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

Shared Library Constructor (rocprofv3 deadlock fix) (#599)

Browse Source 
* Moved tests/apps to tests/bin

* Renamed cmake project in tests/bin

* Update samples

- Use ROCPROFILER_DEFAULT_FAIL_REGEX
- tweaks to stdout messages

* Update tests

- Use ROCPROFILER_DEFAULT_FAIL_REGEX

* Add tests/lib

- libraries with HIP code

* Update PTL submodule

- remove atexit delete of thread_id_map

* Update cmake/rocprofiler_options.cmake

- Set ROCPROFILER_DEFAULT_FAIL_REGEX

* Update common lib: env + logging

- improved customization of logging settings
- default to disabling logging to files
- install failure handler for rocprofv3
- set_env support in environment.*

* Add lib/rocprofiler-sdk/shared_library.cpp

- shared library constructor

* Update lib/rocprofiler-sdk-tool/tool.cpp

- destructor thread safety
- convert callback_name_info and buffered_name_info to pointers
- install failure handler for logging

* Add tests/bin/hip-in-libraries

- hip-in-libraries is an exe which uses two shared libraries where each shared library contains HIP kernels
  - used for testing deadlocking within __hipRegisterFatBinary

* Update bin/rocprofv3

- reorganized the env variables
- use exec to launch command
- set ROCPROFILER_LIBRARY_CTOR=1

* Add tests/rocprofv3/tracing-hip-in-libraries

- uses hip-in-libraries exe for exe which uses shared libraries to launch HIP kernels

* Update bin/rocprofv3

- fix counter collection (no exec)

* Update lib/rocprofiler-sdk-tool/tool.cpp

- replace "Kernel-Name" with "Kernel_Name"

* Update lib/rocprofiler-sdk/registration.cpp

Use RTLD_LOCAL instead of RTLD_GLOBAL for env libraries

* Update tests/rocprofv3

- replace "Kernel-Name" with "Kernel_Name"

* Update tests

- vector-ops (bin) stream syncs + runs with 4 queues per device
- improve counter-collection/input1 validation
- rocprofv3/tracing-hip-in-libraries does not do sys-trace
- improved validation script for tracing-hip-in-libraries
- updated dispatch_callback in json-tool.cpp following reworking of prototypes for counter collection

* Update samples/counter_collection

- updated dispatch_callback(s) and record_callback(s) following reworking of prototypes

* Update bin/rocprofv3

- reorganized help menu
- added options for sub-HSA tables
- added --hip-runtime-trace
- changed --hip-trace to include --hip-compiler-trace

* Update lib/rocprofiler-sdk-tool

- improved kernel filtering
- removed arch_vgpr, accum_vgpr, sgpr code (in rocprofiler-sdk)
- fixed issue with counter-collection w/o tracing
- added support for fine grained HSA API tracing
- removed directly linking to HSA-runtime

* Update lib/rocprofiler-sdk/agent.cpp

- rocp_agents != hsa_agents is non-fatal when ROCPROFILER_BUILD_CI=OFF (CMake option)

* GPR (vector and scalar) info in kernel symbol data

- rocprofiler_callback_tracing_code_object_kernel_symbol_register_data_t contains general purpose register info

* Header include order fix

- Include repo headers first
- Third party library headers next
- standard library headers last

* Update dispatch profiling public API

- introduce rocprofiler_profile_counting_dispatch_data_t
- change signature of rocprofiler_profile_counting_dispatch_callback_t and rocprofiler_profile_counting_record_callback_t
- provide rocprofiler_user_data_t pointer in dispatch callback
- provide rocprofiler_user_data_t value (from dispatch cb) in record callback

* Update tests/bin/CMakeLists.txt

- fix add_subdirectory(hip-in-libraries) order

* Update VERSION

- bump to 0.2.0 in prep for AFAR
This commit is contained in:
Jonathan R. Madsen
2024-03-07 22:21:26 -06:00
committed by GitHub
parent 665c546e65
commit 7b6d3c70bd
85 changed files with 2497 additions and 856 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/tools/json-tool.cpp
+11 -13
View File
@@ -526,19 +526,16 @@ counter_collection_buffered(rocprofiler_context_id_t, /*context*/
}
void
dispatch_callback(rocprofiler_queue_id_t, /*queue_id*/
const rocprofiler_agent_t* agent,
rocprofiler_correlation_id_t, /*correlation_id*/
const hsa_kernel_dispatch_packet_t*, /*dispatch_packet*/
uint64_t, /*kernel_id*/
void* /*callback_data_args*/,
rocprofiler_profile_config_id_t* config)
dispatch_callback(rocprofiler_profile_counting_dispatch_data_t dispatch_data,
rocprofiler_profile_config_id_t* config,
rocprofiler_user_data_t* /*user_data*/,
void* /*callback_data_args*/)
{
static std::shared_mutex m_mutex = {};
static std::unordered_map<uint64_t, rocprofiler_profile_config_id_t> profile_cache = {};
auto search_cache = [&]() {
if(auto pos = profile_cache.find(agent->id.handle); pos != profile_cache.end())
if(auto pos = profile_cache.find(dispatch_data.agent_id.handle); pos != profile_cache.end())
{
*config = pos->second;
return true;
@@ -565,7 +562,7 @@ dispatch_callback(rocprofiler_queue_id_t, /*queue_id*/
// Iterate through the agents and get the counters available on that agent
ROCPROFILER_CALL(rocprofiler_iterate_agent_supported_counters(
agent->id,
dispatch_data.agent_id,
[]([[maybe_unused]] rocprofiler_agent_id_t id,
rocprofiler_counter_id_t* counters,
size_t num_counters,
@@ -600,11 +597,12 @@ dispatch_callback(rocprofiler_queue_id_t, /*queue_id*/
// Create a colleciton profile for the counters
rocprofiler_profile_config_id_t profile;
ROCPROFILER_CALL(rocprofiler_create_profile_config(
agent->id, collect_counters.data(), collect_counters.size(), &profile),
"Could not construct profile cfg");
ROCPROFILER_CALL(
rocprofiler_create_profile_config(
dispatch_data.agent_id, collect_counters.data(), collect_counters.size(), &profile),
"Could not construct profile cfg");
profile_cache.emplace(agent->id.handle, profile);
profile_cache.emplace(dispatch_data.agent_id.handle, profile);
// Return the profile to collect those counters for this dispatch
*config = profile;
}