Async memory copy tracing (#317)

* Update samples/api_buffered_tracing/client.cpp

- support ROCPROFILER_BUFFER_TRACING_MEMORY_COPY

* Update include/rocprofiler-sdk/{buffer_tracing,fwd}.h

- update rocprofiler_buffer_tracing_memory_copy_record_t
- add ROCPROFILER_BUFFER_TRACING_MEMORY_COPY_HOST_TO_HOST to rocprofiler_memory_copy_operation_t

* Update lib/rocprofiler-sdk/context/context.*

- get_registered_contexts functions (local copy)

* Update tests/apps/reproducible-runtime/reproducible-runtime.cpp

- include some memory allocations and memory copies for better testing

* Update tests/common/serialization.hpp

- update serialization save function for rocprofiler_buffer_tracing_memory_copy_record_t

* Update lib/rocprofiler-sdk/hsa/hsa.*

- remove stale set_callback / activity_functor_t code
- forward decl hsa_api_meta
- template struct hsa_api_func for getting function return type and args

* Update tests/kernel-tracing/validate.py

- enforce memory_copies data size
- test timestamps in memory copies data
- improve internal and external correlation id validation

* Update lib/rocprofiler-sdk/hsa/defines.hpp

- HSA_API_META_DEFINITION macro

* Update lib/rocprofiler/hsa/rocprofiler-sdk/hsa/hsa.def.cpp

- HSA_API_META_DEFINITION specializations for async copy functions

* Add lib/rocprofiler-sdk/hsa/async_copy.{hpp,cpp}

- implements buffer memory tracing

* Update lib/rocprofiler-sdk/registration.cpp

- invoke rocprofiler::hsa::async_copy_init

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- logging improvements
- improve hsa <-> rocp agent mapping

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- load original signal in async signal handler before store_screlease

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- use store_relaxed instead of store_screlease

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- logging

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- logging

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- misc changes

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- misc changes

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- misc changes

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- return function pointer instead of lambda

* Update reproducible-runtime.cpp

- device sync

* Update tests/apps/reproducible-runtime/reproducible-runtime.cpp

- use *Async variants of hipMalloc and hipMemcpy

* Update lib/rocprofiler-sdk/hsa/async_copy.cpp

- populate async data properly

* Update tests/kernel-tracing/validate.py

- verification of async copy direction

* Update tests/apps/reproducible-runtime/reproducible-runtime.cpp

- temporarily disable async memcpy functions

* Create tests/tools

- directory containing tool libraries used for collecting data in integration tests

* Update tests/kernel-tracing

- remove kernel-tracing-test-tool library (now rocprofiler-sdk-json-tool)
- update cmake, validate.py, conftest.py accordingly

* Add tests/async-copy-tracing

- integration test validating async copy tracing in transpose example

* Update tests/CMakeLists.txt

- updates for restructuring

* Revert tests/apps/reproducible-runtime

- restore code to semi-original state (no memory copying)

* Update tests/async-copy-tracing/validate.py

- fix comment in test_async_copy_direction

* Fix building tests against installation
This commit is contained in:
Jonathan R. Madsen
2024-01-09 11:34:46 -06:00
کامیت شده توسط GitHub
والد 0dd0cad2e4
کامیت 936816f762
23فایلهای تغییر یافته به همراه1015 افزوده شده و 79 حذف شده
@@ -34,6 +34,7 @@
#include "client.hpp"
#include <rocprofiler-sdk/buffer.h>
#include <rocprofiler-sdk/buffer_tracing.h>
#include <rocprofiler-sdk/callback_tracing.h>
#include <rocprofiler-sdk/external_correlation.h>
#include <rocprofiler-sdk/fwd.h>
@@ -309,6 +310,29 @@ tool_tracing_callback(rocprofiler_context_id_t context,
static_cast<call_stack_t*>(user_data)->emplace_back(
source_location{__FUNCTION__, __FILE__, __LINE__, info.str()});
}
else if(header->category == ROCPROFILER_BUFFER_CATEGORY_TRACING &&
header->kind == ROCPROFILER_BUFFER_TRACING_MEMORY_COPY)
{
auto* record =
static_cast<rocprofiler_buffer_tracing_memory_copy_record_t*>(header->payload);
auto info = std::stringstream{};
info << "src_agent_id=" << record->src_agent_id.handle
<< ", dst_agent_id=" << record->dst_agent_id.handle
<< ", direction=" << record->operation << ", context=" << context.handle
<< ", buffer_id=" << buffer_id.handle
<< ", cid=" << record->correlation_id.internal
<< ", extern_cid=" << record->correlation_id.external.value
<< ", kind=" << record->kind << ", start=" << record->start_timestamp
<< ", stop=" << record->end_timestamp;
if(record->start_timestamp > record->end_timestamp)
throw std::runtime_error("memory copy: start > end");
static_cast<call_stack_t*>(user_data)->emplace_back(
source_location{__FUNCTION__, __FILE__, __LINE__, info.str()});
}
else
{
throw std::runtime_error{"unexpected rocprofiler_record_header_t category + kind"};