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

Add support for external network.

Browse Source 
Dynamically load external network from libnccl-net.so.
Add init function in networks.
Move PCI scoring to net.cu, only ask transport to provide a path.
Simplify CUDA PCI path detection.
Add dummy external network


[ROCm/rccl commit: 0d3a20f96d]
This commit is contained in:
Sylvain Jeaugey
2018-11-13 10:37:20 -08:00
parent b458b3b337
commit 378a103248
29 changed files with 437 additions and 361 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/net.h
+3 -2
View File
@@ -10,6 +10,7 @@
#include "nccl.h"
#include "nccl_net.h"
extern ncclNet_t* ncclNet;
typedef char ncclNetHandle_t[NCCL_NET_HANDLE_MAXSIZE];
/* Socket Interface Selection type */
@@ -19,7 +20,8 @@ typedef enum { findSubnetIf = -1,
// 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 ncclNetDevices(int* ndev) { NCCLCHECK(ncclNet->devices(ndev)); return ncclSuccess; }
static ncclResult_t ncclNetPciPath(int dev, char** path) { NCCLCHECK(ncclNet->pciPath(dev, path)); 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; }
@@ -32,7 +34,6 @@ static ncclResult_t ncclNetCloseSend(void* sendComm) { NCCLCHECK(ncclNet->closeS
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;