mirror of
https://github.com/wolfpld/tracy.git
synced 2024-11-25 23:44:35 +00:00
Save 2 bytes per PlotItem.
Memory savings: android 2614 MB -> 2487 MB (95%) chicken 1932 MB -> 1852 MB (95%) mem 6067 MB -> 5747 MB (94%) q3bsp-mt 5059 MB -> 5017 MB (99%) q3bsp-st 1211 MB -> 1171 MB (96%)
This commit is contained in:
Bartosz Taudul
parent
29dcc5c8bc
commit
d9c3238462
@ -386,6 +386,15 @@ struct MessageData
|
||||
|
||||
enum { MessageDataSize = sizeof( MessageData ) };
|
||||
|
||||
|
||||
struct PlotItem
|
||||
{
|
||||
Int48 time;
|
||||
double val;
|
||||
};
|
||||
|
||||
enum { PlotItemSize = sizeof( PlotItem ) };
|
||||
|
||||
#pragma pack()
|
||||
|
||||
|
||||
@ -452,12 +461,6 @@ struct LockHighlight
|
||||
bool blocked;
|
||||
};
|
||||
|
||||
struct PlotItem
|
||||
{
|
||||
int64_t time;
|
||||
double val;
|
||||
};
|
||||
|
||||
enum class PlotType : uint8_t
|
||||
{
|
||||
User,
|
||||
|
||||
@ -4812,8 +4812,8 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
ImGui::Text( "Plot \"%s\"", txt );
|
||||
ImGui::Separator();
|
||||
|
||||
const auto first = v->data.front().time;
|
||||
const auto last = v->data.back().time;
|
||||
const auto first = v->data.front().time.Val();
|
||||
const auto last = v->data.back().time.Val();
|
||||
const auto activity = last - first;
|
||||
const auto traceLen = m_worker.GetLastTime();
|
||||
|
||||
@ -4829,8 +4829,8 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
TextFocused( "Max value:", FormatPlotValue( v->max, v->type ) );
|
||||
TextFocused( "Data/second:", RealToString( double( v->data.size() ) / activity * 1000000000ll, true ) );
|
||||
|
||||
const auto it = std::lower_bound( v->data.begin(), v->data.end(), last - 1000000000ll * 10, [] ( const auto& l, const auto& r ) { return l.time < r; } );
|
||||
const auto tr10 = last - it->time;
|
||||
const auto it = std::lower_bound( v->data.begin(), v->data.end(), last - 1000000000ll * 10, [] ( const auto& l, const auto& r ) { return l.time.Val() < r; } );
|
||||
const auto tr10 = last - it->time.Val();
|
||||
if( tr10 != 0 )
|
||||
{
|
||||
TextFocused( "D/s (10s):", RealToString( double( std::distance( it, v->data.end() ) ) / tr10 * 1000000000ll, true ) );
|
||||
@ -4896,8 +4896,8 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
}
|
||||
}
|
||||
|
||||
auto it = std::lower_bound( vec.begin(), vec.end(), m_vd.zvStart - m_worker.GetDelay(), [] ( const auto& l, const auto& r ) { return l.time < r; } );
|
||||
auto end = std::lower_bound( it, vec.end(), m_vd.zvEnd + m_worker.GetResolution(), [] ( const auto& l, const auto& r ) { return l.time < r; } );
|
||||
auto it = std::lower_bound( vec.begin(), vec.end(), m_vd.zvStart - m_worker.GetDelay(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r; } );
|
||||
auto end = std::lower_bound( it, vec.end(), m_vd.zvEnd + m_worker.GetResolution(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r; } );
|
||||
|
||||
if( end != vec.end() ) end++;
|
||||
if( it != vec.begin() ) it--;
|
||||
@ -4915,42 +4915,11 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
auto tmp = it;
|
||||
++tmp;
|
||||
const auto sz = end - tmp;
|
||||
#ifdef __AVX2__
|
||||
__m256d vmin = _mm256_set1_pd( min );
|
||||
__m256d vmax = vmin;
|
||||
const auto ssz = sz / 4;
|
||||
for( ptrdiff_t i=0; i<ssz; i++ )
|
||||
{
|
||||
__m256d v0 = _mm256_loadu_pd( (const double*)(tmp+0) );
|
||||
__m256d v1 = _mm256_loadu_pd( (const double*)(tmp+2) );
|
||||
__m256d v = _mm256_unpackhi_pd( v0, v1 );
|
||||
vmin = _mm256_min_pd( vmin, v );
|
||||
vmax = _mm256_max_pd( vmax, v );
|
||||
tmp += 4;
|
||||
}
|
||||
__m256d min0 = _mm256_shuffle_pd( vmin, vmin, 5 );
|
||||
__m256d max0 = _mm256_shuffle_pd( vmax, vmax, 5 );
|
||||
__m256d min1 = _mm256_min_pd( vmin, min0 );
|
||||
__m256d max1 = _mm256_max_pd( vmax, max0 );
|
||||
__m256d min2 = _mm256_permute4x64_pd( min1, _MM_SHUFFLE( 0, 0, 2, 2 ) );
|
||||
__m256d max2 = _mm256_permute4x64_pd( max1, _MM_SHUFFLE( 0, 0, 2, 2 ) );
|
||||
__m256d min3 = _mm256_min_pd( min1, min2 );
|
||||
__m256d max3 = _mm256_max_pd( max1, max2 );
|
||||
min = _mm256_cvtsd_f64( min3 );
|
||||
max = _mm256_cvtsd_f64( max3 );
|
||||
const auto lsz = sz % 4;
|
||||
for( ptrdiff_t i=0; i<lsz; i++ )
|
||||
{
|
||||
min = tmp[i].val < min ? tmp[i].val : min;
|
||||
max = tmp[i].val > max ? tmp[i].val : max;
|
||||
}
|
||||
#else
|
||||
for( ptrdiff_t i=0; i<sz; i++ )
|
||||
{
|
||||
min = tmp[i].val < min ? tmp[i].val : min;
|
||||
max = tmp[i].val > max ? tmp[i].val : max;
|
||||
}
|
||||
#endif
|
||||
}
|
||||
if( min == max )
|
||||
{
|
||||
@ -4987,7 +4956,7 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
|
||||
if( it == vec.begin() )
|
||||
{
|
||||
const auto x = ( it->time - m_vd.zvStart ) * pxns;
|
||||
const auto x = ( it->time.Val() - m_vd.zvStart ) * pxns;
|
||||
const auto y = PlotHeight - ( it->val - min ) * revrange * PlotHeight;
|
||||
DrawPlotPoint( wpos, x, y, offset, 0xFF44DDDD, hover, false, it, 0, false, v->type, PlotHeight );
|
||||
}
|
||||
@ -4998,8 +4967,8 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
ptrdiff_t skip = 0;
|
||||
while( it < end )
|
||||
{
|
||||
const auto x0 = ( prevx->time - m_vd.zvStart ) * pxns;
|
||||
const auto x1 = ( it->time - m_vd.zvStart ) * pxns;
|
||||
const auto x0 = ( prevx->time.Val() - m_vd.zvStart ) * pxns;
|
||||
const auto x1 = ( it->time.Val() - m_vd.zvStart ) * pxns;
|
||||
const auto y0 = PlotHeight - ( prevy->val - min ) * revrange * PlotHeight;
|
||||
const auto y1 = PlotHeight - ( it->val - min ) * revrange * PlotHeight;
|
||||
|
||||
@ -5007,7 +4976,7 @@ int View::DrawPlots( int offset, double pxns, const ImVec2& wpos, bool hover, fl
|
||||
|
||||
const auto rx = skip == 0 ? 2.0 : ( skip == 1 ? 2.5 : 4.0 );
|
||||
|
||||
auto range = std::upper_bound( it, end, int64_t( it->time + nspx * rx ), [] ( const auto& l, const auto& r ) { return l < r.time; } );
|
||||
auto range = std::upper_bound( it, end, int64_t( it->time.Val() + nspx * rx ), [] ( const auto& l, const auto& r ) { return l < r.time.Val(); } );
|
||||
assert( range > it );
|
||||
const auto rsz = std::distance( it, range );
|
||||
if( rsz == 1 )
|
||||
@ -5174,8 +5143,8 @@ void View::DrawPlotPoint( const ImVec2& wpos, float x, float y, int offset, uint
|
||||
const MemEvent* ev = nullptr;
|
||||
if( change > 0 )
|
||||
{
|
||||
auto it = std::lower_bound( mem.data.begin(), mem.data.end(), item->time, [] ( const auto& lhs, const auto& rhs ) { return lhs.TimeAlloc() < rhs; } );
|
||||
if( it != mem.data.end() && it->TimeAlloc() == item->time )
|
||||
auto it = std::lower_bound( mem.data.begin(), mem.data.end(), item->time.Val(), [] ( const auto& lhs, const auto& rhs ) { return lhs.TimeAlloc() < rhs; } );
|
||||
if( it != mem.data.end() && it->TimeAlloc() == item->time.Val() )
|
||||
{
|
||||
ev = it;
|
||||
}
|
||||
@ -5183,8 +5152,8 @@ void View::DrawPlotPoint( const ImVec2& wpos, float x, float y, int offset, uint
|
||||
else
|
||||
{
|
||||
const auto& data = mem.data;
|
||||
auto it = std::lower_bound( mem.frees.begin(), mem.frees.end(), item->time, [&data] ( const auto& lhs, const auto& rhs ) { return data[lhs].TimeFree() < rhs; } );
|
||||
if( it != mem.frees.end() && data[*it].TimeFree() == item->time )
|
||||
auto it = std::lower_bound( mem.frees.begin(), mem.frees.end(), item->time.Val(), [&data] ( const auto& lhs, const auto& rhs ) { return data[lhs].TimeFree() < rhs; } );
|
||||
if( it != mem.frees.end() && data[*it].TimeFree() == item->time.Val() )
|
||||
{
|
||||
ev = &data[*it];
|
||||
}
|
||||
|
||||
@ -1062,7 +1062,7 @@ Worker::Worker( FileRead& f, EventType::Type eventMask, bool bgTasks )
|
||||
int64_t refTime = 0;
|
||||
for( uint64_t j=0; j<psz; j++ )
|
||||
{
|
||||
pd->data[j].time = ReadTimeOffset( f, refTime );
|
||||
pd->data[j].time.SetVal( ReadTimeOffset( f, refTime ) );
|
||||
f.Read( pd->data[j].val );
|
||||
}
|
||||
}
|
||||
@ -1071,16 +1071,19 @@ Worker::Worker( FileRead& f, EventType::Type eventMask, bool bgTasks )
|
||||
int64_t refTime = -m_data.baseTime;
|
||||
for( uint64_t j=0; j<psz; j++ )
|
||||
{
|
||||
pd->data[j].time = ReadTimeOffset( f, refTime );
|
||||
pd->data[j].time.SetVal( ReadTimeOffset( f, refTime ) );
|
||||
f.Read( pd->data[j].val );
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
f.Read( pd->data.data(), psz * sizeof( PlotItem ) );
|
||||
for( uint64_t j=0; j<psz; j++ )
|
||||
{
|
||||
pd->data[j].time -= m_data.baseTime;
|
||||
uint64_t t;
|
||||
f.Read( t );
|
||||
t -= m_data.baseTime;
|
||||
pd->data[j].time.SetVal( t );
|
||||
f.Read( pd->data[j].val );
|
||||
}
|
||||
}
|
||||
m_data.plots.Data().push_back_no_space_check( pd );
|
||||
@ -1100,7 +1103,7 @@ Worker::Worker( FileRead& f, EventType::Type eventMask, bool bgTasks )
|
||||
}
|
||||
uint64_t psz;
|
||||
f.Read( psz );
|
||||
f.Skip( psz * sizeof( PlotItem ) );
|
||||
f.Skip( psz * ( sizeof( uint64_t ) + sizeof( double ) ) );
|
||||
}
|
||||
}
|
||||
|
||||
@ -3223,13 +3226,13 @@ void Worker::InsertPlot( PlotData* plot, int64_t time, double val )
|
||||
{
|
||||
plot->min = val;
|
||||
plot->max = val;
|
||||
plot->data.push_back( { time, val } );
|
||||
plot->data.push_back( { Int48( time ), val } );
|
||||
}
|
||||
else if( plot->data.back().time < time )
|
||||
else if( plot->data.back().time.Val() < time )
|
||||
{
|
||||
if( plot->min > val ) plot->min = val;
|
||||
else if( plot->max < val ) plot->max = val;
|
||||
plot->data.push_back_non_empty( { time, val } );
|
||||
plot->data.push_back_non_empty( { Int48( time ), val } );
|
||||
}
|
||||
else
|
||||
{
|
||||
@ -3238,11 +3241,11 @@ void Worker::InsertPlot( PlotData* plot, int64_t time, double val )
|
||||
if( plot->postpone.empty() )
|
||||
{
|
||||
plot->postponeTime = std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now().time_since_epoch() ).count();
|
||||
plot->postpone.push_back( { time, val } );
|
||||
plot->postpone.push_back( { Int48( time ), val } );
|
||||
}
|
||||
else
|
||||
{
|
||||
plot->postpone.push_back_non_empty( { time, val } );
|
||||
plot->postpone.push_back_non_empty( { Int48( time ), val } );
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -3253,7 +3256,7 @@ void Worker::HandlePlotName( uint64_t name, char* str, size_t sz )
|
||||
m_data.plots.StringDiscovered( name, sl, m_data.strings, [this] ( PlotData* dst, PlotData* src ) {
|
||||
for( auto& v : src->data )
|
||||
{
|
||||
InsertPlot( dst, v.time, v.val );
|
||||
InsertPlot( dst, v.time.Val(), v.val );
|
||||
}
|
||||
} );
|
||||
}
|
||||
@ -3277,16 +3280,16 @@ void Worker::HandlePostponedPlots()
|
||||
if( std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now().time_since_epoch() ).count() - plot->postponeTime < 100 ) continue;
|
||||
auto& dst = plot->data;
|
||||
#ifdef MY_LIBCPP_SUCKS
|
||||
pdqsort_branchless( src.begin(), src.end(), [] ( const auto& l, const auto& r ) { return l.time < r.time; } );
|
||||
pdqsort_branchless( src.begin(), src.end(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r.time.Val(); } );
|
||||
#else
|
||||
std::sort( std::execution::par_unseq, src.begin(), src.end(), [] ( const auto& l, const auto& r ) { return l.time < r.time; } );
|
||||
std::sort( std::execution::par_unseq, src.begin(), src.end(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r.time.Val(); } );
|
||||
#endif
|
||||
const auto ds = std::lower_bound( dst.begin(), dst.end(), src.front().time, [] ( const auto& l, const auto& r ) { return l.time < r; } );
|
||||
const auto ds = std::lower_bound( dst.begin(), dst.end(), src.front().time.Val(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r; } );
|
||||
const auto dsd = std::distance( dst.begin(), ds ) ;
|
||||
const auto de = std::lower_bound( ds, dst.end(), src.back().time, [] ( const auto& l, const auto& r ) { return l.time < r; } );
|
||||
const auto de = std::lower_bound( ds, dst.end(), src.back().time.Val(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r; } );
|
||||
const auto ded = std::distance( dst.begin(), de );
|
||||
dst.insert( de, src.begin(), src.end() );
|
||||
std::inplace_merge( dst.begin() + dsd, dst.begin() + ded, dst.begin() + ded + src.size(), [] ( const auto& l, const auto& r ) { return l.time < r.time; } );
|
||||
std::inplace_merge( dst.begin() + dsd, dst.begin() + ded, dst.begin() + ded + src.size(), [] ( const auto& l, const auto& r ) { return l.time.Val() < r.time.Val(); } );
|
||||
src.clear();
|
||||
}
|
||||
}
|
||||
@ -4596,7 +4599,7 @@ void Worker::ProcessSysTime( const QueueSysTime& ev )
|
||||
else
|
||||
{
|
||||
assert( !m_sysTimePlot->data.empty() );
|
||||
assert( m_sysTimePlot->data.back().time <= time );
|
||||
assert( m_sysTimePlot->data.back().time.Val() <= time );
|
||||
if( m_sysTimePlot->min > val ) m_sysTimePlot->min = val;
|
||||
else if( m_sysTimePlot->max < val ) m_sysTimePlot->max = val;
|
||||
m_sysTimePlot->data.push_back_non_empty( { time, val } );
|
||||
@ -4741,7 +4744,7 @@ void Worker::MemAllocChanged( int64_t time )
|
||||
else
|
||||
{
|
||||
assert( !m_data.memory.plot->data.empty() );
|
||||
assert( m_data.memory.plot->data.back().time <= time );
|
||||
assert( m_data.memory.plot->data.back().time.Val() <= time );
|
||||
if( m_data.memory.plot->min > val ) m_data.memory.plot->min = val;
|
||||
else if( m_data.memory.plot->max < val ) m_data.memory.plot->max = val;
|
||||
m_data.memory.plot->data.push_back_non_empty( { time, val } );
|
||||
@ -5576,7 +5579,7 @@ void Worker::Write( FileWrite& f )
|
||||
f.Write( &sz, sizeof( sz ) );
|
||||
for( auto& v : plot->data )
|
||||
{
|
||||
WriteTimeOffset( f, refTime, v.time );
|
||||
WriteTimeOffset( f, refTime, v.time.Val() );
|
||||
f.Write( &v.val, sizeof( v.val ) );
|
||||
}
|
||||
}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user