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

Correlation ID Retirement + misc (#527)

Browse Source 
* 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
This commit is contained in:
Jonathan R. Madsen
2024-02-23 10:30:33 -06:00
committed by GitHub
parent 2d71520953
commit 875f53b608
27 changed files with 1217 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/lib/rocprofiler-sdk/marker/marker.cpp
+31 -22
View File
@@ -191,7 +191,6 @@ roctx_api_impl<TableIdx, OpIdx>::functor(Args&&... args)
auto thr_id = common::get_tid();
auto callback_contexts = std::vector<callback_context_data>{};
auto buffered_contexts = std::vector<buffered_context_data>{};
auto has_pc_sampling = false;
populate_contexts(info_type::callback_domain_idx,
info_type::buffered_domain_idx,
@@ -208,7 +207,7 @@ roctx_api_impl<TableIdx, OpIdx>::functor(Args&&... args)
return 0;
}
auto ref_count = (has_pc_sampling) ? 4 : 2;
auto ref_count = 2;
auto buffer_record = common::init_public_api_struct(buffer_marker_api_record_t{});
auto tracer_data = callback_api_data_t{.size = sizeof(callback_api_data_t)};
auto* corr_id = correlation_service::construct(ref_count);
@@ -274,7 +273,7 @@ roctx_api_impl<TableIdx, OpIdx>::functor(Args&&... args)
}
// decrement the reference count before invoking
corr_id->ref_count.fetch_sub(1);
corr_id->sub_ref_count();
auto _ret = exec(info_type::get_table_func(), std::forward<Args>(args)...);
@@ -326,7 +325,7 @@ roctx_api_impl<TableIdx, OpIdx>::functor(Args&&... args)
}
// decrement the reference count after usage in the callback/buffers
corr_id->ref_count.fetch_sub(1);
corr_id->sub_ref_count();
context::pop_latest_correlation_id(corr_id);
@@ -420,6 +419,7 @@ iterate_args(const uint32_t id,
user_data);
if(ret != 0) break;
}
return;
}
if constexpr(sizeof...(OpIdxTail) > 0)
iterate_args<TableIdx>(id, data, func, user_data, std::index_sequence<OpIdxTail...>{});
@@ -451,7 +451,7 @@ should_wrap_functor(rocprofiler_callback_tracing_kind_t _callback_domain,
template <size_t TableIdx, typename Tp, size_t OpIdx>
void
copy_table(Tp* _orig, std::integral_constant<size_t, OpIdx>)
copy_table(Tp* _orig, uint64_t _tbl_instance, std::integral_constant<size_t, OpIdx>)
{
using table_type = typename roctx_table_lookup<TableIdx>::type;
@@ -459,24 +459,34 @@ copy_table(Tp* _orig, std::integral_constant<size_t, OpIdx>)
{
auto _info = roctx_api_info<TableIdx, OpIdx>{};
LOG(INFO) << "copying table entry for " << _info.name;
// make sure we don't access a field that doesn't exist in input table
if(_info.offset() >= _orig->size) return;
// 1. get the sub-table containing the function pointer in original table
// 2. get reference to function pointer in sub-table in original table
auto& _table = _info.get_table(_orig);
auto& _func = _info.get_table_func(_table);
auto& _orig_table = _info.get_table(_orig);
auto& _orig_func = _info.get_table_func(_orig_table);
// 3. get the sub-table containing the function pointer in saved table
// 4. get reference to function pointer in sub-table in saved table
// 5. save the original function in the saved table
auto& _saved = _info.get_table(*get_table<TableIdx>());
auto& _ofunc = _info.get_table_func(_saved);
_ofunc = _func;
}
auto& _copy_table = _info.get_table(*get_table<TableIdx>());
auto& _copy_func = _info.get_table_func(_copy_table);
(void) _orig;
LOG_IF(FATAL, _copy_func && _tbl_instance == 0)
<< _info.name << " has non-null function pointer " << _copy_func
<< " despite this being the first instance of the library being copies";
if(!_copy_func)
{
LOG(INFO) << "copying table entry for " << _info.name;
_copy_func = _orig_func;
}
else
{
LOG(INFO) << "skipping copying table entry for " << _info.name
<< " from table instance " << _tbl_instance;
}
}
}
template <size_t TableIdx, typename Tp, size_t OpIdx>
@@ -507,17 +517,15 @@ update_table(Tp* _orig, std::integral_constant<size_t, OpIdx>)
auto& _func = _info.get_table_func(_table);
_func = _info.get_functor(_func);
}
(void) _orig;
}
template <size_t TableIdx, typename Tp, size_t OpIdx, size_t... OpIdxTail>
void
copy_table(Tp* _orig, std::index_sequence<OpIdx, OpIdxTail...>)
copy_table(Tp* _orig, uint64_t _tbl_instance, std::index_sequence<OpIdx, OpIdxTail...>)
{
copy_table<TableIdx>(_orig, std::integral_constant<size_t, OpIdx>{});
copy_table<TableIdx>(_orig, _tbl_instance, std::integral_constant<size_t, OpIdx>{});
if constexpr(sizeof...(OpIdxTail) > 0)
copy_table<TableIdx>(_orig, std::index_sequence<OpIdxTail...>{});
copy_table<TableIdx>(_orig, _tbl_instance, std::index_sequence<OpIdxTail...>{});
}
template <size_t TableIdx, typename Tp, size_t OpIdx, size_t... OpIdxTail>
@@ -585,11 +593,12 @@ iterate_args(uint32_t id,
template <typename TableT>
void
copy_table(TableT* _orig)
copy_table(TableT* _orig, uint64_t _tbl_instance)
{
constexpr auto TableIdx = roctx_table_id_lookup<TableT>::value;
if(_orig)
copy_table<TableIdx>(_orig, std::make_index_sequence<roctx_domain_info<TableIdx>::last>{});
copy_table<TableIdx>(
_orig, _tbl_instance, std::make_index_sequence<roctx_domain_info<TableIdx>::last>{});
}
template <typename TableT>
@@ -606,7 +615,7 @@ using iterate_args_data_t = rocprofiler_callback_tracing_marker_api_data_t;
using iterate_args_cb_t = rocprofiler_callback_tracing_operation_args_cb_t;
#define INSTANTIATE_MARKER_TABLE_FUNC(TABLE_TYPE, TABLE_IDX) \
template void copy_table<TABLE_TYPE>(TABLE_TYPE * _tbl); \
template void copy_table<TABLE_TYPE>(TABLE_TYPE * _tbl, uint64_t _instv); \
template void update_table<TABLE_TYPE>(TABLE_TYPE * _tbl); \
template const char* name_by_id<TABLE_IDX>(uint32_t); \
template uint32_t id_by_name<TABLE_IDX>(const char*); \