Correlation ID Retirement + misc (#527)
* Correlation ID Retirement
- include/rocprofiler-sdk/buffer_tracing.h
- add rocprofiler_buffer_tracing_correlation_id_retirement_record_t
- include/rocprofiler-sdk/fwd.h
- ROCPROFILER_BUFFER_TRACING_CORRELATION_ID_RETIREMENT
- lib/rocprofiler-sdk/buffer_tracing.cpp
- kind string for correlation id retirement
- lib/rocprofiler-sdk/buffer.hpp
- emplace returns bool
- lib/rocprofiler-sdk/registration.cpp
- pass lib_instance to copy_table functions
- lib/rocprofiler-sdk/context/context.*
- update correlation_id struct
- make ref_count private
- {get,add,sub}_ref_count() functions
- sub_ref_count() performs correlation id retirement
- use stack for "latest" thread-local correlation id
- lib/rocprofiler-sdk/hip/hip.*
- migrate to new {get,add,sub}_ref_count() for correlation ids
- return in iterate_args
- handle table instance in copy_table
- lib/rocprofiler-sdk/hsa/hsa.*
- migrate to new {get,add,sub}_ref_count() for correlation ids
- return in iterate_args
- handle table instance in copy_table
- lib/rocprofiler-sdk/marker/marker.*
- migrate to new {get,add,sub}_ref_count() for correlation ids
- return in iterate_args
- handle table instance in copy_table
- lib/rocprofiler-sdk/hsa/async_copy.cpp
- migrate to new {get,add,sub}_ref_count() for correlation ids
- handle table instance in async_copy_init / async_copy_save
- lib/rocprofiler-sdk/hsa/queue.cpp
- migrate to new {get,add,sub}_ref_count() for correlation ids
- tweak to external correlation id mapping in WriteInterceptor
- tests/async-copy-tracing/validate.py
- check retired_correlation_ids
- tests/common/serialization.hpp
- support rocprofiler_buffer_tracing_correlation_id_retirement_record_t
- tests/kernel-tracing/validate.py
- check retired_correlation_ids
- tests/common/CMakeLists.txt
- perfetto external project
- tests/common/perfetto.hpp
- perfetto categories + aliases
- add_perfetto_annotation
- metaprogramming helpers
- tests/tools/CMakeLists.txt
- link to tests-perfetto
- tests/tools/json-tool.cpp
- demangling functions
- serialization of marker API callback args
- reduce parallel bottleneck in tool_tracing_callback
- support correlation id retirement
- Multiple threads for buffers
- Support ROCPROFILER_TOOL_CONTEXTS_EXCLUDE env variable
- write_perfetto() function
* Update tests/rocprofv3/tracing/validate.py
- tweak test_hsa_api_trace
* Update PTL submodule
- fixes for data race during destruction of task
* Update lib/rocprofiler-sdk/buffer.*
- unique_buffer_vec_t uses std::unique_ptr instead of allocator::unique_static_ptr_t
* Reduce timeouts in counter collection samples [skip ci]
* Update tests/tools/json-tool.cpp
- tweak demangle(string_view, int*) -> demangle(string_view, int&)
* Update lib/rocprofiler-sdk/hsa/async_copy.cpp
- move sub_ref_count() to later in async_copy_handler to delay retirement slightly more
Tento commit je obsažen v:
@@ -181,10 +181,10 @@ AsyncSignalHandler(hsa_signal_value_t /*signal_v*/, void* data)
|
||||
|
||||
if(_corr_id)
|
||||
{
|
||||
LOG_IF(FATAL, _corr_id->ref_count.load() == 0)
|
||||
LOG_IF(FATAL, _corr_id->get_ref_count() == 0)
|
||||
<< "reference counter for correlation id " << _corr_id->internal << " from thread "
|
||||
<< _corr_id->thread_idx << " has no reference count";
|
||||
_corr_id->ref_count.fetch_sub(1);
|
||||
_corr_id->sub_ref_count();
|
||||
}
|
||||
|
||||
queue_info_session.queue.async_complete();
|
||||
@@ -257,24 +257,17 @@ WriteInterceptor(const void* packets,
|
||||
auto* corr_id = context::get_latest_correlation_id();
|
||||
|
||||
// use thread-local value to reuse allocation
|
||||
static thread_local auto extern_corr_ids_tl =
|
||||
Queue::queue_info_session_t::external_corr_id_map_t{};
|
||||
auto extern_corr_ids = Queue::queue_info_session_t::external_corr_id_map_t{};
|
||||
|
||||
// increase the reference count to denote that this correlation id is being used in a kernel
|
||||
if(corr_id)
|
||||
{
|
||||
extern_corr_ids_tl.clear(); // clear it so that it only contains the current contexts
|
||||
extern_corr_ids_tl.reserve(ctxs.size()); // reserve for performance
|
||||
extern_corr_ids.clear(); // clear it so that it only contains the current contexts
|
||||
extern_corr_ids.reserve(ctxs.size()); // reserve for performance
|
||||
for(const auto* ctx : ctxs)
|
||||
extern_corr_ids_tl.emplace(ctx,
|
||||
ctx->correlation_tracer.external_correlator.get(thr_id));
|
||||
corr_id->ref_count.fetch_add(1);
|
||||
extern_corr_ids.emplace(ctx, ctx->correlation_tracer.external_correlator.get(thr_id));
|
||||
}
|
||||
|
||||
// move to local variable
|
||||
auto extern_corr_ids = std::move(extern_corr_ids_tl);
|
||||
|
||||
// hsa_ext_amd_aql_pm4_packet_t
|
||||
const auto* packets_arr = static_cast<const rocprofiler_packet*>(packets);
|
||||
auto transformed_packets = std::vector<rocprofiler_packet>{};
|
||||
|
||||
@@ -375,6 +368,8 @@ WriteInterceptor(const void* packets,
|
||||
LOG_IF(FATAL, packet_type != HSA_PACKET_TYPE_KERNEL_DISPATCH)
|
||||
<< "get_kernel_id below might need to be updated";
|
||||
|
||||
if(corr_id) corr_id->add_ref_count();
|
||||
|
||||
// Enqueue the signal into the handler. Will call completed_cb when
|
||||
// signal completes.
|
||||
queue.async_started();
|
||||
|
||||
Odkázat v novém úkolu
Zablokovat Uživatele