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

SWDEV-492625 memory free functions (#11)

Browse Source 
* SWDEV-492625: Track free memory HSA functions to help determine total amount of memory allocated on the system at any one time

* Minor fixes to address comments

* Update allocation size description

* Moved get function back to specialization, minor typo fixes

* Removed memory_operation_type field, removed memory_pool allocation enum, converted starting address to hex string for json format.

* Made conversion to hex_string a function, changed address to use union rocprofiler_address_t type, changed VMEM descriptors

* Removed as_hex from the global namespace

* Formatting

* Removed TRACK_EVENT for memory allocation, now TRACK_COUNTER for memory allocation is being performed

* Check if address was recorded before retrieving allocation size in generate Perfetto

* Formatting

* Update source/lib/output/generatePerfetto.cpp

* Explicitly disable app-abort tests

* Remove excluding app-abort test from workflow CI

- redundant bc these tests are explicitly marked as disabled now

---------

Co-authored-by: Madsen, Jonathan <Jonathan.Madsen@amd.com>
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>

[ROCm/rocprofiler-sdk commit: 79006bb896]
This commit is contained in:
Trowbridge, Ian
2024-12-06 00:05:30 -06:00
committed by GitHub
parent a79f8a0198
commit 792329fefd
23 changed files with 564 additions and 248 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-sdk/source/lib/output/generateOTF2.cpp
+21 -7
View File
@@ -486,17 +486,25 @@ write_otf2(
{
for(auto& [agent, evt] : itr)
{
const auto* _agent = _get_agent(agent);
auto _type_name = std::string_view{"UNK"};
if(_agent->type == ROCPROFILER_AGENT_TYPE_CPU)
// Free functions do not track agent information. Below handles case where
// null rocprof agent id is passed to generate OTF2
constexpr auto null_rocp_agent_id =
rocprofiler_agent_id_t{.handle = std::numeric_limits<uint64_t>::max()};
const rocprofiler_agent_t* _agent = nullptr;
if(agent != null_rocp_agent_id)
{
_agent = _get_agent(agent);
}
auto _type_name = std::string_view{"UNK"};
if(_agent != nullptr && _agent->type == ROCPROFILER_AGENT_TYPE_CPU)
_type_name = "CPU";
else if(_agent->type == ROCPROFILER_AGENT_TYPE_GPU)
else if(_agent != nullptr && _agent->type == ROCPROFILER_AGENT_TYPE_GPU)
_type_name = "GPU";
evt.name = fmt::format("Thread {}, Memory Allocation at {} {}",
evt.name = fmt::format("Thread {}, Memory Operation at {} {}",
tid,
_type_name,
_agent->logical_node_type_id);
_agent == nullptr ? 0 : _agent->logical_node_type_id);
}
}
@@ -860,6 +868,12 @@ write_otf2(
for(auto& [agent, evt] : itr)
{
auto _hash = get_hash_id(evt.name);
// Using max numeric limits results in an out-of-bound runtime error for OTF2
// and perfetto for agent ids. Setting handle to 0 for free functions.
constexpr auto null_rocp_agent_id =
rocprofiler_agent_id_t{.handle = std::numeric_limits<uint64_t>::max()};
auto handle = agent.handle;
if(agent == null_rocp_agent_id) handle = 0;
add_write_string(_hash, evt.name);
OTF2_CHECK(OTF2_GlobalDefWriter_WriteLocation(global_def_writer,
@@ -867,7 +881,7 @@ write_otf2(
_hash,
OTF2_LOCATION_TYPE_ACCELERATOR_STREAM,
2 * evt.event_count, // # events
agent.handle // location group
handle // location group
));
}
}