2
0
Ficheiros
rocm-systems/samples/pc_sampling/common.h
T
Jonathan R. Madsen d3eaacd610 Contexts, tracing, include reorg, registration, thread-pool (#65)
* Update scripts/update-doxygen.sh

- ensure build-docs folder exists

* Update scripts/run-ci.py

- exclude files in details subdirectory from code coverage

* Update scripts/thread-sanitizer-suppr.txt

- exclude races in glog

* Update docs/rocprofiler.dox.in

- exclude defines in include/rocprofiler/defines.h from doxygen
- Tweak EXCLUDE_PATTERNS and EXAMPLE_PATTERNS

* Update docs workflow

- trigger workflow whenever there is a change to the public headers (which may be doxygen comments)

* Update include/rocprofiler (reorg and overhaul)

- rocprofiler_status_t additions
  - CONTEXT_NOT_FOUND
  - CONTEXT_ERROR
  - INVALID_CONTEXT_ID
  - INVALID_CONTEXT
  - BUFFER_BUSY
- rocprofiler_context_is_active func
- rocprofiler_context_is_valid func
- rocprofiler_service_callback_tracing_kind_t update
  - remove ROCPROFILER_SERVICE_CALLBACK_TRACING_HELPER_THREAD
- Remove rocprofiler_tracing_helper_thread_operation_t
- Remove rocprofiler_helper_thread_callback_tracer_data_t
- Added rocprofiler_internal_thread_library_t
- Added rocprofiler_at_internal_thread_create
- split rocprofiler.h into several smaller headers
- reworked rocprofiler_status_t values
- added doxygen comments for enums
- replaced rocprofiler_trace_record_operation_kind_t with rocprofiler_trace_operation_t
- use @ instead of / in doxygen comment in rocprofiler_plugin.h
- fix ref to ROCPROFILER_SERVICE_CALLBACK_TRACING_MARKER_API
- end group in fwd.h
- remove PROFILE_COUNTING group in dispatch_profile.h
- remove premature group close in callback_tracing.h
- hsa.h: remove rocprofiler_hsa_trace_data_t
- fwd.h: remove rocprofiler_tracer_callback_data_t
- rename rocprofiler_correlation_id_t.handle to rocprofiler_correlation_id_t.id (consistency)
- fwd.h: add rocprofiler_callback_tracing_record_t
- callback_tracing.h: update rocprofiler_hsa_api_callback_tracer_data_t
- callback_tracing.h: add size fields
- simplify rocprofiler_tracer_callback_t
- removed ROCPROFILER_NONNULL from rocprofiler_get_version
- added rocprofiler_get_timestamp
- ROCPROFILER_STATUS_ERROR_CONFIGURATION_LOCKED in rocprofiler_status_t
- add ROCPROFILER_STATUS_ERROR_THREAD_NOT_FOUND rocprofiler_status_t
- add rocprofiler_buffer_category_t
- rocprofiler_trace_operation_t -> rocprofiler_tracing_operation_t
- rocprofiler_user_data_t union
- tweak rocprofiler_callback_tracing_record_t
  - make external_correlation_id non-pointer
  - add rocprofiler_user_data_t data field
- tweak rocprofiler_record_header_t
  - instead of single uint64_t kind field, have union for category + kind (two u32) with u64 hash
- API extensions for kind id <-> kind string
- API extensions for operation id <-> operation string
- rocprofiler_callback_trace_kind_name_cb_t
- rocprofiler_callback_trace_operation_name_cb_t
- rocprofiler_iterate_callback_trace_kind_names
- rocprofiler_iterate_callback_trace_kind_operation_names
- modify rocprofiler_hsa_api_callback_tracer_data_t data members (remove pointers)
- add rocprofiler_callback_trace_operation_args_cb_t function pointer typedef
- add rocprofiler_iterate_callback_trace_operation_args function
- fixed inconsistent use of *_trace_* vs. *_tracing_* (opting for tracing)
- removed rocprofiler_query_callback_trace_kind_name
- removed rocprofiler_query_callback_kind_operation_name
- Add include/rocprofiler/registration.h
  - header dedicated to registering a tool/client with rocprofiler
  - this header is not intended to be included by rocprofiler.h
  - rocprofiler_client_id_t
    - identifier for client tool
  - rocprofiler_client_finalize_t
    - function pointer prototype for tool-initiated finalization
  - rocprofiler_tool_initialize_t
    - function pointer prototype for tool initialization (i.e. configuration)
  - rocprofiler_tool_finalize_t
    - function pointer prototype for tool finalization
  - rocprofiler_tool_configure_result_t
    - struct returned by tool/client to rocprofiler
  - rocprofiler_is_initialized
    - function for querying whether tool-induced initialization is possible
  - rocprofiler_is_finalized
    - function for querying whether rocprofiler has been finalized
  - rocprofiler_configure prototype
    - this is the function tools implement
    - prototype is always marked as having default visibility
    - no implementation in rocprofiler
  - added typedef for rocprofiler_configure function pointer
  - added rocprofiler_force_configure to explicitly invoke rocprofiler_configure instead of relying on lazy init
- made callback typedef names more consistent (_cb_t suffix)
- typedef for rocprofiler_internal_thread_library_cb_t function pointer
- added rocprofiler_at_internal_thread_create function
- added rocprofiler_callback_thread_t struct
- added rocprofiler_create_callback_thread function
- added rocprofiler_assign_callback_thread function
- removed rocprofiler_buffer_tracing_record_header_t in favor of kind and correlation id in each record type
- added rocprofiler_buffer_tracing_kind_name_cb_t typedef
- added rocprofiler_buffer_tracing_operation_name_cb_t typedef
- added rocprofiler_iterate_buffer_tracing_kind_names function
- added rocprofiler_iterate_buffer_tracing_kind_operation_names function
- removed rocprofiler_query_buffer_trace_kind_name function
- removed rocprofiler_query_buffer_kind_operation_name function

* Update lib/common/container/stable_vector.hpp

- include limits header
- reserve_size struct
- overload stable_vector constructor to support reserving as part of construction

* Update lib/common/container/record_header_buffer.{hpp,cpp}

- add emplace member function accepting category and kind (two u32 variables) instead of one u64 kind
- use std::shared_mutex to prevent data-race when reading m_headers
- record_header_buffer is now multiple writer, single reader
- add read_lock member function (shared)
- add read_unlock member function (shared)
- lock member function gets exclusive lock
- unlock member function releases exclusive lock

* Rename "config" to "context" + restructure + implement

- Restructure config files + license
  - move config files into lib/rocprofiler/config subfolder
  - rename some files
  - add license to some files which were missing it
- Rename config/helpers.hpp
  - rename to allocator.hpp
  - remove get_domain_max_ops
- Create config/domain.{hpp,cpp}
  - structures for handling tracing domains and ops
- Update config/config.{hpp,cpp}
  - buffer_instance struct
  - callback_tracing_service struct
  - buffer_tracing_service struct
  - config struct
  - allocate_{config,buffer} func
  - {validate,start,stop}_config funcs
  - get_registered_configs func
  - get_active_configs func
  - get_buffers func
- Update rocprofiler.cpp
  - Implement rocprofiler_create_context
  - Implement rocprofiler_start_context
  - Implement rocprofiler_stop_context
  - Implement rocprofiler_context_is_active
  - Implement rocprofiler_context_is_valid
  - Implement rocprofiler_flush_buffer
  - Implement rocprofiler_destroy_buffer
  - Implement rocprofiler_create_buffer
- Update lib/rocprofiler/hsa
  - use rocprofiler_tracer_activity_domain_t instead of rocprofiler_tracer_activity_domain_t
  - remove ROCPROFILER_TRACER_ACTIVITY_DOMAIN_HSA_API fromHSA_API_INFO_DEFINITION_* macros
- Update lib/rocprofiler/context/domain.*
  - fixes for domain_info (i.e. use correct enums)
  - update rocprofiler_status_t codes
  - fix template instantiations
- Update lib/rocprofiler/context/context.*
  - use rocprofiler_service_callback_tracing_kind_t instead of rocprofiler_tracer_activity_domain_t
  - rename correlation_context to correlation_tracing_service
  - fix domains in callback_tracing_service and buffer_tracing_service
  - unique_ptr for callback_tracer and buffered_tracer in context
- Update lib/rocprofiler/rocprofiler.cpp
  - implement rocprofiler_configure_callback_tracing_service
- Update lib/rocprofiler/hsa/ostream.hpp
  - include rocprofiler.h instead of tracer.hpp
- Update lib/rocprofiler/hsa
  - migration to use rocprofiler_hsa_api_callback_tracer_data_t instead of rocprofiler_hsa_trace_data_t
  - restructure hsa_api_impl<Idx>
    - remove phase_enter and phase_exit
    - add set_data_args (partial replacement for phase_enter)
    - functor handles the contexts
- Update lib/rocprofiler/rocprofiler.cpp
  - implement rocprofiler_get_version
- Update lib/rocprofiler/hsa/hsa.{hpp,cpp}
  - remove hsa_api_ prefix for functions already in hsa namespace
- Update lib/rocprofiler/context/context.{hpp,cpp}
  - add client_idx to context struct (tool identifier)
  - add push_client function to set client_idx before context is allocated
  - add pop_client function to remove client identifier from future context creations
  - implemented {registered,active}_contexts and buffers to use new container::reserve_size overload to stable_vector
  - fix implementation of start_context
  - fix implementation of stop_context
- Update lib/rocprofiler/rocprofiler.cpp
  - prevent context creation, buffer creation, pc sampling config, etc. after initialization
  - add nullptr checks to rocprofiler_context_is_valid
  - fix rocprofiler_configure_callback_tracing_service
    - was checking size of buffers, not registered context
  - implement rocprofiler_iterate_callback_trace_kind_names
  - implement rocprofiler_iterate_callback_trace_kind_operation_names
- Update lib/rocprofiler/CMakeLists.txt
  - add registration.{hpp,cpp} to rocprofiler-library target sources
- Update lib/rocprofiler/hsa/utils.hpp
  - fix using fmt::formt with const char* strings
  - remove join functions (no longer used)
- Update lib/rocprofiler/hsa/hsa.{hpp,cpp}
  - remove args_string function
  - remove named_args_string function
  - update iterate_args function
    - change callback type
    - accept user data
  - rework the hsa_api_impl<Idx>::functor function
    - save the rocprofiler_callback_tracing_record_t between callbacks
  - update update_table function
    - check buffered_tracer domains
  - remove comments
- Update lib/rocprofiler/hsa/defines.hpp
  - remove MEMBER_<N> macros
  - add ADDR_MEMBER_<N> macros
  - remove doxygen comments for GET_MEMBER_FIELDS
  - add GET_ADDR_MEMBER_FIELDS
  - update HSA_API_INFO_DEFINITION_{0,V}
    - rename domain_idx to callback_domain_idx
    - add buffered_domain_idx
    - add as_arg_addr function
- Update lib/rocprofiler/rocprofiler.cpp
  - implement rocprofiler_iterate_callback_trace_operation_args
- Remove lib/rocprofiler/tracing.{hpp,cpp} and lib/rocprofiler/CMakeLists.txt
  - unused
- Update lib/rocprofiler/hsa/hsa.{hpp,cpp}
  - support buffered tracing in hsa_api_impl<Idx>::functor
  - rocprofiler_callback_trace_operation_args_cb_t -> rocprofiler_callback_tracing_operation_args_cb_t
    - i.e. trace -> tracing
- Update lib/rocprofiler/context/context.{hpp,cpp}
  - removed buffer_instance struct
  - removed allocate_buffer function
  - removed get_buffers function
  - changed buffer_tracing_service::buffer_array_t
- Update lib/rocprofiler/hsa: hsa.cpp, ostream.hpp, details folder
  - move ostream.hpp into details folder to prevent from contributing to code coverage
  - update cmake build system for new directory

* Add lib/rocprofiler/registration.{hpp,cpp}

- implements rocprofiler_set_api_table (called by rocprofiler-register)
- miscellaneous functions for client configure/initialize/finalize
- functions for querying the init/fini status
- relocated OnLoad HSA workaround to this file
  - at present, this is used to workaround ROCr not having rocprofiler-register integration yet
- implement rocprofiler_force_configure function
- implement rocprofiler_is_initialized function
- implement rocprofiler_is_finalized function
- ensure configure functions only invoked once
- ensure internal thread creation notification functions are invoked
- get_status is pair of atomics
- fix heap-use-after-free in init_logging
- update finalize
  - invoke hsa_shut_down
  - set all active contexts to null pointers

* Add lib/rocprofiler/buffer_tracing.cpp

- contains implementations of buffer_tracing (i.e. rocprofiler/buffer_tracing.h)
- previous implementation may have been moved out of lib/rocprofiler/rocprofiler.cpp

* Add lib/rocprofiler/buffer.{hpp,cpp}

- contains implementations of buffer (i.e. rocprofiler/buffer.h) and misc internal access functions
- previous implementation may have been moved out of lib/rocprofiler/rocprofiler.cpp and lib/rocprofiler/context/context.{hpp,cpp}

* Add lib/rocprofiler/callback_tracing.cpp

- contains implementations of callback_tracing (i.e. rocprofiler/callback_tracing.h)
- previous implementation may have been moved out of lib/rocprofiler/rocprofiler.cpp

* Add lib/rocprofiler/context.cpp

- contains implementations of context public API functions (i.e. rocprofiler/context.h)
- previous implementation may have been moved out of lib/rocprofiler/rocprofiler.cpp

* Add lib/rocprofiler/internal_threading.{hpp,cpp}

- contains implementations of internal_threading (i.e. rocprofiler/internal_threading.h)
- also contains implementations of internal access functions
- update finalize function
  - join all task groups and destroy all thread pools first, then reset unique_ptr

* Update lib/rocprofiler/rocprofiler.cpp

- rocprofiler_get_version returns status
- implement rocprofiler_get_timestamp
- remove misc implementations that were split into other files

* Update lib/rocprofiler/CMakeLists.txt

- compile new implementation files
  - buffer.cpp
  - buffer_tracing.cpp
  - callback_tracing.cpp
  - context.cpp
  - internal_threading.cpp

* Update lib/tests/buffering/buffering-*.cpp

- update to reflect changes to rocprofiler_record_header_t

* Update CMakeLists.txt

- increase minimum cmake version to 3.21 which added HIP support as a language

* Add samples/apps/transpose

- simple HIP application for testing

* Add samples/api_callback_tracing

- HIP application and tool library
- This effectively demos how to setup HSA API tracing
  - For each function called in tool, it stores the func/file/line and prints it during finalization
- client.hpp and client.cpp are the tool library
- Implement use of rocprofiler_iterate_callback_trace_operation_args
- add demo of using rocprofiler_get_version
- add_test
  - remove PASS_REGULAR_EXPRESSION
    - causing false passes during memcheck
  - add ROCPROFILER_MEMCHECK_PRELOAD_ENV to environment
- check if rocprofiler is initialized before stopping context

* Add samples/api_buffered_tracing

- Sample demonstrating tracing the HSA API via buffering
- demo rocprofiler_record_header_compute_hash
- throw exceptions for unexpected buffer data
- add_test
  - remove PASS_REGULAR_EXPRESSION
    - causing false passes during memcheck
  - add ROCPROFILER_MEMCHECK_PRELOAD_ENV to environment

* Update samples/CMakeLists.txt

- add subdirectory for api_callback_tracing
- add subdirectory api_buffered_tracing

* Update samples/pc_sampling/common.h

- fix processing of headers

* Update lib/rocprofiler/hsa/details/ostream.hpp

- fix data race on HSA_depth_max_cnt and recursion
- HSA_depth_max_cnt and recursion is now thread-local static instead of global static
- replace std::string usage with std::string_view

* Actions update

- add dependabot.yml
- use actions/checkout@v4
- install latest libasan and libtsan in sanitizer containers

* Add PTL (Parallel Tasking Library) submodule
2023-09-20 19:32:02 -05:00

128 linhas
5.1 KiB
C++

#ifndef PC_SAMPLING_COMMON_H
#define PC_SAMPLING_COMMON_H
#include <rocprofiler/rocprofiler.h>
#include <cstddef>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <string>
#include <string_view>
constexpr size_t BUFFER_SIZE_BYTES = 4096;
constexpr size_t WATERMARK = (BUFFER_SIZE_BYTES / 2);
const std::string_view MI200_NAME = "gfx90a";
#define ROCPROFILER_CALL(result, msg) \
{ \
rocprofiler_status_t CHECKSTATUS = result; \
if(CHECKSTATUS != ROCPROFILER_STATUS_SUCCESS) \
{ \
puts(#result " failed"); \
} \
}
// We might want to test the calls that fails
// e.g. calling `rocprofiler_configure_pc_sampling_service `
// after previous initialization.
#define ROCPROFILER_CALL_FAILS(result, msg) \
{ \
rocprofiler_status_t CHECKSTATUS = result; \
if(CHECKSTATUS == ROCPROFILER_STATUS_SUCCESS) \
{ \
puts(#result " succeeded"); \
} \
}
static rocprofiler_status_t
find_first_gpu_agent_impl(rocprofiler_agent_t** agents, size_t num_agents, void* data)
{
// data is required
if(!data) return ROCPROFILER_STATUS_ERROR;
auto* _out_agent = static_cast<rocprofiler_agent_t*>(data);
// find the first GPU agent
for(size_t i = 0; i < num_agents; i++)
{
if(agents[i]->type == ROCPROFILER_AGENT_TYPE_GPU)
{
*_out_agent = *agents[i];
printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
__FUNCTION__,
_out_agent->name,
_out_agent->id.handle,
_out_agent->type,
_out_agent->pc_sampling_configs.size);
return ROCPROFILER_STATUS_SUCCESS;
}
else
{
printf("[%s] %s :: id=%zu, type=%i, num pc sample configs=%zu\n",
__FUNCTION__,
agents[i]->name,
agents[i]->id.handle,
agents[i]->type,
agents[i]->pc_sampling_configs.size);
}
}
return ROCPROFILER_STATUS_ERROR;
}
static rocprofiler_agent_t
find_first_gpu_agent()
{
// This function returns the first gpu agent it encounters.
// TODO: write the better function querying information about the agent,
// and return if the agent is MI200.
rocprofiler_agent_t gpu_agent;
ROCPROFILER_CALL(rocprofiler_query_available_agents(&find_first_gpu_agent_impl,
sizeof(rocprofiler_agent_t),
static_cast<void*>(&gpu_agent)),
"Could not query GPU agents");
return gpu_agent;
}
static void
rocprofiler_pc_sampling_callback(rocprofiler_context_id_t /*context_id*/,
rocprofiler_buffer_id_t /*buffer_id*/,
rocprofiler_record_header_t** headers,
size_t num_headers,
void* /*data*/,
uint64_t drop_count)
{
// Vladimir: I am not sure if this is the right way of iterating over PC sampling records.
printf(
"The number of delivered samples is: %zu, while the number of dropped samples is: %lu.\n",
num_headers,
drop_count);
for(size_t i = 0; i < num_headers; i++)
{
auto* cur_header = headers[i];
if(cur_header->category == ROCPROFILER_BUFFER_CATEGORY_PC_SAMPLING)
{
auto* pc_sample = static_cast<rocprofiler_pc_sampling_record_t*>(cur_header->payload);
printf("--- pc: %lx, dispatch_id: %lx, timestamp: %lu, hardware_id: %lu\n",
pc_sample->pc,
pc_sample->dispatch_id,
pc_sample->timestamp,
pc_sample->hardware_id);
// Vladimir: How to parse the remaining part of the `rocprofiler_pc_sampling_record_t`
// struct?
}
}
// Vladimr: We might want to add somewhere in the documentation that headars actually contain PC
// samples.
}
static void
run_HIP_app()
{
// TODO: provide the simple HIP app
}
#endif