From 15143b1cfbc6294592bf7046985838ae40dc638e Mon Sep 17 00:00:00 2001 From: Wenkai Du <43822138+wenkaidu@users.noreply.github.com> Date: Tue, 26 Oct 2021 10:30:20 -0700 Subject: [PATCH] Query XGMI link count through rocm_smi_lib API (#442) [ROCm/rccl commit: 14a184eb67ee14492a015aef90345e1393b97f92] --- projects/rccl/src/graph/xml.cc | 6 +- projects/rccl/src/include/rocm_smi_wrap.h | 3 +- projects/rccl/src/misc/rocm_smi_wrap.cc | 18 +++- .../tools/topo_expl/models/topo_16p1h.xml | 96 +++++++++---------- 4 files changed, 67 insertions(+), 56 deletions(-) diff --git a/projects/rccl/src/graph/xml.cc b/projects/rccl/src/graph/xml.cc index 3aca65c1c6..b9cba9953d 100644 --- a/projects/rccl/src/graph/xml.cc +++ b/projects/rccl/src/graph/xml.cc @@ -671,8 +671,8 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm for (int i=0; i= 5 + uint64_t min_bw = 0, max_bw = 0; + rsmi_version_t version; + ROCMSMICHECK(rsmi_version_get(&version)); + if (version.major >= 5) + ROCMSMICHECK(rsmi_minmax_bandwidth_get(srcIndex, dstIndex, &min_bw, &max_bw)); + if (max_bw && min_bw) + *count = max_bw/min_bw; +#endif + } return ncclSuccess; } diff --git a/projects/rccl/tools/topo_expl/models/topo_16p1h.xml b/projects/rccl/tools/topo_expl/models/topo_16p1h.xml index 3389967c79..44a0451414 100644 --- a/projects/rccl/tools/topo_expl/models/topo_16p1h.xml +++ b/projects/rccl/tools/topo_expl/models/topo_16p1h.xml @@ -2,19 +2,19 @@ - + - + - + - + - + @@ -23,33 +23,33 @@ - + - + - + - + - - + + - + @@ -58,20 +58,20 @@ - + - - + + - + - + @@ -79,33 +79,33 @@ - + - + - + - + - - + + - + @@ -114,19 +114,19 @@ - + - + - + - + - + @@ -135,33 +135,33 @@ - + - + - + - + - - + + - + @@ -170,54 +170,54 @@ - + - - + + - + - + - + - + - + - + - - + + - +