def merge_nodes_detailed(nodes, first_child, nchild):
'''Merges nodes into len(first_child) parent nodes, using
the provided arrays to determine the index of the first
child of each parent, and how many children there are.'''
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
gpu_nodes = ga.to_gpu(nodes)
gpu_first_child = ga.to_gpu(first_child.astype(np.int32))
gpu_nchild = ga.to_gpu(nchild.astype(np.int32))
nparent = len(first_child)
gpu_parent_nodes = ga.empty(shape=nparent, dtype=ga.vec.uint4)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(nparent, nthreads_per_block, max_blocks=10000):
bvh_funcs.make_parents_detailed(np.uint32(first_index),
np.uint32(elements_this_iter),
gpu_nodes,
gpu_parent_nodes,
gpu_first_child,
gpu_nchild,
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
return gpu_parent_nodes.get()
def __init__(self, photons, ncopies=1, max_time=4.):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
module = get_cu_module('propagate_hit.cu', options=cuda_options)
propagate_hit_kernel = module.get_function('propagate_hit')
propagate_hit_kernel.prepare('iiPPPPPPPPPPPiiiPPP')
self.propagate_hit_kernel = propagate_hit_kernel
self.gpu_funcs = GPUFuncs(module)
self.max_time = max_time
self.ncopies = ncopies
self.true_nphotons = len(photons)
self.marshall_photons(photons, ncopies)
def concatenate_layers(layers):
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
# Put 0 at beginning of list
layer_bounds = np.insert(np.cumsum(map(len, layers)), 0, 0)
nodes = ga.empty(shape=int(layer_bounds[-1]), dtype=ga.vec.uint4)
nthreads_per_block = 256
for layer_start, layer_end, layer in zip(layer_bounds[:-1],
layer_bounds[1:], layers):
if layer_end == layer_bounds[-1]:
# leaf nodes need no offset
child_offset = 0
else:
child_offset = layer_end
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(layer_end-layer_start, nthreads_per_block,
max_blocks=10000):
bvh_funcs.copy_and_offset(np.uint32(first_index),
np.uint32(elements_this_iter),
np.uint32(child_offset),
cuda.In(layer),
nodes[layer_start:],
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
return nodes.get(), layer_bounds
def merge_nodes_detailed(nodes, first_child, nchild):
'''Merges nodes into len(first_child) parent nodes, using
the provided arrays to determine the index of the first
child of each parent, and how many children there are.'''
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
gpu_nodes = ga.to_gpu(nodes)
gpu_first_child = ga.to_gpu(first_child.astype(np.int32))
gpu_nchild = ga.to_gpu(nchild.astype(np.int32))
nparent = len(first_child)
gpu_parent_nodes = ga.empty(shape=nparent, dtype=ga.vec.uint4)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(nparent, nthreads_per_block, max_blocks=10000):
bvh_funcs.make_parents_detailed(np.uint32(first_index),
np.uint32(elements_this_iter),
gpu_nodes,
gpu_parent_nodes,
gpu_first_child,
gpu_nchild,
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
return gpu_parent_nodes.get()
def concatenate_layers(layers):
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
# Put 0 at beginning of list
layer_bounds = np.insert(np.cumsum(map(len, layers)), 0, 0)
nodes = ga.empty(shape=int(layer_bounds[-1]), dtype=ga.vec.uint4)
nthreads_per_block = 256
for layer_start, layer_end, layer in zip(layer_bounds[:-1],
layer_bounds[1:],
layers):
if layer_end == layer_bounds[-1]:
# leaf nodes need no offset
child_offset = 0
else:
child_offset = layer_end
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(layer_end-layer_start, nthreads_per_block,
max_blocks=10000):
bvh_funcs.copy_and_offset(np.uint32(first_index),
np.uint32(elements_this_iter),
np.uint32(child_offset),
cuda.In(layer),
nodes[layer_start:],
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
return nodes.get(), layer_bounds
class GPURays(object):
"""The GPURays class holds arrays of ray positions and directions
on the GPU that are used to render a geometry."""
def __init__(self, pos, dir, max_alpha_depth=10, nblocks=64):
self.pos = ga.to_gpu(to_float3(pos))
self.dir = ga.to_gpu(to_float3(dir))
self.max_alpha_depth = max_alpha_depth
self.nblocks = nblocks
transform_module = get_cu_module('transform.cu', options=cuda_options)
self.transform_funcs = GPUFuncs(transform_module)
render_module = get_cu_module('render.cu', options=cuda_options)
self.render_funcs = GPUFuncs(render_module)
self.dx = ga.empty(max_alpha_depth*self.pos.size, dtype=np.float32)
self.color = ga.empty(self.dx.size, dtype=ga.vec.float4)
self.dxlen = ga.zeros(self.pos.size, dtype=np.uint32)
def rotate(self, phi, n):
"Rotate by an angle phi around the axis `n`."
self.transform_funcs.rotate(np.int32(self.pos.size), self.pos, np.float32(phi), ga.vec.make_float3(*n), block=(self.nblocks,1,1), grid=(self.pos.size//self.nblocks+1,1))
self.transform_funcs.rotate(np.int32(self.dir.size), self.dir, np.float32(phi), ga.vec.make_float3(*n), block=(self.nblocks,1,1), grid=(self.dir.size//self.nblocks+1,1))
def rotate_around_point(self, phi, n, point):
""""Rotate by an angle phi around the axis `n` passing through
the point `point`."""
self.transform_funcs.rotate_around_point(np.int32(self.pos.size), self.pos, np.float32(phi), ga.vec.make_float3(*n), ga.vec.make_float3(*point), block=(self.nblocks,1,1), grid=(self.pos.size//self.nblocks+1,1))
self.transform_funcs.rotate(np.int32(self.dir.size), self.dir, np.float32(phi), ga.vec.make_float3(*n), block=(self.nblocks,1,1), grid=(self.dir.size//self.nblocks+1,1))
def translate(self, v):
"Translate the ray positions by the vector `v`."
self.transform_funcs.translate(np.int32(self.pos.size), self.pos, ga.vec.make_float3(*v), block=(self.nblocks,1,1), grid=(self.pos.size//self.nblocks+1,1))
def render(self, gpu_geometry, pixels, alpha_depth=10,
keep_last_render=False):
"""Render `gpu_geometry` and fill the GPU array `pixels` with pixel
colors."""
if not keep_last_render:
self.dxlen.fill(0)
if alpha_depth > self.max_alpha_depth:
raise Exception('alpha_depth > max_alpha_depth')
if not isinstance(pixels, ga.GPUArray):
raise TypeError('`pixels` must be a %s instance.' % ga.GPUArray)
if pixels.size != self.pos.size:
raise ValueError('`pixels`.size != number of rays')
self.render_funcs.render(np.int32(self.pos.size), self.pos, self.dir, gpu_geometry.gpudata, np.uint32(alpha_depth), pixels, self.dx, self.dxlen, self.color, block=(self.nblocks,1,1), grid=(self.pos.size//self.nblocks+1,1))
def snapshot(self, gpu_geometry, alpha_depth=10):
"Render `gpu_geometry` and return a numpy array of pixel colors."
pixels = ga.empty(self.pos.size, dtype=np.uint32)
self.render(gpu_geometry, pixels, alpha_depth)
return pixels.get()
def __init__(self, gpu_detector, ndaq=1):
self.earliest_time_gpu = ga.empty(gpu_detector.nchannels*ndaq, dtype=np.float32)
self.earliest_time_int_gpu = ga.empty(gpu_detector.nchannels*ndaq, dtype=np.uint32)
self.channel_history_gpu = ga.zeros_like(self.earliest_time_int_gpu)
self.channel_q_int_gpu = ga.zeros_like(self.earliest_time_int_gpu)
self.channel_q_gpu = ga.zeros(len(self.earliest_time_int_gpu), dtype=np.float32)
self.detector_gpu = gpu_detector.detector_gpu
self.solid_id_map_gpu = gpu_detector.solid_id_map
self.solid_id_to_channel_index_gpu = gpu_detector.solid_id_to_channel_index_gpu
self.module = get_cu_module('daq.cu', options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
self.ndaq = ndaq
self.stride = gpu_detector.nchannels
def area_sort_nodes(gpu_geometry, layer_bounds):
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
bounds = zip(layer_bounds[:-1], layer_bounds[1:])[:-1]
bounds.reverse()
nthreads_per_block = 256
for start, end in bounds:
bvh_funcs.area_sort_child(np.uint32(start),
np.uint32(end),
gpu_geometry,
block=(nthreads_per_block,1,1),
grid=(120,1))
return gpu_geometry.nodes.get()
def area_sort_nodes(gpu_geometry, layer_bounds):
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
bounds = zip(layer_bounds[:-1], layer_bounds[1:])[:-1]
bounds.reverse()
nthreads_per_block = 256
for start, end in bounds:
bvh_funcs.area_sort_child(np.uint32(start),
np.uint32(end),
gpu_geometry,
block=(nthreads_per_block, 1, 1),
grid=(120, 1))
return gpu_geometry.nodes.get()
def __init__(self, pos, dir, max_alpha_depth=10, nblocks=64):
self.pos = ga.to_gpu(to_float3(pos))
self.dir = ga.to_gpu(to_float3(dir))
self.max_alpha_depth = max_alpha_depth
self.nblocks = nblocks
transform_module = get_cu_module('transform.cu', options=cuda_options)
self.transform_funcs = GPUFuncs(transform_module)
render_module = get_cu_module('render.cu', options=cuda_options)
self.render_funcs = GPUFuncs(render_module)
self.dx = ga.empty(max_alpha_depth * self.pos.size, dtype=np.float32)
self.color = ga.empty(self.dx.size, dtype=ga.vec.float4)
self.dxlen = ga.zeros(self.pos.size, dtype=np.uint32)
def __init__(self, pos, dir, max_alpha_depth=10, nblocks=64):
self.pos = ga.to_gpu(to_float3(pos))
self.dir = ga.to_gpu(to_float3(dir))
self.max_alpha_depth = max_alpha_depth
self.nblocks = nblocks
transform_module = get_cu_module('transform.cu', options=cuda_options)
self.transform_funcs = GPUFuncs(transform_module)
render_module = get_cu_module('render.cu', options=cuda_options)
self.render_funcs = GPUFuncs(render_module)
self.dx = ga.empty(max_alpha_depth*self.pos.size, dtype=np.float32)
self.color = ga.empty(self.dx.size, dtype=ga.vec.float4)
self.dxlen = ga.zeros(self.pos.size, dtype=np.uint32)
def collapse_chains(nodes, layer_bounds):
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
gpu_nodes = ga.to_gpu(nodes)
bounds = zip(layer_bounds[:-1], layer_bounds[1:])[:-1]
bounds.reverse()
nthreads_per_block = 256
for start, end in bounds:
bvh_funcs.collapse_child(np.uint32(start),
np.uint32(end),
gpu_nodes,
block=(nthreads_per_block,1,1),
grid=(120,1))
return gpu_nodes.get()
def collapse_chains(nodes, layer_bounds):
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
gpu_nodes = ga.to_gpu(nodes)
bounds = zip(layer_bounds[:-1], layer_bounds[1:])[:-1]
bounds.reverse()
nthreads_per_block = 256
for start, end in bounds:
bvh_funcs.collapse_child(np.uint32(start),
np.uint32(end),
gpu_nodes,
block=(nthreads_per_block, 1, 1),
grid=(120, 1))
return gpu_nodes.get()
def __init__(self, photons, ncopies=1):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
nphotons = len(photons)
self.pos = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.dir = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.pol = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.wavelengths = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
self.t = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
self.last_hit_triangles = ga.empty(shape=nphotons*ncopies, dtype=np.int32)
self.flags = ga.empty(shape=nphotons*ncopies, dtype=np.uint32)
self.weights = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
# Assign the provided photons to the beginning (possibly
# the entire array if ncopies is 1
self.pos[:nphotons].set(to_float3(photons.pos))
self.dir[:nphotons].set(to_float3(photons.dir))
self.pol[:nphotons].set(to_float3(photons.pol))
self.wavelengths[:nphotons].set(photons.wavelengths.astype(np.float32))
self.t[:nphotons].set(photons.t.astype(np.float32))
self.last_hit_triangles[:nphotons].set(photons.last_hit_triangles.astype(np.int32))
self.flags[:nphotons].set(photons.flags.astype(np.uint32))
self.weights[:nphotons].set(photons.weights.astype(np.float32))
module = get_cu_module('propagate.cu', options=cuda_options)
self.gpu_funcs = GPUFuncs(module)
# Replicate the photons to the rest of the slots if needed
if ncopies > 1:
max_blocks = 1024
nthreads_per_block = 64
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.photon_duplicate(np.int32(first_photon), np.int32(photons_this_round),
self.pos, self.dir, self.wavelengths, self.pol, self.t,
self.flags, self.last_hit_triangles, self.weights,
np.int32(ncopies-1),
np.int32(nphotons),
block=(nthreads_per_block,1,1), grid=(blocks, 1))
# Save the duplication information for the iterate_copies() method
self.true_nphotons = nphotons
self.ncopies = ncopies
def __init__(self, pos, dir, pol, wavelengths, t, last_hit_triangles,
flags, weights):
'''Create new object using slices of GPUArrays from an instance
of GPUPhotons. NOTE THESE ARE NOT CPU ARRAYS!'''
self.pos = pos
self.dir = dir
self.pol = pol
self.wavelengths = wavelengths
self.t = t
self.last_hit_triangles = last_hit_triangles
self.flags = flags
self.weights = weights
module = get_cu_module('propagate.cu', options=cuda_options)
self.gpu_funcs = GPUFuncs(module)
self.true_nphotons = len(pos)
self.ncopies = 1
class GPUPDF(object):
def __init__(self):
self.module = get_cu_module('pdf.cu', options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
def setup_pdf(self, nchannels, tbins, trange, qbins, qrange):
"""Setup GPU arrays to hold PDF information.
nchannels: int, number of channels
tbins: number of time bins
trange: tuple of (min, max) time in PDF
qbins: number of charge bins
qrange: tuple of (min, max) charge in PDF
"""
self.events_in_histogram = 0
self.hitcount_gpu = ga.zeros(nchannels, dtype=np.uint32)
self.pdf_gpu = ga.zeros(shape=(nchannels, tbins, qbins),
dtype=np.uint32)
self.tbins = tbins
self.trange = trange
self.qbins = qbins
self.qrange = qrange
def clear_pdf(self):
"""Rezero the PDF counters."""
self.hitcount_gpu.fill(0)
self.pdf_gpu.fill(0)
def add_hits_to_pdf(self, gpuchannels, nthreads_per_block=64):
self.gpu_funcs.bin_hits(np.int32(len(self.hitcount_gpu)),
gpuchannels.q,
gpuchannels.t,
self.hitcount_gpu,
np.int32(self.tbins),
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.int32(self.qbins),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.pdf_gpu,
block=(nthreads_per_block,1,1),
grid=(len(gpuchannels.t)//nthreads_per_block+1,1))
self.events_in_histogram += 1
def get_pdfs(self):
"""Returns the 1D hitcount array and the 3D [channel, time, charge]
histogram."""
return self.hitcount_gpu.get(), self.pdf_gpu.get()
def setup_pdf_eval(self, event_hit, event_time, event_charge, min_twidth,
trange, min_qwidth, qrange, min_bin_content=10,
time_only=True):
"""Setup GPU arrays to compute PDF values for the given event.
The pdf_eval calculation allows the PDF to be evaluated at a
single point for each channel as the Monte Carlo is run. The
effective bin size will be as small as (`min_twidth`,
`min_qwidth`) around the point of interest, but will be large
enough to ensure that `min_bin_content` Monte Carlo events
fall into the bin.
event_hit: ndarray
Hit or not-hit status for each channel in the detector.
event_time: ndarray
Hit time for each channel in the detector. If channel
not hit, the time will be ignored.
event_charge: ndarray
Integrated charge for each channel in the detector.
If channel not hit, the charge will be ignored.
min_twidth: float
Minimum bin size in the time dimension
trange: (float, float)
Range of time dimension in PDF
min_qwidth: float
Minimum bin size in charge dimension
qrange: (float, float)
Range of charge dimension in PDF
min_bin_content: int
The bin will be expanded to include at least this many events
time_only: bool
If True, only the time observable will be used in the PDF.
"""
self.event_nhit = count_nonzero(event_hit)
# Define a mapping from an array of len(event_hit) to an array of length event_nhit
self.map_hit_offset_to_channel_id = np.where(event_hit)[0].astype(np.uint32)
self.map_hit_offset_to_channel_id_gpu = ga.to_gpu(self.map_hit_offset_to_channel_id)
self.map_channel_id_to_hit_offset = np.maximum(0, event_hit.cumsum() - 1).astype(np.uint32)
self.map_channel_id_to_hit_offset_gpu = ga.to_gpu(self.map_channel_id_to_hit_offset)
self.event_hit_gpu = ga.to_gpu(event_hit.astype(np.uint32))
self.event_time_gpu = ga.to_gpu(event_time.astype(np.float32))
self.event_charge_gpu = ga.to_gpu(event_charge.astype(np.float32))
self.eval_hitcount_gpu = ga.zeros(len(event_hit), dtype=np.uint32)
self.eval_bincount_gpu = ga.zeros(len(event_hit), dtype=np.uint32)
self.nearest_mc_gpu = ga.empty(shape=self.event_nhit * min_bin_content,
dtype=np.float32)
self.nearest_mc_gpu.fill(1e9)
self.min_twidth = min_twidth
self.trange = trange
self.min_qwidth = min_qwidth
self.qrange = qrange
self.min_bin_content = min_bin_content
assert time_only # Only support time right now
self.time_only = time_only
def clear_pdf_eval(self):
"Reset PDF evaluation counters to start accumulating new Monte Carlo."
self.eval_hitcount_gpu.fill(0)
self.eval_bincount_gpu.fill(0)
self.nearest_mc_gpu.fill(1e9)
@profile_if_possible
def accumulate_pdf_eval(self, gpuchannels, nthreads_per_block=64, max_blocks=10000):
"Add the most recent results of run_daq() to the PDF evaluation."
self.work_queues = ga.empty(shape=self.event_nhit * (gpuchannels.ndaq+1), dtype=np.uint32)
self.work_queues.fill(1)
self.gpu_funcs.accumulate_bincount(np.int32(self.event_hit_gpu.size),
np.int32(gpuchannels.ndaq),
self.event_hit_gpu,
self.event_time_gpu,
gpuchannels.t,
self.eval_hitcount_gpu,
self.eval_bincount_gpu,
np.float32(self.min_twidth),
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.int32(self.min_bin_content),
self.map_channel_id_to_hit_offset_gpu,
self.work_queues,
block=(nthreads_per_block,1,1),
grid=(self.event_hit_gpu.size//nthreads_per_block+1,1))
cuda.Context.get_current().synchronize()
self.gpu_funcs.accumulate_nearest_neighbor_block(np.int32(self.event_nhit),
np.int32(gpuchannels.ndaq),
self.map_hit_offset_to_channel_id_gpu,
self.work_queues,
self.event_time_gpu,
gpuchannels.t,
self.nearest_mc_gpu,
np.int32(self.min_bin_content),
block=(nthreads_per_block,1,1),
grid=(self.event_nhit,1))
cuda.Context.get_current().synchronize()
def get_pdf_eval(self):
evhit = self.event_hit_gpu.get().astype(bool)
hitcount = self.eval_hitcount_gpu.get()
bincount = self.eval_bincount_gpu.get()
pdf_value = np.zeros(len(hitcount), dtype=float)
pdf_frac_uncert = np.zeros_like(pdf_value)
# PDF value for high stats bins
high_stats = (bincount >= self.min_bin_content)
if high_stats.any():
if self.time_only:
pdf_value[high_stats] = bincount[high_stats].astype(float) / hitcount[high_stats] / self.min_twidth
else:
assert Exception('Unimplemented 2D (time,charge) mode!')
pdf_frac_uncert[high_stats] = 1.0/np.sqrt(bincount[high_stats])
# PDF value for low stats bins
low_stats = ~high_stats & (hitcount > 0) & evhit
nearest_mc_by_hit = self.nearest_mc_gpu.get().reshape((self.event_nhit, self.min_bin_content))
nearest_mc = np.empty(shape=(len(hitcount), self.min_bin_content), dtype=np.float32)
nearest_mc.fill(1e9)
nearest_mc[self.map_hit_offset_to_channel_id,:] = nearest_mc_by_hit
# Deal with the case where we did not even get min_bin_content events
# in the PDF but also clamp the lower range to ensure we don't index
# by a negative number in 2 lines
last_valid_entry = np.maximum(0, (nearest_mc < 1e9).astype(int).sum(axis=1) - 1)
distance = nearest_mc[np.arange(len(last_valid_entry)),last_valid_entry]
if low_stats.any():
if self.time_only:
pdf_value[low_stats] = (last_valid_entry[low_stats] + 1).astype(float) / hitcount[low_stats] / distance[low_stats] / 2.0
else:
assert Exception('Unimplemented 2D (time,charge) mode!')
pdf_frac_uncert[low_stats] = 1.0/np.sqrt(last_valid_entry[low_stats] + 1)
# PDFs with no stats got zero by default during array creation
print 'high_stats:', high_stats.sum(), 'low_stats', low_stats.sum()
return hitcount, pdf_value, pdf_value * pdf_frac_uncert
class GPUDaq(object):
def __init__(self, gpu_detector, ndaq=1):
self.earliest_time_gpu = ga.empty(gpu_detector.nchannels*ndaq, dtype=np.float32)
self.earliest_time_int_gpu = ga.empty(gpu_detector.nchannels*ndaq, dtype=np.uint32)
self.channel_history_gpu = ga.zeros_like(self.earliest_time_int_gpu)
self.channel_q_int_gpu = ga.zeros_like(self.earliest_time_int_gpu)
self.channel_q_gpu = ga.zeros(len(self.earliest_time_int_gpu), dtype=np.float32)
self.detector_gpu = gpu_detector.detector_gpu
self.solid_id_map_gpu = gpu_detector.solid_id_map
self.solid_id_to_channel_index_gpu = gpu_detector.solid_id_to_channel_index_gpu
self.module = get_cu_module('daq.cu', options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
self.ndaq = ndaq
self.stride = gpu_detector.nchannels
def begin_acquire(self, nthreads_per_block=64):
self.gpu_funcs.reset_earliest_time_int(np.float32(1e9), np.int32(len(self.earliest_time_int_gpu)), self.earliest_time_int_gpu, block=(nthreads_per_block,1,1), grid=(len(self.earliest_time_int_gpu)//nthreads_per_block+1,1))
self.channel_q_int_gpu.fill(0)
self.channel_q_gpu.fill(0)
self.channel_history_gpu.fill(0)
def acquire(self, gpuphotons, rng_states, nthreads_per_block=64, max_blocks=1024, start_photon=None, nphotons=None, weight=1.0):
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = len(gpuphotons.pos) - start_photon
if self.ndaq == 1:
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.run_daq(rng_states, np.uint32(0x1 << 2),
np.int32(start_photon+first_photon), np.int32(photons_this_round), gpuphotons.t,
gpuphotons.flags, gpuphotons.last_hit_triangles, gpuphotons.weights,
self.solid_id_map_gpu,
self.detector_gpu,
self.earliest_time_int_gpu,
self.channel_q_int_gpu, self.channel_history_gpu,
np.float32(weight),
block=(nthreads_per_block,1,1), grid=(blocks,1))
else:
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, 1, max_blocks):
self.gpu_funcs.run_daq_many(rng_states, np.uint32(0x1 << 2),
np.int32(start_photon+first_photon), np.int32(photons_this_round), gpuphotons.t,
gpuphotons.flags, gpuphotons.last_hit_triangles, gpuphotons.weights,
self.solid_id_map_gpu,
self.detector_gpu,
self.earliest_time_int_gpu,
self.channel_q_int_gpu, self.channel_history_gpu,
np.int32(self.ndaq), np.int32(self.stride),
np.float32(weight),
block=(nthreads_per_block,1,1), grid=(blocks,1))
cuda.Context.get_current().synchronize()
def end_acquire(self, nthreads_per_block=64):
self.gpu_funcs.convert_sortable_int_to_float(np.int32(len(self.earliest_time_int_gpu)), self.earliest_time_int_gpu, self.earliest_time_gpu, block=(nthreads_per_block,1,1), grid=(len(self.earliest_time_int_gpu)//nthreads_per_block+1,1))
self.gpu_funcs.convert_charge_int_to_float(self.detector_gpu, self.channel_q_int_gpu, self.channel_q_gpu, block=(nthreads_per_block,1,1), grid=(len(self.channel_q_int_gpu)//nthreads_per_block+1,1))
cuda.Context.get_current().synchronize()
return GPUChannels(self.earliest_time_gpu, self.channel_q_gpu, self.channel_history_gpu, self.ndaq, self.stride)
class GPURays(object):
"""The GPURays class holds arrays of ray positions and directions
on the GPU that are used to render a geometry."""
def __init__(self, pos, dir, max_alpha_depth=10, nblocks=64):
self.pos = ga.to_gpu(to_float3(pos))
self.dir = ga.to_gpu(to_float3(dir))
self.max_alpha_depth = max_alpha_depth
self.nblocks = nblocks
transform_module = get_cu_module('transform.cu', options=cuda_options)
self.transform_funcs = GPUFuncs(transform_module)
render_module = get_cu_module('render.cu', options=cuda_options)
self.render_funcs = GPUFuncs(render_module)
self.dx = ga.empty(max_alpha_depth * self.pos.size, dtype=np.float32)
self.color = ga.empty(self.dx.size, dtype=ga.vec.float4)
self.dxlen = ga.zeros(self.pos.size, dtype=np.uint32)
def rotate(self, phi, n):
"Rotate by an angle phi around the axis `n`."
self.transform_funcs.rotate(np.int32(self.pos.size),
self.pos,
np.float32(phi),
ga.vec.make_float3(*n),
block=(self.nblocks, 1, 1),
grid=(self.pos.size // self.nblocks + 1,
1))
self.transform_funcs.rotate(np.int32(self.dir.size),
self.dir,
np.float32(phi),
ga.vec.make_float3(*n),
block=(self.nblocks, 1, 1),
grid=(self.dir.size // self.nblocks + 1,
1))
def rotate_around_point(self, phi, n, point):
""""Rotate by an angle phi around the axis `n` passing through
the point `point`."""
self.transform_funcs.rotate_around_point(
np.int32(self.pos.size),
self.pos,
np.float32(phi),
ga.vec.make_float3(*n),
ga.vec.make_float3(*point),
block=(self.nblocks, 1, 1),
grid=(self.pos.size // self.nblocks + 1, 1))
self.transform_funcs.rotate(np.int32(self.dir.size),
self.dir,
np.float32(phi),
ga.vec.make_float3(*n),
block=(self.nblocks, 1, 1),
grid=(self.dir.size // self.nblocks + 1,
1))
def translate(self, v):
"Translate the ray positions by the vector `v`."
self.transform_funcs.translate(np.int32(self.pos.size),
self.pos,
ga.vec.make_float3(*v),
block=(self.nblocks, 1, 1),
grid=(self.pos.size // self.nblocks + 1,
1))
def render(self,
gpu_geometry,
pixels,
alpha_depth=10,
keep_last_render=False):
"""Render `gpu_geometry` and fill the GPU array `pixels` with pixel
colors."""
if not keep_last_render:
self.dxlen.fill(0)
if alpha_depth > self.max_alpha_depth:
raise Exception('alpha_depth > max_alpha_depth')
if not isinstance(pixels, ga.GPUArray):
raise TypeError('`pixels` must be a %s instance.' % ga.GPUArray)
if pixels.size != self.pos.size:
raise ValueError('`pixels`.size != number of rays')
self.render_funcs.render(np.int32(self.pos.size),
self.pos,
self.dir,
gpu_geometry.gpudata,
np.uint32(alpha_depth),
pixels,
self.dx,
self.dxlen,
self.color,
block=(self.nblocks, 1, 1),
grid=(self.pos.size // self.nblocks + 1, 1))
def snapshot(self, gpu_geometry, alpha_depth=10):
"Render `gpu_geometry` and return a numpy array of pixel colors."
pixels = ga.empty(self.pos.size, dtype=np.uint32)
self.render(gpu_geometry, pixels, alpha_depth)
return pixels.get()
def create_leaf_nodes(mesh, morton_bits=16, round_to_multiple=1):
'''Compute the leaf nodes surrounding a triangle mesh.
``mesh``: chroma.geometry.Mesh
Triangles to box
``morton_bits``: int
Number of bits to use per dimension when computing Morton code.
``round_to_multiple``: int
Round the number of nodes created up to multiple of this number
Extra nodes will be all zero.
Returns (world_coords, nodes, morton_codes), where
``world_coords``: chroma.bvh.WorldCoords
Defines the fixed point coordinate system
``nodes``: ndarray(shape=len(mesh.triangles), dtype=uint4)
List of leaf nodes. Child IDs will be set to triangle offsets.
``morton_codes``: ndarray(shape=len(mesh.triangles), dtype=np.uint64)
Morton codes for each triangle, using ``morton_bits`` per axis.
Must be <= 16 bits.
'''
# Load GPU functions
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
# compute world coordinates
world_origin = mesh.vertices.min(axis=0)
world_scale = np.max((mesh.vertices.max(axis=0) - world_origin)) \
/ (2**16 - 2)
world_coords = WorldCoords(world_origin=world_origin,
world_scale=world_scale)
# Put triangles and vertices in mapped host memory
triangles = mapped_empty(shape=len(mesh.triangles),
dtype=ga.vec.uint3,
write_combined=True)
triangles[:] = to_uint3(mesh.triangles)
vertices = mapped_empty(shape=len(mesh.vertices),
dtype=ga.vec.float3,
write_combined=True)
vertices[:] = to_float3(mesh.vertices)
# Call GPU to compute nodes
nodes = ga.zeros(shape=round_up_to_multiple(len(triangles),
round_to_multiple),
dtype=ga.vec.uint4)
morton_codes = ga.empty(shape=len(triangles), dtype=np.uint64)
# Convert world coords to GPU-friendly types
world_origin = ga.vec.make_float3(*world_origin)
world_scale = np.float32(world_scale)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(len(triangles), nthreads_per_block,
max_blocks=30000):
bvh_funcs.make_leaves(np.uint32(first_index),
np.uint32(elements_this_iter),
Mapped(triangles),
Mapped(vertices),
world_origin,
world_scale,
nodes,
morton_codes,
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
morton_codes_host = morton_codes.get() >> (16 - morton_bits)
return world_coords, nodes.get(), morton_codes_host
def optimize_layer(orig_nodes):
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
nodes = ga.to_gpu(orig_nodes)
n = len(nodes)
areas = ga.empty(shape=n / 2, dtype=np.uint64)
nthreads_per_block = 128
min_areas = ga.empty(shape=int(np.ceil(n / float(nthreads_per_block))),
dtype=np.uint64)
min_index = ga.empty(shape=min_areas.shape, dtype=np.uint32)
update = 10000
skip_size = 1
flag = mapped_empty(shape=skip_size, dtype=np.uint32)
i = 0
skips = 0
swaps = 0
while i < n / 2 - 1:
# How are we doing?
if i % update == 0:
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(n/2, nthreads_per_block, max_blocks=10000):
bvh_funcs.pair_area(np.uint32(first_index),
np.uint32(elements_this_iter),
nodes,
areas,
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
areas_host = areas.get()
#print nodes.get(), areas_host.astype(float)
print 'Area of parent layer so far (%d): %1.12e' % (
i * 2, areas_host.astype(float).sum())
print 'Skips: %d, Swaps: %d' % (skips, swaps)
test_index = i * 2
blocks = 0
look_forward = min(8192 * 50, n - test_index - 2)
skip_this_round = min(skip_size, n - test_index - 1)
flag[:] = 0
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(look_forward, nthreads_per_block, max_blocks=10000):
bvh_funcs.min_distance_to(np.uint32(first_index + test_index + 2),
np.uint32(elements_this_iter),
np.uint32(test_index),
nodes,
np.uint32(blocks),
min_areas,
min_index,
Mapped(flag),
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter,
skip_this_round))
blocks += nblocks_this_iter
#print i, first_index, nblocks_this_iter, look_forward
cuda.Context.get_current().synchronize()
if flag[0] == 0:
flag_nonzero = flag.nonzero()[0]
if len(flag_nonzero) == 0:
no_swap_required = skip_size
else:
no_swap_required = flag_nonzero[0]
i += no_swap_required
skips += no_swap_required
continue
min_areas_host = min_areas[:blocks].get()
min_index_host = min_index[:blocks].get()
best_block = min_areas_host.argmin()
better_i = min_index_host[best_block]
swaps += 1
#print 'swap', test_index+1, better_i
assert 0 < better_i < len(nodes)
assert 0 < test_index + 1 < len(nodes)
bvh_funcs.swap(np.uint32(test_index + 1),
np.uint32(better_i),
nodes,
block=(1, 1, 1),
grid=(1, 1))
cuda.Context.get_current().synchronize()
i += 1
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(n/2, nthreads_per_block, max_blocks=10000):
bvh_funcs.pair_area(np.uint32(first_index),
np.uint32(elements_this_iter),
nodes,
areas,
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
areas_host = areas.get()
print 'Final area of parent layer: %1.12e' % areas_host.sum()
print 'Skips: %d, Swaps: %d' % (skips, swaps)
return nodes.get()
class GPUKernelPDF(object):
def __init__(self):
self.module = get_cu_module('pdf.cu', options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
def setup_moments(self, nchannels, trange, qrange, time_only=True):
"""Setup GPU arrays to accumulate moments and eventually
compute a kernel estimate of PDF values for each hit channel.
trange: (float, float)
Range of time dimension in PDF
qrange: (float, float)
Range of charge dimension in PDF
time_only: bool
If True, only the time observable will be used in the PDF.
"""
self.hitcount_gpu = ga.zeros(nchannels, dtype=np.uint32)
self.tmom1_gpu = ga.zeros(nchannels, dtype=np.float32)
self.tmom2_gpu = ga.zeros(nchannels, dtype=np.float32)
self.qmom1_gpu = ga.zeros(nchannels, dtype=np.float32)
self.qmom2_gpu = ga.zeros(nchannels, dtype=np.float32)
self.trange = trange
self.qrange = qrange
self.time_only = time_only
def clear_moments(self):
"Reset PDF evaluation counters to start accumulating new Monte Carlo."
self.hitcount_gpu.fill(0)
self.tmom1_gpu.fill(0.0)
self.tmom2_gpu.fill(0.0)
self.qmom1_gpu.fill(0.0)
self.qmom2_gpu.fill(0.0)
def accumulate_moments(self, gpuchannels, nthreads_per_block=64):
"""Add the most recent results of run_daq() to the accumulate of
moments for future bandwidth calculation."""
self.gpu_funcs.accumulate_moments(np.int32(self.time_only),
np.int32(len(gpuchannels.t)),
gpuchannels.t,
gpuchannels.q,
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.hitcount_gpu,
self.tmom1_gpu,
self.tmom2_gpu,
self.qmom1_gpu,
self.qmom2_gpu,
block=(nthreads_per_block,1,1),
grid=(len(gpuchannels.t)//nthreads_per_block+1,1))
def compute_bandwidth(self, event_hit, event_time, event_charge,
scale_factor=1.0):
"""Use the MC information accumulated by accumulate_moments() to
estimate the best bandwidth to use when kernel estimating."""
rho = 1.0
hitcount = self.hitcount_gpu.get()
mom0 = np.maximum(hitcount, 1)
tmom1 = self.tmom1_gpu.get()
tmom2 = self.tmom2_gpu.get()
tmean = tmom1 / mom0
tvar = np.maximum(tmom2 / mom0 - tmean**2, 0.0) # roundoff can go neg
trms = tvar**0.5
if self.time_only:
d = 1
else:
d = 2
dimensionality_factor = ((4.0/(d+2)) / (mom0/scale_factor))**(-1.0/(d+4))
gaussian_density = np.minimum(1.0/trms, (1.0/np.sqrt(2.0*np.pi)) * np.exp(-0.5*((event_time - tmean)/trms)) / trms)
time_bandwidths = dimensionality_factor / gaussian_density * rho
inv_time_bandwidths = np.zeros_like(time_bandwidths)
inv_time_bandwidths[time_bandwidths > 0] = time_bandwidths[time_bandwidths > 0] ** -1
# precompute inverse to speed up GPU evaluation
self.inv_time_bandwidths_gpu = ga.to_gpu(
inv_time_bandwidths.astype(np.float32)
)
# Compute charge bandwidths if needed
if self.time_only:
self.inv_charge_bandwidths_gpu = ga.empty_like(
self.inv_time_bandwidths_gpu
)
self.inv_charge_bandwidths_gpu.fill(0.0)
else:
qmom1 = self.qmom1_gpu.get()
qmom2 = self.qmom2_gpu.get()
qmean = qmom1 / mom0
qrms = (qmom2 / mom0 - qmean**2)**0.5
gaussian_density = np.minimum(1.0/qrms, (1.0/np.sqrt(2.0*np.pi)) * np.exp(-0.5*((event_charge - qmean)/qrms)) / qrms)
charge_bandwidths = dimensionality_factor / gaussian_density * rho
# precompute inverse to speed up GPU evaluation
self.inv_charge_bandwidths_gpu = ga.to_gpu(
(charge_bandwidths**-1).astype(np.float32)
)
def setup_kernel(self, event_hit, event_time, event_charge):
"""Setup GPU arrays to accumulate moments and eventually
compute a kernel estimate of PDF values for each hit channel.
event_hit: ndarray
Hit or not-hit status for each channel in the detector.
event_time: ndarray
Hit time for each channel in the detector. If channel
not hit, the time will be ignored.
event_charge: ndarray
Integrated charge for each channel in the detector.
If channel not hit, the charge will be ignored.
"""
self.event_hit_gpu = ga.to_gpu(event_hit.astype(np.uint32))
self.event_time_gpu = ga.to_gpu(event_time.astype(np.float32))
self.event_charge_gpu = ga.to_gpu(event_charge.astype(np.float32))
self.hitcount_gpu.fill(0)
self.time_pdf_values_gpu = ga.zeros(len(event_hit), dtype=np.float32)
self.charge_pdf_values_gpu = ga.zeros(len(event_hit), dtype=np.float32)
def clear_kernel(self):
self.hitcount_gpu.fill(0)
self.time_pdf_values_gpu.fill(0.0)
self.charge_pdf_values_gpu.fill(0.0)
def accumulate_kernel(self, gpuchannels, nthreads_per_block=64):
"Add the most recent results of run_daq() to the kernel PDF evaluation."
self.gpu_funcs.accumulate_kernel_eval(np.int32(self.time_only),
np.int32(len(self.event_hit_gpu)),
self.event_hit_gpu,
self.event_time_gpu,
self.event_charge_gpu,
gpuchannels.t,
gpuchannels.q,
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.inv_time_bandwidths_gpu,
self.inv_charge_bandwidths_gpu,
self.hitcount_gpu,
self.time_pdf_values_gpu,
self.charge_pdf_values_gpu,
block=(nthreads_per_block,1,1),
grid=(len(gpuchannels.t)//nthreads_per_block+1,1))
def get_kernel_eval(self):
hitcount = self.hitcount_gpu.get()
hit = self.event_hit_gpu.get().astype(bool)
time_pdf_values = self.time_pdf_values_gpu.get()
time_pdf_values /= np.maximum(1, hitcount) # avoid divide by zero
charge_pdf_values = self.charge_pdf_values_gpu.get()
charge_pdf_values /= np.maximum(1, hitcount) # avoid divide by zero
if self.time_only:
pdf_values = time_pdf_values
else:
pdf_values = time_pdf_values * charge_pdf_values
return hitcount, pdf_values, np.zeros_like(pdf_values)
def __init__(self):
self.module = get_cu_module('pdf.cu', options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
def __init__(self,
photons,
ncopies=1,
copy_flags=True,
copy_triangles=True,
copy_weights=True):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
nphotons = len(photons)
self.pos = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.dir = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.pol = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.wavelengths = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.t = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.last_hit_triangles = ga.empty(shape=nphotons * ncopies,
dtype=np.int32)
if not copy_triangles:
self.last_hit_triangles.fill(-1)
if not copy_flags:
self.flags = ga.zeros(shape=nphotons * ncopies, dtype=np.uint32)
else:
self.flags = ga.empty(shape=nphotons * ncopies, dtype=np.uint32)
if not copy_weights:
self.weights = ga.ones_like(self.last_hit_triangles,
dtype=np.float32)
else:
self.weights = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.evidx = ga.empty(shape=nphotons, dtype=np.uint32)
# Assign the provided photons to the beginning (possibly
# the entire array if ncopies is 1
self.pos[:nphotons].set(to_float3(photons.pos))
self.dir[:nphotons].set(to_float3(photons.dir))
self.pol[:nphotons].set(to_float3(photons.pol))
self.wavelengths[:nphotons].set(photons.wavelengths.astype(np.float32))
self.t[:nphotons].set(photons.t.astype(np.float32))
if copy_triangles:
self.last_hit_triangles[:nphotons].set(
photons.last_hit_triangles.astype(np.int32))
if copy_flags:
self.flags[:nphotons].set(photons.flags.astype(np.uint32))
if copy_weights:
self.weights[:nphotons].set(photons.weights.astype(np.float32))
self.evidx[:nphotons].set(photons.evidx.astype(np.uint32))
module = get_cu_module('propagate.cu', options=cuda_options)
self.gpu_funcs = GPUFuncs(module)
# Replicate the photons to the rest of the slots if needed
if ncopies > 1:
max_blocks = 1024
nthreads_per_block = 64
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.photon_duplicate(np.int32(first_photon),
np.int32(photons_this_round),
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
np.int32(ncopies - 1),
np.int32(nphotons),
block=(nthreads_per_block, 1,
1),
grid=(blocks, 1))
# Save the duplication information for the iterate_copies() method
self.true_nphotons = nphotons
self.ncopies = ncopies
class GPUKernelPDF(object):
def __init__(self):
self.module = get_cu_module('pdf.cu',
options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
def setup_moments(self, nchannels, trange, qrange, time_only=True):
"""Setup GPU arrays to accumulate moments and eventually
compute a kernel estimate of PDF values for each hit channel.
trange: (float, float)
Range of time dimension in PDF
qrange: (float, float)
Range of charge dimension in PDF
time_only: bool
If True, only the time observable will be used in the PDF.
"""
self.hitcount_gpu = ga.zeros(nchannels, dtype=np.uint32)
self.tmom1_gpu = ga.zeros(nchannels, dtype=np.float32)
self.tmom2_gpu = ga.zeros(nchannels, dtype=np.float32)
self.qmom1_gpu = ga.zeros(nchannels, dtype=np.float32)
self.qmom2_gpu = ga.zeros(nchannels, dtype=np.float32)
self.trange = trange
self.qrange = qrange
self.time_only = time_only
def clear_moments(self):
"Reset PDF evaluation counters to start accumulating new Monte Carlo."
self.hitcount_gpu.fill(0)
self.tmom1_gpu.fill(0.0)
self.tmom2_gpu.fill(0.0)
self.qmom1_gpu.fill(0.0)
self.qmom2_gpu.fill(0.0)
def accumulate_moments(self, gpuchannels, nthreads_per_block=64):
"""Add the most recent results of run_daq() to the accumulate of
moments for future bandwidth calculation."""
self.gpu_funcs.accumulate_moments(
np.int32(self.time_only),
np.int32(len(gpuchannels.t)),
gpuchannels.t,
gpuchannels.q,
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.hitcount_gpu,
self.tmom1_gpu,
self.tmom2_gpu,
self.qmom1_gpu,
self.qmom2_gpu,
block=(nthreads_per_block, 1, 1),
grid=(len(gpuchannels.t) // nthreads_per_block + 1, 1))
def compute_bandwidth(self,
event_hit,
event_time,
event_charge,
scale_factor=1.0):
"""Use the MC information accumulated by accumulate_moments() to
estimate the best bandwidth to use when kernel estimating."""
rho = 1.0
hitcount = self.hitcount_gpu.get()
mom0 = np.maximum(hitcount, 1)
tmom1 = self.tmom1_gpu.get()
tmom2 = self.tmom2_gpu.get()
tmean = tmom1 / mom0
tvar = np.maximum(tmom2 / mom0 - tmean**2, 0.0) # roundoff can go neg
trms = tvar**0.5
if self.time_only:
d = 1
else:
d = 2
dimensionality_factor = ((4.0 / (d + 2)) /
(mom0 / scale_factor))**(-1.0 / (d + 4))
gaussian_density = np.minimum(
1.0 / trms, (1.0 / np.sqrt(2.0 * np.pi)) *
np.exp(-0.5 * ((event_time - tmean) / trms)) / trms)
time_bandwidths = dimensionality_factor / gaussian_density * rho
inv_time_bandwidths = np.zeros_like(time_bandwidths)
inv_time_bandwidths[time_bandwidths > 0] = time_bandwidths[
time_bandwidths > 0]**-1
# precompute inverse to speed up GPU evaluation
self.inv_time_bandwidths_gpu = ga.to_gpu(
inv_time_bandwidths.astype(np.float32))
# Compute charge bandwidths if needed
if self.time_only:
self.inv_charge_bandwidths_gpu = ga.empty_like(
self.inv_time_bandwidths_gpu)
self.inv_charge_bandwidths_gpu.fill(0.0)
else:
qmom1 = self.qmom1_gpu.get()
qmom2 = self.qmom2_gpu.get()
qmean = qmom1 / mom0
qrms = (qmom2 / mom0 - qmean**2)**0.5
gaussian_density = np.minimum(
1.0 / qrms, (1.0 / np.sqrt(2.0 * np.pi)) *
np.exp(-0.5 * ((event_charge - qmean) / qrms)) / qrms)
charge_bandwidths = dimensionality_factor / gaussian_density * rho
# precompute inverse to speed up GPU evaluation
self.inv_charge_bandwidths_gpu = ga.to_gpu(
(charge_bandwidths**-1).astype(np.float32))
def setup_kernel(self, event_hit, event_time, event_charge):
"""Setup GPU arrays to accumulate moments and eventually
compute a kernel estimate of PDF values for each hit channel.
event_hit: ndarray
Hit or not-hit status for each channel in the detector.
event_time: ndarray
Hit time for each channel in the detector. If channel
not hit, the time will be ignored.
event_charge: ndarray
Integrated charge for each channel in the detector.
If channel not hit, the charge will be ignored.
"""
self.event_hit_gpu = ga.to_gpu(event_hit.astype(np.uint32))
self.event_time_gpu = ga.to_gpu(event_time.astype(np.float32))
self.event_charge_gpu = ga.to_gpu(event_charge.astype(np.float32))
self.hitcount_gpu.fill(0)
self.time_pdf_values_gpu = ga.zeros(len(event_hit), dtype=np.float32)
self.charge_pdf_values_gpu = ga.zeros(len(event_hit), dtype=np.float32)
def clear_kernel(self):
self.hitcount_gpu.fill(0)
self.time_pdf_values_gpu.fill(0.0)
self.charge_pdf_values_gpu.fill(0.0)
def accumulate_kernel(self, gpuchannels, nthreads_per_block=64):
"Add the most recent results of run_daq() to the kernel PDF evaluation."
self.gpu_funcs.accumulate_kernel_eval(
np.int32(self.time_only),
np.int32(len(self.event_hit_gpu)),
self.event_hit_gpu,
self.event_time_gpu,
self.event_charge_gpu,
gpuchannels.t,
gpuchannels.q,
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.inv_time_bandwidths_gpu,
self.inv_charge_bandwidths_gpu,
self.hitcount_gpu,
self.time_pdf_values_gpu,
self.charge_pdf_values_gpu,
block=(nthreads_per_block, 1, 1),
grid=(len(gpuchannels.t) // nthreads_per_block + 1, 1))
def get_kernel_eval(self):
hitcount = self.hitcount_gpu.get()
hit = self.event_hit_gpu.get().astype(bool)
time_pdf_values = self.time_pdf_values_gpu.get()
time_pdf_values /= np.maximum(1, hitcount) # avoid divide by zero
charge_pdf_values = self.charge_pdf_values_gpu.get()
charge_pdf_values /= np.maximum(1, hitcount) # avoid divide by zero
if self.time_only:
pdf_values = time_pdf_values
else:
pdf_values = time_pdf_values * charge_pdf_values
return hitcount, pdf_values, np.zeros_like(pdf_values)
def merge_nodes(nodes, degree, max_ratio=None):
bvh_module = get_cu_module('bvh.cu',
options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
nparent = len(nodes) / degree
if len(nodes) % degree != 0:
nparent += 1
if nparent == 1:
nparent_pad = nparent
else:
nparent_pad = round_up_to_multiple(nparent, 1) #degree)
gpu_parent_nodes = ga.zeros(shape=nparent_pad, dtype=ga.vec.uint4)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(nparent, nthreads_per_block, max_blocks=10000):
bvh_funcs.make_parents(np.uint32(first_index),
np.uint32(elements_this_iter),
np.uint32(degree),
gpu_parent_nodes,
cuda.In(nodes),
np.uint32(0),
np.uint32(len(nodes)),
block=(nthreads_per_block, 1, 1),
grid=(nblocks_this_iter, 1))
parent_nodes = gpu_parent_nodes.get()
if max_ratio is not None:
areas = node_areas(parent_nodes)
child_areas = node_areas(nodes)
excessive_area = np.zeros(shape=len(areas), dtype=bool)
for i, parent_area in enumerate(areas):
nchild = parent_nodes['w'][i] >> CHILD_BITS
child_index = parent_nodes['w'][i] & ~NCHILD_MASK
child_area = child_areas[child_index:child_index + nchild].sum()
#if parent_area > 1e9:
# print i, 'Children: %e, Parent: %e' % (child_area, parent_area)
if child_area / parent_area < 0.3:
excessive_area[i] = True
#print i, 'Children: %e, Parent: %e' % (child_area, parent_area)
extra_slots = round_up_to_multiple(
(degree - 1) * np.count_nonzero(excessive_area), 1)
print 'Extra slots:', extra_slots
new_parent_nodes = np.zeros(shape=len(parent_nodes) + extra_slots,
dtype=parent_nodes.dtype)
new_parent_nodes[:len(parent_nodes)] = parent_nodes
offset = 0
for count, index in enumerate(np.argwhere(excessive_area)):
index = index[0] + offset
nchild = new_parent_nodes['w'][index] >> CHILD_BITS
child_index = new_parent_nodes['w'][index] & ~NCHILD_MASK
new_parent_nodes[index] = nodes[child_index]
#new_parent_nodes['w'][index] = 1 << CHILD_BITS | child_index
tmp_nchild = new_parent_nodes['w'][index] >> CHILD_BITS
tmp_child_index = new_parent_nodes['w'][index] & ~NCHILD_MASK
new_parent_nodes['w'][index] = tmp_nchild << CHILD_BITS | (
tmp_child_index + len(nodes))
if nchild == 1:
continue
# slide everyone over
#print index, nchild, len(new_parent_nodes)
new_parent_nodes[index + nchild:] = new_parent_nodes[index +
1:-nchild + 1]
offset += nchild - 1
for sibling in xrange(nchild - 1):
new_parent_index = index + 1 + sibling
new_parent_nodes[new_parent_index] = nodes[child_index +
sibling + 1]
if new_parent_nodes['x'][new_parent_index] != 0:
tmp_nchild = new_parent_nodes['w'][
new_parent_index] >> CHILD_BITS
tmp_child_index = new_parent_nodes['w'][
new_parent_index] & ~NCHILD_MASK
new_parent_nodes['w'][
new_parent_index] = tmp_nchild << CHILD_BITS | (
tmp_child_index + len(nodes))
#new_parent_nodes['w'][new_parent_index] = 1 << CHILD_BITS | (child_index + sibling + 1)
#print 'intermediate: %e' % node_areas(new_parent_nodes).max()
print 'old: %e' % node_areas(parent_nodes).max()
print 'new: %e' % node_areas(new_parent_nodes).max()
if len(new_parent_nodes) < len(nodes):
# Only adopt new set of parent nodes if it actually reduces the
# total number of nodes at this level by 1.
parent_nodes = new_parent_nodes
return parent_nodes
class GPUPhotons(object):
def __init__(self, photons, ncopies=1):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
nphotons = len(photons)
self.pos = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.dir = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.pol = ga.empty(shape=nphotons*ncopies, dtype=ga.vec.float3)
self.wavelengths = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
self.t = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
self.last_hit_triangles = ga.empty(shape=nphotons*ncopies, dtype=np.int32)
self.flags = ga.empty(shape=nphotons*ncopies, dtype=np.uint32)
self.weights = ga.empty(shape=nphotons*ncopies, dtype=np.float32)
# Assign the provided photons to the beginning (possibly
# the entire array if ncopies is 1
self.pos[:nphotons].set(to_float3(photons.pos))
self.dir[:nphotons].set(to_float3(photons.dir))
self.pol[:nphotons].set(to_float3(photons.pol))
self.wavelengths[:nphotons].set(photons.wavelengths.astype(np.float32))
self.t[:nphotons].set(photons.t.astype(np.float32))
self.last_hit_triangles[:nphotons].set(photons.last_hit_triangles.astype(np.int32))
self.flags[:nphotons].set(photons.flags.astype(np.uint32))
self.weights[:nphotons].set(photons.weights.astype(np.float32))
module = get_cu_module('propagate.cu', options=cuda_options)
self.gpu_funcs = GPUFuncs(module)
# Replicate the photons to the rest of the slots if needed
if ncopies > 1:
max_blocks = 1024
nthreads_per_block = 64
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.photon_duplicate(np.int32(first_photon), np.int32(photons_this_round),
self.pos, self.dir, self.wavelengths, self.pol, self.t,
self.flags, self.last_hit_triangles, self.weights,
np.int32(ncopies-1),
np.int32(nphotons),
block=(nthreads_per_block,1,1), grid=(blocks, 1))
# Save the duplication information for the iterate_copies() method
self.true_nphotons = nphotons
self.ncopies = ncopies
def get(self):
pos = self.pos.get().view(np.float32).reshape((len(self.pos),3))
dir = self.dir.get().view(np.float32).reshape((len(self.dir),3))
pol = self.pol.get().view(np.float32).reshape((len(self.pol),3))
wavelengths = self.wavelengths.get()
t = self.t.get()
last_hit_triangles = self.last_hit_triangles.get()
flags = self.flags.get()
weights = self.weights.get()
return event.Photons(pos, dir, pol, wavelengths, t, last_hit_triangles, flags, weights)
def get_hits(self, gpu_detector, target_flag=(0x1<<2), nthreads_per_block=64, max_blocks=1024,
start_photon=None, nphotons=None):
'''Return a map of GPUPhoton objects containing only photons that
have a particular bit set in their history word and were detected by
a channel.'''
cuda.Context.get_current().synchronize()
index_counter_gpu = ga.zeros(shape=1, dtype=np.uint32)
cuda.Context.get_current().synchronize()
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = self.pos.size - start_photon
# First count how much space we need
for first_photon, photons_this_round, blocks in chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.count_photon_hits(np.int32(start_photon+first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
self.flags,
gpu_detector.solid_id_map,
self.last_hit_triangles,
gpu_detector.detector_gpu,
index_counter_gpu,
block=(nthreads_per_block,1,1),
grid=(blocks, 1))
cuda.Context.get_current().synchronize()
reduced_nphotons = int(index_counter_gpu.get()[0])
# Then allocate new storage space
pos = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=reduced_nphotons, dtype=np.float32)
t = ga.empty(shape=reduced_nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=reduced_nphotons, dtype=np.int32)
flags = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
weights = ga.empty(shape=reduced_nphotons, dtype=np.float32)
channels = ga.empty(shape=reduced_nphotons, dtype=np.int32)
# And finaly copy hits, if there are any
if reduced_nphotons > 0:
index_counter_gpu.fill(0)
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photon_hits(np.int32(start_photon+first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
gpu_detector.solid_id_map,
gpu_detector.detector_gpu,
index_counter_gpu,
self.pos, self.dir, self.wavelengths, self.pol, self.t, self.flags, self.last_hit_triangles, self.weights,
pos, dir, wavelengths, pol, t, flags, last_hit_triangles, weights, channels,
block=(nthreads_per_block,1,1),
grid=(blocks, 1))
assert index_counter_gpu.get()[0] == reduced_nphotons
pos = pos.get().view(np.float32).reshape((len(pos),3))
dir = dir.get().view(np.float32).reshape((len(dir),3))
pol = pol.get().view(np.float32).reshape((len(pol),3))
wavelengths = wavelengths.get()
t = t.get()
last_hit_triangles = last_hit_triangles.get()
flags = flags.get()
weights = weights.get()
channels = channels.get()
hitmap = {}
for chan in np.unique(channels):
mask = (channels == chan).astype(bool)
hitmap[chan] = event.Photons(pos[mask], dir[mask], pol[mask], wavelengths[mask], t[mask], last_hit_triangles[mask], flags[mask], weights[mask])
return hitmap
def iterate_copies(self):
'''Returns an iterator that yields GPUPhotonsSlice objects
corresponding to the event copies stored in ``self``.'''
for i in xrange(self.ncopies):
window = slice(self.true_nphotons*i, self.true_nphotons*(i+1))
yield GPUPhotonsSlice(pos=self.pos[window],
dir=self.dir[window],
pol=self.pol[window],
wavelengths=self.wavelengths[window],
t=self.t[window],
last_hit_triangles=self.last_hit_triangles[window],
flags=self.flags[window],
weights=self.weights[window])
@profile_if_possible
def propagate(self, gpu_geometry, rng_states, nthreads_per_block=64,
max_blocks=1024, max_steps=10, use_weights=False,
scatter_first=0):
"""Propagate photons on GPU to termination or max_steps, whichever
comes first.
May be called repeatedly without reloading photon information if
single-stepping through photon history.
..warning::
`rng_states` must have at least `nthreads_per_block`*`max_blocks`
number of curandStates.
"""
nphotons = self.pos.size
step = 0
input_queue = np.empty(shape=nphotons+1, dtype=np.uint32)
input_queue[0] = 0
# Order photons initially in the queue to put the clones next to each other
for copy in xrange(self.ncopies):
input_queue[1+copy::self.ncopies] = np.arange(self.true_nphotons, dtype=np.uint32) + copy * self.true_nphotons
input_queue_gpu = ga.to_gpu(input_queue)
output_queue = np.zeros(shape=nphotons+1, dtype=np.uint32)
output_queue[0] = 1
output_queue_gpu = ga.to_gpu(output_queue)
while step < max_steps:
# Just finish the rest of the steps if the # of photons is low
if nphotons < nthreads_per_block * 16 * 8 or use_weights:
nsteps = max_steps - step
else:
nsteps = 1
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.propagate(np.int32(first_photon), np.int32(photons_this_round), input_queue_gpu[1:], output_queue_gpu, rng_states, self.pos, self.dir, self.wavelengths, self.pol, self.t, self.flags, self.last_hit_triangles, self.weights, np.int32(nsteps), np.int32(use_weights), np.int32(scatter_first), gpu_geometry.gpudata, block=(nthreads_per_block,1,1), grid=(blocks, 1))
step += nsteps
scatter_first = 0 # Only allow non-zero in first pass
if step < max_steps:
temp = input_queue_gpu
input_queue_gpu = output_queue_gpu
output_queue_gpu = temp
# Assign with a numpy array of length 1 to silence
# warning from PyCUDA about setting array with different strides/storage orders.
output_queue_gpu[:1].set(np.ones(shape=1, dtype=np.uint32))
nphotons = input_queue_gpu[:1].get()[0] - 1
if ga.max(self.flags).get() & (1 << 31):
print >>sys.stderr, "WARNING: ABORTED PHOTONS"
cuda.Context.get_current().synchronize()
@profile_if_possible
def select(self, target_flag, nthreads_per_block=64, max_blocks=1024,
start_photon=None, nphotons=None):
'''Return a new GPUPhoton object containing only photons that
have a particular bit set in their history word.'''
cuda.Context.get_current().synchronize()
index_counter_gpu = ga.zeros(shape=1, dtype=np.uint32)
cuda.Context.get_current().synchronize()
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = self.pos.size - start_photon
# First count how much space we need
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.count_photons(np.int32(start_photon+first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu, self.flags,
block=(nthreads_per_block,1,1),
grid=(blocks, 1))
cuda.Context.get_current().synchronize()
reduced_nphotons = int(index_counter_gpu.get()[0])
# Then allocate new storage space
pos = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=reduced_nphotons, dtype=np.float32)
t = ga.empty(shape=reduced_nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=reduced_nphotons, dtype=np.int32)
flags = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
weights = ga.empty(shape=reduced_nphotons, dtype=np.float32)
# And finaly copy photons, if there are any
if reduced_nphotons > 0:
index_counter_gpu.fill(0)
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photons(np.int32(start_photon+first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu,
self.pos, self.dir, self.wavelengths, self.pol, self.t, self.flags, self.last_hit_triangles, self.weights,
pos, dir, wavelengths, pol, t, flags, last_hit_triangles, weights,
block=(nthreads_per_block,1,1),
grid=(blocks, 1))
assert index_counter_gpu.get()[0] == reduced_nphotons
return GPUPhotonsSlice(pos, dir, pol, wavelengths, t, last_hit_triangles, flags, weights)
def __del__(self):
del self.pos
del self.dir
del self.pol
del self.wavelengths
del self.t
del self.flags
del self.last_hit_triangles
# Free up GPU memory quickly if now available
gc.collect()
def __len__(self):
return self.pos.size
def merge_nodes(nodes, degree, max_ratio=None):
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
nparent = len(nodes) / degree
if len(nodes) % degree != 0:
nparent += 1
if nparent == 1:
nparent_pad = nparent
else:
nparent_pad = round_up_to_multiple(nparent, 1)#degree)
gpu_parent_nodes = ga.zeros(shape=nparent_pad, dtype=ga.vec.uint4)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(nparent, nthreads_per_block, max_blocks=10000):
bvh_funcs.make_parents(np.uint32(first_index),
np.uint32(elements_this_iter),
np.uint32(degree),
gpu_parent_nodes,
cuda.In(nodes),
np.uint32(0),
np.uint32(len(nodes)),
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
parent_nodes = gpu_parent_nodes.get()
if max_ratio is not None:
areas = node_areas(parent_nodes)
child_areas = node_areas(nodes)
excessive_area = np.zeros(shape=len(areas), dtype=bool)
for i, parent_area in enumerate(areas):
nchild = parent_nodes['w'][i] >> CHILD_BITS
child_index = parent_nodes['w'][i] & ~NCHILD_MASK
child_area = child_areas[child_index:child_index+nchild].sum()
#if parent_area > 1e9:
# print i, 'Children: %e, Parent: %e' % (child_area, parent_area)
if child_area/parent_area < 0.3:
excessive_area[i] = True
#print i, 'Children: %e, Parent: %e' % (child_area, parent_area)
extra_slots = round_up_to_multiple((degree - 1) * np.count_nonzero(excessive_area), 1)
print 'Extra slots:', extra_slots
new_parent_nodes = np.zeros(shape=len(parent_nodes) + extra_slots,
dtype=parent_nodes.dtype)
new_parent_nodes[:len(parent_nodes)] = parent_nodes
offset = 0
for count, index in enumerate(np.argwhere(excessive_area)):
index = index[0] + offset
nchild = new_parent_nodes['w'][index] >> CHILD_BITS
child_index = new_parent_nodes['w'][index] & ~NCHILD_MASK
new_parent_nodes[index] = nodes[child_index]
#new_parent_nodes['w'][index] = 1 << CHILD_BITS | child_index
tmp_nchild = new_parent_nodes['w'][index] >> CHILD_BITS
tmp_child_index = new_parent_nodes['w'][index] & ~NCHILD_MASK
new_parent_nodes['w'][index] = tmp_nchild << CHILD_BITS | (tmp_child_index + len(nodes))
if nchild == 1:
continue
# slide everyone over
#print index, nchild, len(new_parent_nodes)
new_parent_nodes[index+nchild:] = new_parent_nodes[index+1:-nchild+1]
offset += nchild - 1
for sibling in xrange(nchild - 1):
new_parent_index = index + 1 + sibling
new_parent_nodes[new_parent_index] = nodes[child_index + sibling + 1]
if new_parent_nodes['x'][new_parent_index] != 0:
tmp_nchild = new_parent_nodes['w'][new_parent_index] >> CHILD_BITS
tmp_child_index = new_parent_nodes['w'][new_parent_index] & ~NCHILD_MASK
new_parent_nodes['w'][new_parent_index] = tmp_nchild << CHILD_BITS | (tmp_child_index + len(nodes))
#new_parent_nodes['w'][new_parent_index] = 1 << CHILD_BITS | (child_index + sibling + 1)
#print 'intermediate: %e' % node_areas(new_parent_nodes).max()
print 'old: %e' % node_areas(parent_nodes).max()
print 'new: %e' % node_areas(new_parent_nodes).max()
if len(new_parent_nodes) < len(nodes):
# Only adopt new set of parent nodes if it actually reduces the
# total number of nodes at this level by 1.
parent_nodes = new_parent_nodes
return parent_nodes
class GPUPhotons(object):
def __init__(self,
photons,
ncopies=1,
copy_flags=True,
copy_triangles=True,
copy_weights=True):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
nphotons = len(photons)
self.pos = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.dir = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.pol = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.wavelengths = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.t = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.last_hit_triangles = ga.empty(shape=nphotons * ncopies,
dtype=np.int32)
if not copy_triangles:
self.last_hit_triangles.fill(-1)
if not copy_flags:
self.flags = ga.zeros(shape=nphotons * ncopies, dtype=np.uint32)
else:
self.flags = ga.empty(shape=nphotons * ncopies, dtype=np.uint32)
if not copy_weights:
self.weights = ga.ones_like(self.last_hit_triangles,
dtype=np.float32)
else:
self.weights = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.evidx = ga.empty(shape=nphotons, dtype=np.uint32)
# Assign the provided photons to the beginning (possibly
# the entire array if ncopies is 1
self.pos[:nphotons].set(to_float3(photons.pos))
self.dir[:nphotons].set(to_float3(photons.dir))
self.pol[:nphotons].set(to_float3(photons.pol))
self.wavelengths[:nphotons].set(photons.wavelengths.astype(np.float32))
self.t[:nphotons].set(photons.t.astype(np.float32))
if copy_triangles:
self.last_hit_triangles[:nphotons].set(
photons.last_hit_triangles.astype(np.int32))
if copy_flags:
self.flags[:nphotons].set(photons.flags.astype(np.uint32))
if copy_weights:
self.weights[:nphotons].set(photons.weights.astype(np.float32))
self.evidx[:nphotons].set(photons.evidx.astype(np.uint32))
module = get_cu_module('propagate.cu', options=cuda_options)
self.gpu_funcs = GPUFuncs(module)
# Replicate the photons to the rest of the slots if needed
if ncopies > 1:
max_blocks = 1024
nthreads_per_block = 64
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.photon_duplicate(np.int32(first_photon),
np.int32(photons_this_round),
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
np.int32(ncopies - 1),
np.int32(nphotons),
block=(nthreads_per_block, 1,
1),
grid=(blocks, 1))
# Save the duplication information for the iterate_copies() method
self.true_nphotons = nphotons
self.ncopies = ncopies
def get(self):
pos = self.pos.get().view(np.float32).reshape((len(self.pos), 3))
dir = self.dir.get().view(np.float32).reshape((len(self.dir), 3))
pol = self.pol.get().view(np.float32).reshape((len(self.pol), 3))
wavelengths = self.wavelengths.get()
t = self.t.get()
last_hit_triangles = self.last_hit_triangles.get()
flags = self.flags.get()
weights = self.weights.get()
evidx = self.evidx.get()
return event.Photons(pos, dir, pol, wavelengths, t, last_hit_triangles,
flags, weights, evidx)
def get_hits(self, *args, **kwargs):
'''Return a map of GPUPhoton objects containing only photons that
have a particular bit set in their history word and were detected by
a channel.'''
flat_hits = self.get_flat_hits(*args, **kwargs)
hitmap = {}
for chan in np.unique(flat_hits.channel):
mask = (flat_hits.channel == chan).astype(bool)
hitmap[int(chan)] = flat_hits[mask]
return hitmap
def get_flat_hits(self,
gpu_detector,
target_flag=(0x1 << 2),
nthreads_per_block=64,
max_blocks=1024,
start_photon=None,
nphotons=None,
no_map=False):
'''GPUPhoton objects containing only photons that
have a particular bit set in their history word and were detected by
a channel.'''
cuda.Context.get_current().synchronize()
index_counter_gpu = ga.zeros(shape=1, dtype=np.uint32)
cuda.Context.get_current().synchronize()
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = self.pos.size - start_photon
# First count how much space we need
for first_photon, photons_this_round, blocks in chunk_iterator(
nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.count_photon_hits(np.int32(start_photon +
first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
self.flags,
gpu_detector.solid_id_map,
self.last_hit_triangles,
gpu_detector.detector_gpu,
index_counter_gpu,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
cuda.Context.get_current().synchronize()
reduced_nphotons = int(index_counter_gpu.get()[0])
# Then allocate new storage space
pos = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=reduced_nphotons, dtype=np.float32)
t = ga.empty(shape=reduced_nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=reduced_nphotons, dtype=np.int32)
flags = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
weights = ga.empty(shape=reduced_nphotons, dtype=np.float32)
evidx = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
channels = ga.empty(shape=reduced_nphotons, dtype=np.int32)
# And finaly copy hits, if there are any
if reduced_nphotons > 0:
index_counter_gpu.fill(0)
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photon_hits(
np.int32(start_photon + first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
gpu_detector.solid_id_map,
gpu_detector.detector_gpu,
index_counter_gpu,
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
pos,
dir,
wavelengths,
pol,
t,
flags,
last_hit_triangles,
weights,
evidx,
channels,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
assert index_counter_gpu.get()[0] == reduced_nphotons
pos = pos.get().view(np.float32).reshape((len(pos), 3))
dir = dir.get().view(np.float32).reshape((len(dir), 3))
pol = pol.get().view(np.float32).reshape((len(pol), 3))
wavelengths = wavelengths.get()
t = t.get()
last_hit_triangles = last_hit_triangles.get()
flags = flags.get()
weights = weights.get()
evidx = evidx.get()
channels = channels.get()
hitmap = {}
return event.Photons(pos, dir, pol, wavelengths, t, last_hit_triangles,
flags, weights, evidx, channels)
def iterate_copies(self):
'''Returns an iterator that yields GPUPhotonsSlice objects
corresponding to the event copies stored in ``self``.'''
for i in range(self.ncopies):
window = slice(self.true_nphotons * i,
self.true_nphotons * (i + 1))
yield GPUPhotonsSlice(
pos=self.pos[window],
dir=self.dir[window],
pol=self.pol[window],
wavelengths=self.wavelengths[window],
t=self.t[window],
last_hit_triangles=self.last_hit_triangles[window],
flags=self.flags[window],
weights=self.weights[window],
evidx=self.evidx[window])
@profile_if_possible
def propagate(self,
gpu_geometry,
rng_states,
nthreads_per_block=64,
max_blocks=1024,
max_steps=10,
use_weights=False,
scatter_first=0,
track=False):
"""Propagate photons on GPU to termination or max_steps, whichever
comes first.
May be called repeatedly without reloading photon information if
single-stepping through photon history.
..warning::
`rng_states` must have at least `nthreads_per_block`*`max_blocks`
number of curandStates.
"""
nphotons = self.pos.size
step = 0
input_queue = np.empty(shape=nphotons + 1, dtype=np.uint32)
input_queue[0] = 0
# Order photons initially in the queue to put the clones next to each other
for copy in range(self.ncopies):
input_queue[1 + copy::self.ncopies] = np.arange(
self.true_nphotons,
dtype=np.uint32) + copy * self.true_nphotons
input_queue_gpu = ga.to_gpu(input_queue)
output_queue = np.zeros(shape=nphotons + 1, dtype=np.uint32)
output_queue[0] = 1
output_queue_gpu = ga.to_gpu(output_queue)
if track:
step_photon_ids = []
step_photons = []
#save the first step for all photons in the input queue
step_photon_ids.append(input_queue_gpu[1:nphotons + 1].get())
step_photons.append(
self.copy_queue(input_queue_gpu[1:], nphotons).get())
while step < max_steps:
# Just finish the rest of the steps if the # of photons is low and not tracking
if not track and (nphotons < nthreads_per_block * 16 * 8
or use_weights):
nsteps = max_steps - step
else:
nsteps = 1
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.propagate(np.int32(first_photon),
np.int32(photons_this_round),
input_queue_gpu[1:],
output_queue_gpu,
rng_states,
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
np.int32(nsteps),
np.int32(use_weights),
np.int32(scatter_first),
gpu_geometry.gpudata,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
if track: #save the next step for all photons in the input queue
step_photon_ids.append(input_queue_gpu[1:nphotons + 1].get())
step_photons.append(
self.copy_queue(input_queue_gpu[1:], nphotons).get())
step += nsteps
scatter_first = 0 # Only allow non-zero in first pass
if step < max_steps:
temp = input_queue_gpu
input_queue_gpu = output_queue_gpu
output_queue_gpu = temp
# Assign with a numpy array of length 1 to silence
# warning from PyCUDA about setting array with different strides/storage orders.
output_queue_gpu[:1].set(np.ones(shape=1, dtype=np.uint32))
nphotons = input_queue_gpu[:1].get()[0] - 1
if nphotons == 0:
break
if ga.max(self.flags).get() & (1 << 31):
print("WARNING: ABORTED PHOTONS", file=sys.stderr)
cuda.Context.get_current().synchronize()
if track:
return step_photon_ids, step_photons
@profile_if_possible
def copy_queue(self,
queue_gpu,
nphotons,
nthreads_per_block=64,
max_blocks=1024,
start_photon=0):
# Allocate new storage space
pos = ga.empty(shape=nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=nphotons, dtype=np.float32)
t = ga.empty(shape=nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=nphotons, dtype=np.int32)
flags = ga.empty(shape=nphotons, dtype=np.uint32)
weights = ga.empty(shape=nphotons, dtype=np.float32)
evidx = ga.empty(shape=nphotons, dtype=np.uint32)
# And finaly copy photons, if there are any
if nphotons > 0:
for first_photon, photons_this_round, blocks in chunk_iterator(
nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photon_queue(
np.int32(start_photon + first_photon),
np.int32(photons_this_round),
queue_gpu,
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
pos,
dir,
wavelengths,
pol,
t,
flags,
last_hit_triangles,
weights,
evidx,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
return GPUPhotonsSlice(pos, dir, pol, wavelengths, t,
last_hit_triangles, flags, weights, evidx)
@profile_if_possible
def select(self,
target_flag,
nthreads_per_block=64,
max_blocks=1024,
start_photon=None,
nphotons=None):
'''Return a new GPUPhoton object containing only photons that
have a particular bit set in their history word.'''
cuda.Context.get_current().synchronize()
index_counter_gpu = ga.zeros(shape=1, dtype=np.uint32)
cuda.Context.get_current().synchronize()
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = self.pos.size - start_photon
# First count how much space we need
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.count_photons(np.int32(start_photon + first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu,
self.flags,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
cuda.Context.get_current().synchronize()
reduced_nphotons = int(index_counter_gpu.get()[0])
# Then allocate new storage space
pos = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=reduced_nphotons, dtype=np.float32)
t = ga.empty(shape=reduced_nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=reduced_nphotons, dtype=np.int32)
flags = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
weights = ga.empty(shape=reduced_nphotons, dtype=np.float32)
evidx = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
# And finaly copy photons, if there are any
if reduced_nphotons > 0:
index_counter_gpu.fill(0)
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photons(np.int32(start_photon +
first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu,
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
self.evidx,
pos,
dir,
wavelengths,
pol,
t,
flags,
last_hit_triangles,
weights,
evidx,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
assert index_counter_gpu.get()[0] == reduced_nphotons
return GPUPhotonsSlice(pos, dir, pol, wavelengths, t,
last_hit_triangles, flags, weights, evidx)
def __del__(self):
del self.pos
del self.dir
del self.pol
del self.wavelengths
del self.t
del self.flags
del self.last_hit_triangles
# Free up GPU memory quickly if now available
gc.collect()
def __len__(self):
return self.pos.size
def create_leaf_nodes(mesh, morton_bits=16, round_to_multiple=1):
'''Compute the leaf nodes surrounding a triangle mesh.
``mesh``: chroma.geometry.Mesh
Triangles to box
``morton_bits``: int
Number of bits to use per dimension when computing Morton code.
``round_to_multiple``: int
Round the number of nodes created up to multiple of this number
Extra nodes will be all zero.
Returns (world_coords, nodes, morton_codes), where
``world_coords``: chroma.bvh.WorldCoords
Defines the fixed point coordinate system
``nodes``: ndarray(shape=len(mesh.triangles), dtype=uint4)
List of leaf nodes. Child IDs will be set to triangle offsets.
``morton_codes``: ndarray(shape=len(mesh.triangles), dtype=np.uint64)
Morton codes for each triangle, using ``morton_bits`` per axis.
Must be <= 16 bits.
'''
# Load GPU functions
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
# compute world coordinates
world_origin = mesh.vertices.min(axis=0)
world_scale = np.max((mesh.vertices.max(axis=0) - world_origin)) \
/ (2**16 - 2)
world_coords = WorldCoords(world_origin=world_origin,
world_scale=world_scale)
# Put triangles and vertices in mapped host memory
triangles = mapped_empty(shape=len(mesh.triangles), dtype=ga.vec.uint3,
write_combined=True)
triangles[:] = to_uint3(mesh.triangles)
vertices = mapped_empty(shape=len(mesh.vertices), dtype=ga.vec.float3,
write_combined=True)
vertices[:] = to_float3(mesh.vertices)
# Call GPU to compute nodes
nodes = ga.zeros(shape=round_up_to_multiple(len(triangles),
round_to_multiple),
dtype=ga.vec.uint4)
morton_codes = ga.empty(shape=len(triangles), dtype=np.uint64)
# Convert world coords to GPU-friendly types
world_origin = ga.vec.make_float3(*world_origin)
world_scale = np.float32(world_scale)
nthreads_per_block = 256
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(len(triangles), nthreads_per_block,
max_blocks=30000):
bvh_funcs.make_leaves(np.uint32(first_index),
np.uint32(elements_this_iter),
Mapped(triangles), Mapped(vertices),
world_origin, world_scale,
nodes, morton_codes,
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
morton_codes_host = morton_codes.get() >> (16 - morton_bits)
return world_coords, nodes.get(), morton_codes_host
class GPUPDF(object):
def __init__(self):
self.module = get_cu_module('pdf.cu',
options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
def setup_pdf(self, nchannels, tbins, trange, qbins, qrange):
"""Setup GPU arrays to hold PDF information.
nchannels: int, number of channels
tbins: number of time bins
trange: tuple of (min, max) time in PDF
qbins: number of charge bins
qrange: tuple of (min, max) charge in PDF
"""
self.events_in_histogram = 0
self.hitcount_gpu = ga.zeros(nchannels, dtype=np.uint32)
self.pdf_gpu = ga.zeros(shape=(nchannels, tbins, qbins),
dtype=np.uint32)
self.tbins = tbins
self.trange = trange
self.qbins = qbins
self.qrange = qrange
def clear_pdf(self):
"""Rezero the PDF counters."""
self.hitcount_gpu.fill(0)
self.pdf_gpu.fill(0)
def add_hits_to_pdf(self, gpuchannels, nthreads_per_block=64):
self.gpu_funcs.bin_hits(
np.int32(len(self.hitcount_gpu)),
gpuchannels.q,
gpuchannels.t,
self.hitcount_gpu,
np.int32(self.tbins),
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.int32(self.qbins),
np.float32(self.qrange[0]),
np.float32(self.qrange[1]),
self.pdf_gpu,
block=(nthreads_per_block, 1, 1),
grid=(len(gpuchannels.t) // nthreads_per_block + 1, 1))
self.events_in_histogram += 1
def get_pdfs(self):
"""Returns the 1D hitcount array and the 3D [channel, time, charge]
histogram."""
return self.hitcount_gpu.get(), self.pdf_gpu.get()
def setup_pdf_eval(self,
event_hit,
event_time,
event_charge,
min_twidth,
trange,
min_qwidth,
qrange,
min_bin_content=10,
time_only=True):
"""Setup GPU arrays to compute PDF values for the given event.
The pdf_eval calculation allows the PDF to be evaluated at a
single point for each channel as the Monte Carlo is run. The
effective bin size will be as small as (`min_twidth`,
`min_qwidth`) around the point of interest, but will be large
enough to ensure that `min_bin_content` Monte Carlo events
fall into the bin.
event_hit: ndarray
Hit or not-hit status for each channel in the detector.
event_time: ndarray
Hit time for each channel in the detector. If channel
not hit, the time will be ignored.
event_charge: ndarray
Integrated charge for each channel in the detector.
If channel not hit, the charge will be ignored.
min_twidth: float
Minimum bin size in the time dimension
trange: (float, float)
Range of time dimension in PDF
min_qwidth: float
Minimum bin size in charge dimension
qrange: (float, float)
Range of charge dimension in PDF
min_bin_content: int
The bin will be expanded to include at least this many events
time_only: bool
If True, only the time observable will be used in the PDF.
"""
self.event_nhit = count_nonzero(event_hit)
# Define a mapping from an array of len(event_hit) to an array of length event_nhit
self.map_hit_offset_to_channel_id = np.where(event_hit)[0].astype(
np.uint32)
self.map_hit_offset_to_channel_id_gpu = ga.to_gpu(
self.map_hit_offset_to_channel_id)
self.map_channel_id_to_hit_offset = np.maximum(0,
event_hit.cumsum() -
1).astype(np.uint32)
self.map_channel_id_to_hit_offset_gpu = ga.to_gpu(
self.map_channel_id_to_hit_offset)
self.event_hit_gpu = ga.to_gpu(event_hit.astype(np.uint32))
self.event_time_gpu = ga.to_gpu(event_time.astype(np.float32))
self.event_charge_gpu = ga.to_gpu(event_charge.astype(np.float32))
self.eval_hitcount_gpu = ga.zeros(len(event_hit), dtype=np.uint32)
self.eval_bincount_gpu = ga.zeros(len(event_hit), dtype=np.uint32)
self.nearest_mc_gpu = ga.empty(shape=self.event_nhit * min_bin_content,
dtype=np.float32)
self.nearest_mc_gpu.fill(1e9)
self.min_twidth = min_twidth
self.trange = trange
self.min_qwidth = min_qwidth
self.qrange = qrange
self.min_bin_content = min_bin_content
assert time_only # Only support time right now
self.time_only = time_only
def clear_pdf_eval(self):
"Reset PDF evaluation counters to start accumulating new Monte Carlo."
self.eval_hitcount_gpu.fill(0)
self.eval_bincount_gpu.fill(0)
self.nearest_mc_gpu.fill(1e9)
@profile_if_possible
def accumulate_pdf_eval(self,
gpuchannels,
nthreads_per_block=64,
max_blocks=10000):
"Add the most recent results of run_daq() to the PDF evaluation."
self.work_queues = ga.empty(shape=self.event_nhit *
(gpuchannels.ndaq + 1),
dtype=np.uint32)
self.work_queues.fill(1)
self.gpu_funcs.accumulate_bincount(
np.int32(self.event_hit_gpu.size),
np.int32(gpuchannels.ndaq),
self.event_hit_gpu,
self.event_time_gpu,
gpuchannels.t,
self.eval_hitcount_gpu,
self.eval_bincount_gpu,
np.float32(self.min_twidth),
np.float32(self.trange[0]),
np.float32(self.trange[1]),
np.int32(self.min_bin_content),
self.map_channel_id_to_hit_offset_gpu,
self.work_queues,
block=(nthreads_per_block, 1, 1),
grid=(self.event_hit_gpu.size // nthreads_per_block + 1, 1))
cuda.Context.get_current().synchronize()
self.gpu_funcs.accumulate_nearest_neighbor_block(
np.int32(self.event_nhit),
np.int32(gpuchannels.ndaq),
self.map_hit_offset_to_channel_id_gpu,
self.work_queues,
self.event_time_gpu,
gpuchannels.t,
self.nearest_mc_gpu,
np.int32(self.min_bin_content),
block=(nthreads_per_block, 1, 1),
grid=(self.event_nhit, 1))
cuda.Context.get_current().synchronize()
def get_pdf_eval(self):
evhit = self.event_hit_gpu.get().astype(bool)
hitcount = self.eval_hitcount_gpu.get()
bincount = self.eval_bincount_gpu.get()
pdf_value = np.zeros(len(hitcount), dtype=float)
pdf_frac_uncert = np.zeros_like(pdf_value)
# PDF value for high stats bins
high_stats = (bincount >= self.min_bin_content)
if high_stats.any():
if self.time_only:
pdf_value[high_stats] = bincount[high_stats].astype(
float) / hitcount[high_stats] / self.min_twidth
else:
assert Exception('Unimplemented 2D (time,charge) mode!')
pdf_frac_uncert[high_stats] = 1.0 / np.sqrt(bincount[high_stats])
# PDF value for low stats bins
low_stats = ~high_stats & (hitcount > 0) & evhit
nearest_mc_by_hit = self.nearest_mc_gpu.get().reshape(
(self.event_nhit, self.min_bin_content))
nearest_mc = np.empty(shape=(len(hitcount), self.min_bin_content),
dtype=np.float32)
nearest_mc.fill(1e9)
nearest_mc[self.map_hit_offset_to_channel_id, :] = nearest_mc_by_hit
# Deal with the case where we did not even get min_bin_content events
# in the PDF but also clamp the lower range to ensure we don't index
# by a negative number in 2 lines
last_valid_entry = np.maximum(
0, (nearest_mc < 1e9).astype(int).sum(axis=1) - 1)
distance = nearest_mc[np.arange(len(last_valid_entry)),
last_valid_entry]
if low_stats.any():
if self.time_only:
pdf_value[low_stats] = (
last_valid_entry[low_stats] + 1).astype(float) / hitcount[
low_stats] / distance[low_stats] / 2.0
else:
assert Exception('Unimplemented 2D (time,charge) mode!')
pdf_frac_uncert[low_stats] = 1.0 / np.sqrt(
last_valid_entry[low_stats] + 1)
# PDFs with no stats got zero by default during array creation
print 'high_stats:', high_stats.sum(), 'low_stats', low_stats.sum()
return hitcount, pdf_value, pdf_value * pdf_frac_uncert
def optimize_layer(orig_nodes):
bvh_module = get_cu_module('bvh.cu', options=cuda_options,
include_source_directory=True)
bvh_funcs = GPUFuncs(bvh_module)
nodes = ga.to_gpu(orig_nodes)
n = len(nodes)
areas = ga.empty(shape=n/2, dtype=np.uint64)
nthreads_per_block = 128
min_areas = ga.empty(shape=int(np.ceil(n/float(nthreads_per_block))), dtype=np.uint64)
min_index = ga.empty(shape=min_areas.shape, dtype=np.uint32)
update = 10000
skip_size = 1
flag = mapped_empty(shape=skip_size, dtype=np.uint32)
i = 0
skips = 0
swaps = 0
while i < n/2 - 1:
# How are we doing?
if i % update == 0:
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(n/2, nthreads_per_block, max_blocks=10000):
bvh_funcs.pair_area(np.uint32(first_index),
np.uint32(elements_this_iter),
nodes,
areas,
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
areas_host = areas.get()
#print nodes.get(), areas_host.astype(float)
print 'Area of parent layer so far (%d): %1.12e' % (i*2, areas_host.astype(float).sum())
print 'Skips: %d, Swaps: %d' % (skips, swaps)
test_index = i * 2
blocks = 0
look_forward = min(8192*50, n - test_index - 2)
skip_this_round = min(skip_size, n - test_index - 1)
flag[:] = 0
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(look_forward, nthreads_per_block, max_blocks=10000):
bvh_funcs.min_distance_to(np.uint32(first_index + test_index + 2),
np.uint32(elements_this_iter),
np.uint32(test_index),
nodes,
np.uint32(blocks),
min_areas,
min_index,
Mapped(flag),
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter, skip_this_round))
blocks += nblocks_this_iter
#print i, first_index, nblocks_this_iter, look_forward
cuda.Context.get_current().synchronize()
if flag[0] == 0:
flag_nonzero = flag.nonzero()[0]
if len(flag_nonzero) == 0:
no_swap_required = skip_size
else:
no_swap_required = flag_nonzero[0]
i += no_swap_required
skips += no_swap_required
continue
min_areas_host = min_areas[:blocks].get()
min_index_host = min_index[:blocks].get()
best_block = min_areas_host.argmin()
better_i = min_index_host[best_block]
swaps += 1
#print 'swap', test_index+1, better_i
assert 0 < better_i < len(nodes)
assert 0 < test_index + 1 < len(nodes)
bvh_funcs.swap(np.uint32(test_index+1), np.uint32(better_i),
nodes, block=(1,1,1), grid=(1,1))
cuda.Context.get_current().synchronize()
i += 1
for first_index, elements_this_iter, nblocks_this_iter in \
chunk_iterator(n/2, nthreads_per_block, max_blocks=10000):
bvh_funcs.pair_area(np.uint32(first_index),
np.uint32(elements_this_iter),
nodes,
areas,
block=(nthreads_per_block,1,1),
grid=(nblocks_this_iter,1))
areas_host = areas.get()
print 'Final area of parent layer: %1.12e' % areas_host.sum()
print 'Skips: %d, Swaps: %d' % (skips, swaps)
return nodes.get()
def __init__(self):
self.module = get_cu_module('pdf.cu',
options=cuda_options,
include_source_directory=True)
self.gpu_funcs = GPUFuncs(self.module)
class GPUPhotonsHit(object):
def __init__(self, photons, ncopies=1, max_time=4.):
"""Load ``photons`` onto the GPU, replicating as requested.
Args:
- photons: chroma.Event.Photons
Photon state information to load onto GPU
- ncopies: int, *optional*
Number of times to replicate the photons
on the GPU. This is used if you want
to propagate the same event many times,
for example in a likelihood calculation.
The amount of GPU storage will be proportionally
larger if ncopies > 1, so be careful.
"""
module = get_cu_module('propagate_hit.cu', options=cuda_options)
propagate_hit_kernel = module.get_function('propagate_hit')
propagate_hit_kernel.prepare('iiPPPPPPPPPPPiiiPPP')
self.propagate_hit_kernel = propagate_hit_kernel
self.gpu_funcs = GPUFuncs(module)
self.max_time = max_time
self.ncopies = ncopies
self.true_nphotons = len(photons)
self.marshall_photons(photons, ncopies)
def marshall_photons_npl(self, npl):
pass
def marshall_photons(self, photons, ncopies):
"""
Assign the provided photons to the beginning (possibly
the entire array if ncopies is 1
"""
nphotons = len(photons)
self.pos = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.dir = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.pol = ga.empty(shape=nphotons * ncopies, dtype=ga.vec.float3)
self.wavelengths = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.t = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.last_hit_triangles = ga.empty(shape=nphotons * ncopies,
dtype=np.int32)
self.flags = ga.empty(shape=nphotons * ncopies, dtype=np.uint32)
self.weights = ga.empty(shape=nphotons * ncopies, dtype=np.float32)
self.pos[:nphotons].set(to_float3(photons.pos))
self.dir[:nphotons].set(to_float3(photons.dir))
self.pol[:nphotons].set(to_float3(photons.pol))
self.wavelengths[:nphotons].set(photons.wavelengths.astype(np.float32))
self.t[:nphotons].set(photons.t.astype(np.float32))
self.last_hit_triangles[:nphotons].set(
photons.last_hit_triangles.astype(np.int32))
self.flags[:nphotons].set(photons.flags.astype(np.uint32))
self.weights[:nphotons].set(photons.weights.astype(np.float32))
# Replicate the photons to the rest of the slots if needed
if ncopies > 1:
max_blocks = 1024
nthreads_per_block = 64
block = (nthreads_per_block, 1, 1)
for first_photon, photons_this_round, blocks in chunk_iterator(
nphotons, nthreads_per_block, max_blocks):
pass
grid = (blocks, 1)
args = (
np.int32(first_photon),
np.int32(photons_this_round),
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
np.int32(ncopies - 1),
np.int32(nphotons),
)
self.gpu_funcs.photon_duplicate(*args, block=block, grid=grid)
pass
pass
def get(self, npl=0, hit=0):
log.info("get npl:%d hit:%d " % (npl, hit))
pos = self.pos.get().view(np.float32).reshape((len(self.pos), 3))
dir = self.dir.get().view(np.float32).reshape((len(self.dir), 3))
pol = self.pol.get().view(np.float32).reshape((len(self.pol), 3))
wavelengths = self.wavelengths.get()
t = self.t.get()
last_hit_triangles = self.last_hit_triangles.get()
flags = self.flags.get()
weights = self.weights.get()
if npl:
nall = len(pos)
a = np.zeros((nall, 4, 4), dtype=np.float32)
a[:, 0, :3] = pos
a[:, 0, 3] = t
a[:, 1, :3] = dir
a[:, 1, 3] = wavelengths
a[:, 2, :3] = pol
a[:, 2, 3] = weights
assert len(last_hit_triangles) == len(flags)
pmtid = np.zeros(nall, dtype=np.int32)
# a kludge setting of pmtid into lht using the map argument of propagate_hit.cu
SURFACE_DETECT = 0x1 << 2
detected = np.where(flags & SURFACE_DETECT)
pmtid[detected] = last_hit_triangles[
detected] # sparsely populate, leaving zeros for undetected
a[:, 3, 0] = np.arange(nall,
dtype=np.int32).view(a.dtype) # photon_id
a[:, 3, 1] = 0 # used in comparison againt vbo prop
a[:, 3, 2] = flags.view(a.dtype) # history flags
a[:, 3, 3] = pmtid.view(a.dtype) # channel_id ie PmtId
if hit:
return a[pmtid > 0].view(NPY)
else:
return a.view(NPY)
pass
else: # the old way
return event.Photons(pos, dir, pol, wavelengths, t,
last_hit_triangles, flags, weights)
def iterate_copies(self):
'''Returns an iterator that yields GPUPhotonsSlice objects
corresponding to the event copies stored in ``self``.'''
for i in xrange(self.ncopies):
window = slice(self.true_nphotons * i,
self.true_nphotons * (i + 1))
yield GPUPhotonsSlice(
pos=self.pos[window],
dir=self.dir[window],
pol=self.pol[window],
wavelengths=self.wavelengths[window],
t=self.t[window],
last_hit_triangles=self.last_hit_triangles[window],
flags=self.flags[window],
weights=self.weights[window])
def upload_queues(self, nwork):
"""
# Order photons initially in the queue to put the clones next to each other
#. input_queue starts as [0,0,1,2,3,.....,nwork]
#. output_queue starts as [1,0,0,0,0,....]
"""
input_queue = np.empty(shape=nwork + 1, dtype=np.uint32)
input_queue[0] = 0
for copy in xrange(self.ncopies):
input_queue[1 + copy::self.ncopies] = np.arange(
self.true_nphotons,
dtype=np.uint32) + copy * self.true_nphotons
output_queue = np.zeros(shape=nwork + 1, dtype=np.uint32)
output_queue[0] = 1
self.input_queue_gpu = ga.to_gpu(input_queue)
self.output_queue_gpu = ga.to_gpu(output_queue)
def swap_queues(self):
"""
Swaps queues and returns photons remaining to propagate
#. output_queue[0] = 1 initially, this avoids enqueued photon_id
stomping on output_queue[0] as atomicAdd returns the non-incremented::
230 // Not done, put photon in output queue
231 if ((p.history & (NO_HIT | BULK_ABSORB | SURFACE_DETECT | SURFACE_ABSORB | NAN_ABORT)) == 0)
232 {
// pulling queue ticket
233 int out_idx = atomicAdd(output_queue, 1); // atomic add 1 to slot zero value, returns non-incremented original value
234 output_queue[out_idx] = photon_id;
235 }
#. At kernel tail non-completed photon threads enqueue their photon_id
into a slot in the output_queue. The slot to use is obtained
by atomic incrementing output_queue[0], ensuring orderly queue.
#. after kernel completes output_queue[0] contains the
number of photon_id enqued in output_queue[1:]
"""
temp = self.input_queue_gpu
self.input_queue_gpu = self.output_queue_gpu
self.output_queue_gpu = temp
self.output_queue_gpu[:1].set(np.ones(shape=1, dtype=np.uint32))
slot0minus1 = self.input_queue_gpu[:1].get(
)[0] - 1 # which was just now the output_queue before swap
log.debug("swap_queues slot0minus1 %s " % slot0minus1)
return slot0minus1
@profile_if_possible
def propagate_hit(self, gpu_geometry, rng_states, parameters):
"""Propagate photons on GPU to termination or max_steps, whichever
comes first.
May be called repeatedly without reloading photon information if
single-stepping through photon history.
..warning::
`rng_states` must have at least `nthreads_per_block`*`max_blocks`
number of curandStates.
got one abort::
In [1]: a = ph("hhMOCK")
In [9]: f = a[:,3,2].view(np.uint32)
In [12]: np.where( f & 1<<31 )
Out[12]: (array([279]),)
failed to just mock that one::
RANGE=279:280 MockNuWa MOCK
"""
nphotons = self.pos.size
nwork = nphotons
nthreads_per_block = parameters['threads_per_block']
max_blocks = parameters['max_blocks']
max_steps = parameters['max_steps']
use_weights = False
scatter_first = 0
self.upload_queues(nwork)
solid_id_map_gpu = gpu_geometry.solid_id_map
solid_id_to_channel_id_gpu = gpu_geometry.solid_id_to_channel_id_gpu
small_remainder = nthreads_per_block * 16 * 8
block = (nthreads_per_block, 1, 1)
results = {}
results['name'] = "propagate_hit"
results['nphotons'] = nphotons
results['nwork'] = nwork
results['nsmall'] = small_remainder
results['COLUMNS'] = "name:s,nphotons:i,nwork:i,nsmall:i"
step = 0
times = []
npass = 0
nabort = 0
while step < max_steps:
npass += 1
if nwork < small_remainder or use_weights:
nsteps = max_steps - step # Just finish the rest of the steps if the # of photons is low
log.debug(
"increase nsteps for stragglers: small_remainder %s nwork %s nsteps %s max_steps %s "
% (small_remainder, nwork, nsteps, max_steps))
else:
nsteps = 1
pass
log.info("nphotons %s nwork %s step %s max_steps %s nsteps %s " %
(nphotons, nwork, step, max_steps, nsteps))
abort = False
for first_photon, photons_this_round, blocks in chunk_iterator(
nwork, nthreads_per_block, max_blocks):
if abort:
nabort += 1
else:
grid = (blocks, 1)
args = (
np.int32(first_photon),
np.int32(photons_this_round),
self.input_queue_gpu[1:].gpudata,
self.output_queue_gpu.gpudata,
rng_states,
self.pos.gpudata,
self.dir.gpudata,
self.wavelengths.gpudata,
self.pol.gpudata,
self.t.gpudata,
self.flags.gpudata,
self.last_hit_triangles.gpudata,
self.weights.gpudata,
np.int32(nsteps),
np.int32(use_weights),
np.int32(scatter_first),
gpu_geometry.gpudata,
solid_id_map_gpu.gpudata,
solid_id_to_channel_id_gpu.gpudata,
)
log.info(
"propagate_hit_kernel.prepared_timed_call grid %s block %s first_photon %s photons_this_round %s "
% (repr(grid), repr(block), first_photon,
photons_this_round))
get_time = self.propagate_hit_kernel.prepared_timed_call(
grid, block, *args)
t = get_time()
times.append(t)
if t > self.max_time:
abort = True
log.warn(
"kernel launch time %s > max_time %s : ABORTING " %
(t, self.max_time))
pass
pass
pass
log.info("step %s propagate_hit_kernel times %s " %
(step, repr(times)))
pass
step += nsteps
scatter_first = 0 # Only allow non-zero in first pass
if step < max_steps:
nwork = self.swap_queues()
pass
pass
log.info("calling max ")
if ga.max(self.flags).get() & (1 << 31):
log.warn("ABORTED PHOTONS")
log.info("done calling max ")
cuda.Context.get_current().synchronize()
results['npass'] = npass
results['nabort'] = nabort
results['nlaunch'] = len(times)
results['tottime'] = sum(times)
results['maxtime'] = max(times)
results['mintime'] = min(times)
results[
'COLUMNS'] += ",npass:i,nabort:i,nlaunch:i,tottime:f,maxtime:f,mintime:f"
return results
@profile_if_possible
def select(self,
target_flag,
nthreads_per_block=64,
max_blocks=1024,
start_photon=None,
nphotons=None):
'''Return a new GPUPhoton object containing only photons that
have a particular bit set in their history word.'''
cuda.Context.get_current().synchronize()
index_counter_gpu = ga.zeros(shape=1, dtype=np.uint32)
cuda.Context.get_current().synchronize()
if start_photon is None:
start_photon = 0
if nphotons is None:
nphotons = self.pos.size - start_photon
# First count how much space we need
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.count_photons(np.int32(start_photon + first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu,
self.flags,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
cuda.Context.get_current().synchronize()
reduced_nphotons = int(index_counter_gpu.get()[0])
# Then allocate new storage space
pos = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
dir = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
pol = ga.empty(shape=reduced_nphotons, dtype=ga.vec.float3)
wavelengths = ga.empty(shape=reduced_nphotons, dtype=np.float32)
t = ga.empty(shape=reduced_nphotons, dtype=np.float32)
last_hit_triangles = ga.empty(shape=reduced_nphotons, dtype=np.int32)
flags = ga.empty(shape=reduced_nphotons, dtype=np.uint32)
weights = ga.empty(shape=reduced_nphotons, dtype=np.float32)
# And finaly copy photons, if there are any
if reduced_nphotons > 0:
index_counter_gpu.fill(0)
for first_photon, photons_this_round, blocks in \
chunk_iterator(nphotons, nthreads_per_block, max_blocks):
self.gpu_funcs.copy_photons(np.int32(start_photon +
first_photon),
np.int32(photons_this_round),
np.uint32(target_flag),
index_counter_gpu,
self.pos,
self.dir,
self.wavelengths,
self.pol,
self.t,
self.flags,
self.last_hit_triangles,
self.weights,
pos,
dir,
wavelengths,
pol,
t,
flags,
last_hit_triangles,
weights,
block=(nthreads_per_block, 1, 1),
grid=(blocks, 1))
assert index_counter_gpu.get()[0] == reduced_nphotons
return GPUPhotonsHitSlice(pos, dir, pol, wavelengths, t,
last_hit_triangles, flags, weights)
def __del__(self):
del self.pos
del self.dir
del self.pol
del self.wavelengths
del self.t
del self.flags
del self.last_hit_triangles
# Free up GPU memory quickly if now available
gc.collect()
def __len__(self):
return self.pos.size