std::string CProfiler2::ThreadStorage::GetBuffer()
{
// Called from an arbitrary thread (not the one writing to the buffer).
//
// See comments on m_BufferPos0 etc.
shared_ptr<u8> buffer(new u8[BUFFER_SIZE], ArrayDeleter());
u32 pos1 = m_BufferPos1;
COMPILER_FENCE; // must read m_BufferPos1 before m_Buffer
memcpy(buffer.get(), m_Buffer, BUFFER_SIZE);
COMPILER_FENCE; // must read m_BufferPos0 after m_Buffer
u32 pos0 = m_BufferPos0;
// The range [pos1, pos0) modulo BUFFER_SIZE is invalid, so concatenate the rest of the buffer
if (pos1 <= pos0) // invalid range is in the middle of the buffer
return std::string(buffer.get()+pos0, buffer.get()+BUFFER_SIZE) + std::string(buffer.get(), buffer.get()+pos1);
else // invalid wrap is wrapped around the end/start buffer
return std::string(buffer.get()+pos0, buffer.get()+pos1);
}
void ProcessFrames()
{
while (!m_Frames.empty())
{
SFrame& frame = m_Frames.front();
// Queries don't become available in order, so check them all before
// trying to read the results from any
for (size_t j = 0; j < m_QueryTypes.size(); ++j)
{
size_t size = m_QueryTypes[j].counterBufferSize;
shared_ptr<char> buf(new char[size], ArrayDeleter());
for (size_t i = 0; i < frame.events.size(); ++i)
{
if (!frame.events[i].isEnter)
continue;
GLuint length = 0;
pglGetPerfQueryDataINTEL(frame.events[i].queries[j], INTEL_PERFQUERIES_NONBLOCK, size, buf.get(), &length);
ogl_WarnIfError();
if (length == 0)
return;
}
}
double lastTime = frame.timeStart;
std::stack<double> endTimes;
m_Storage.RecordFrameStart(frame.timeStart);
for (size_t i = 0; i < frame.events.size(); ++i)
{
if (frame.events[i].isEnter)
{
m_Storage.Record(CProfiler2::ITEM_ENTER, lastTime, frame.events[i].id);
if (i == 0)
m_Storage.RecordAttributePrintf("%u", frame.num);
double elapsed = 0.0;
for (size_t j = 0; j < m_QueryTypes.size(); ++j)
{
GLuint length;
char* buf = new char[m_QueryTypes[j].counterBufferSize];
pglGetPerfQueryDataINTEL(frame.events[i].queries[j], INTEL_PERFQUERIES_BLOCK, m_QueryTypes[j].counterBufferSize, buf, &length);
ogl_WarnIfError();
ENSURE(length == m_QueryTypes[j].counterBufferSize);
m_Storage.RecordAttributePrintf("-- %s --", m_QueryTypes[j].name.c_str());
for (size_t k = 0; k < m_QueryTypes[j].counters.size(); ++k)
{
SPerfCounter& counter = m_QueryTypes[j].counters[k];
if (counter.type == INTEL_PERFQUERIES_TYPE_UNSIGNED_INT)
{
ENSURE(counter.size == 4);
GLuint value;
memcpy(&value, buf + counter.offset, counter.size);
m_Storage.RecordAttributePrintf("%s: %u", counter.name.c_str(), value);
}
else if (counter.type == INTEL_PERFQUERIES_TYPE_UNSIGNED_INT64)
{
ENSURE(counter.size == 8);
GLuint64 value;
memcpy(&value, buf + counter.offset, counter.size);
m_Storage.RecordAttributePrintf("%s: %.0f", counter.name.c_str(), (double)value);
if (counter.name == "TotalTime")
elapsed = (double)value / 1e6;
}
else if (counter.type == INTEL_PERFQUERIES_TYPE_FLOAT)
{
ENSURE(counter.size == 4);
GLfloat value;
memcpy(&value, buf + counter.offset, counter.size);
m_Storage.RecordAttributePrintf("%s: %f", counter.name.c_str(), value);
}
else if (counter.type == INTEL_PERFQUERIES_TYPE_BOOL)
{
ENSURE(counter.size == 4);
GLuint value;
memcpy(&value, buf + counter.offset, counter.size);
ENSURE(value == 0 || value == 1);
m_Storage.RecordAttributePrintf("%s: %u", counter.name.c_str(), value);
}
else
{
debug_warn(L"unrecognised Intel performance counter type");
}
}
delete[] buf;
}
endTimes.push(lastTime + elapsed);
}
else
{
lastTime = endTimes.top();
endTimes.pop();
m_Storage.Record(CProfiler2::ITEM_LEAVE, lastTime, frame.events[i].id);
}
}
PopFrontFrame();
}
}
