Files
rocm-systems/projects/rocprofiler-systems/source/lib/omnitrace/library/rocm_smi.cpp
T
Jonathan R. Madsen a1b11b94f0 Dynamic expansion of thread data (#294)
* Tests for exceeding OMNITRACE_MAX_THREADS

- tests which exceeds OMNITRACE_MAX_THREADS value for thread creation

* CMake Formatting.cmake update

- include source files in /tests/source directory

* Add unknown-hash= to OMNITRACE_ABORT_FAIL_REGEX

- fail if a timemory hash is not resolved to a name

* Tests for exceeding OMNITRACE_MAX_THREADS

- update

* omnitrace-sample update

- remove env disabling of critical-trace and process-sampling

* core library update

- make_unique in concepts.hpp
- add OMNITRACE_USE_ROCM_SMI to "process_sampling" category
- remove forced disabling of critical-trace in sampling mode
- parentheses for OMNITRACE_PREFER
- use tim::get_hash_id instead of tim::get_combined_hash_id

* core library update (containers)

- added aligned_static_vector.hpp
  - similar to static_vector.hpp but attempts to align to cache line size
- alignment template parameter for stable_vector
- added missing aliases in static_vector
  - consistent with aligned_static_vector aliases

* thread_info update

- track the peak number of threads created
- thread_info::get_peak_num_threads() returns the peak number of threads

* thread_data update

- generic thread_data inherits from base_thread_data
- thread_data reworked to support dynamic expansion
- base_thread_data updated to invoke private_instance() function
- thread_data<optional<T>> uses stable_vector aligned to cache line width
- thread_data<identity<T>> uses stable_vector aligned to cache line width
- thread_data for optional and identity provide private private_instance function + friend to base_thread_data
- component_bundle_cache<T> is now thread_data<component_bundle_cache_impl<T>>

* causal update

- thread_data<T>::instances -> thread_data<T>::instance(construct_on_thread{ ... })
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()
- tim::get_combined_hash_id -> tim::get_hash_id
- update progress_bundle usage to new thread_data API

* backtrace/backtrace_metrics component update

- backtrace_metrics update
  - update to new thead_data API
  - add thread CPU time row in perfetto
  - fix potential bug when rusage categories are disabled
  - fix bug in operator-= not subtracting cpu time of rhs
- backtrace update
  - skip all child call-stack below 'tim::openmp::' if sampling_keep_internal = false

* pthread_gotcha component update

- pthread_gotcha::shutdown() invokes pthread_create_gotcha::shutdown()

* pthread_create_gotcha component update

- minor tweak to {start,stop}_bundle functions: pass in thread id
- update to new thread_data API
- track native handles of internal threads
- implement system with pthread_kill to stop dangling bundles

* rocprofiler/roctracer component update

- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()

* critical trace (library) update

- update to new thread_data API
- tim::get_combined_hash_id -> tim::get_hash_id

* coverage update

- update to new thread_data API

* tasking update

- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()

* roctracer update

- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()

* rocm_smi update

- update to new thread_data API

* runtime.cpp update

- update to new thread_data API

* sampling.cpp update

- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()

* ompt.cpp update

- invoke pthread_gotcha::shutdown before invoking OMPT finalize function
  - this prevents signals from being delivered to OpenMP threads

* tracing.hpp and tracing.cpp update

- replace get_timemory_hash_{ids,aliases} functions with copy_timemory_hash_ids function
- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()
- tim::get_combined_hash_id -> tim::get_hash_id
- improvements to + error checking in thread_init function

* library.cpp update

- move copying timemory hash id/aliases to tracing.cpp
- update to new thread_data API
- loop over max_supported_threads (constexpr) -> loop over thread_info::get_peak_num_threads()

* Update BuildSettings.cmake

- add -Wno-interference-size to suppress warning about use of std::hardware_destructive_interference

* Update fork example

- improve scheme for waiting on child processes via waitpid instead of wait
- support running main routine multiple times
- push/pop regions in child process

* Update lib/common/defines.h.in

- allow use to specify misc values via -D <name>=<value>
  - OMNITRACE_CACHELINE_SIZE
  - OMNITRACE_CACHELINE_SIZE_MIN
  - OMNITRACE_ROCM_MAX_COUNTERS
- remove unused defines
  - OMNITRACE_ROCM_LOOK_AHEAD
  - OMNITRACE_MAX_ROCM_QUEUES

* Update rocprofiler.hpp

- OMNITRACE_MAX_ROCM_COUNTERS -> OMNITRACE_ROCM_MAX_COUNTERS

* Update aligned_static_vector

- set cacheline_align_v from max of OMNITRACE_CACHELINE_SIZE and OMNITRACE_CACHELINE_SIZE_MIN

* Update tracing.cpp

- acquire locks for updating main hash ids/aliases
- only propagate ids/aliases when finalizing

* Update pthread_create_gotcha.cpp

- make sure hash for "start_thread" exists on main thread

* Update causal end to end tests

- if OMNITRACE_BUILD_NUMBER is 1, set OMNITRACE_VERBOSE=0

[ROCm/rocprofiler-systems commit: 518c83e0f9]
2023-10-16 18:04:47 -05:00

466 строки
15 KiB
C++

// Copyright (c) 2018 Advanced Micro Devices, Inc. All Rights Reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// with the Software without restriction, including without limitation the
// rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
// sell copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// * Redistributions of source code must retain the above copyright notice,
// this list of conditions and the following disclaimers.
//
// * Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimers in the
// documentation and/or other materials provided with the distribution.
//
// * Neither the names of Advanced Micro Devices, Inc. nor the names of its
// contributors may be used to endorse or promote products derived from
// this Software without specific prior written permission.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
// THE SOFTWARE.
#if defined(NDEBUG)
# undef NDEBUG
#endif
#include "library/rocm_smi.hpp"
#include "core/common.hpp"
#include "core/components/fwd.hpp"
#include "core/config.hpp"
#include "core/debug.hpp"
#include "core/gpu.hpp"
#include "core/perfetto.hpp"
#include "core/state.hpp"
#include "library/critical_trace.hpp"
#include "library/runtime.hpp"
#include "library/thread_info.hpp"
#include <timemory/backends/threading.hpp>
#include <timemory/components/timing/backends.hpp>
#include <timemory/units.hpp>
#include <timemory/utility/locking.hpp>
#include <rocm_smi/rocm_smi.h>
#include <cassert>
#include <chrono>
#include <ios>
#include <sstream>
#include <stdexcept>
#include <string>
#include <sys/resource.h>
#include <thread>
#define OMNITRACE_ROCM_SMI_CALL(ERROR_CODE) \
::omnitrace::rocm_smi::check_error(ERROR_CODE, __FILE__, __LINE__)
namespace omnitrace
{
namespace rocm_smi
{
using bundle_t = std::deque<data>;
using sampler_instances = thread_data<bundle_t, category::rocm_smi>;
namespace
{
bool&
is_initialized()
{
static bool _v = false;
return _v;
}
void
check_error(rsmi_status_t _code, const char* _file, int _line)
{
if(_code == RSMI_STATUS_SUCCESS) return;
const char* _msg = nullptr;
auto _err = rsmi_status_string(_code, &_msg);
if(_err != RSMI_STATUS_SUCCESS)
OMNITRACE_THROW("rsmi_status_string failed. No error message available. "
"Error code %i originated at %s:%i\n",
static_cast<int>(_code), _file, _line);
OMNITRACE_THROW("[%s:%i] Error code %i :: %s", _file, _line, static_cast<int>(_code),
_msg);
}
std::atomic<State>&
get_state()
{
static std::atomic<State> _v{ State::PreInit };
return _v;
}
} // namespace
//--------------------------------------------------------------------------------------//
size_t data::device_count = 0;
std::set<uint32_t> data::device_list = {};
std::unique_ptr<data::promise_t> data::polling_finished = {};
data::data(uint32_t _dev_id) { sample(_dev_id); }
void
data::sample(uint32_t _dev_id)
{
auto _ts = tim::get_clock_real_now<size_t, std::nano>();
assert(_ts < std::numeric_limits<int64_t>::max());
auto _state = get_state().load();
if(_state != State::Active) return;
m_dev_id = _dev_id;
m_ts = _ts;
#define OMNITRACE_RSMI_GET(FUNCTION, ...) \
try \
{ \
OMNITRACE_ROCM_SMI_CALL(FUNCTION(__VA_ARGS__)); \
} catch(std::runtime_error & _e) \
{ \
OMNITRACE_VERBOSE_F( \
0, "[%s] Exception: %s. Disabling future samples from rocm-smi...\n", \
#FUNCTION, _e.what()); \
get_state().store(State::Disabled); \
}
OMNITRACE_RSMI_GET(rsmi_dev_busy_percent_get, _dev_id, &m_busy_perc);
OMNITRACE_RSMI_GET(rsmi_dev_temp_metric_get, _dev_id, RSMI_TEMP_TYPE_EDGE,
RSMI_TEMP_CURRENT, &m_temp);
OMNITRACE_RSMI_GET(rsmi_dev_power_ave_get, _dev_id, 0, &m_power);
OMNITRACE_RSMI_GET(rsmi_dev_memory_usage_get, _dev_id, RSMI_MEM_TYPE_VRAM,
&m_mem_usage);
#undef OMNITRACE_RSMI_GET
}
void
data::print(std::ostream& _os) const
{
std::stringstream _ss{};
_ss << "device: " << m_dev_id << ", busy = " << m_busy_perc << "%, temp = " << m_temp
<< ", power = " << m_power << ", memory usage = " << m_mem_usage;
_os << _ss.str();
}
namespace
{
std::vector<unique_ptr_t<bundle_t>*> _bundle_data{};
}
void
config()
{
_bundle_data.resize(data::device_count, nullptr);
for(size_t i = 0; i < data::device_count; ++i)
{
if(data::device_list.count(i) > 0)
{
_bundle_data.at(i) = &sampler_instances::get()->at(i);
if(!*_bundle_data.at(i))
*_bundle_data.at(i) = unique_ptr_t<bundle_t>{ new bundle_t{} };
}
}
data::get_initial().resize(data::device_count);
for(auto itr : data::device_list)
data::get_initial().at(itr).sample(itr);
}
void
sample()
{
for(auto itr : data::device_list)
{
if(rocm_smi::get_state() != State::Active) continue;
OMNITRACE_DEBUG_F("Polling rocm-smi for device %u...\n", itr);
auto& _data = *_bundle_data.at(itr);
if(!_data) continue;
_data->emplace_back(data{ itr });
OMNITRACE_DEBUG_F(" %s\n", TIMEMORY_JOIN("", _data->back()).c_str());
}
}
void
set_state(State _v)
{
rocm_smi::get_state().store(_v);
}
std::vector<data>&
data::get_initial()
{
static std::vector<data> _v{};
return _v;
}
bool
data::setup()
{
perfetto_counter_track<data>::init();
rocm_smi::set_state(State::PreInit);
return true;
}
bool
data::shutdown()
{
OMNITRACE_DEBUG("Shutting down rocm-smi...\n");
rocm_smi::set_state(State::Finalized);
return true;
}
#define GPU_METRIC(COMPONENT, ...) \
if constexpr(tim::trait::is_available<COMPONENT>::value) \
{ \
auto* _val = _v.get<COMPONENT>(); \
if(_val) \
{ \
_val->set_value(itr.__VA_ARGS__); \
_val->set_accum(itr.__VA_ARGS__); \
} \
}
void
data::post_process(uint32_t _dev_id)
{
using component::sampling_gpu_busy;
using component::sampling_gpu_memory;
using component::sampling_gpu_power;
using component::sampling_gpu_temp;
if(device_count < _dev_id) return;
auto& _rocm_smi_v = sampler_instances::get()->at(_dev_id);
auto _rocm_smi = (_rocm_smi_v) ? *_rocm_smi_v : std::deque<rocm_smi::data>{};
const auto& _thread_info = thread_info::get(0, InternalTID);
OMNITRACE_VERBOSE(1, "Post-processing %zu rocm-smi samples from device %u\n",
_rocm_smi.size(), _dev_id);
OMNITRACE_CI_THROW(!_thread_info, "Missing thread info for thread 0");
if(!_thread_info) return;
auto _process_perfetto = [&]() {
for(auto& itr : _rocm_smi)
{
using counter_track = perfetto_counter_track<data>;
if(itr.m_dev_id != _dev_id) continue;
if(!counter_track::exists(_dev_id))
{
auto addendum = [&](const char* _v) {
return JOIN(" ", "GPU", _v, JOIN("", '[', _dev_id, ']'), "(S)");
};
counter_track::emplace(_dev_id, addendum("Busy"), "%");
counter_track::emplace(_dev_id, addendum("Temperature"), "deg C");
counter_track::emplace(_dev_id, addendum("Power"), "watts");
counter_track::emplace(_dev_id, addendum("Memory Usage"), "megabytes");
}
uint64_t _ts = itr.m_ts;
if(!_thread_info->is_valid_time(_ts)) continue;
double _busy = itr.m_busy_perc;
double _temp = itr.m_temp / 1.0e3;
double _power = itr.m_power / 1.0e6;
double _usage = itr.m_mem_usage / static_cast<double>(units::megabyte);
TRACE_COUNTER("device_busy", counter_track::at(_dev_id, 0), _ts, _busy);
TRACE_COUNTER("device_temp", counter_track::at(_dev_id, 1), _ts, _temp);
TRACE_COUNTER("device_power", counter_track::at(_dev_id, 2), _ts, _power);
TRACE_COUNTER("device_memory_usage", counter_track::at(_dev_id, 3), _ts,
_usage);
}
};
if(get_use_perfetto()) _process_perfetto();
if(!get_use_timemory()) return;
#if !defined(TIMEMORY_USE_MPI)
// timemory + MPI here causes hangs for some reason. it is unclear why
using samp_bundle_t = tim::lightweight_tuple<sampling_gpu_busy, sampling_gpu_temp,
sampling_gpu_power, sampling_gpu_memory>;
using entry_t = critical_trace::entry;
auto _gpu_entries = critical_trace::get_entries(
[](const entry_t& _e) { return (_e.device == critical_trace::Device::GPU); });
for(auto& itr : _rocm_smi)
{
auto _ts = itr.m_ts;
if(!_thread_info->is_valid_time(_ts)) continue;
auto _entries = std::vector<std::pair<std::string_view, const entry_t*>>{};
for(const auto& eitr : _gpu_entries)
{
if(_ts >= eitr.second.begin_ns && _ts <= eitr.second.end_ns)
_entries.emplace_back(std::string_view{ eitr.first }, &eitr.second);
}
std::vector<samp_bundle_t> _tc{};
_tc.reserve(_entries.size());
for(auto& eitr : _entries)
{
auto& _v = _tc.emplace_back(eitr.first);
_v.push();
_v.start();
_v.stop();
GPU_METRIC(sampling_gpu_busy, m_busy_perc)
GPU_METRIC(sampling_gpu_temp, m_temp / 1.0e3) // provided in milli-degree C
GPU_METRIC(sampling_gpu_power,
m_power * units::microwatt / static_cast<double>(units::watt))
GPU_METRIC(sampling_gpu_memory,
m_mem_usage / static_cast<double>(units::megabyte))
_v.pop();
}
}
#endif
}
//--------------------------------------------------------------------------------------//
void
setup()
{
auto_lock_t _lk{ type_mutex<category::rocm_smi>() };
if(is_initialized() || !get_use_rocm_smi()) return;
OMNITRACE_SCOPED_SAMPLING_ON_CHILD_THREADS(false);
// assign the data value to determined by rocm-smi
data::device_count = device_count();
auto _devices_v = get_sampling_gpus();
for(auto& itr : _devices_v)
itr = tolower(itr);
if(_devices_v == "off")
_devices_v = "none";
else if(_devices_v == "on")
_devices_v = "all";
bool _all_devices = _devices_v.find("all") != std::string::npos || _devices_v.empty();
bool _no_devices = _devices_v.find("none") != std::string::npos;
std::set<uint32_t> _devices = {};
auto _emplace = [&_devices](auto idx) {
if(idx < data::device_count) _devices.emplace(idx);
};
if(_all_devices)
{
for(uint32_t i = 0; i < data::device_count; ++i)
_emplace(i);
}
else if(!_no_devices)
{
auto _enabled = tim::delimit(_devices_v, ",; \t");
for(auto&& itr : _enabled)
{
if(itr.find_first_not_of("0123456789-") != std::string::npos)
{
OMNITRACE_THROW("Invalid GPU specification: '%s'. Only numerical values "
"(e.g., 0) or ranges (e.g., 0-7) are permitted.",
itr.c_str());
}
if(itr.find('-') != std::string::npos)
{
auto _v = tim::delimit(itr, "-");
OMNITRACE_CONDITIONAL_THROW(_v.size() != 2,
"Invalid GPU range specification: '%s'. "
"Required format N-M, e.g. 0-4",
itr.c_str());
for(auto i = std::stoul(_v.at(0)); i < std::stoul(_v.at(1)); ++i)
_emplace(i);
}
else
{
_emplace(std::stoul(itr));
}
}
}
data::device_list = _devices;
try
{
for(auto itr : _devices)
{
uint16_t dev_id = 0;
OMNITRACE_ROCM_SMI_CALL(rsmi_dev_id_get(itr, &dev_id));
// dev_id holds the device ID of device i, upon a successful call
}
is_initialized() = true;
data::setup();
} catch(std::runtime_error& _e)
{
OMNITRACE_VERBOSE(0, "Exception thrown when initializing rocm-smi: %s\n",
_e.what());
data::device_list = {};
}
}
void
shutdown()
{
auto_lock_t _lk{ type_mutex<category::rocm_smi>() };
if(!is_initialized()) return;
try
{
if(data::shutdown())
{
OMNITRACE_ROCM_SMI_CALL(rsmi_shut_down());
}
} catch(std::runtime_error& _e)
{
OMNITRACE_VERBOSE(0, "Exception thrown when shutting down rocm-smi: %s\n",
_e.what());
}
is_initialized() = false;
}
void
post_process()
{
for(auto itr : data::device_list)
data::post_process(itr);
}
uint32_t
device_count()
{
return gpu::rsmi_device_count();
}
} // namespace rocm_smi
} // namespace omnitrace
OMNITRACE_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, omnitrace::component::backtrace_gpu_busy>), true,
double)
OMNITRACE_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, omnitrace::component::backtrace_gpu_temp>), true,
double)
OMNITRACE_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, omnitrace::component::backtrace_gpu_power>), true,
double)
OMNITRACE_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, omnitrace::component::backtrace_gpu_memory>), true,
double)