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

2.3.5-5

Browse Source 
Add support for inter-node communication using sockets and InfiniBand/RoCE.
Improve latency.
Add support for aggregation.
Improve LL/regular tuning.
Remove tests as those are now at github.com/nvidia/nccl-tests .


[ROCm/rccl commit: f93fe9bfd9]
This commit is contained in:
Sylvain Jeaugey
2018-09-24 16:06:59 -07:00
parent 63ab6df5b3
commit 8ffcfac437
132 changed files with 12424 additions and 9415 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/net.h
+40
View File
@@ -0,0 +1,40 @@
/*************************************************************************
* Copyright (c) 2016-2018, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_INT_NET_H_
#define NCCL_INT_NET_H_
#include "nccl.h"
#include "nccl_net.h"
typedef char ncclNetHandle_t[NCCL_NET_HANDLE_MAXSIZE];
/* Socket Interface Selection type */
typedef enum { findSubnetIf = -1,
dontCareIf = -2
} ncclSocketIfSl_t;
// Translation to external API
static const char* ncclNetName() { return ncclNet->name; }
static ncclResult_t ncclNetDevices(int* ndev, int** scores) { NCCLCHECK(ncclNet->devices(ndev, scores)); return ncclSuccess; }
static ncclResult_t ncclNetPtrSupport(int dev, int* supportedTypes) { NCCLCHECK(ncclNet->ptrSupport(dev, supportedTypes)); return ncclSuccess; }
static ncclResult_t ncclNetListen(int dev, void* handle, void** listenComm) { NCCLCHECK(ncclNet->listen(dev, handle, listenComm)); return ncclSuccess; }
static ncclResult_t ncclNetConnect(int dev, void* handle, void** sendComm) { NCCLCHECK(ncclNet->connect(dev, handle, sendComm)); return ncclSuccess; }
static ncclResult_t ncclNetAccept(void* listenComm, void** recvComm) { NCCLCHECK(ncclNet->accept(listenComm, recvComm)); return ncclSuccess; }
static ncclResult_t ncclNetIsend(void* sendComm, void* data, int size, int type, void** request) { NCCLCHECK(ncclNet->isend(sendComm, data, size, type, request)); return ncclSuccess; }
static ncclResult_t ncclNetIrecv(void* recvComm, void* data, int size, int type, void** request) { NCCLCHECK(ncclNet->irecv(recvComm, data, size, type, request)); return ncclSuccess; }
static ncclResult_t ncclNetFlush(void* recvComm, void* data, int size) { NCCLCHECK(ncclNet->flush(recvComm, data, size)); return ncclSuccess; }
static ncclResult_t ncclNetTest(void* request, int* done, int* size) { NCCLCHECK(ncclNet->test(request, done, size)); return ncclSuccess; }
static ncclResult_t ncclNetCloseSend(void* sendComm) { NCCLCHECK(ncclNet->closeSend(sendComm)); return ncclSuccess; }
static ncclResult_t ncclNetCloseRecv(void* recvComm) { NCCLCHECK(ncclNet->closeRecv(recvComm)); return ncclSuccess; }
static ncclResult_t ncclNetCloseListen(void* listenComm) { NCCLCHECK(ncclNet->closeListen(listenComm)); return ncclSuccess; }
extern bool ncclIbSupport();
extern ncclResult_t ncclSocketCreateHandle(void* opaqueHandle, const char* str);
extern ncclNet_t ncclNetIb;
extern ncclNet_t ncclNetSocket;
#endif