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

Split ROCTx API tables and update intercept table API (#421)

Browse Source 
* Update include/rocprofiler-sdk

- buffer_tracing.h
  - fix doxygen for rocprofiler_buffer_tracing_hip_api_record_t
  - update doxygen for rocprofiler_buffer_tracing_marker_api_record_t
    - remove unused marker_id field
- fwd.h
  - Split ROCPROFILER_CALLBACK_TRACING_MARKER_API into ROCPROFILER_CALLBACK_TRACING_MARKER_{CORE,CONTROL,NAME}_API
  - Split ROCPROFILER_BUFFER_TRACING_MARKER_API into ROCPROFILER_BUFFER_TRACING_MARKER_{CORE,CONTROL,NAME}_API
  - split rocprofiler_runtime_library_t into rocprofiler_runtime_library_t and rocprofiler_intercept_table_t
    - after split of ROCTx into 3 tables, specifying rocprofiler_at_internal_thread_create became confusing

* Update include/rocprofiler-sdk-roctx/api_trace.h

- Split into three tables: core, control, and name
  - core: what it sounds like
  - control: functions for controling the profiler
  - name: functions for giving resources names

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

- modifications following split into multiple tables

* Update lib/rocprofiler-sdk/marker/*

- modifications following split of ROCTx API into multiple intercept tables

* Update lib/rocprofiler-sdk/tests

- common.hpp
  - add enums to get_callback_tracing_names() and get_buffer_tracing_names()
- intercept_table.cpp
  - update test to use rocprofiler_intercept_table_t (and enums) instead of rocproifler_runtime_library_t
  - update OR combos tested
- roctx.cpp
  - updates following split of ROCTx API table into multiple tables
  - use simplified specification of control API

* Update lib/rocprofiler-sdk

- buffer_tracing.cpp
  - Updates for ROCPROFILER_BUFFER_TRACING_MARKER_{CORE,CONTROL,NAME}_API enum values
- callback_tracing.cpp
  - Updates for ROCPROFILER_CALLBACK_TRACING_MARKER_{CORE,CONTROL,NAME}_API enum values
- intercept_table.hpp
  - notify_runtime_api_registration -> notify_intercept_table_registration
- intercept_table.cpp
  - updates for new rocprofiler_intercept_table_t enum and new ROCTx tables
- registration.cpp
  - updates for new rocprofiler_intercept_table_t enum and new ROCTx tables
  - updates for notify_runtime_api_registration -> notify_intercept_table_registration

* Update lib/rocprofiler-sdk-tool

- helper.cpp
  - Updates for new enums in get_callback_id_names() and get_buffer_id_names()
- tool.cpp
  - migrate to new enums for split ROCTx tables
  - use simplified split for control table vs. core+name tables

* Update samples/{api_callback_tracing,intercept_table}

- intercept_table/client.cpp
  - rocprofiler_runtime_library_t -> rocprofiler_intercept_table_t
- api_callback_tracing/client.cpp
  - Updates for new enums in get_callback_id_names()
  - use simplified split for control table vs. core+name tables
  - migrate to new enums for split ROCTx tables

* Update tests

- rocprofv3/tracing/validate.py
  - handle new marker domain names
- tools/json-tool.cpp
  - Updates for new enums in get_callback_id_names() and get_buffer_id_names()
  - use simplified split for control table vs. core+name tables
  - migrate to new enums for split ROCTx tables

* Update tests/rocprofv3/tracing/CMakeLists.txt

- fix FAIL_REGULAR_EXPRESSION for rocprofv3-test-trace-execute

* Update lib/rocprofiler-sdk-tool/{output_file,tool}.*

- logging in output_file dtor
- support stdout/stderr

* Update lib/common/container/record_header_buffer.hpp

- reduce probability of is_empty() returning true while emplace is happening

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

- logging for buffered_tracing_callback
- counter collection uses CSV encoder

* Update bin/rocprofv3

- remove -i flag from help menu

[ROCm/rocprofiler-sdk commit: 9efafc4d23]
This commit is contained in:
Jonathan R. Madsen
2024-01-26 13:56:15 -06:00
committed by GitHub
parent 24c8d296ba
commit bc67b1e823
32 changed files with 881 additions and 628 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-sdk/source/lib/common/container/record_header_buffer.hpp
+31 -14
View File
@@ -134,11 +134,12 @@ private:
void write_unlock();
private:
std::atomic<int64_t> m_locked = {0};
std::atomic<size_t> m_index = {};
std::shared_mutex m_shared = {};
base_buffer_t m_buffer = {};
record_vec_t m_headers = {};
std::atomic<int64_t> m_requested = {0};
std::atomic<int64_t> m_locked = {0};
std::atomic<size_t> m_index = {};
std::shared_mutex m_shared = {};
base_buffer_t m_buffer = {};
record_vec_t m_headers = {};
};
inline bool
@@ -218,7 +219,7 @@ record_header_buffer::free() const
inline auto
record_header_buffer::is_empty() const
{
return m_buffer.is_empty() || m_headers.empty();
return (m_buffer.is_empty() && m_requested.load() == 0) || m_headers.empty();
}
inline auto
@@ -235,6 +236,9 @@ record_header_buffer::emplace(uint64_t _hash, Tp& _v)
constexpr auto request_size = sizeof(Tp);
// notify there was a request
m_requested.fetch_add(1);
// in theory, we shouldn't need to lock here but the thread sanitizer says there is a race.
// the lock will be short-lived so hopefully, it will scale fine
write_lock();
@@ -244,20 +248,25 @@ record_header_buffer::emplace(uint64_t _hash, Tp& _v)
read_lock();
if(_addr)
{
// placement new
new(_addr) Tp{_v};
// if there is space in the buffer, atomically get an index
// for where the header record should be placed.
// NOTE: m_headers was resized to be large enough to accomodate
// sizeof(Tp) == 1 for every entry in buffer
auto idx = m_index.fetch_add(1, std::memory_order_release);
auto idx = m_index.fetch_add(1, std::memory_order_release);
// placement new
new(_addr) Tp{_v};
rocprofiler_record_header_t record = {};
record.hash = _hash;
record.payload = _addr;
m_headers.at(idx) = record;
}
read_unlock();
// remove notification of request
m_requested.fetch_sub(1);
return (_addr != nullptr);
}
@@ -269,6 +278,9 @@ record_header_buffer::emplace(uint32_t _category, uint32_t _kind, Tp& _v)
constexpr auto request_size = sizeof(Tp);
// notify there was a request
m_requested.fetch_add(1);
// in theory, we shouldn't need to lock here but the thread sanitizer says there is a race.
// the lock will be short-lived so hopefully, it will scale fine
write_lock();
@@ -278,20 +290,25 @@ record_header_buffer::emplace(uint32_t _category, uint32_t _kind, Tp& _v)
read_lock();
if(_addr)
{
// placement new
new(_addr) Tp{_v};
// if there is space in the buffer, atomically get an index
// for where the header record should be placed.
// NOTE: m_headers was resized to be large enough to accomodate
// sizeof(Tp) == 1 for every entry in buffer
auto idx = m_index.fetch_add(1, std::memory_order_release);
auto idx = m_index.fetch_add(1, std::memory_order_release);
// placement new
new(_addr) Tp{_v};
m_headers.at(idx) = rocprofiler_record_header_t{};
m_headers.at(idx).category = _category;
m_headers.at(idx).kind = _kind;
m_headers.at(idx).payload = _addr;
}
read_unlock();
// remove notification of request
m_requested.fetch_sub(1);
return (_addr != nullptr);
}