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

[Rocprofiler-system]: Fix GPU event enumeration for rocprof-sys-avail and CLI option for parsing GPU HW Counters (#2476)

Browse Source 
## Motivation

The `rocprof-sys-avail -H -c GPU` command is returning blank output which is expected to display a list of available GPU hardware counters instead.
The `rocprof-sys-sample` and `rocprof-sys-run` is missing the `--gpu-events` option for specifying GPU counter events during profiling.

## Technical Details

The initialize_event_info() function had a logic bug where it only called set_agents() if the agent_manager was empty, but the actual issue was that the gpu_agents and cpu_agents vectors were empty even when agents were discovered.
Fixed the conditional logic to properly call set_agents() when gpu_agents and cpu_agents are empty, regardless of the agent_manager state.

Added the `--gpu-events (-G)` option which sets the `ROCPROFSYS_ROCM_EVENTS` environment variable to the specified values.

Fixes an issue where unsupported GPU/APU arch is being skipped gracefully - more details about this issue in the below comment.
This commit is contained in:
Sajina PK
2026-01-09 11:59:45 -05:00
committed by GitHub
parent ebe22b5907
commit b3f59a37e4
5 changed files with 118 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-systems/source/lib/rocprof-sys/library/rocprofiler-sdk.cpp
+39 -4
View File
@@ -213,6 +213,18 @@ create_agent_profile(rocprofiler_agent_id_t agent_id,
auto counters_v = counter_vec_t{};
const auto* tool_agent_v = data->get_gpu_tool_agent(agent_id);
// Check if agent info is available (may not be for unsupported architectures)
auto agent_info_it = data->agent_counter_info.find(agent_id);
if(agent_info_it == data->agent_counter_info.end())
{
ROCPROFSYS_WARNING_F(0,
"Skipping GPU agent %lu (device %lu) due to unsupported "
"architecture or missing counter info\n",
agent_id.handle, tool_agent_v->device_id);
data->agent_counter_profiles.emplace(agent_id, profile);
return counter_vec_t{};
}
constexpr auto device_qualifier = std::string_view{ ":device=" };
for(const auto& itr : counters)
{
@@ -263,7 +275,7 @@ create_agent_profile(rocprofiler_agent_id_t agent_id,
}
// search the gpu agent counter info for a counter with a matching name
for(const auto& citr : data->agent_counter_info.at(agent_id))
for(const auto& citr : agent_info_it->second)
{
if(name_v == std::string_view{ citr.name })
{
@@ -280,11 +292,34 @@ create_agent_profile(rocprofiler_agent_id_t agent_id,
auto found_counters =
timemory::join::join(timemory::join::array_config{ ", ", "", "" }, found_v);
ROCPROFSYS_ABORT_F(
"Unable to find all counters for agent %i (gpu-%li, %s) in %s. Found: %s\n",
// Determine which counters were not found
auto missing_counters = std::vector<std::string>{};
for(const auto& counter : counters)
{
if(std::find(found_v.begin(), found_v.end(), counter) == found_v.end())
missing_counters.emplace_back(counter);
}
auto missing_counters_str = timemory::join::join(
timemory::join::array_config{ ", ", "", "" }, missing_counters);
// In production, warn and continue with available counters
ROCPROFSYS_WARNING_F(0,
"Unable to find all counters for agent %i (gpu-%li, %s). "
"Requested: %s. Found: %s. Missing: %s. Continuing with "
"available counters.\n",
tool_agent_v->agent->node_id, tool_agent_v->device_id,
tool_agent_v->agent->name.c_str(),
requested_counters.c_str(), found_counters.c_str(),
missing_counters_str.c_str());
// In CI, throw to catch issues early
ROCPROFSYS_CI_THROW(
true,
"Unable to find all counters for agent %i (gpu-%li, %s). Requested: %s. "
"Found: %s. Missing: %s",
tool_agent_v->agent->node_id, tool_agent_v->device_id,
tool_agent_v->agent->name.c_str(), requested_counters.c_str(),
found_counters.c_str());
found_counters.c_str(), missing_counters_str.c_str());
}
if(!counters_v.empty())