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

Disable SDMA ganging on non-APU multi-partition modes

Browse Source 
Work around SDMA hang in non-SPX modes for non-APU devices by disabling
ganging.
Root cause of hang not found.
non-APU xGMI modes have only 1 link between socket devices anyways so
there's likely no real system level gain in ganging intra-socket.

Change-Id: Ia4eda2f85cbf25151d3dbcf50cc45b8b775c60e2


[ROCm/ROCR-Runtime commit: ed462035fa]
This commit is contained in:
Jonathan Kim
2024-02-28 14:18:27 -05:00
committed by David Yat Sin
parent 1e46d28ee6
commit f9a6578b6b
1 changed files with 12 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_topology.cpp
+12 -1
View File
@@ -367,6 +367,17 @@ void BuildTopology() {
}
const_cast<Flag&>(core::Runtime::runtime_singleton_->flag()).parse_masks(maxGpu, maxCu);
// Temporary work-around, disable SDMA ganging on non-APUs in non-SPX modes
// Check xGMI APU status
bool isXgmiApu = false;
auto& firstCpu = core::Runtime::runtime_singleton_->cpu_agents()[0];
for (auto& peer_gpu : core::Runtime::runtime_singleton_->gpu_agents()) {
auto linfo = core::Runtime::runtime_singleton_->GetLinkInfo(firstCpu->node_id(),
peer_gpu->node_id());
isXgmiApu = linfo.info.link_type == HSA_AMD_LINK_INFO_TYPE_XGMI;
if (isXgmiApu) break;
}
// Register destination agents that can SDMA gang copy for source agents
for (auto& src_gpu : core::Runtime::runtime_singleton_->gpu_agents()) {
uint32_t src_id = src_gpu->node_id();
@@ -383,7 +394,7 @@ void BuildTopology() {
// Weight of 41 - Inter-socket GPU link in multi-partition mode
if (linfo.info.link_type == HSA_AMD_LINK_INFO_TYPE_XGMI) {
if (linfo.info.numa_distance == 13 || linfo.info.numa_distance == 41)
gang_factor = 2;
gang_factor = isXgmiApu ? 2 : 1;
else if (linfo.info.numa_distance == 15 && linfo.info.min_bandwidth)
gang_factor = linfo.info.max_bandwidth/linfo.info.min_bandwidth;
else gang_factor = 1;