ilikecats/stk-code_catmod
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'fix-profiler' into improve-xbox-gamepad-support

Browse Source
This commit is contained in:
hiker 2017-08-09 17:56:40 +10:00
commit 8be5443f8c
4 changed files with 170 additions and 75 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
src/audio/sfx_manager.cpp
View File

@ -24,6 +24,7 @@
#include "io/file_manager.hpp"
#include "modes/world.hpp"
#include "race/race_manager.hpp"
#include "utils/profiler.hpp"
#include "utils/vs.hpp"
#include <pthread.h>
@ -46,6 +47,17 @@
# endif
#endif
// Define this if the profiler should also collect data of the sfx manager
#undef ENABLE_PROFILING_FOR_SFX_MANAGER
#ifndef ENABLE_PROFILING_FOR_SFX_MANAGER
// Otherwise ignore the profiler push/pop events
// Use undef to remove preprocessor warning
# undef PROFILER_PUSH_CPU_MARKER
# undef PROFILER_POP_CPU_MARKER
# define PROFILER_PUSH_CPU_MARKER(name, r, g, b)
# define PROFILER_POP_CPU_MARKER()
#endif
SFXManager *SFXManager::m_sfx_manager;
// ----------------------------------------------------------------------------
@ -302,6 +314,7 @@ void* SFXManager::mainLoop(void *obj)
while (me->m_sfx_commands.getData().empty() ||
me->m_sfx_commands.getData().front()->m_command!=SFX_EXIT)
{
PROFILER_PUSH_CPU_MARKER("Wait", 255, 0, 0);
bool empty = me->m_sfx_commands.getData().empty();
// Wait in cond_wait for a request to arrive. The 'while' is necessary
@ -321,6 +334,8 @@ void* SFXManager::mainLoop(void *obj)
break;
}
me->m_sfx_commands.unlock();
PROFILER_POP_CPU_MARKER();
PROFILER_PUSH_CPU_MARKER("Execute", 0, 255, 0);
switch (current->m_command)
{
case SFX_PLAY: current->m_sfx->reallyPlayNow(); break;
@ -383,6 +398,8 @@ void* SFXManager::mainLoop(void *obj)
}
delete current;
current = NULL;
PROFILER_POP_CPU_MARKER();
PROFILER_PUSH_CPU_MARKER("yield", 0, 0, 255);
// We access the size without lock, doesn't matter if we
// should get an incorrect value because of concurrent read/writes
if (me->m_sfx_commands.getData().size() == 0)
@ -395,7 +412,7 @@ void* SFXManager::mainLoop(void *obj)
me->queue(SFX_UPDATE, (SFXBase*)NULL, float(t));
}
me->m_sfx_commands.lock();
PROFILER_POP_CPU_MARKER();
} // while
// Signal that the sfx manager can now be deleted.

3
src/utils/debug.cpp
View File

@ -346,8 +346,7 @@ bool handleContextMenuAction(s32 cmd_id)
irr_driver->toggleBoundingBoxesViz();
break;
case DEBUG_PROFILER:
UserConfigParams::m_profiler_enabled =
!UserConfigParams::m_profiler_enabled;
profiler.toggleStatus();
break;
case DEBUG_PROFILER_WRITE_REPORT:
profiler.writeToFile();

201
src/utils/profiler.cpp
View File

@ -26,6 +26,7 @@
#include "graphics/irr_driver.hpp"
#include "guiengine/scalable_font.hpp"
#include "io/file_manager.hpp"
#include "utils/string_utils.hpp"
#include "utils/vs.hpp"
#include <algorithm>
@ -113,18 +114,18 @@ Profiler::Profiler()
m_time_between_sync = 0.0;
m_freeze_state = UNFROZEN;
// By limiting the number of threads that can be created, we avoid the
// problem that all access to m_all_event_data need to be locked
// (otherwise adding a thread to m_all_event_data can trigger a
// reallocate, which makes concurrent access invalid)
m_max_frames = int( UserConfigParams::m_profiler_buffer_duration
* UserConfigParams::m_max_fps );
m_current_frame = 0;
m_has_wrapped_around = false;
m_threads_used = 0;
const int MAX_THREADS = 10;
m_all_threads_data.resize(MAX_THREADS);
m_thread_mapping.getData().resize(MAX_THREADS);
m_thread_mapping.resize(MAX_THREADS);
// Add this thread to the thread mapping
m_thread_mapping[0] = pthread_self();
m_threads_used = 1;
m_gpu_times.resize(Q_LAST*m_max_frames);
} // Profile
@ -134,27 +135,28 @@ Profiler::~Profiler()
} // ~Profiler
//-----------------------------------------------------------------------------
/** Returns a unique index for a thread. If the calling thread is not yet in
* the mapping, it will assign a new unique id to this thread. This function
* is NOT thread-safe and must be called from a properly protected code
* section. */
int Profiler::getThreadID()
{
m_thread_mapping.lock();
pthread_t thread = pthread_self();
int i = 0;
while(i < m_threads_used)
{
if (memcmp( &m_thread_mapping.getData()[i],
if (memcmp( &m_thread_mapping[i],
&thread,
sizeof(thread)) ==0 )
{
m_thread_mapping.unlock();
return i;
}
i++;
} // for i <m_threads_used
assert(m_threads_used < (int)m_thread_mapping.getData().size());
m_thread_mapping.getData()[m_threads_used] = thread;
assert(m_threads_used < (int)m_thread_mapping.size());
m_thread_mapping[m_threads_used] = thread;
m_threads_used++;
m_thread_mapping.unlock();
return m_threads_used - 1;
} // getThreadID
@ -163,15 +165,19 @@ int Profiler::getThreadID()
/// Push a new marker that starts now
void Profiler::pushCPUMarker(const char* name, const video::SColor& colour)
{
// Don't do anything when frozen
if (m_freeze_state == FROZEN || m_freeze_state == WAITING_FOR_UNFREEZE)
// Don't do anything when disabled or frozen
if (!UserConfigParams::m_profiler_enabled ||
m_freeze_state == FROZEN || m_freeze_state == WAITING_FOR_UNFREEZE )
return;
double start = getTimeMilliseconds() - m_time_last_sync;
// We need to look before getting the thread id (since this might
// be a new thread which changes the structure).
m_lock.lock();
int thread_id = getThreadID();
ThreadData &td = m_all_threads_data[thread_id];
AllEventData::iterator i = td.m_all_event_data.find(name);
double start = getTimeMilliseconds() - m_time_last_sync;
if (i != td.m_all_event_data.end())
{
i->second.setStart(m_current_frame, start, td.m_event_stack.size());
@ -181,31 +187,63 @@ void Profiler::pushCPUMarker(const char* name, const video::SColor& colour)
EventData ed(colour, m_max_frames);
ed.setStart(m_current_frame, start, td.m_event_stack.size());
td.m_all_event_data[name] = ed;
// Ordered headings is used to determine the order in which the
// bar graph is drawn. Outer profiling events will be added first,
// so they will be drawn first, which gives the proper nested
// displayed of events.
td.m_ordered_headings.push_back(name);
}
td.m_event_stack.push_back(name);
m_lock.unlock();
} // pushCPUMarker
//-----------------------------------------------------------------------------
/// Stop the last pushed marker
void Profiler::popCPUMarker()
{
// Don't do anything when frozen
if(m_freeze_state == FROZEN || m_freeze_state == WAITING_FOR_UNFREEZE)
// Don't do anything when disabled or frozen
if( !UserConfigParams::m_profiler_enabled ||
m_freeze_state == FROZEN || m_freeze_state == WAITING_FOR_UNFREEZE )
return;
double now = getTimeMilliseconds();
m_lock.lock();
int thread_id = getThreadID();
ThreadData &td = m_all_threads_data[thread_id];
// When the profiler gets enabled (which happens in the middle of the
// main loop), there can be some pops without matching pushes (for one
// frame) - ignore those events.
if (td.m_event_stack.size() == 0)
{
m_lock.unlock();
return;
}
assert(td.m_event_stack.size() > 0);
const std::string &name = td.m_event_stack.back();
td.m_all_event_data[name].setEnd(m_current_frame,
getTimeMilliseconds() - m_time_last_sync);
td.m_all_event_data[name].setEnd(m_current_frame, now - m_time_last_sync);
td.m_event_stack.pop_back();
m_lock.unlock();
} // popCPUMarker
//-----------------------------------------------------------------------------
/** Switches the profiler either on or off.
*/
void Profiler::toggleStatus()
{
UserConfigParams::m_profiler_enabled = !UserConfigParams::m_profiler_enabled;
// If the profiler would immediately enabled, calls that have started but
// not finished would not be registered correctly. So set the state to
// waiting, so the unfreeze started at the next sync frame (which is
// outside of the main loop, i.e. all profiling events inside of the main
// loop will work as expected.
if (m_freeze_state == UNFROZEN)
m_freeze_state = WAITING_FOR_UNFREEZE;
} // toggleStatus
//-----------------------------------------------------------------------------
/** Saves all data for the current frame, and starts the next frame in the
* circular buffer. Any events that are currently active (e.g. in a separate
@ -215,13 +253,14 @@ void Profiler::popCPUMarker()
void Profiler::synchronizeFrame()
{
// Don't do anything when frozen
if(m_freeze_state == FROZEN)
if(!UserConfigParams::m_profiler_enabled || m_freeze_state == FROZEN)
return;
// Avoid using several times getTimeMilliseconds(),
// which would yield different results
double now = getTimeMilliseconds();
m_lock.lock();
// Set index to next frame
int next_frame = m_current_frame+1;
if (next_frame >= m_max_frames)
@ -240,23 +279,39 @@ void Profiler::synchronizeFrame()
for(unsigned int j=0; j<td.m_event_stack.size(); j++)
{
EventData &ed = td.m_all_event_data[td.m_event_stack[j]];
ed.setEnd(m_current_frame, now);
ed.setStart(next_frame, now, j);
}
ed.setEnd(m_current_frame, now-m_time_last_sync);
ed.setStart(next_frame, 0, j);
} // for j in event stack
} // for i in threads
if (m_has_wrapped_around)
{
// The new entries for the circular buffer need to be cleared
// to make sure the new values are not accumulated on top of
// the data from a previous frame.
for (int i = 0; i < m_threads_used; i++)
{
ThreadData &td = m_all_threads_data[i];
AllEventData &aed = td.m_all_event_data;
AllEventData::iterator k;
for (k = aed.begin(); k != aed.end(); ++k)
k->second.getMarker(next_frame).clear();
}
} // is has wrapped around
}
m_current_frame = next_frame;
// Remember the date of last synchronization
m_time_between_sync = now - m_time_last_sync;
m_time_last_sync = now;
m_time_last_sync = now;
// Freeze/unfreeze as needed
if(m_freeze_state == WAITING_FOR_FREEZE)
m_freeze_state = FROZEN;
else if(m_freeze_state == WAITING_FOR_UNFREEZE)
m_freeze_state = UNFROZEN;
m_lock.unlock();
} // synchronizeFrame
//-----------------------------------------------------------------------------
@ -269,8 +324,10 @@ void Profiler::draw()
// Current frame points to the frame in which currently data is
// being accumulated. Draw the previous (i.e. complete) frame.
m_lock.lock();
int indx = m_current_frame - 1;
if (indx < 0) indx = m_max_frames - 1;
m_lock.unlock();
drawBackground();
@ -289,19 +346,20 @@ void Profiler::draw()
// Compute start end end time for this frame
double start = 99999.0f;
double end = -1.0f;
for (int i = 0; i < m_threads_used; i++)
{
AllEventData &aed = m_all_threads_data[i].m_all_event_data;
AllEventData::iterator j;
for (j = aed.begin(); j != aed.end(); ++j)
{
const Marker &marker = j->second.getMarker(indx);
start = std::min(start, marker.getStart() );
end = std::max(end, marker.getEnd() );
} // for j in events
} // for i in threads
// Use this thread (thread 0) to compute start and end time. All other
// threads might have 'unfinished' events, or multiple identical events
// in this frame (i.e. start time would be incorrect(.
AllEventData &aed = m_all_threads_data[0].m_all_event_data;
AllEventData::iterator j;
for (j = aed.begin(); j != aed.end(); ++j)
{
const Marker &marker = j->second.getMarker(indx);
start = std::min(start, marker.getStart());
end = std::max(end, marker.getEnd());
} // for j in events
const double duration = end - start;
const double factor = profiler_width / duration;
@ -313,14 +371,25 @@ void Profiler::draw()
{
ThreadData &td = m_all_threads_data[i];
AllEventData &aed = td.m_all_event_data;
AllEventData::iterator j;
for (j = aed.begin(); j != aed.end(); ++j)
// Thread 1 has 'proper' start and end events (assuming that each
// event is at most called once). But all other threads might have
// multiple start and end events, so the recorder start time is only
// of the last event and so can not be used to draw the bar graph
double start_xpos = 0;
for(int k=0; k<(int)td.m_ordered_headings.size(); k++)
{
AllEventData::iterator j = aed.find(td.m_ordered_headings[k]);
const Marker &marker = j->second.getMarker(indx);
core::rect<s32> pos((s32)(x_offset + factor*marker.getStart()),
if (i == 0)
start_xpos = factor*marker.getStart();
core::rect<s32> pos((s32)(x_offset + start_xpos),
(s32)(y_offset + i*line_height),
(s32)(x_offset + factor*marker.getEnd()),
(s32)(y_offset + (i + 1)*line_height) );
(s32)(x_offset + start_xpos
+ factor*marker.getDuration()),
(s32)(y_offset + (i + 1)*line_height) );
if (i != 0)
start_xpos += factor*marker.getDuration();
// Reduce vertically the size of the markers according to their layer
pos.UpperLeftCorner.Y += 2 * marker.getLayer();
@ -336,7 +405,7 @@ void Profiler::draw()
} // for j in AllEventdata
} // for i in threads
// GPU profiler
QueryPerf hovered_gpu_marker = Q_LAST;
long hovered_gpu_marker_elapsed = 0;
@ -350,7 +419,7 @@ void Profiler::draw()
total += n;
#endif
}
static video::SColor colors[] = {
video::SColor(255, 255, 0, 0),
video::SColor(255, 0, 255, 0),
@ -391,7 +460,7 @@ void Profiler::draw()
{
s32 x_sync = (s32)(x_offset + factor*m_time_between_sync);
s32 y_up_sync = (s32)(MARGIN_Y*screen_size.Height);
s32 y_down_sync = (s32)( (MARGIN_Y + (2+m_threads_used)*LINE_HEIGHT)
s32 y_down_sync = (s32)( (MARGIN_Y + (2+m_threads_used)*LINE_HEIGHT)
* screen_size.Height );
GL32_draw2DRectangle(video::SColor(0xFF, 0x00, 0x00, 0x00),
@ -492,43 +561,33 @@ void Profiler::drawBackground()
*/
void Profiler::writeToFile()
{
std::string base_name =
m_lock.lock();
std::string base_name =
file_manager->getUserConfigFile(file_manager->getStdoutName());
// First CPU data
std::ofstream f(base_name + ".profile-cpu");
for (int thread_id = 0; thread_id < m_threads_used; thread_id++)
{
std::ofstream f(base_name + ".profile-cpu-" +
StringUtils::toString(thread_id) );
ThreadData &td = m_all_threads_data[thread_id];
AllEventData::iterator j;
std::vector<std::string> new_headings;
for (j = td.m_all_event_data.begin(); j != td.m_all_event_data.end(); j++)
{
std::vector<std::string>::iterator f =
std::find(m_all_event_names.begin(),
m_all_event_names.end(), j->first);
if(f==m_all_event_names.end())
new_headings.push_back(j->first);
}
std::sort(new_headings.begin(), new_headings.end());
f << "# \"Thread(1)\" ";
for (unsigned int i = 0; i < new_headings.size(); i++)
f << "\"" << new_headings[i] << "(" << i+2 <<")\" ";
f << "# ";
for (unsigned int i = 0; i < td.m_ordered_headings.size(); i++)
f << "\"" << td.m_ordered_headings[i] << "(" << i+1 <<")\" ";
f << std::endl;
int start = m_has_wrapped_around ? m_current_frame + 1 : 0;
if (start > m_max_frames) start -= m_max_frames;
while (start != m_current_frame)
{
f << "t" << thread_id << " ";
for (unsigned int i = 0; i < new_headings.size(); i++)
for (unsigned int i = 0; i < td.m_ordered_headings.size(); i++)
{
const EventData &ed = td.m_all_event_data[new_headings[i]];
const EventData &ed = td.m_all_event_data[td.m_ordered_headings[i]];
f << int(ed.getMarker(start).getDuration()*1000) << " ";
} // for i i new_headings
f << std::endl;
start = (start + 1) % m_max_frames;
}
} // for
f.close();
} // while start != m_current_frame
f.close();
} // for all thread_ids
std::ofstream f_gpu(base_name + ".profile-gpu");
f_gpu << "# ";
@ -550,5 +609,7 @@ void Profiler::writeToFile()
f_gpu << std::endl;
start = (start + 1) % m_max_frames;
}
f.close();
f_gpu.close();
m_lock.unlock();
} // writeFile

22
src/utils/profiler.hpp
View File

@ -144,10 +144,14 @@ private:
// --------------------------------------------------------------------
size_t getLayer() const { return m_layer; }
// --------------------------------------------------------------------
/** Called when an entry in the cyclic buffer is reused. Makes sure
* that time for a new event can be accumulated. */
void clear() { m_duration = 0; }
// --------------------------------------------------------------------
/** Sets start time and layer for this event. */
void setStart(double start, size_t layer = 0)
{
m_start = start; m_duration = 0; m_layer = layer;
m_start = start; m_layer = layer;
} // setStart
// --------------------------------------------------------------------
/** Sets the end time of this event. */
@ -194,6 +198,7 @@ private:
} // setEnd
// --------------------------------------------------------------------
const Marker& getMarker(int n) const { return m_all_markers[n]; }
Marker& getMarker(int n) { return m_all_markers[n]; }
// --------------------------------------------------------------------
/** Returns the colour for this event. */
video::SColor getColour() const { return m_colour; }
@ -209,6 +214,12 @@ private:
/** Stack of events to detect nesting. */
std::vector< std::string > m_event_stack;
/** This stores the event names in the order in which they occur.
* This means that 'outer' events occur here before any child
* events. This list is then used to determine the order in which the
* bar graphs are drawn, which results in the proper nesting of events.*/
std::vector<std::string> m_ordered_headings;
AllEventData m_all_event_data;
}; // class ThreadData
@ -219,7 +230,7 @@ private:
std::vector< ThreadData> m_all_threads_data;
/** A mapping of thread_t pointers to a unique integer (starting from 0).*/
Synchronised< std::vector<pthread_t> > m_thread_mapping;
std::vector<pthread_t> m_thread_mapping;
/** Buffer for the GPU times (in ms). */
std::vector<int> m_gpu_times;
@ -230,6 +241,12 @@ private:
/** Index of the current frame in the buffer. */
int m_current_frame;
/** We don't need the bool, but easiest way to get a lock for the whole
* instance (since we need to avoid that a synch is done which changes
* the current frame while another threaded uses this variable, or
* while a new thread is added. */
Synchronised<bool> m_lock;
/** True if the circular buffer has wrapped around. */
bool m_has_wrapped_around;
@ -271,6 +288,7 @@ public:
void pushCPUMarker(const char* name="N/A",
const video::SColor& color=video::SColor());
void popCPUMarker();
void toggleStatus();
void synchronizeFrame();
void draw();
void onClick(const core::vector2di& mouse_pos);