def _extract_photons_from_tracking_action(self, sort=True):
n = self.tracking_action.GetNumPhotons()
pos = np.zeros(shape=(n, 3), dtype=np.float32)
pos[:, 0] = self.tracking_action.GetX()
pos[:, 1] = self.tracking_action.GetY()
pos[:, 2] = self.tracking_action.GetZ()
dir = np.zeros(shape=(n, 3), dtype=np.float32)
dir[:, 0] = self.tracking_action.GetDirX()
dir[:, 1] = self.tracking_action.GetDirY()
dir[:, 2] = self.tracking_action.GetDirZ()
pol = np.zeros(shape=(n, 3), dtype=np.float32)
pol[:, 0] = self.tracking_action.GetPolX()
pol[:, 1] = self.tracking_action.GetPolY()
pol[:, 2] = self.tracking_action.GetPolZ()
wavelengths = self.tracking_action.GetWavelength().astype(np.float32)
t0 = self.tracking_action.GetT0().astype(np.float32)
if sort:
reorder = argsort_direction(dir)
pos = pos[reorder]
dir = dir[reorder]
pol = pol[reorder]
wavelengths = wavelengths[reorder]
t0 = t0[reorder]
return Photons(pos, dir, pol, wavelengths, t0)
def intersect(gpu_geometry,
number=100,
nphotons=500000,
nthreads_per_block=64,
max_blocks=1024):
"Returns the average number of ray intersections per second."
distances_gpu = ga.empty(nphotons, dtype=np.float32)
module = gpu.get_cu_module('mesh.h', options=('--use_fast_math', ))
gpu_funcs = gpu.GPUFuncs(module)
run_times = []
for i in tools.progress(list(range(number))):
pos = ga.zeros(nphotons, dtype=ga.vec.float3)
dir = sample.uniform_sphere(nphotons)
reorder = tools.argsort_direction(dir)
dir = ga.to_gpu(gpu.to_float3(dir[reorder]))
t0 = time.time()
gpu_funcs.distance_to_mesh(np.int32(pos.size),
pos,
dir,
gpu_geometry.gpudata,
distances_gpu,
block=(nthreads_per_block, 1, 1),
grid=(pos.size // nthreads_per_block + 1,
1))
cuda.Context.get_current().synchronize()
elapsed = time.time() - t0
if i > 0:
# first kernel call incurs some driver overhead
run_times.append(elapsed)
return nphotons / ufloat((np.mean(run_times), np.std(run_times)))
def propagate(gpu_detector,
number=10,
nphotons=500000,
nthreads_per_block=64,
max_blocks=1024):
"Returns the average number of photons propagated on the GPU per second."
rng_states = gpu.get_rng_states(nthreads_per_block * max_blocks)
run_times = []
for i in tools.progress(list(range(number))):
pos = np.zeros((nphotons, 3))
dir = sample.uniform_sphere(nphotons)
reorder = tools.argsort_direction(dir)
dir = dir[reorder]
pol = normalize(np.cross(sample.uniform_sphere(nphotons), dir))
wavelengths = np.random.uniform(400, 800, size=nphotons)
photons = event.Photons(pos, dir, pol, wavelengths)
gpu_photons = gpu.GPUPhotons(photons)
t0 = time.time()
gpu_photons.propagate(gpu_detector, rng_states, nthreads_per_block,
max_blocks)
cuda.Context.get_current().synchronize()
elapsed = time.time() - t0
if i > 0:
# first kernel call incurs some driver overhead
run_times.append(elapsed)
return nphotons / ufloat((np.mean(run_times), np.std(run_times)))
def _extract_photons_from_tracking_action(self, sort=True):
n = self.tracking_action.GetNumPhotons()
pos = np.zeros(shape=(n,3), dtype=np.float32)
pos[:,0] = self.tracking_action.GetX()
pos[:,1] = self.tracking_action.GetY()
pos[:,2] = self.tracking_action.GetZ()
dir = np.zeros(shape=(n,3), dtype=np.float32)
dir[:,0] = self.tracking_action.GetDirX()
dir[:,1] = self.tracking_action.GetDirY()
dir[:,2] = self.tracking_action.GetDirZ()
pol = np.zeros(shape=(n,3), dtype=np.float32)
pol[:,0] = self.tracking_action.GetPolX()
pol[:,1] = self.tracking_action.GetPolY()
pol[:,2] = self.tracking_action.GetPolZ()
wavelengths = self.tracking_action.GetWavelength().astype(np.float32)
t0 = self.tracking_action.GetT0().astype(np.float32)
if sort:
reorder = argsort_direction(dir)
pos = pos[reorder]
dir = dir[reorder]
pol = pol[reorder]
wavelengths = wavelengths[reorder]
t0 = t0[reorder]
return Photons(pos, dir, pol, wavelengths, t0)
