/*-------------------------------------------------------------------------
* drawElements Quality Program OpenGL (ES) Module
* -----------------------------------------------
*
* Copyright 2014 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*//*!
* \file
* \brief Calibration tools.
*//*--------------------------------------------------------------------*/
#include "glsCalibration.hpp"
#include "tcuTestLog.hpp"
#include "tcuVectorUtil.hpp"
#include "deStringUtil.hpp"
#include "deMath.h"
#include "deClock.h"
#include <algorithm>
#include <limits>
using std::string;
using std::vector;
using tcu::Vec2;
using tcu::TestLog;
using tcu::TestNode;
using namespace glu;
namespace deqp
{
namespace gls
{
// Reorders input arbitrarily, linear complexity and no allocations
template<typename T>
float destructiveMedian (vector<T>& data)
{
const typename vector<T>::iterator mid = data.begin()+data.size()/2;
std::nth_element(data.begin(), mid, data.end());
if (data.size()%2 == 0) // Even number of elements, need average of two centermost elements
return (*mid + *std::max_element(data.begin(), mid))*0.5f; // Data is partially sorted around mid, mid is half an item after center
else
return *mid;
}
LineParameters theilSenLinearRegression (const std::vector<tcu::Vec2>& dataPoints)
{
const float epsilon = 1e-6f;
const int numDataPoints = (int)dataPoints.size();
vector<float> pairwiseCoefficients;
vector<float> pointwiseOffsets;
LineParameters result (0.0f, 0.0f);
// Compute the pairwise coefficients.
for (int i = 0; i < numDataPoints; i++)
{
const Vec2& ptA = dataPoints[i];
for (int j = 0; j < i; j++)
{
const Vec2& ptB = dataPoints[j];
if (de::abs(ptA.x() - ptB.x()) > epsilon)
pairwiseCoefficients.push_back((ptA.y() - ptB.y()) / (ptA.x() - ptB.x()));
}
}
// Find the median of the pairwise coefficients.
// \note If there are no data point pairs with differing x values, the coefficient variable will stay zero as initialized.
if (!pairwiseCoefficients.empty())
result.coefficient = destructiveMedian(pairwiseCoefficients);
// Compute the offsets corresponding to the median coefficient, for all data points.
for (int i = 0; i < numDataPoints; i++)
pointwiseOffsets.push_back(dataPoints[i].y() - result.coefficient*dataPoints[i].x());
// Find the median of the offsets.
// \note If there are no data points, the offset variable will stay zero as initialized.
if (!pointwiseOffsets.empty())
result.offset = destructiveMedian(pointwiseOffsets);
return result;
}
// Sample from given values using linear interpolation at a given position as if values were laid to range [0, 1]
template <typename T>
static float linearSample (const std::vector<T>& values, float position)
{
DE_ASSERT(position >= 0.0f);
DE_ASSERT(position <= 1.0f);
const int maxNdx = (int)values.size() - 1;
const float floatNdx = maxNdx * position;
const int lowerNdx = (int)deFloatFloor(floatNdx);
const int higherNdx = lowerNdx + (lowerNdx == maxNdx ? 0 : 1); // Use only last element if position is 1.0
const float interpolationFactor = floatNdx - (float)lowerNdx;
DE_ASSERT(lowerNdx >= 0 && lowerNdx < (int)values.size());
DE_ASSERT(higherNdx >= 0 && higherNdx < (int)values.size());
DE_ASSERT(interpolationFactor >= 0 && interpolationFactor < 1.0f);
return tcu::mix((float)values[lowerNdx], (float)values[higherNdx], interpolationFactor);
}
LineParametersWithConfidence theilSenSiegelLinearRegression (const std::vector<tcu::Vec2>& dataPoints, float reportedConfidence)
{
DE_ASSERT(!dataPoints.empty());
// Siegel's variation
const float epsilon = 1e-6f;
const int numDataPoints = (int)dataPoints.size();
std::vector<float> medianSlopes;
std::vector<float> pointwiseOffsets;
LineParametersWithConfidence result;
// Compute the median slope via each element
for (int i = 0; i < numDataPoints; i++)
{
const tcu::Vec2& ptA = dataPoints[i];
std::vector<float> slopes;
slopes.reserve(numDataPoints);
for (int j = 0; j < numDataPoints; j++)
{
const tcu::Vec2& ptB = dataPoints[j];
if (de::abs(ptA.x() - ptB.x()) > epsilon)
slopes.push_back((ptA.y() - ptB.y()) / (ptA.x() - ptB.x()));
}
// Add median of slopes through point i
medianSlopes.push_back(destructiveMedian(slopes));
}
DE_ASSERT(!medianSlopes.empty());
// Find the median of the pairwise coefficients.
std::sort(medianSlopes.begin(), medianSlopes.end());
result.coefficient = linearSample(medianSlopes, 0.5f);
// Compute the offsets corresponding to the median coefficient, for all data points.
for (int i = 0; i < numDataPoints; i++)
pointwiseOffsets.push_back(dataPoints[i].y() - result.coefficient*dataPoints[i].x());
// Find the median of the offsets.
std::sort(pointwiseOffsets.begin(), pointwiseOffsets.end());
result.offset = linearSample(pointwiseOffsets, 0.5f);
// calculate confidence intervals
result.coefficientConfidenceLower = linearSample(medianSlopes, 0.5f - reportedConfidence*0.5f);
result.coefficientConfidenceUpper = linearSample(medianSlopes, 0.5f + reportedConfidence*0.5f);
result.offsetConfidenceLower = linearSample(pointwiseOffsets, 0.5f - reportedConfidence*0.5f);
result.offsetConfidenceUpper = linearSample(pointwiseOffsets, 0.5f + reportedConfidence*0.5f);
result.confidence = reportedConfidence;
return result;
}
bool MeasureState::isDone (void) const
{
return (int)frameTimes.size() >= maxNumFrames || (frameTimes.size() >= 2 &&
frameTimes[frameTimes.size()-2] >= (deUint64)frameShortcutTime &&
frameTimes[frameTimes.size()-1] >= (deUint64)frameShortcutTime);
}
deUint64 MeasureState::getTotalTime (void) const
{
deUint64 time = 0;
for (int i = 0; i < (int)frameTimes.size(); i++)
time += frameTimes[i];
return time;
}
void MeasureState::clear (void)
{
maxNumFrames = 0;
frameShortcutTime = std::numeric_limits<float>::infinity();
numDrawCalls = 0;
frameTimes.clear();
}
void MeasureState::start (int maxNumFrames_, float frameShortcutTime_, int numDrawCalls_)
{
frameTimes.clear();
frameTimes.reserve(maxNumFrames_);
maxNumFrames = maxNumFrames_;
frameShortcutTime = frameShortcutTime_;
numDrawCalls = numDrawCalls_;
}
TheilSenCalibrator::TheilSenCalibrator (void)
: m_params (1 /* initial calls */, 10 /* calibrate iter frames */, 2000.0f /* calibrate iter shortcut threshold */, 31 /* max calibration iterations */,
1000.0f/30.0f /* target frame time */, 1000.0f/60.0f /* frame time cap */, 1000.0f /* target measure duration */)
, m_state (INTERNALSTATE_LAST)
{
clear();
}
TheilSenCalibrator::TheilSenCalibrator (const CalibratorParameters& params)
: m_params (params)
, m_state (INTERNALSTATE_LAST)
{
clear();
}
TheilSenCalibrator::~TheilSenCalibrator()
{
}
void TheilSenCalibrator::clear (void)
{
m_measureState.clear();
m_calibrateIterations.clear();
m_state = INTERNALSTATE_CALIBRATING;
}
void TheilSenCalibrator::clear (const CalibratorParameters& params)
{
m_params = params;
clear();
}
TheilSenCalibrator::State TheilSenCalibrator::getState (void) const
{
if (m_state == INTERNALSTATE_FINISHED)
return STATE_FINISHED;
else
{
DE_ASSERT(m_state == INTERNALSTATE_CALIBRATING || !m_measureState.isDone());
return m_measureState.isDone() ? STATE_RECOMPUTE_PARAMS : STATE_MEASURE;
}
}
void TheilSenCalibrator::recordIteration (deUint64 iterationTime)
{
DE_ASSERT((m_state == INTERNALSTATE_CALIBRATING || m_state == INTERNALSTATE_RUNNING) && !m_measureState.isDone());
m_measureState.frameTimes.push_back(iterationTime);
if (m_state == INTERNALSTATE_RUNNING && m_measureState.isDone())
m_state = INTERNALSTATE_FINISHED;
}
void TheilSenCalibrator::recomputeParameters (void)
{
DE_ASSERT(m_state == INTERNALSTATE_CALIBRATING);
DE_ASSERT(m_measureState.isDone());
// Minimum and maximum acceptable frame times.
const float minGoodFrameTimeUs = m_params.targetFrameTimeUs * 0.95f;
const float maxGoodFrameTimeUs = m_params.targetFrameTimeUs * 1.15f;
const int numIterations = (int)m_calibrateIterations.size();
// Record frame time.
if (numIterations > 0)
{
m_calibrateIterations.back().frameTime = (float)((double)m_measureState.getTotalTime() / (double)m_measureState.frameTimes.size());
// Check if we're good enough to stop calibrating.
{
bool endCalibration = false;
// Is the maximum calibration iteration limit reached?
endCalibration = endCalibration || (int)m_calibrateIterations.size() >= m_params.maxCalibrateIterations;
// Do a few past iterations have frame time in acceptable range?
{
const int numRelevantPastIterations = 2;
if (!endCalibration && (int)m_calibrateIterations.size() >= numRelevantPastIterations)
{
const CalibrateIteration* const past = &m_calibrateIterations[m_calibrateIterations.size() - numRelevantPastIterations];
bool allInGoodRange = true;
for (int i = 0; i < numRelevantPastIterations && allInGoodRange; i++)
{
const float frameTimeUs = past[i].frameTime;
if (!de::inRange(frameTimeUs, minGoodFrameTimeUs, maxGoodFrameTimeUs))
allInGoodRange = false;
}
endCalibration = endCalibration || allInGoodRange;
}
}
// Do a few past iterations have similar-enough call counts?
{
const int numRelevantPastIterations = 3;
if (!endCalibration && (int)m_calibrateIterations.size() >= numRelevantPastIterations)
{
const CalibrateIteration* const past = &m_calibrateIterations[m_calibrateIterations.size() - numRelevantPastIterations];
int minCallCount = std::numeric_limits<int>::max();
int maxCallCount = std::numeric_limits<int>::min();
for (int i = 0; i < numRelevantPastIterations; i++)
{
minCallCount = de::min(minCallCount, past[i].numDrawCalls);
maxCallCount = de::max(maxCallCount, past[i].numDrawCalls);
}
if ((float)(maxCallCount - minCallCount) <= (float)minCallCount * 0.1f)
endCalibration = true;
}
}
// Is call count just 1, and frame time still way too high?
endCalibration = endCalibration || (m_calibrateIterations.back().numDrawCalls == 1 && m_calibrateIterations.back().frameTime > m_params.targetFrameTimeUs*2.0f);
if (endCalibration)
{
const int minFrames = 10;
const int maxFrames = 60;
int numMeasureFrames = deClamp32(deRoundFloatToInt32(m_params.targetMeasureDurationUs / m_calibrateIterations.back().frameTime), minFrames, maxFrames);
m_state = INTERNALSTATE_RUNNING;
m_measureState.start(numMeasureFrames, m_params.calibrateIterationShortcutThreshold, m_calibrateIterations.back().numDrawCalls);
return;
}
}
}
DE_ASSERT(m_state == INTERNALSTATE_CALIBRATING);
// Estimate new call count.
{
int newCallCount;
if (numIterations == 0)
newCallCount = m_params.numInitialCalls;
else
{
vector<Vec2> dataPoints;
for (int i = 0; i < numIterations; i++)
{
if (m_calibrateIterations[i].numDrawCalls == 1 || m_calibrateIterations[i].frameTime > m_params.frameTimeCapUs*1.05f) // Only account for measurements not too near the cap.
dataPoints.push_back(Vec2((float)m_calibrateIterations[i].numDrawCalls, m_calibrateIterations[i].frameTime));
}
if (numIterations == 1)
dataPoints.push_back(Vec2(0.0f, 0.0f)); // If there's just one measurement so far, this will help in getting the next estimate.
{
const float targetFrameTimeUs = m_params.targetFrameTimeUs;
const float coeffEpsilon = 0.001f; // Coefficient must be large enough (and positive) to be considered sensible.
const LineParameters estimatorLine = theilSenLinearRegression(dataPoints);
int prevMaxCalls = 0;
// Find the maximum of the past call counts.
for (int i = 0; i < numIterations; i++)
prevMaxCalls = de::max(prevMaxCalls, m_calibrateIterations[i].numDrawCalls);
if (estimatorLine.coefficient < coeffEpsilon) // Coefficient not good for sensible estimation; increase call count enough to get a reasonably different value.
newCallCount = 2*prevMaxCalls;
else
{
// Solve newCallCount such that approximately targetFrameTime = offset + coefficient*newCallCount.
newCallCount = (int)((targetFrameTimeUs - estimatorLine.offset) / estimatorLine.coefficient + 0.5f);
// We should generally prefer FPS counts below the target rather than above (i.e. higher frame times rather than lower).
if (estimatorLine.offset + estimatorLine.coefficient*(float)newCallCount < minGoodFrameTimeUs)
newCallCount++;
}
// Make sure we have at least minimum amount of calls, and don't allow increasing call count too much in one iteration.
newCallCount = de::clamp(newCallCount, 1, prevMaxCalls*10);
}
}
m_measureState.start(m_params.maxCalibrateIterationFrames, m_params.calibrateIterationShortcutThreshold, newCallCount);
m_calibrateIterations.push_back(CalibrateIteration(newCallCount, 0.0f));
}
}
void logCalibrationInfo (tcu::TestLog& log, const TheilSenCalibrator& calibrator)
{
const CalibratorParameters& params = calibrator.getParameters();
const std::vector<CalibrateIteration>& calibrateIterations = calibrator.getCalibrationInfo();
// Write out default calibration info.
log << TestLog::Section("CalibrationInfo", "Calibration Info")
<< TestLog::Message << "Target frame time: " << params.targetFrameTimeUs << " us (" << 1000000 / params.targetFrameTimeUs << " fps)" << TestLog::EndMessage;
for (int iterNdx = 0; iterNdx < (int)calibrateIterations.size(); iterNdx++)
{
log << TestLog::Message << " iteration " << iterNdx << ": " << calibrateIterations[iterNdx].numDrawCalls << " calls => "
<< de::floatToString(calibrateIterations[iterNdx].frameTime, 2) << " us ("
<< de::floatToString(1000000.0f / calibrateIterations[iterNdx].frameTime, 2) << " fps)" << TestLog::EndMessage;
}
log << TestLog::Integer("CallCount", "Calibrated call count", "", QP_KEY_TAG_NONE, calibrator.getMeasureState().numDrawCalls)
<< TestLog::Integer("FrameCount", "Calibrated frame count", "", QP_KEY_TAG_NONE, (int)calibrator.getMeasureState().frameTimes.size());
log << TestLog::EndSection;
}
} // gls
} // deqp