Merge remote-tracking branch 'nccl/master' into develop

[ROCm/rccl commit: 84081064a0]
This commit is contained in:
BertanDogancay
2024-10-02 09:29:22 -05:00
54 changed files with 2158 additions and 964 deletions
+18 -2
View File
@@ -36,7 +36,6 @@ union ncclProxyOpSpecifics {
struct ncclProxyOp {
struct ncclProxyConnection* connection;
void* buffer;
ssize_t nbytes;
uint64_t opCount;
int root:30;
@@ -53,6 +52,11 @@ struct ncclProxyOp {
uint8_t /*ncclPattern_t*/ pattern;
uint8_t protocol;
uint8_t reg;
// collnet buffer reg handles
void* sendMhandle;
void* recvMhandle;
uint8_t* sendbuff;
uint8_t* recvbuff;
union ncclProxyOpSpecifics specifics;
@@ -62,8 +66,14 @@ struct ncclProxyOp {
struct ncclProxySubArgs {
struct ncclProxyConnection* connection;
int reg;
void* buffer;
// p2p mhandle
void* mhandle;
// collnet handles
void* sendMhandle;
void* recvMhandle;
uint8_t* sendbuff;
uint8_t* recvbuff;
size_t offset;
int channelId;
int nsteps;
ssize_t nbytes;
@@ -97,6 +107,10 @@ struct ncclProxyArgs {
int sliceSteps;
int chunkSteps;
int chunkSize;
size_t totalSendSize;
size_t totalRecvSize;
size_t sendSizePerRound;
size_t recvSizePerRound;
uint8_t /*ncclDataType_t*/ dtype;
uint8_t /*ncclDevRedOp_t*/ redOp;
uint8_t /*ncclPattern_t*/ pattern;
@@ -312,6 +326,8 @@ enum ncclProxyMsgType {
ncclProxyMsgAbort = 7,
ncclProxyMsgStop = 8,
ncclProxyMsgGetFd = 9, // cuMem API support (UDS)
ncclProxyMsgRegister = 10,
ncclProxyMsgDeregister = 11
};
// This function is called by a client of the proxy that needs to invoke any of the non-progress proxyOp types