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

[ROCm/rocprofiler-sdk commit: 875f53b608]
Tento commit je obsažen v:
Jonathan R. Madsen
2024-02-23 10:30:33 -06:00
odevzdal GitHub
rodič a651f1e215
revize a360de4550
27 změnil soubory, kde provedl 1217 přidání a 184 odebrání
+45
Zobrazit soubor
@@ -44,6 +44,7 @@ def test_data_structure(input_data):
node_exists("hsa_api_traces", sdk_data["buffer_records"])
node_exists("hip_api_traces", sdk_data["buffer_records"], 0)
node_exists("marker_api_traces", sdk_data["buffer_records"])
node_exists("retired_correlation_ids", sdk_data["buffer_records"])
def test_timestamps(input_data):
@@ -123,6 +124,50 @@ def test_internal_correlation_ids(input_data):
assert max(api_corr_ids_sorted) == len_corr_id_unq
def test_retired_correlation_ids(input_data):
data = input_data
sdk_data = data["rocprofiler-sdk-json-tool"]
def _sort_dict(inp):
return dict(sorted(inp.items()))
api_corr_ids = {}
for titr in ["hsa_api_traces", "marker_api_traces", "hip_api_traces"]:
for itr in sdk_data["buffer_records"][titr]:
corr_id = itr["correlation_id"]["internal"]
assert corr_id not in api_corr_ids.keys()
api_corr_ids[corr_id] = itr
async_corr_ids = {}
for titr in ["kernel_dispatches", "memory_copies"]:
for itr in sdk_data["buffer_records"][titr]:
corr_id = itr["correlation_id"]["internal"]
assert corr_id not in async_corr_ids.keys()
async_corr_ids[corr_id] = itr
retired_corr_ids = {}
for itr in sdk_data["buffer_records"]["retired_correlation_ids"]:
corr_id = itr["internal_correlation_id"]
assert corr_id not in retired_corr_ids.keys()
retired_corr_ids[corr_id] = itr
api_corr_ids = _sort_dict(api_corr_ids)
async_corr_ids = _sort_dict(async_corr_ids)
retired_corr_ids = _sort_dict(retired_corr_ids)
for cid, itr in async_corr_ids.items():
assert cid in retired_corr_ids.keys()
ts = retired_corr_ids[cid]["timestamp"]
assert ts > itr["end_timestamp"]
for cid, itr in api_corr_ids.items():
assert cid in retired_corr_ids.keys()
ts = retired_corr_ids[cid]["timestamp"]
assert ts > itr["end_timestamp"]
assert len(api_corr_ids.keys()) == (len(retired_corr_ids.keys()))
def test_external_correlation_ids(input_data):
data = input_data
sdk_data = data["rocprofiler-sdk-json-tool"]