Support multiple tuning tables (#522)

* Support multiple tuning tables

* [UnitTests] Skip managed memory testing

[ROCm/rccl commit: bbe780ca6c]
This commit is contained in:
Wenkai Du
2022-03-31 17:09:21 -07:00
committed by GitHub
parent 828f3d11a0
commit 3332cdff07
7 changed files with 133 additions and 51 deletions
+2 -2
View File
@@ -100,7 +100,7 @@ void ncclDebugInit() {
void ncclDebugLog(ncclDebugLogLevel level, unsigned long flags, const char *filefunc, int line, const char *fmt, ...) {
if (ncclDebugLevel == -1) ncclDebugInit();
if (level == NCCL_LOG_TRACE && ncclDebugLevel != NCCL_LOG_TRACE) return;
if (ncclDebugLevel < level || ((flags & (NCCL_INIT|NCCL_GRAPH)) == 0)) return;
if (ncclDebugLevel < level || ((flags & (NCCL_INIT|NCCL_GRAPH|NCCL_TUNING)) == 0)) return;
char buffer[1024];
size_t len = 0;
@@ -913,7 +913,7 @@ collnet_cleanup:
TRACE(NCCL_INIT, "rank %d nranks %d - CONNECTED %d RINGS AND TREES", rank, nranks, comm->nChannels);
// Compute time models for algorithm and protocol combinations
//NCCLCHECK(ncclTopoTuneModel(comm, minCompCap, maxCompCap, &treeGraph, &ringGraph, &collNetGraph));
NCCLCHECK(ncclTopoTuneModel(comm, 1, 1, &treeGraph, &ringGraph, &collNetGraph));
// Compute nChannels per peer for p2p
NCCLCHECK(ncclTopoComputeP2pChannels(comm));