print "Timing Info saved in: cuda_profile_1.log \n\n"
sys.exit()
###########################################################################
###########################################################################
changeInitial = True
savePos = False
start = cuda.Event()
end = cuda.Event()
occupancyOld, occupancy = 0, 0
launchCounter = 0
start.record()
while occupancyOld < maxTimeIndx:
occupancy = sum(timesOccupancy_d.get()>=nParticles)
if occupancy>occupancyOld or occupancy==0: printProgressTime( occupancy, maxTimeIndx, start.time_till(end.record().synchronize())*1e-3 )
occupancyOld = occupancy
mainKernel(np.uint8(usingAnimation), np.int32(nParticles), np.int32(collisionsPerRun), np.int32(nCircles), circlesCaract_d, np.int32(nLines), linesCaract_d,
initialPosX_d, initialPosY_d, initialVelX_d, initialVelY_d, initialRegionX_d, initialRegionY_d,
outPosX_d, outPosY_d, times_d,
np.float32(deltaTime_anim), timesIdx_anim_d,
np.float32( deltaTime_radius ), timesIdx_rad_d, timesOccupancy_d, radiusAll_d,
np.uint8(savePos), np.int32(particlesForPlot), np.uint8(changeInitial),
np.intp(0), grid=grid, block=block)
launchCounter += 1
#if runNumber%5==0 and plotting: plotData( runNumber )
print "\n\nFinished in : {0:.4f} sec\n".format( float( start.time_till(end.record().synchronize())*1e-3 ) )
#Get the results
#Save data
dataDir = currentDirectory + "/data"
print " Total particles: ", nParticles*nProc
print ' Particles per gpu: ', nParticles
print ' nSteps: ', totalSteps
print ''
MPIcomm.Barrier()
computeTime = 0
transferTime = 0
saveTime = 0
step = 0
start, end = cuda.Event(), cuda.Event()
startT = cuda.Event()
startT.record()
for stepCounter in range(totalSteps):
if stepCounter%1==0 and pId==0: printProgressTime( stepCounter, totalSteps, computeTime + transferTime + saveTime )
if stepCounter%4==0:
#Save positions
start.record()
posHD.create_dataset( '{0:03}'.format(stepCounter/4), data=posSend_h )
end.record(), end.synchronize()
saveTime += start.time_till(end)*1e-3
timeStep()
if pId == 0:
end.record(), end.synchronize()
totalTimeReal = startT.time_till(end)*1e-3
totalTime = computeTime + transferTime + saveTime
h, m, s = timeSplit( totalTime )
print '\n\nTotal time: {0}:{1:02}:{2:02} '.format( h, m, s )
pAnim.updateFunc = animationUpdate
pAnim.keyboard = keyboard
pAnim.startAnimation()
print ' nSteps: {0} \n'.format( totalSteps )
print " Output: {0} \n".format( outDataFile )
start, end = cuda.Event(), cuda.Event()
for stepCounter in range(totalSteps):
start.record()
if stepCounter%1 == 0: printProgressTime( stepCounter, totalSteps, runningTime )
getEnergy( stepCounter, simulationTime, energyList )
moveParticles( cudaPre(dt), posX_d, posY_d, posZ_d, velX_d, velY_d, velZ_d, accelX_d, accelY_d, accelZ_d )
getAcccelKernel( np.int32(nParticles), GMass, np.int32(usingAnimation),
posX_d, posY_d, posZ_d, posX_d, posY_d, posZ_d,
velX_d, velY_d, velZ_d, accelX_d, accelY_d, accelZ_d,
cudaPre( dt ), cudaPre(epsilon), np.int32(1),
np.intp(0), grid=grid, block=block )
simulationTime += dt
end.record()
end.synchronize()
runningTime += start.time_till(end)*1e-3
h, m, s = timeSplit( runningTime )
volumeRender.animate()
#Open outFile
snapsDir = dataDir + 'snapshots/'
ensureDirectory( snapsDir )
n = len( [ f for f in os.listdir( snapsDir ) if f.find(timeDirection)>0 ] )
snapsFileName = snapsDir + 'psi_{0}_{1}.h5'.format( timeDirection, n )
snapsFile = h5.File( snapsFileName, 'w')
print "\nnPoints: {0}x{1}x{2}".format(nWidth, nHeight, nWidth )
print 'Snapshots File: ', snapsFileName
print "Starting Real Dynamics: {0} timeUnits, dt: {1:1.2e} \n".format( endTime, dtReal )
printCounter = 0
start, end = cuda.Event(), cuda.Event()
start.record()
for n in range( nSnapshots ):
end.record().synchronize()
printProgressTime( n, nSnapshots, start.time_till(end)*1e-3 )
saveSnapshot( n, snapsFile, simulationTime, psi_d.get() )
[rk4_texture_iteration() for i in range(nPartialIter)]
simulationTime += dtReal*nPartialIter
printProgressTime( nSnapshots, nSnapshots, start.time_till(end)*1e-3 )
saveSnapshot( nSnapshots, snapsFile, simulationTime, psi_d.get() )
end.record().synchronize()
print "\n\nEnd Real Dynamics "
print 'Time: {0:.0f}\n'.format( start.time_till(end)*1e-3 )
volumeRender.specialKeys = specialKeyboardFunc
volumeRender.stepFunc = stepFuntion
#run volumeRender animation
volumeRender.animate()
print "nPoints: {0}x{1}x{2}".format(nWidth, nHeight, nWidth )
print "Starting Imaginary Dynamics: {0} timeUnits, dt: {1:1.2e} \n".format( timeRelax, dtImag )
simulationTime = -timeRelax
printCounter = 0
start, end = cuda.Event(), cuda.Event()
start.record()
while simulationTime < 0:
if timeRelax+simulationTime >= printCounter*timeRelax/500.:
end.record().synchronize()
printProgressTime( timeRelax + simulationTime, timeRelax, start.time_till(end)*1e-3 )
printCounter += 1
imaginaryStep()
simulationTime += dtImag
print "\nEnd Imaginary Dynamics\n "
saveState( fileName='initial/converged/psiConverged.h5')
#applyTransition = True
#timeTransition()
##cuda.memset_d8(activity_d.ptr, 1, nBlocks3D )
#endTime = 1
#nIterations = int( endTime/dtReal )
#print "Starting Real Dynamics: {0} timeUnits ".format( endTime )
#[ rk4_iteration() for i in range(nIterations) ]