // [RCCL] Adding rootPid
ncclResult_t bootstrapAllGather(void* commState, void* allData, int size);
ncclResult_t bootstrapSend(void* commState, int peer, void* data, int size);
ncclResult_t bootstrapRecv(void* commState, int peer, void* data, int size);
+9 -2
Vedi File
@@ -10,6 +10,9 @@
#include "transport.h"
#include "p2p.h"
// [RCCL]
#include "clique/CliqueManager.h"
// [/RCCL]
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#else
@@ -143,8 +146,12 @@ struct ncclComm {
//list of async p2p operation queued in a group semantics
struct ncclP2Plist p2plist;
// RCCL AllToAll/Scatter/Gather API
bool alltoallDisable;
// [RCCL]
bool alltoallDisable; // RCCL AllToAll/Scatter/Gather API
CliqueManager* cliqueManager; // CliqueManager handles pointer collection / distribution for clique-based kernels
int rootPid; // Process ID of root
// [/RCCL]
};
#endif