tracy/client/TracyProfiler.cpp

343 lines
8.9 KiB
C++
Raw Normal View History

2017-09-14 17:25:16 +00:00
#ifdef _MSC_VER
# include <winsock2.h>
#else
# include <sys/time.h>
#endif
2017-09-23 19:33:05 +00:00
#include <atomic>
2017-09-10 15:43:56 +00:00
#include <assert.h>
2017-09-23 19:09:46 +00:00
#include <chrono>
2017-09-14 17:25:16 +00:00
#include <limits>
2017-09-21 22:36:36 +00:00
#include <memory>
2017-09-19 00:19:20 +00:00
#include <string.h>
2017-09-10 15:43:56 +00:00
#include "../common/TracyProtocol.hpp"
2017-09-11 20:51:47 +00:00
#include "../common/TracySocket.hpp"
#include "../common/TracySystem.hpp"
#include "concurrentqueue.h"
2017-09-24 14:02:09 +00:00
#include "TracyScoped.hpp"
2017-09-10 15:43:56 +00:00
#include "TracyProfiler.hpp"
#ifdef _DEBUG
# define DISABLE_LZ4
#endif
2017-09-10 15:43:56 +00:00
namespace tracy
{
2017-09-24 13:59:53 +00:00
enum { QueuePrealloc = 256 * 1024 };
static moodycamel::ConcurrentQueue<QueueItem> s_queue( QueueItemSize * QueuePrealloc );
static moodycamel::ProducerToken& GetToken()
{
static thread_local moodycamel::ProducerToken token( s_queue );
return token;
}
#ifndef TRACY_DISABLE
Profiler s_profiler;
#endif
2017-09-10 15:43:56 +00:00
static Profiler* s_instance = nullptr;
Profiler::Profiler()
2017-09-23 19:33:05 +00:00
: m_mainThread( GetThreadHandle() )
2017-09-10 18:14:16 +00:00
, m_shutdown( false )
2017-09-10 18:08:42 +00:00
, m_id( 0 )
, m_stream( LZ4_createStream() )
, m_buffer( new char[TargetFrameSize*3] )
, m_bufferOffset( 0 )
2017-09-10 15:43:56 +00:00
{
assert( !s_instance );
s_instance = this;
2017-09-23 19:33:05 +00:00
CalibrateTimer();
2017-09-24 14:02:09 +00:00
CalibrateDelay();
2017-09-23 19:33:05 +00:00
m_timeBegin = GetTime();
2017-09-10 15:43:56 +00:00
m_thread = std::thread( [this] { Worker(); } );
2017-09-10 15:46:20 +00:00
SetThreadName( m_thread, "Tracy Profiler" );
2017-09-10 15:43:56 +00:00
}
Profiler::~Profiler()
{
m_shutdown.store( true, std::memory_order_relaxed );
m_thread.join();
delete[] m_buffer;
LZ4_freeStream( m_stream );
2017-09-10 15:43:56 +00:00
assert( s_instance );
s_instance = nullptr;
}
2017-09-10 18:08:42 +00:00
uint64_t Profiler::GetNewId()
{
return s_instance->m_id.fetch_add( 1, std::memory_order_relaxed );
}
uint64_t Profiler::ZoneBegin( QueueZoneBegin&& data )
2017-09-10 18:09:14 +00:00
{
auto id = GetNewId();
QueueItem item;
item.hdr.type = QueueType::ZoneBegin;
item.hdr.id = id;
2017-09-10 18:09:14 +00:00
item.zoneBegin = std::move( data );
s_queue.enqueue( GetToken(), std::move( item ) );
return id;
2017-09-10 18:09:14 +00:00
}
void Profiler::ZoneEnd( uint64_t id, QueueZoneEnd&& data )
2017-09-10 18:09:14 +00:00
{
QueueItem item;
item.hdr.type = QueueType::ZoneEnd;
item.hdr.id = id;
2017-09-10 18:09:14 +00:00
item.zoneEnd = std::move( data );
s_queue.enqueue( GetToken(), std::move( item ) );
}
2017-09-15 22:30:27 +00:00
void Profiler::FrameMark()
{
QueueItem item;
item.hdr.type = QueueType::FrameMark;
item.hdr.id = (uint64_t)GetTime();
s_queue.enqueue( GetToken(), std::move( item ) );
2017-09-10 18:09:14 +00:00
}
bool Profiler::ShouldExit()
{
return s_instance->m_shutdown.load( std::memory_order_relaxed );
}
2017-09-10 15:43:56 +00:00
void Profiler::Worker()
{
2017-09-11 22:49:38 +00:00
enum { BulkSize = TargetFrameSize / QueueItemSize };
moodycamel::ConsumerToken token( s_queue );
2017-09-11 20:51:47 +00:00
ListenSocket listen;
listen.Listen( "8086", 8 );
2017-09-10 15:43:56 +00:00
for(;;)
{
2017-09-11 20:51:47 +00:00
for(;;)
{
2017-09-11 20:51:47 +00:00
if( m_shutdown.load( std::memory_order_relaxed ) ) return;
m_sock = listen.Accept();
if( m_sock ) break;
}
2017-09-11 20:51:47 +00:00
{
WelcomeMessage welcome;
#ifdef DISABLE_LZ4
// notify client that lz4 compression is disabled (too slow in debug builds)
welcome.lz4 = 0;
#else
welcome.lz4 = 1;
#endif
welcome.timerMul = m_timerMul;
welcome.timeBegin = m_timeBegin;
welcome.delay = m_delay;
m_sock->Send( &welcome, sizeof( welcome ) );
}
2017-09-11 20:51:47 +00:00
LZ4_resetStream( m_stream );
2017-09-11 20:51:47 +00:00
for(;;)
{
2017-09-11 20:51:47 +00:00
if( m_shutdown.load( std::memory_order_relaxed ) ) return;
QueueItem item[BulkSize];
const auto sz = s_queue.try_dequeue_bulk( token, item, BulkSize );
2017-09-11 20:51:47 +00:00
if( sz > 0 )
{
auto buf = m_buffer + m_bufferOffset;
auto ptr = buf;
2017-09-22 20:05:51 +00:00
for( size_t i=0; i<sz; i++ )
2017-09-11 23:14:04 +00:00
{
const auto dsz = QueueDataSize[item[i].hdr.idx];
2017-09-11 23:14:04 +00:00
memcpy( ptr, item+i, dsz );
ptr += dsz;
}
if( !SendData( buf, ptr - buf ) ) break;
m_bufferOffset += ptr - buf;
if( m_bufferOffset > TargetFrameSize * 2 ) m_bufferOffset = 0;
2017-09-11 20:51:47 +00:00
}
else
{
std::this_thread::sleep_for( std::chrono::milliseconds( 10 ) );
}
2017-09-14 17:25:16 +00:00
while( m_sock->HasData() )
{
if( !HandleServerQuery() ) break;
2017-09-14 17:25:16 +00:00
}
}
2017-09-10 15:43:56 +00:00
}
}
bool Profiler::SendData( const char* data, size_t len )
{
#ifdef DISABLE_LZ4
if( m_sock->Send( data, len ) == -1 ) return false;
#else
char lz4[LZ4Size + sizeof( lz4sz_t )];
const lz4sz_t lz4sz = LZ4_compress_fast_continue( m_stream, data, lz4 + sizeof( lz4sz_t ), len, LZ4Size, 1 );
memcpy( lz4, &lz4sz, sizeof( lz4sz ) );
if( m_sock->Send( lz4, lz4sz + sizeof( lz4sz_t ) ) == -1 ) return false;
#endif
return true;
}
bool Profiler::SendString( uint64_t str, const char* ptr, QueueType type )
{
2017-09-21 23:55:02 +00:00
assert( type == QueueType::StringData || type == QueueType::ThreadName );
QueueHeader hdr;
hdr.type = type;
hdr.id = str;
auto buf = m_buffer + m_bufferOffset;
memcpy( buf, &hdr, sizeof( hdr ) );
auto len = strlen( ptr );
assert( len < TargetFrameSize - sizeof( hdr ) - sizeof( uint16_t ) );
assert( len <= std::numeric_limits<uint16_t>::max() );
uint16_t l16 = len;
memcpy( buf + sizeof( hdr ), &l16, sizeof( l16 ) );
memcpy( buf + sizeof( hdr ) + sizeof( l16 ), ptr, l16 );
m_bufferOffset += sizeof( hdr ) + sizeof( l16 ) + l16;
if( m_bufferOffset > TargetFrameSize * 2 ) m_bufferOffset = 0;
return SendData( buf, sizeof( hdr ) + sizeof( l16 ) + l16 );
}
void Profiler::SendSourceLocation( uint64_t ptr )
{
auto srcloc = (const SourceLocation*)ptr;
QueueItem item;
item.hdr.type = QueueType::SourceLocation;
item.hdr.id = ptr;
item.srcloc.file = (uint64_t)srcloc->file;
item.srcloc.function = (uint64_t)srcloc->function;
item.srcloc.line = srcloc->line;
item.srcloc.color = srcloc->color;
s_queue.enqueue( GetToken(), std::move( item ) );
}
bool Profiler::HandleServerQuery()
{
timeval tv;
tv.tv_sec = 0;
tv.tv_usec = 10000;
uint8_t type;
if( !m_sock->Read( &type, sizeof( type ), &tv, ShouldExit ) ) return false;
uint64_t ptr;
if( !m_sock->Read( &ptr, sizeof( ptr ), &tv, ShouldExit ) ) return false;
switch( type )
{
case ServerQueryString:
SendString( ptr, (const char*)ptr, QueueType::StringData );
break;
2017-09-21 23:55:02 +00:00
case ServerQueryThreadString:
2017-09-22 23:38:26 +00:00
if( ptr == m_mainThread )
{
SendString( ptr, "Main thread", QueueType::ThreadName );
}
else
{
SendString( ptr, GetThreadName( ptr ), QueueType::ThreadName );
}
2017-09-21 23:55:02 +00:00
break;
case ServerQuerySourceLocation:
SendSourceLocation( ptr );
break;
default:
assert( false );
break;
}
return true;
}
2017-09-23 19:33:05 +00:00
void Profiler::CalibrateTimer()
{
#if defined _MSC_VER || defined __CYGWIN__
unsigned int ui;
std::atomic_signal_fence( std::memory_order_acq_rel );
const auto t0 = std::chrono::high_resolution_clock::now();
const auto r0 = __rdtscp( &ui );
std::atomic_signal_fence( std::memory_order_acq_rel );
std::this_thread::sleep_for( std::chrono::milliseconds( 100 ) );
std::atomic_signal_fence( std::memory_order_acq_rel );
const auto t1 = std::chrono::high_resolution_clock::now();
const auto r1 = __rdtscp( &ui );
std::atomic_signal_fence( std::memory_order_acq_rel );
const auto dt = std::chrono::duration_cast<std::chrono::nanoseconds>( t1 - t0 ).count();
const auto dr = r1 - r0;
m_timerMul = double( dt ) / double( dr );
#else
m_timerMul = 1.;
2017-09-23 19:33:05 +00:00
#endif
}
2017-09-24 14:02:09 +00:00
class FakeZone
{
public:
FakeZone( const SourceLocation* srcloc ) {}
2017-09-24 14:02:09 +00:00
~FakeZone() {}
private:
uint64_t m_id;
};
void Profiler::CalibrateDelay()
{
enum { Iterations = 50000 };
enum { Events = Iterations * 2 }; // start + end
static_assert( Events * 2 < QueuePrealloc, "Delay calibration loop will allocate memory in queue" );
for( int i=0; i<Iterations; i++ )
{
static const tracy::SourceLocation __tracy_source_location { __FUNCTION__, __FILE__, __LINE__, 0 };
ScopedZone ___tracy_scoped_zone( &__tracy_source_location );
2017-09-24 14:02:09 +00:00
}
const auto f0 = GetTime();
for( int i=0; i<Iterations; i++ )
{
static const tracy::SourceLocation __tracy_source_location { __FUNCTION__, __FILE__, __LINE__, 0 };
FakeZone ___tracy_scoped_zone( &__tracy_source_location );
2017-09-24 14:02:09 +00:00
}
const auto t0 = GetTime();
for( int i=0; i<Iterations; i++ )
{
static const tracy::SourceLocation __tracy_source_location { __FUNCTION__, __FILE__, __LINE__, 0 };
ScopedZone ___tracy_scoped_zone( &__tracy_source_location );
2017-09-24 14:02:09 +00:00
}
const auto t1 = GetTime();
const auto dt = t1 - t0;
const auto df = t0 - f0;
m_delay = ( dt - df ) / Events;
enum { Bulk = 1000 };
moodycamel::ConsumerToken token( s_queue );
int left = Events * 2;
QueueItem item[Bulk];
while( left != 0 )
{
const auto sz = s_queue.try_dequeue_bulk( token, item, std::min( left, (int)Bulk ) );
assert( sz > 0 );
left -= sz;
}
}
2017-09-10 15:43:56 +00:00
}