def pdf(gpu_detector, npdfs=10, nevents=100, nreps=16, ndaq=1,
nthreads_per_block=64, max_blocks=1024):
"""
Returns the average number of 100 MeV events per second that can be
histogrammed per second.
Args:
- gpu_instance, chroma.gpu.GPU
The GPU instance passed to the GPUDetector constructor.
- npdfs, int
The number of pdf generations to average.
- nevents, int
The number of 100 MeV events to generate for each PDF.
- nreps, int
The number of times to propagate each event and add to PDF
- ndaq, int
The number of times to run the DAQ simulation on the propagated
event and add it to the PDF.
"""
rng_states = gpu.get_rng_states(nthreads_per_block*max_blocks)
gpu_daq = gpu.GPUDaq(gpu_detector)
gpu_pdf = gpu.GPUPDF()
gpu_pdf.setup_pdf(gpu_detector.nchannels, 100, (-0.5, 999.5), 10, (-0.5, 9.5))
run_times = []
for i in tools.progress(range(npdfs)):
t0 = time.time()
gpu_pdf.clear_pdf()
vertex_gen = generator.vertex.constant_particle_gun('e-', (0,0,0),
(1,0,0), 100)
vertex_iter = itertools.islice(vertex_gen, nevents)
for ev in g4generator.generate_events(vertex_iter):
gpu_photons = gpu.GPUPhotons(ev.photons_beg, ncopies=nreps)
gpu_photons.propagate(gpu_detector, rng_states,
nthreads_per_block, max_blocks)
for gpu_photon_slice in gpu_photons.iterate_copies():
for idaq in xrange(ndaq):
gpu_daq.begin_acquire()
gpu_daq.acquire(gpu_photon_slice, rng_states,
nthreads_per_block, max_blocks)
gpu_channels = gpu_daq.end_acquire()
gpu_pdf.add_hits_to_pdf(gpu_channels, nthreads_per_block)
hitcount, pdf = gpu_pdf.get_pdfs()
elapsed = time.time() - t0
if i > 0:
# first kernel call incurs some driver overhead
run_times.append(elapsed)
return nevents*nreps*ndaq/ufloat((np.mean(run_times),np.std(run_times)))
def __init__(self,
detector,
seed=None,
cuda_device=None,
particle_tracking=False,
photon_tracking=False,
geant4_processes=4,
nthreads_per_block=64,
max_blocks=1024):
self.detector = detector
self.nthreads_per_block = nthreads_per_block
self.max_blocks = max_blocks
self.photon_tracking = photon_tracking
if seed is None:
self.seed = pick_seed()
else:
self.seed = seed
# We have three generators to seed: numpy.random, GEANT4, and CURAND.
# The latter two are done below.
np.random.seed(self.seed)
if geant4_processes > 0:
self.photon_generator = generator.photon.G4ParallelGenerator(
geant4_processes,
detector.detector_material,
base_seed=self.seed,
tracking=particle_tracking)
else:
self.photon_generator = None
self.context = gpu.create_cuda_context(cuda_device)
if hasattr(detector, 'num_channels'):
self.gpu_geometry = gpu.GPUDetector(detector)
self.gpu_daq = gpu.GPUDaq(self.gpu_geometry)
self.gpu_pdf = gpu.GPUPDF()
self.gpu_pdf_kernel = gpu.GPUKernelPDF()
else:
self.gpu_geometry = gpu.GPUGeometry(detector)
self.rng_states = gpu.get_rng_states(self.nthreads_per_block *
self.max_blocks,
seed=self.seed)
self.pdf_config = None
def testGPUPDF(self):
'''Create a hit count and (q,t) PDF for 10 MeV events in MicroLBNE'''
g4generator = G4ParallelGenerator(1, water)
context = gpu.create_cuda_context()
gpu_geometry = gpu.GPUDetector(self.detector)
nthreads_per_block, max_blocks = 64, 1024
rng_states = gpu.get_rng_states(nthreads_per_block * max_blocks)
gpu_daq = gpu.GPUDaq(gpu_geometry)
gpu_pdf = gpu.GPUPDF()
gpu_pdf.setup_pdf(self.detector.num_channels(), 100, (-0.5, 999.5), 10,
(-0.5, 9.5))
gpu_pdf.clear_pdf()
for ev in g4generator.generate_events(
itertools.islice(self.vertex_gen, 10)):
gpu_photons = gpu.GPUPhotons(ev.photons_beg)
gpu_photons.propagate(gpu_geometry, rng_states, nthreads_per_block,
max_blocks)
gpu_daq.begin_acquire()
gpu_daq.acquire(gpu_photons, rng_states, nthreads_per_block,
max_blocks)
gpu_channels = gpu_daq.end_acquire()
gpu_pdf.add_hits_to_pdf(gpu_channels)
hitcount, pdf = gpu_pdf.get_pdfs()
self.assertTrue((hitcount > 0).any())
self.assertTrue((pdf > 0).any())
# Consistency checks
for i, nhits in enumerate(hitcount):
self.assertEqual(nhits, pdf[i].sum())
context.pop()
def pdf_eval(gpu_detector,
npdfs=10,
nevents=25,
nreps=16,
ndaq=128,
nthreads_per_block=64,
max_blocks=1024):
"""
Returns the average number of 100 MeV events that can be
histogrammed per second.
Args:
- gpu_instance, chroma.gpu.GPU
The GPU instance passed to the GPUDetector constructor.
- npdfs, int
The number of pdf generations to average.
- nevents, int
The number of 100 MeV events to generate for each PDF.
- nreps, int
The number of times to propagate each event and add to PDF
- ndaq, int
The number of times to run the DAQ simulation on the propagated
event and add it to the PDF.
"""
rng_states = gpu.get_rng_states(nthreads_per_block * max_blocks)
# Make data event
data_ev = next(
g4generator.generate_events(
itertools.islice(
generator.vertex.constant_particle_gun('e-', (0, 0, 0),
(1, 0, 0), 100), 1)))
gpu_photons = gpu.GPUPhotons(data_ev.photons_beg)
gpu_photons.propagate(gpu_detector, rng_states, nthreads_per_block,
max_blocks)
gpu_daq = gpu.GPUDaq(gpu_detector)
gpu_daq.begin_acquire()
gpu_daq.acquire(gpu_photons, rng_states, nthreads_per_block,
max_blocks).get()
data_ev_channels = gpu_daq.end_acquire()
# Setup PDF evaluation
gpu_daq = gpu.GPUDaq(gpu_detector, ndaq=ndaq)
gpu_pdf = gpu.GPUPDF()
gpu_pdf.setup_pdf_eval(data_ev_channels.hit,
data_ev_channels.t,
data_ev_channels.q,
0.05, (-0.5, 999.5),
1.0, (-0.5, 20),
min_bin_content=20,
time_only=True)
run_times = []
for i in tools.progress(list(range(npdfs))):
t0 = time.time()
gpu_pdf.clear_pdf_eval()
vertex_gen = generator.vertex.constant_particle_gun(
'e-', (0, 0, 0), (1, 0, 0), 100)
vertex_iter = itertools.islice(vertex_gen, nevents)
for ev in g4generator.generate_events(vertex_iter):
gpu_photons = gpu.GPUPhotons(ev.photons_beg, ncopies=nreps)
gpu_photons.propagate(gpu_detector, rng_states, nthreads_per_block,
max_blocks)
for gpu_photon_slice in gpu_photons.iterate_copies():
gpu_daq.begin_acquire()
gpu_photon_slice = gpu_photon_slice.select(
event.SURFACE_DETECT)
gpu_daq.acquire(gpu_photon_slice, rng_states,
nthreads_per_block, max_blocks)
gpu_channels = gpu_daq.end_acquire()
gpu_pdf.accumulate_pdf_eval(gpu_channels, nthreads_per_block)
cuda.Context.get_current().synchronize()
elapsed = time.time() - t0
if i > 0:
# first kernel call incurs some driver overhead
run_times.append(elapsed)
return nevents * nreps * ndaq / ufloat(
(np.mean(run_times), np.std(run_times)))