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

2.19.1-1

Browse Source 
Add local user buffer registration for NVLink SHARP.
Add tuning plugin support.
Increase net API to v7 to allow for device-side packet reordering;
remove support for v4 plugins.
Add support for RoCE ECE.
Add support for C2C links.
Better detect SHM allocation failures to avoid crash with Bus Error.
Fix missing thread unlocks in bootstrap (Fixes #936).
Disable network flush by default on H100.
Move device code from src/collectives/device to src/device.
This commit is contained in:
Sylvain Jeaugey
2023-09-26 05:47:28 -07:00
parent 559b70f86c
commit f9c3dc251e
108 changed files with 4749 additions and 2059 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/info.h
+14 -1
View File
@@ -8,7 +8,7 @@
#define NCCL_INFO_H_
#include "nccl.h"
#include "devcomm.h"
#include "device.h"
#include "collectives.h"
#include "core.h"
#include "utils.h"
@@ -54,6 +54,8 @@ struct ncclInfo {
int nChannels;
int nThreads;
size_t nBytes;
size_t sendbuffSize;
size_t recvbuffSize;
int nstepsPerLoop;
int nchunksPerLoop;
int chunkSize;
@@ -67,6 +69,17 @@ inline ncclResult_t ncclInfoSetDerived(struct ncclInfo* info, int nRanks) {
info->datatype = ncclInt8;
}
if (info->coll == ncclFuncAllGather || info->coll == ncclFuncReduceScatter) info->nBytes *= nRanks; // count is per rank
/* compute buffer size for NVLS buffer registration */
if (info->coll == ncclFuncAllGather) {
info->sendbuffSize = info->count * ncclTypeSize(info->datatype);
info->recvbuffSize = info->sendbuffSize * nRanks;
} else if (info->coll == ncclFuncReduceScatter) {
info->recvbuffSize = info->count * ncclTypeSize(info->datatype);
info->sendbuffSize = info->recvbuffSize * nRanks;
} else {
info->sendbuffSize = info->recvbuffSize = info->count * ncclTypeSize(info->datatype);
}
return ncclSuccess;
}