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

API Tracing Overhaul (#437)

Telusuri Sumber 
* Update include/rocprofiler-sdk/hsa/*

- split HSA API IDs into separate enumerations
- add support for finalize ext table

* Update include/rocprofiler-sdk/hip/*

- remove compiler_api_args.h
- rocprofiler_hip_api_args_t contains all for HIP runtime and HIP compiler
- ROCPROFILER_HIP_API_ID_ -> ROCPROFILER_HIP_RUNTIME_API_ID_

* Update include/rocprofiler-sdk/marker/table_api_id.h

- ROCPROFILER_MARKER_API_TABLE_ID_ -> ROCPROFILER_MARKER_TABLE_ID_

* Update include/rocprofiler-sdk/*/table_api_id.h

- table_api_id.h -> table_id.h

* Update include/rocprofiler-sdk/*/table_api_id.h

- table_api_id.h -> table_id.h

* Update include/rocprofiler-sdk/fwd.h

- ROCPROFILER_CALLBACK_TRACING_HSA_API split into 4 enum values:
  - ROCPROFILER_CALLBACK_TRACING_HSA_CORE_API
  - ROCPROFILER_CALLBACK_TRACING_HSA_AMD_EXT_API
  - ROCPROFILER_CALLBACK_TRACING_HSA_IMAGE_EXT_API
  - ROCPROFILER_CALLBACK_TRACING_HSA_FINALIZE_EXT_API
- ROCPROFILER_BUFFER_TRACING_HSA_API split into 4 enum values:
  - ROCPROFILER_BUFFER_TRACING_HSA_CORE_API
  - ROCPROFILER_BUFFER_TRACING_HSA_AMD_EXT_API
  - ROCPROFILER_BUFFER_TRACING_HSA_IMAGE_EXT_API
  - ROCPROFILER_BUFFER_TRACING_HSA_FINALIZE_EXT_API
- rocprofiler_callback_tracing_code_object_operation_t renamed to rocprofiler_code_object_operation_t (more consistent)
- doxygen updates

* Update include/rocprofiler-sdk/buffer_tracing.h

- improved doxygen comments
- removed unused rocprofiler_buffer_tracing_queue_scheduling_record_t
- removed unused rocprofiler_buffer_tracing_correlation_record_t

* Update include/rocprofiler-sdk/callback_tracing.h

- removed rocprofiler_callback_tracing_hip_compiler_api_data_t
  - rocprofiler_hip_api_args_t and rocprofiler_hip_compiler_api_args_t were combined
  - rocprofiler_hsa_api_retval_t and rocprofiler_hsa_compiler_api_retval_t were combined

* Update lib/rocprofiler-sdk/hsa/*

- utils.hpp
  - formatters for hsa_ext_program_t and hsa_ext_control_directives_t
- defines.hpp
  - removed variadic macros from lib/common/defines.hpp
  - HSA_API_META_DEFINITION, HSA_API_INFO_DEFINITION_0, HSA_API_INFO_DEFINITION_V specialize on table id
- async_copy.cpp
  - ROCPROFILER_HSA_API_ID_* -> ROCPROFILER_HSA_AMD_EXT_API_ID_*
  - add table id to templates
  - improve async_copy_fini
- hsa.hpp
  - add hsa_table_id_lookup
  - add hsa_domain_info
  - add table id to templates
  - add copy_table function
- hsa.cpp
  - add table id to templates
  - require hsa tables to be trivial and standard layout
  - remove set_data_args specialization for hsa_amd_memory_async_copy_rect
  - implement copy_table function
- hsa.def.cpp
  - update enums

* Update lib/rocprofiler-sdk/hip/*

- defines.hpp
  - use lib/common/defines.hpp
  - add hip_table_id_lookup to HIP_API_TABLE_LOOKUP_DEFINITION
- hip.hpp
  - hip_table_id_lookup
  - template iterate_args on table id
  - templated copy_table and update_table
- hip.cpp
  - replaced api_id_bounds with hip_domain_info
  - templated iterate_args on table id
  - templated copy_table and update_table

* Update lib/rocprofiler-sdk/marker/*

- defines.hpp
  - use lib/common/defines.hpp
- marker.cpp
  - updated enums
- marker.def.cpp
  - updated enums

* Update lib/rocprofiler-sdk/tests

- common.hpp
  - ROCPROFILER_CALL_EXPECT
  - callback_data_ext
  - update get_callback_tracing_names with new enums
  - update get_buffer_tracing_names with new enums
- external_correlation.cpp
  - support new HSA API enums
- intercept_table.cpp
  - use test/common.hpp
  - update to new HSA API enums
- registration.cpp
  - support new HSA API enums
- naming.cpp
  - validation for all get_ids(), get_names(), name_by_id(), id_by_name(), etc.

* Update lib/common

- defines.hpp
  - Move IMPL_DETAIL_FOR_EACH_NARG, GET_ADDR_MEMBER_FIELDS, and GET_NAMED_MEMBER_FIELDS here
    - used by HSA, HIP, and Marker
- static_object.hpp
  - is_trivial_standard_layout static constexpr member function
  - suppress register_static_dtor when is_trivial_standard_layout

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

- name_by_id
- id_by_name
- get_names
- get_ids

* Update lib/rocprofiler-sdk/registration.cpp

- Update rocprofiler_set_api_table for HSA

* Update lib/rocprofiler-sdk/callback_tracing.cpp

- Update for new HSA enums
- Rework to use switch statement
  - rocprofiler_query_callback_tracing_kind_operation_name
  - rocprofiler_iterate_callback_tracing_kind_operations
  - rocprofiler_iterate_callback_tracing_kind_operation_args

* Update lib/rocprofiler-sdk/buffer_tracing.cpp

- Update for new HSA enums
- Rework to use switch statement
  - rocprofiler_query_buffer_tracing_kind_operation_name
  - rocprofiler_iterate_buffer_tracing_kind_operations

* Update lib/rocprofiler-sdk-tool

- helper.cpp
  - update get_buffer_id_names with new enums
  - update get_callback_id_names with new enums
- tools.cpp
  - update to use new HSA enums

* Update samples/common

- added call_stack.hpp
  - source_location struct
  - call_stack_t alias
  - print_call_stack function
- added name_info.hpp
  - utils for getting buffer/callback domain and operation names

* Update samples/api_buffered_tracing/client.cpp

- use samples/common/call_stack.hpp
- use samples/common/name_info.hpp
- update for new HSA enums

* Update samples/api_callback_tracing/client.cpp

- use samples/common/call_stack.hpp
- use samples/common/name_info.hpp
- update for new HSA enums

* Update tests/tools/json-tool.cpp

- update for new HSA enums

* Update tests/rocprofv3/tracing/validate.py

- update for new HSA domain names

* Update samples/counter_collection/main.cpp

- reduce number of kernels to 50,000 since 200,000 causes issues with thread sanitizer
This commit is contained in:
Jonathan R. Madsen authored and GitHub committed 2024-01-30 12:14:26 -06:00
1 orang tua 14cd132b89
melakukan 3f39339926
58 files changed
+3733 -2695
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

source/lib/rocprofiler-sdk/hsa/code_object.cpp
+107
Melihat File
@@ -57,6 +57,113 @@ namespace rocprofiler
{
namespace hsa
{
namespace code_object
{
namespace
{
using context_t = context::context;
using context_array_t = common::container::small_vector<const context_t*>;
using external_corr_id_map_t = std::unordered_map<const context_t*, rocprofiler_user_data_t>;
template <size_t OpIdx>
struct code_object_info;
#define SPECIALIZE_CODE_OBJECT_INFO(OPERATION) \
template <> \
struct code_object_info<ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_##OPERATION> \
{ \
static constexpr auto operation_idx = \
ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_##OPERATION; \
static constexpr auto name = #OPERATION; \
};
SPECIALIZE_CODE_OBJECT_INFO(NONE)
SPECIALIZE_CODE_OBJECT_INFO(LOAD)
SPECIALIZE_CODE_OBJECT_INFO(DEVICE_KERNEL_SYMBOL_REGISTER)
#undef SPECIALIZE_CODE_OBJECT_INFO
template <size_t Idx, size_t... IdxTail>
const char*
name_by_id(const uint32_t id, std::index_sequence<Idx, IdxTail...>)
{
if(Idx == id) return code_object_info<Idx>::name;
if constexpr(sizeof...(IdxTail) > 0)
return name_by_id(id, std::index_sequence<IdxTail...>{});
else
return nullptr;
}
template <size_t Idx, size_t... IdxTail>
uint32_t
id_by_name(const char* name, std::index_sequence<Idx, IdxTail...>)
{
if(std::string_view{code_object_info<Idx>::name} == std::string_view{name})
return code_object_info<Idx>::operation_idx;
if constexpr(sizeof...(IdxTail) > 0)
return id_by_name(name, std::index_sequence<IdxTail...>{});
else
return ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_NONE;
}
template <size_t... Idx>
void
get_ids(std::vector<uint32_t>& _id_list, std::index_sequence<Idx...>)
{
auto _emplace = [](auto& _vec, uint32_t _v) {
if(_v < static_cast<uint32_t>(ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST))
_vec.emplace_back(_v);
};
(_emplace(_id_list, code_object_info<Idx>::operation_idx), ...);
}
template <size_t... Idx>
void
get_names(std::vector<const char*>& _name_list, std::index_sequence<Idx...>)
{
auto _emplace = [](auto& _vec, const char* _v) {
if(_v != nullptr && strnlen(_v, 1) > 0) _vec.emplace_back(_v);
};
(_emplace(_name_list, code_object_info<Idx>::name), ...);
}
} // namespace
// check out the assembly here... this compiles to a switch statement
const char*
name_by_id(uint32_t id)
{
return name_by_id(id,
std::make_index_sequence<ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST>{});
}
uint32_t
id_by_name(const char* name)
{
return id_by_name(name,
std::make_index_sequence<ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST>{});
}
std::vector<uint32_t>
get_ids()
{
auto _data = std::vector<uint32_t>{};
_data.reserve(ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST);
get_ids(_data, std::make_index_sequence<ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST>{});
return _data;
}
std::vector<const char*>
get_names()
{
auto _data = std::vector<const char*>{};
_data.reserve(ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST);
get_names(_data, std::make_index_sequence<ROCPROFILER_CALLBACK_TRACING_CODE_OBJECT_LAST>{});
return _data;
}
} // namespace code_object
namespace
{
using hsa_loader_table_t = hsa_ven_amd_loader_1_01_pfn_t;