Files
rocm-systems/source/lib/omnitrace/library/critical_trace.hpp
T
Jonathan R. Madsen 9ff4b6b624 Dramatic improvement of post-processing critical trace data (#185)
- several orders of magnitude faster
2022-10-21 09:23:52 -05:00

369 строки
12 KiB
C++

// MIT License
//
// Copyright (c) 2022 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
// in 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:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
//
// 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
// AUTHORS 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 IN THE
// SOFTWARE.
#pragma once
#include "library/common.hpp"
#include "library/config.hpp"
#include "library/defines.hpp"
#include "library/runtime.hpp"
#include "library/thread_data.hpp"
#include <timemory/backends/process.hpp>
#include <timemory/backends/threading.hpp>
#include <timemory/hash/types.hpp>
#include <timemory/macros/language.hpp>
#include <timemory/tpls/cereal/cereal.hpp>
#include <timemory/utility/demangle.hpp>
#include <timemory/utility/utility.hpp>
#include <cstdint>
#include <cstdlib>
#include <mutex>
#include <ostream>
#include <string>
#include <vector>
namespace omnitrace
{
namespace critical_trace
{
enum class Device : uint8_t
{
NONE = 0,
CPU,
GPU,
ANY,
};
enum class Phase : uint8_t
{
NONE = 0,
BEGIN,
END,
DELTA,
};
struct OMNITRACE_ATTRIBUTE(packed) entry
{
entry() = default;
~entry() = default;
entry(const entry&) = default;
entry(entry&&) noexcept = default;
entry& operator=(const entry&) = default;
entry& operator=(entry&&) noexcept = default;
Device device = Device::CPU; /// which device it executed on
Phase phase = Phase::NONE; /// start / stop / unspecified
uint16_t priority = 0; /// priority value (for sorting)
uint32_t depth = 0; /// call-stack depth
int32_t devid = 0; /// device id
int32_t pid = 0; /// process id
int32_t tid = 0; /// thread id it was registered on
uint64_t cpu_cid = 0; /// CPU correlation id
uint64_t gpu_cid = 0; /// GPU correlation id
uint64_t parent_cid = 0; /// parent CPU correlation id
int64_t begin_ns = 0; /// timestamp of start
int64_t end_ns = 0; /// timestamp of end
uintptr_t queue_id = 0; /// stream id (GPU) or mutex id
size_t hash = 0; /// hash for name
bool operator==(const entry& rhs) const;
bool operator!=(const entry& rhs) const { return !(*this == rhs); }
bool operator<(const entry& rhs) const;
bool operator>(const entry& rhs) const;
bool operator<=(const entry& rhs) const { return !(*this > rhs); }
bool operator>=(const entry& rhs) const { return !(*this < rhs); }
entry& operator+=(const entry& rhs);
size_t get_hash() const;
int64_t get_timestamp() const;
int64_t get_cost() const;
bool is_bounded(const entry& rhs) const;
int64_t get_overlap(const entry& rhs) const;
int64_t get_independent(const entry& rhs) const;
int64_t get_overlap(const entry& rhs, int32_t _devid, int32_t _pid,
int64_t _tid) const;
int64_t get_independent(const entry& rhs, int32_t _devid, int32_t _pid,
int64_t _tid) const;
bool is_bounded(const entry& rhs, int32_t _devid, int32_t _pid, int64_t _tid) const;
void write(std::ostream& _os) const;
static bool is_delta(const entry&, const std::string_view&);
friend std::ostream& operator<<(std::ostream& _os, const entry& _v)
{
_v.write(_os);
return _os;
}
template <typename Archive>
void save(Archive& ar, unsigned int) const;
template <typename Archive>
void load(Archive& ar, unsigned int);
};
template <typename Archive>
void
entry::save(Archive& ar, unsigned int) const
{
namespace cereal = tim::cereal;
#define SAVE_PACKED_ENTRY_FIELD(VAR) \
{ \
auto _val = VAR; \
ar(cereal::make_nvp(#VAR, _val)); \
}
SAVE_PACKED_ENTRY_FIELD(priority);
SAVE_PACKED_ENTRY_FIELD(device);
SAVE_PACKED_ENTRY_FIELD(phase);
SAVE_PACKED_ENTRY_FIELD(depth);
SAVE_PACKED_ENTRY_FIELD(devid);
SAVE_PACKED_ENTRY_FIELD(pid);
SAVE_PACKED_ENTRY_FIELD(tid);
SAVE_PACKED_ENTRY_FIELD(cpu_cid);
SAVE_PACKED_ENTRY_FIELD(gpu_cid);
SAVE_PACKED_ENTRY_FIELD(parent_cid);
SAVE_PACKED_ENTRY_FIELD(begin_ns);
SAVE_PACKED_ENTRY_FIELD(end_ns);
SAVE_PACKED_ENTRY_FIELD(queue_id);
SAVE_PACKED_ENTRY_FIELD(hash);
#undef SAVE_PACKED_ENTRY_FIELD
std::string _name{};
auto _hash = hash;
if(_hash > 0) _name = tim::get_hash_identifier(_hash);
ar(cereal::make_nvp("name", _name),
cereal::make_nvp("demangled_name", tim::demangle(_name)));
}
template <typename Archive>
void
entry::load(Archive& ar, unsigned int)
{
namespace cereal = tim::cereal;
#define LOAD_PACKED_ENTRY_FIELD(VAR) \
{ \
auto _val = VAR; \
ar(cereal::make_nvp(#VAR, _val)); \
VAR = _val; \
}
LOAD_PACKED_ENTRY_FIELD(priority);
LOAD_PACKED_ENTRY_FIELD(device);
LOAD_PACKED_ENTRY_FIELD(phase);
LOAD_PACKED_ENTRY_FIELD(depth);
LOAD_PACKED_ENTRY_FIELD(devid);
LOAD_PACKED_ENTRY_FIELD(pid);
LOAD_PACKED_ENTRY_FIELD(tid);
LOAD_PACKED_ENTRY_FIELD(cpu_cid);
LOAD_PACKED_ENTRY_FIELD(gpu_cid);
LOAD_PACKED_ENTRY_FIELD(parent_cid);
LOAD_PACKED_ENTRY_FIELD(begin_ns);
LOAD_PACKED_ENTRY_FIELD(end_ns);
LOAD_PACKED_ENTRY_FIELD(queue_id);
LOAD_PACKED_ENTRY_FIELD(hash);
#undef LOAD_PACKED_ENTRY_FIELD
std::string _name{};
std::string _demangled_name{};
ar(cereal::make_nvp("name", _name),
cereal::make_nvp("demangled_name", _demangled_name));
auto _hash = hash;
tim::get_hash_ids()->emplace(_hash, _name);
}
struct call_chain : private std::vector<entry>
{
using base_type = std::vector<entry>;
using base_type::at;
using base_type::back;
using base_type::begin;
using base_type::cbegin;
using base_type::cend;
using base_type::clear;
using base_type::emplace_back;
using base_type::empty;
using base_type::end;
using base_type::erase;
using base_type::front;
using base_type::pop_back;
using base_type::push_back;
using base_type::rbegin;
using base_type::rend;
using base_type::reserve;
using base_type::size;
size_t get_hash() const;
int64_t get_cost(int64_t _tid = -1) const;
int64_t get_overlap(int32_t _devid, int32_t _pid, int64_t _tid = -1) const;
int64_t get_independent(int32_t _devid, int32_t _pid, int64_t _tid = -1) const;
static std::vector<call_chain>& get_top_chains();
bool operator==(const call_chain& rhs) const;
bool operator!=(const call_chain& rhs) const { return !(*this == rhs); }
friend std::ostream& operator<<(std::ostream& _os, const call_chain& _v)
{
size_t _n = 0;
for(const auto& itr : _v)
_os << " [" << _n++ << "] " << itr << "\n";
return _os;
}
template <typename Archive>
void serialize(Archive& ar, unsigned int)
{
namespace cereal = tim::cereal;
ar(cereal::make_nvp("call_chain", static_cast<base_type&>(*this)));
}
template <Device DevT>
void generate_perfetto(std::set<entry>& _used) const;
template <bool BoolV = true, typename FuncT>
bool query(FuncT&&) const;
};
template <bool BoolV, typename FuncT>
bool
call_chain::query(FuncT&& _func) const
{
for(const auto& itr : *this)
{
if(std::forward<FuncT>(_func)(itr)) return BoolV;
}
return !BoolV;
}
using hash_ids = std::unordered_set<std::string>;
uint64_t
get_update_frequency();
unique_ptr_t<call_chain>&
get(int64_t _tid = threading::get_id());
size_t
add_hash_id(const std::string& _label);
void
add_hash_id(const hash_ids&);
void
update(int64_t _tid = threading::get_id());
void
compute(int64_t _tid = threading::get_id());
std::vector<std::pair<std::string, entry>>
get_entries(const std::function<bool(const entry&)>& _eval = [](const entry&) {
return true;
});
struct id
{};
} // namespace critical_trace
template <critical_trace::Device DevID, critical_trace::Phase PhaseID,
bool UpdateStack = true>
inline void
add_critical_trace(int32_t _targ_tid, size_t _cpu_cid, size_t _gpu_cid,
size_t _parent_cid, int64_t _ts_beg, int64_t _ts_val, int32_t _devid,
uintptr_t _queue, size_t _hash, uint32_t _depth, uint16_t _prio = 0)
{
// clang-format off
// these are used to create unique type mutexes
struct critical_insert {};
struct cpu_cid_stack {};
// clang-format on
OMNITRACE_SCOPED_THREAD_STATE(ThreadState::Internal);
static constexpr auto num_mutexes = max_supported_threads;
static auto _update_freq = critical_trace::get_update_frequency();
static auto _pid = process::get_id();
auto _self_tid = threading::get_id();
if constexpr(PhaseID != critical_trace::Phase::NONE)
{
auto& _self_mtx =
type_mutex<critical_insert, project::omnitrace, num_mutexes>(_self_tid);
auto_lock_t _self_lk{ _self_mtx, std::defer_lock };
// unique lock per thread
if(!_self_lk.owns_lock()) _self_lk.lock();
auto& _critical_trace = critical_trace::get(_self_tid);
_critical_trace->emplace_back(critical_trace::entry{
DevID, PhaseID, _prio, _depth, _devid, _pid, _targ_tid, _cpu_cid, _gpu_cid,
_parent_cid, _ts_beg, _ts_val, _queue, _hash });
}
if constexpr(UpdateStack)
{
auto& _self_mtx = get_cpu_cid_stack_lock(_self_tid);
auto& _targ_mtx = get_cpu_cid_stack_lock(_targ_tid);
auto_lock_t _self_lk{ _self_mtx, std::defer_lock };
auto_lock_t _targ_lk{ _targ_mtx, std::defer_lock };
// unique lock per thread
auto _lock = [&_self_lk, &_targ_lk, _self_tid, _targ_tid]() {
if(!_self_lk.owns_lock() && _self_tid != _targ_tid) _self_lk.lock();
if(!_targ_lk.owns_lock()) _targ_lk.lock();
};
if constexpr(PhaseID == critical_trace::Phase::NONE)
{
_lock();
get_cpu_cid_stack(_targ_tid)->emplace_back(_cpu_cid);
}
else if constexpr(PhaseID == critical_trace::Phase::BEGIN)
{
_lock();
get_cpu_cid_stack(_targ_tid)->emplace_back(_cpu_cid);
}
else if constexpr(PhaseID == critical_trace::Phase::END)
{
_lock();
get_cpu_cid_stack(_targ_tid)->pop_back();
if(_gpu_cid == 0 && _cpu_cid % _update_freq == (_update_freq - 1))
critical_trace::update(_targ_tid);
}
tim::consume_parameters(_lock);
}
tim::consume_parameters(_pid, _targ_tid, _cpu_cid, _gpu_cid, _parent_cid, _ts_beg,
_ts_val, _devid, _queue, _hash, _depth, _prio, num_mutexes);
}
} // namespace omnitrace