float GetProgressUnitSize(int NumDataPoints, int num_cfft, long ac_fft_len) {
#ifdef USE_MANUAL_CALLSTACK
call_stack.enter("GetProgressUnitSize()");
#endif
// A ProgressUnit is defined as the computation time of a
// detection algorithm on the entire data block at a
// given chirpfft pair.
// Spike detection takes place for every FFT length and at
// any slew rate. PoT detctors, on the other hand, may opt
// to not execute if slew rate and/or FFT length limits are
// not met.
int i, ThisPoTLen, ThisPulsePoTLen, DummyOverlap;
double NumProgressUnits;
double LastChirpRate = 0.0f;
double TotalSpikeProgressUnits = 0.0;
double TotalGaussianProgressUnits = 0.0;
double TotalPulseProgressUnits = 0.0;
double TotalTripletProgressUnits = 0.0;
double TotalChirpProgressUnits = 0.0;
double rv;
NumProgressUnits = 0.0;
for (i = 0; i < num_cfft; i++) {
// FFTs and spike finding
TotalSpikeProgressUnits += SpikeProgressUnits(ChirpFftPairs[i].FftLen);
// Autocorr FFTs and finding
if ((long)ChirpFftPairs[i].FftLen == ac_fft_len) {
TotalSpikeProgressUnits += SpikeProgressUnits(ChirpFftPairs[i].FftLen);
}
// Chirping
if(ChirpFftPairs[i].ChirpRate != LastChirpRate) {
TotalChirpProgressUnits += ChirpProgressUnits();
LastChirpRate = ChirpFftPairs[i].ChirpRate;
}
ThisPoTLen = NumDataPoints / ChirpFftPairs[i].FftLen;
// Gaussians....
if (ChirpFftPairs[i].GaussFit) {
TotalGaussianProgressUnits += GaussianProgressUnits();
}
// Pulses and Triplets....
GetPulsePoTLen(ThisPoTLen, &ThisPulsePoTLen, &DummyOverlap);
#ifdef USE_PULSE
if(ChirpFftPairs[i].PulseFind)
TotalPulseProgressUnits += PulseProgressUnits(ThisPulsePoTLen, ChirpFftPairs[i].FftLen - 1);
#endif
#ifdef USE_TRIPLET
if(ThisPulsePoTLen >= PoTInfo.TripletMin && ThisPulsePoTLen <= PoTInfo.TripletMax)
TotalTripletProgressUnits += TripletProgressUnits();
#endif
}
NumProgressUnits = TotalChirpProgressUnits +
TotalSpikeProgressUnits +
TotalGaussianProgressUnits +
TotalPulseProgressUnits +
TotalTripletProgressUnits;
/*
fprintf(stderr, "%f ChirpProgressUnits (%f\%)\n", TotalChirpProgressUnits, TotalChirpProgressUnits/NumProgressUnits);
fprintf(stderr, "%f SpikeProgressUnits (%f\%)\n", TotalSpikeProgressUnits, TotalSpikeProgressUnits/NumProgressUnits);
fprintf(stderr, "%f GaussianProgressUnits (%f\%)\n", TotalGaussianProgressUnits, TotalGaussianProgressUnits/NumProgressUnits);
fprintf(stderr, "%f TripletProgressUnits (%f\%)\n", TotalTripletProgressUnits, TotalTripletProgressUnits/NumProgressUnits);
fprintf(stderr, "%f PulseProgressUnits (%f\%)\n", TotalPulseProgressUnits, TotalPulseProgressUnits/NumProgressUnits);
fprintf(stderr, "%f NumProgressUnits\n", NumProgressUnits);
*/
// Add a fudge factor of 0.01% to make sure we do not hit 100% done too soon
rv=(1.0f/(float)(NumProgressUnits + NumProgressUnits * 0.0001));
#ifdef USE_MANUAL_CALLSTACK
call_stack.exit();
#endif
return rv;
}
int analyze_pot(float *PowerSpectrum, int NumDataPoints, ChirpFftPair_t &cfft, int offset)
{
// This function analyses Power over Time for the current data block.
// The PoT array is created by taking an array of power spectra (a
// standard row-major 2D array) and extracting the PoT as column-major
// data. We essentialy turn the array on its side.
int retval = 0,
i,
FftLength=cfft.FftLen, // Current FFT length
ThisPoT, // index of current PoT along the freq axis
PoTLen, // complement of FFT length - determines time res
PulsePoTLen, // length of PoT segment passed to pulse finders
Overlap, // PoT segment overlap in bins
TOffset, // index into ThisPoT of current pulse segment
PulsePoTNum = 0, // the oridinal position of a pulse PoT w/in a full PoT
NumPulsePoTs = 0, // the number of pulse PoTs w/in a full PoT. This is
// constant regardless of FFT or PoT length and is
// determined by slew rate.
AdvanceBy; // the number of bins to advance for the next pulse PoT
float ProgressAddFactor = 0.0, // sum of progress adds for ThisPoT
ProgressPerPulsePoT = 0.0; // for local progress display
bool SkipGauss = false,
SkipPulse = false,
SkipTriplet = false,
TOffsetOK = true;
static float *GaussPoT = NULL,
*PulsePoT = NULL;
#ifdef DEBUG_POT
fprintf(stderr, "========= FftLength = %d =========\n", FftLength);
#endif
PoTLen = NumDataPoints / FftLength; // in bins
GetPulsePoTLen(PoTLen, &PulsePoTLen, &Overlap); // in bins
AdvanceBy = PulsePoTLen - Overlap; // in bins
// Max limits how *slow* the slewrate can be, while Min limits how
// *fast* the slewrate can be. Max is limited only by the client
// memory budget.
if(PulsePoTLen > PoTInfo.TripletMax || PulsePoTLen < PoTInfo.TripletMin)
SkipTriplet = true;
SkipGauss = !(cfft.GaussFit);
SkipPulse = !(cfft.PulseFind);
if(!SkipPulse || !SkipTriplet)
{
// NumPulsePoTs is the number of PoT segments that we pass to the pulse
// detectors per frequency bin. ProgressPerPulsePoT is the inverse of
// number of pulse detection segments in the entire data block, taking
// into account that we skip the first (DC) frequency bin. An assumption
// is made here that minimum pulse/triplet PoT length will always be
// greater than 1. Otherwise, AdvanceBy can become zero and a divide by
// zero can occur. The assumption is also made that FftLength is always
// greater than 1!
NumPulsePoTs = 1 + (PoTLen-PulsePoTLen)/AdvanceBy + ((PoTLen-PulsePoTLen)%AdvanceBy ? 1 : 0);
ProgressPerPulsePoT = (float)1 / ((FftLength - 1) * NumPulsePoTs);
}
#ifdef DEBUG_POT
fprintf(stderr, "SlewRate = %f\n", PoTInfo.SlewRate);
fprintf(stderr, "PoTLen = %d\n", PoTLen);
fprintf(stderr, "MaxPoTLen = %d\n", PoTInfo.MaxPoTLen);
fprintf(stderr, "PoTDuration = %f\n", PoTInfo.WUDuration);
fprintf(stderr, "BeamRate = %f\n", PoTInfo.BeamRate);
fprintf(stderr, "PulsePoTLen = %d\n", PulsePoTLen);
fprintf(stderr, "Overlap = %d\n", Overlap);
fprintf(stderr, "AdvanceBy = %d\n", AdvanceBy);
fprintf(stderr, "min_slew = %f\n", PoTInfo.min_slew);
fprintf(stderr, "max_slew = %f\n", PoTInfo.max_slew);
fprintf(stderr, "PulseOverlapFactor = %f\n", PoTInfo.PulseOverlapFactor);
fprintf(stderr, "PulseBeams = %f\n", PoTInfo.PulseBeams);
fprintf(stderr, "PulseThresh = %f\n", PoTInfo.PulseThresh);
fprintf(stderr, "PulseMax = %d\n", PoTInfo.PulseMax);
fprintf(stderr, "PulseMin = %d\n", PoTInfo.PulseMin);
fprintf(stderr, "PulseFftMax = %d\n", PoTInfo.PulseFftMax);
fprintf(stderr, "TripletThresh = %f\n", PoTInfo.TripletThresh);
fprintf(stderr, "TripletMax = %d\n", PoTInfo.TripletMax);
fprintf(stderr, "TripletMin = %d\n", PoTInfo.TripletMin);
#endif
#ifndef USE_PULSE
SkipPulse = TRUE;
static int doneprintnopulsefind = 0;
if(!doneprintnopulsefind)
{
fprintf(stderr,"SkipPulse is set to TRUE: Not doing Pulsefinds.\n");
doneprintnopulsefind = TRUE;
}
#endif
#ifndef USE_TRIPLET
SkipTriplet = TRUE;
#endif
// Get memory fot the PoT arrays. The PoT array for gausian analysis is
// of set size. The PoT array for pulse analysis is sized to cover
// PulseBeams beams, regardless of whether this violates either the
// triplet or pulse limits on array size.
if(!GaussPoT)
{
GaussPoT = (float *)malloc_a(swi.analysis_cfg.gauss_pot_length * sizeof(float), MEM_ALIGN);
if(GaussPoT == NULL)
{
SETIERROR(MALLOC_FAILED, "GaussPoT == NULL");
}
}
if(!PulsePoT)
{
PulsePoT = (float *)calloc_a(PoTInfo.MaxPoTLen+3, sizeof(float), MEM_ALIGN);
if(PulsePoT == NULL)
{
SETIERROR(MALLOC_FAILED, "PulsePoT == NULL");
}
}
bool b_gaussStarted = false;
// Look for gaussians ---------------------------------------------------
if(!SkipGauss && (analysis_state.PoT_activity == POT_DOING_GAUSS ||
analysis_state.PoT_activity == POT_INACTIVE))
{
#ifdef BOINC_APP_GRAPHICS
if (!nographics())
strcpy(sah_graphics->status, "Searching for Gaussians");
#endif
// If we are back from being interrupted in the middle of gaussian PoT
// analysis, load state and continue. Otherwise start anew, skipping
// the DC (0) bin.
if(analysis_state.PoT_activity == POT_DOING_GAUSS)
{
ThisPoT = analysis_state.PoT_freq_bin;
}
else
{
ThisPoT = 1; // skip the DC bin on start of new cfft pair
}
// Initial display of local Progress / CPU time. Assumes that
// we start ThisPoT at 1 each time in!
#ifdef BOINC_APP_GRAPHICS
if (!nographics())
sah_graphics->local_progress = ((float)ThisPoT-1)/(FftLength-1);
#endif
#ifdef USE_CUDA
//#ifndef CUDAACC_EMULATION
if(gSetiUseCudaDevice)
{
//TODO: remove the autocorr_fftlen test when v6 fully deprecated.
bool noscore = false; //swi.analysis_cfg.autocorr_fftlen!=0 && gaussian_count!=0;
//cudaAcc_Gaussfit(FftLength, best_gauss->score, noscore);
//printf("GaussFitStart\r\n");
b_gaussStarted = true; // started earlier
//cudaAcc_GaussfitStart(FftLength, best_gauss->score, noscore);
if(PoTLen > swi.analysis_cfg.gauss_pot_length)
analysis_state.FLOP_counter+=((double)NumDataPoints+swi.analysis_cfg.gauss_pot_length * (FftLength-1)); // GetFixedPoT
// There are FftLength - 1 fixed PoTs for a chirp/fft pair, and for each fixed PoT full analysis would do
// (1 + (PoTInfo.GaussTOffsetStop - PoTInfo.GaussTOffsetStart)) GetPeak and GetTrueMean operations.
// Use (1 - PoTInfo.GaussSigma*0.09) empirically found to represent fraction of PoTs which don't
// take either of the two early out paths.
double CorrWt = (1.0 - PoTInfo.GaussSigma*0.09) * (FftLength - 1) * (1 + (PoTInfo.GaussTOffsetStop - PoTInfo.GaussTOffsetStart));
analysis_state.FLOP_counter+=6.0*floor(PoTInfo.GaussSigma+0.5) * CorrWt; // GetPeak
analysis_state.FLOP_counter+=(double)(floor(PoTInfo.GaussSigma+0.5) * 3.911+5) * CorrWt; // GetTrueMean
// Estimate one in twenty fit positions look good enough to be checked further.
analysis_state.FLOP_counter+= 0.05 * (20.0*swi.analysis_cfg.gauss_pot_length+5.0) * CorrWt; // GetChiSq / 20
progress += ProgressUnitSize * GaussianProgressUnits();
progress=std::min(progress,1.0); // prevent display of > 100%
fraction_done(progress,remaining);
}
//#endif CUDAACC_EMULATION
else
{
#endif //USE_CUDA
// loop through frequencies
/* for(; ThisPoT < FftLength; ThisPoT++)
{
// Create PowerOfTime array for gaussian fit
retval = GetFixedPoT(
PowerSpectrum,
NumDataPoints,
FftLength,
GaussPoT,
swi.analysis_cfg.gauss_pot_length,
ThisPoT
);
if (retval)
continue;
retval = GaussFit(GaussPoT, FftLength, ThisPoT);
if (retval)
SETIERROR(retval,"from GaussFit");
progress += ProgressUnitSize * GaussianProgressUnits() / (float)(FftLength - 1);
progress=std::min(progress,1.0); // prevent display of > 100%
fraction_done(progress,remaining);
// At the end of each frequency bin we update progress and save state.
#ifdef BOINC_APP_GRAPHICS
if (!nographics())
{
sah_graphics->local_progress = ((float)ThisPoT)/(FftLength-1);
}
#endif
analysis_state.PoT_freq_bin = ThisPoT;
analysis_state.PoT_activity = POT_DOING_GAUSS;
retval = checkpoint();
if (retval)
SETIERROR(retval,"from checkpoint()");
} // end loop through frequencies
*/
#ifdef USE_CUDA
}
#endif //USE_CUDA
analysis_state.PoT_freq_bin = -1;
analysis_state.PoT_activity = POT_INACTIVE;
} // end looking for gaussians
// Look for pulses -------------------------------------------------------
if(!SkipPulse || !SkipTriplet)
{
#ifdef BOINC_APP_GRAPHICS
if (!nographics())
{
strcpy(sah_graphics->status, "Searching for Pulses / Triplets");
// init local progress for pulse search
sah_graphics->local_progress = 0;
}
#endif
// If we are back from being interrupted in the middle of pulse PoT
// analysis, load state and continue. Otherwise start anew, skipping
// the DC (0) bin.
if(analysis_state.PoT_activity == POT_DOING_PULSE)
{
ThisPoT = analysis_state.PoT_freq_bin;
}
else
{
ThisPoT = 1; // skip the DC bin on start of new cfft pair
}
PulsePoTNum = 0;
#ifdef BOINC_APP_GRAPHICS
// Inital display of Local Progress
if(!nographics())
{
sah_graphics->local_progress = (((ThisPoT-1) * NumPulsePoTs) +
PulsePoTNum) *
ProgressPerPulsePoT;
}
#endif
#ifdef USE_CUDA
if(gSetiUseCudaDevice)
{
/*
if(!SkipTriplet || !SkipPulse) // do beforehand on fftstreamX
{
// CUDASYNC;
//printf("CalculateMean\r\n");
cudaAcc_calculate_mean(PulsePoTLen, 0, AdvanceBy, FftLength);
}
if(!SkipPulse)
{
//printf("FindPulses\r\n");
cudaAcc_find_pulses((float) best_pulse->score, PulsePoTLen, AdvanceBy, FftLength);
}
if(!SkipTriplet)
{
//printf("FindTriplets\r\n");
cudaAcc_find_triplets(PulsePoTLen, (float)PoTInfo.TripletThresh, AdvanceBy, FftLength);
}
*/
/*
timespec t1, t2;
t1.tv_sec = 0;
t1.tv_nsec = 5000;
while(cudaEventQuery(summaxDoneEvent) != cudaSuccess)
nanosleep(&t1, &t2);
*/
int gflags = 0;
if(b_gaussStarted)
{
//printf("fetchGaussFitFlags\r\n");
b_gaussStarted = false;
gflags = cudaAcc_fetchGaussfitFlags(FftLength, best_gauss->score);
}
//printf("fetchTripletAndPulseFlags\r\n");
int has_dataT = 0, has_dataP = 0;
if(!SkipTriplet)
has_dataT = cudaAcc_fetchTripletFlags(SkipTriplet, PulsePoTLen, AdvanceBy, FftLength, offset);
if(gflags > 0)
{
//printf("ProcessGaussFit\r\n");
cudaAcc_processGaussFit(FftLength, best_gauss->score);
}
if(!SkipTriplet)
{
if((has_dataT & 1) && !(has_dataT & 4)) // has triplet data and no error in triplet
{
//printf("processTripletResults\r\n");
cudaAcc_processTripletResults(PulsePoTLen, AdvanceBy, FftLength);
}
analysis_state.FLOP_counter+=(10.0 + PulsePoTLen) * NumPulsePoTs * (FftLength - 1);
// ! hard to estimate, so be generous and use 9
analysis_state.FLOP_counter+=810.0; // (10.0*numBinsAboveThreshold*numBinsAboveThreshold);
progress += ProgressUnitSize * TripletProgressUnits();
}
if(!SkipPulse)
has_dataP = cudaAcc_fetchPulseFlags(SkipTriplet, PulsePoTLen, AdvanceBy, FftLength, offset);
if(!SkipPulse)
{
if((has_dataP & 2) && !(has_dataP & 8)) // has pulse data and no error in pulse
{
//printf("processPulseResults\r\n");
cudaAcc_processPulseResults(PulsePoTLen, AdvanceBy, FftLength);
}
analysis_state.FLOP_counter+=(PulsePoTLen*0.1818181818182+400.0)*PulsePoTLen * NumPulsePoTs * (FftLength - 1);
progress += ProgressUnitSize * PulseProgressUnits(PulsePoTLen, FftLength - 1);
}
//#ifndef CUDAACC_EMULATION
//if(!SkipTriplet)
// cudaAcc_fetchTripletAndPulseFlags(SkipTriplet, SkipPulse, PulsePoTLen, AdvanceBy, FftLength);
progress=std::min(progress,1.0); // prevent display of > 100%
fraction_done(progress,remaining);
//if(!SkipTriplet)
// cudaAcc_processTripletResults(PulsePoTLen, AdvanceBy, FftLength);
//#endif //CUDAACC_EMULATION
}
// else
// {
#endif //USE_CUDA
// loop through frequencies
/* for(; ThisPoT < FftLength; ThisPoT++)
{
// loop through time for each frequency. PulsePoTNum is
// used only for progress calculation.
for(TOffset = 0, PulsePoTNum = 1, TOffsetOK = true;
TOffsetOK;
PulsePoTNum++, TOffset += AdvanceBy)
{
// Create PowerOfTime array for pulse detection. If there
// are not enough points left in this PoT, adjust TOffset
// to get the latest possible pulse PoT.
if(TOffset + PulsePoTLen >= PoTLen)
{
TOffsetOK = false;
TOffset = PoTLen - PulsePoTLen;
}
if (use_transposed_pot)
{
memcpy(PulsePoT, &PowerSpectrum[ThisPoT * PoTLen + TOffset], PulsePoTLen*sizeof(float));
}
else
{
for(i = 0; i < PulsePoTLen; i++)
{
PulsePoT[i] = PowerSpectrum[ThisPoT + (TOffset+i) * FftLength];
}
}
if(!SkipTriplet)
{
retval = find_triplets(PulsePoT,
PulsePoTLen,
(float)PoTInfo.TripletThresh,
TOffset,
ThisPoT);
if (retval)
SETIERROR(retval,"from find_triplets()");
}
//#ifndef CUDAACC_EMULATION
if(!SkipPulse)
{
retval = find_pulse(PulsePoT,
PulsePoTLen,
(float)PoTInfo.PulseThresh,
TOffset,
ThisPoT
);
if (retval)
SETIERROR(retval,"from find_pulse()");
}
//#endif //CUDAACC_EMULATION
// At the end of each pulse PoT we update progress. Progress
// will thus get updted several times per frequency bin.
#ifdef BOINC_APP_GRAPHICS
if (!nographics())
{
sah_graphics->local_progress = (((ThisPoT-1) * NumPulsePoTs) +
PulsePoTNum) *
ProgressPerPulsePoT;
}
#endif
if(!SkipTriplet)
{
progress += (ProgressUnitSize * TripletProgressUnits()) /
(float)(FftLength - 1) / NumPulsePoTs;
}
if(!SkipPulse)
{
progress += (ProgressUnitSize * PulseProgressUnits(PulsePoTLen, FftLength - 1)) /
(float)(FftLength - 1) / NumPulsePoTs;
}
progress=std::min(progress,1.0); // prevent display of > 100%
fraction_done(progress,remaining);
} // end loop through time for each frequency
// At the end of each frequency bin we save state.
analysis_state.PoT_activity = POT_DOING_PULSE;
analysis_state.PoT_freq_bin = ThisPoT;
retval = checkpoint();
if (retval)
SETIERROR(retval,"from checkpoint()");
} // end loop through frequencies
analysis_state.PoT_freq_bin = -1;
analysis_state.PoT_activity = POT_INACTIVE;
#ifdef USE_CUDA
}
#endif //USE_CUDA
*/
} // end looking for pulses
if(b_gaussStarted) // process results late
{
//printf("late GaussBlock\r\n");
int flags = cudaAcc_fetchGaussfitFlags(FftLength, best_gauss->score);
if(flags>0)
cudaAcc_processGaussFit(FftLength, best_gauss->score);
}
return (retval); // no error return point
}
