AI/rocm-systems
2
0
Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

Code object, kernel dispatch, and memory copy tracing (#177)

Telusuri Sumber 
* Update samples/api_buffered_tracing

- external correlation id
- support ROCPROFILER_BUFFER_TRACING_KERNEL_DISPATCH

* Update lib/rocprofiler/context.cpp

- update alternative get_active_contexts paradigm

* Update lib/rocprofiler/external_correlation.cpp

- inherit correlation id from main thread

* Update lib/rocprofiler/hsa/queue.*

- typedef changes
- rocprofiler_packet union
- modify Queue::queue_info_session_t
  - use rocprofiler_packet
  - add thread id
  - add kernel id
  - add correlation id
- out of line definitions
- AsyncSignalHandler function update
  - handle kernel dispatch tracing
- Move CreateBarrierPacket and AddVendorSpecificPacket to lambdas
- handle contexts

* Update lib/rocprofiler/hsa/hsa.cpp

- remove unnecessary log function
- use new get_active_contexts paradigm
- use new correlation id updates

* Update AgentCache and kernel dispatch record

- include const rocprofiler_agent_t* in rocprofiler_buffer_tracing_kernel_dispatch_record_t
- AgentCache::get_rocp_agent returns const pointer

* Replace ROCPROFILER_SERVICE_ with ROCPROFILER_

* source formatting

* Code Object Tracing

- include/rocprofiler/callback_tracing.h
  - remove rocprofiler_callback_tracing_code_object_unload_data_t
  - remove rocprofiler_callback_tracing_code_object_kernel_symbol_register_data_t
- include/rocprofiler/fwd.h
  - remove ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_UNLOAD
  - remove ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_DEVICE_KERNEL_SYMBOL_UNREGISTER
- lib/common/utility.hpp
  - assert_public_api_struct_properties()
  - init_public_api_struct(...)
- lib/rocprofiler/registration.cpp
  - invoke hsa::code_object_init
- lib/rocprofiler/hsa/CMakeLists.txt
  - compile code_object code
- lib/rocprofiler/hsa/code_object.{hpp,cpp}
  - tracing code object load/unload
- lib/rocprofiler/hsa/queue.cpp
  - get_kernel_id

* Update lib/rocprofiler/hsa/hsa.cpp

- fix should_wrap_functor logic (which was not handling callback_tracer + buffered_tracer properly)

* Update lib/rocprofiler/hsa/queue.cpp

- fix rocprofiler_buffer_tracing_kernel_dispatch_record_t construction

* Update samples/api_buffered_tracing/client.cpp

- print kernel names

* Move samples/apps to tests/apps

* Update lib/rocprofiler/hsa/code_object.cpp

- ensure unload callbacks when application is exiting
- support user data in between load/unload callbacks

* Update lib/rocprofiler/hsa/queue.{hpp,cpp}

- store contexts and external correlation ids in queue_info_session
- reduce signal_limiter to 96 to fix hangs
- fix support for kernel tracing and async memory copies

* Add lib/common/scope_destructor.hpp

- similar to static_cleanup_wrapper but different

* Update include/rocprofiler/buffer_tracing.h

- update rocprofiler_buffer_tracing_memory_copy_record_t
- remove operation: user can figure that out from correlation id
- add kernel id
- add rocprofiler agent id

* Update include/rocprofiler/callback_tracing.h

- fix data type of load_delta field in code object
- remove rocp_agent from kernel_symbol_register_data_t (known via code_object_id)

* Add samples/code_object_tracing

- sample demonstrating code object tracing

* Update samples

- minor tweak to print_call_stack

* Update lib/rocprofiler/hsa/code_object.cpp

- flip ordering of unload callbacks for code object unloading and kernel symbol deregistering

* clang-tidy fixes

* Update lib/rocprofiler/hsa/code_object.cpp

- fix heap-use-after-free issue with code object

* Update include/rocprofiler/external_correlation.h

- update documentation to include info about default value of external correlation value

* Use common::container::small_vector for contexts

- small_vector<const context*> is an ideal data structure for array of active contexts

* Update context handling for code object unload

- code object unload is only called for contexts which received the load callback

* Update samples

- improve ROCPROFILER_CALL macro to include status string
- api_buffered_tracing handles ROCPROFILER_STATUS_ERROR_BUFFER_BUSY

* Code object shutdown

- ensure code object callbacks are invoked prior to finalizing

* Update lib/common (memory allocators)

- added lib/common/memory folder with allocators

* Add lib/rocprofiler/allocator.*

- rocprofiler::allocator::static_data_allocator
  - special allocator for static data which finalizes before any data gets destroyed
- rocprofiler::allocator::unique_static_ptr_t
  - unique_ptr that uses static data deleter (ensure finalize is called)

* Update lib/rocprofiler/buffer.cpp

- flush checks fini status
- use unique_static_ptr_t

* Update lib/rocprofiler/internal_threading.*

- change meaning of thread_pool_t and task_group_t
- improve finalization to prevent data races and heap-use-after-free

* Update lib/rocprofiler/registration.cpp

- use static_data_allocator for client_library vector

* Update lib/rocprofiler/context/context.*

- use allocator::unique_static_ptr_t

* Update lib/rocprofiler/allocator.cpp

- avoid deadlock in deleter<static_data>::operator()

* Update lib/rocprofiler/registration.cpp

- avoid deadlock in rocprofiler::registration::finalize()

* Update lib/rocprofiler/hsa/code_object.cpp

- suppress duplicate reporting of code-object/kernel-symbol load/unload

* Update leak sanitizer suppressions

- __new_exitfn (via stdlib/cxa_atexit.c leaks

[ROCm/rocprofiler-sdk commit: 3082288a25]
This commit is contained in:
Jonathan R. Madsen
2023-11-13 22:30:15 -06:00
committed by GitHub
orang tua c57bfb22f8
melakukan 6108ea7efd
41 mengubah file dengan 2944 tambahan dan 174 penghapusan
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler-sdk/source/include/rocprofiler/buffer_tracing.h
+5 -8
Melihat File
@@ -86,14 +86,11 @@ typedef struct
uint64_t size;
rocprofiler_service_buffer_tracing_kind_t kind;
rocprofiler_correlation_id_t correlation_id;
/**
* Memory copy operation that can be derived from
* ::rocprofiler_tracing_operation_t
*/
uint32_t operation;
rocprofiler_timestamp_t start_timestamp;
rocprofiler_timestamp_t end_timestamp;
rocprofiler_queue_id_t queue_id;
rocprofiler_timestamp_t start_timestamp;
rocprofiler_timestamp_t end_timestamp;
rocprofiler_agent_id_t agent_id;
rocprofiler_queue_id_t queue_id;
rocprofiler_kernel_id_t kernel_id;
} rocprofiler_buffer_tracing_memory_copy_record_t;
/**
projects/rocprofiler-sdk/source/include/rocprofiler/callback_tracing.h
+6 -7
Melihat File
@@ -97,7 +97,7 @@ typedef struct
///< the code object that is loaded. Note that any non-loaded segments
///< before the first loaded segment are ignored.
uint64_t load_size; ///< The byte size of the loaded code objects contiguous memory allocation.
uint64_t load_delta; ///< The signed byte address difference of the memory address at which the
int64_t load_delta; ///< The signed byte address difference of the memory address at which the
///< code object is loaded minus the virtual address specified in the code
///< object that is loaded.
rocprofiler_code_object_storage_type_t
@@ -131,12 +131,11 @@ typedef struct
*/
typedef struct
{
uint64_t size; ///< size of this struct
uint64_t kernel_id; ///< unique symbol identifier value
uint64_t code_object_id; ///< parent unique code object identifier
rocprofiler_agent_id_t rocp_agent; ///< Agent associated with this symbol
const char* kernel_name; ///< name of the kernel
uint64_t kernel_object; ///< kernel object handle, used in the kernel dispatch packet
uint64_t size; ///< size of this struct
uint64_t kernel_id; ///< unique symbol identifier value
uint64_t code_object_id; ///< parent unique code object identifier
const char* kernel_name; ///< name of the kernel
uint64_t kernel_object; ///< kernel object handle, used in the kernel dispatch packet
uint32_t kernarg_segment_size; ///< size of memory (in bytes) allocated for kernel arguments.
///< Will be multiple of 16
uint32_t kernarg_segment_alignment; ///< Alignment (in bytes) of the buffer used to pass
projects/rocprofiler-sdk/source/include/rocprofiler/external_correlation.h
+7
Melihat File
@@ -37,6 +37,13 @@ ROCPROFILER_EXTERN_C_INIT
/**
* @brief Push default value for `external` field in @ref rocprofiler_correlation_id_t onto stack.
*
* External correlation ids are thread-local values. However, if rocprofiler internally requests an
* external correlation id on a non-main thread and an external correlation id has not been pushed
* for this thread, the external correlation ID will default to the latest external correlation id
* on the main thread -- this allows tools to push an external correlation id once on the main
* thread for, say, the MPI rank or process-wide UUID and this value will be used by all subsequent
* child threads.
*
* @param [in] context Associated context
* @param [in] tid thread identifier. @see rocprofiler_get_thread_id
* @param [in] external_correlation_id User data to place in external field in @ref