def __init__(self, geometry, wavelengths=None, print_usage=False, min_free_gpu_mem=300e6):
if wavelengths is None:
wavelengths = standard_wavelengths
try:
wavelength_step = np.unique(np.diff(wavelengths)).item()
except ValueError:
raise ValueError('wavelengths must be equally spaced apart.')
geometry_source = get_cu_source('geometry_types.h')
material_struct_size = characterize.sizeof('Material', geometry_source)
surface_struct_size = characterize.sizeof('Surface', geometry_source)
geometry_struct_size = characterize.sizeof('Geometry', geometry_source)
self.material_data = []
self.material_ptrs = []
def interp_material_property(wavelengths, property):
# note that it is essential that the material properties be
# interpolated linearly. this fact is used in the propagation
# code to guarantee that probabilities still sum to one.
return np.interp(wavelengths, property[:,0], property[:,1]).astype(np.float32)
for i in range(len(geometry.unique_materials)):
material = geometry.unique_materials[i]
if material is None:
raise Exception('one or more triangles is missing a material.')
refractive_index = interp_material_property(wavelengths, material.refractive_index)
refractive_index_gpu = ga.to_gpu(refractive_index)
absorption_length = interp_material_property(wavelengths, material.absorption_length)
absorption_length_gpu = ga.to_gpu(absorption_length)
scattering_length = interp_material_property(wavelengths, material.scattering_length)
scattering_length_gpu = ga.to_gpu(scattering_length)
reemission_prob = interp_material_property(wavelengths, material.reemission_prob)
reemission_prob_gpu = ga.to_gpu(reemission_prob)
reemission_cdf = interp_material_property(wavelengths, material.reemission_cdf)
reemission_cdf_gpu = ga.to_gpu(reemission_cdf)
self.material_data.append(refractive_index_gpu)
self.material_data.append(absorption_length_gpu)
self.material_data.append(scattering_length_gpu)
self.material_data.append(reemission_prob_gpu)
self.material_data.append(reemission_cdf_gpu)
material_gpu = \
make_gpu_struct(material_struct_size,
[refractive_index_gpu, absorption_length_gpu,
scattering_length_gpu,
reemission_prob_gpu,
reemission_cdf_gpu,
np.uint32(len(wavelengths)),
np.float32(wavelength_step),
np.float32(wavelengths[0])])
self.material_ptrs.append(material_gpu)
self.material_pointer_array = \
make_gpu_struct(8*len(self.material_ptrs), self.material_ptrs)
self.surface_data = []
self.surface_ptrs = []
for i in range(len(geometry.unique_surfaces)):
surface = geometry.unique_surfaces[i]
if surface is None:
# need something to copy to the surface array struct
# that is the same size as a 64-bit pointer.
# this pointer will never be used by the simulation.
self.surface_ptrs.append(np.uint64(0))
continue
detect = interp_material_property(wavelengths, surface.detect)
detect_gpu = ga.to_gpu(detect)
absorb = interp_material_property(wavelengths, surface.absorb)
absorb_gpu = ga.to_gpu(absorb)
reemit = interp_material_property(wavelengths, surface.reemit)
reemit_gpu = ga.to_gpu(reemit)
reflect_diffuse = interp_material_property(wavelengths, surface.reflect_diffuse)
reflect_diffuse_gpu = ga.to_gpu(reflect_diffuse)
reflect_specular = interp_material_property(wavelengths, surface.reflect_specular)
reflect_specular_gpu = ga.to_gpu(reflect_specular)
eta = interp_material_property(wavelengths, surface.eta)
eta_gpu = ga.to_gpu(eta)
k = interp_material_property(wavelengths, surface.k)
k_gpu = ga.to_gpu(k)
reemission_cdf = interp_material_property(wavelengths, surface.reemission_cdf)
reemission_cdf_gpu = ga.to_gpu(reemission_cdf)
self.surface_data.append(detect_gpu)
self.surface_data.append(absorb_gpu)
self.surface_data.append(reemit_gpu)
self.surface_data.append(reflect_diffuse_gpu)
self.surface_data.append(reflect_specular_gpu)
self.surface_data.append(eta_gpu)
self.surface_data.append(k_gpu)
self.surface_data.append(reemission_cdf_gpu)
surface_gpu = \
make_gpu_struct(surface_struct_size,
[detect_gpu, absorb_gpu, reemit_gpu,
reflect_diffuse_gpu,reflect_specular_gpu,
eta_gpu, k_gpu, reemission_cdf_gpu,
np.uint32(surface.model),
np.uint32(len(wavelengths)),
np.uint32(surface.transmissive),
np.float32(wavelength_step),
np.float32(wavelengths[0]),
np.float32(surface.thickness)])
self.surface_ptrs.append(surface_gpu)
self.surface_pointer_array = \
make_gpu_struct(8*len(self.surface_ptrs), self.surface_ptrs)
self.vertices = mapped_empty(shape=len(geometry.mesh.vertices),
dtype=ga.vec.float3,
write_combined=True)
self.triangles = mapped_empty(shape=len(geometry.mesh.triangles),
dtype=ga.vec.uint3,
write_combined=True)
self.vertices[:] = to_float3(geometry.mesh.vertices)
self.triangles[:] = to_uint3(geometry.mesh.triangles)
self.world_origin = ga.vec.make_float3(*geometry.bvh.world_coords.world_origin)
self.world_scale = np.float32(geometry.bvh.world_coords.world_scale)
material_codes = (((geometry.material1_index & 0xff) << 24) |
((geometry.material2_index & 0xff) << 16) |
((geometry.surface_index & 0xff) << 8)).astype(np.uint32)
self.material_codes = ga.to_gpu(material_codes)
colors = geometry.colors.astype(np.uint32)
self.colors = ga.to_gpu(colors)
self.solid_id_map = ga.to_gpu(geometry.solid_id.astype(np.uint32))
# Limit memory usage by splitting BVH into on and off-GPU parts
gpu_free, gpu_total = cuda.mem_get_info()
node_array_usage = geometry.bvh.nodes.nbytes
# Figure out how many elements we can fit on the GPU,
# but no fewer than 100 elements, and no more than the number of actual nodes
n_nodes = len(geometry.bvh.nodes)
split_index = min(
max(int((gpu_free - min_free_gpu_mem) / geometry.bvh.nodes.itemsize),100),
n_nodes
)
self.nodes = ga.to_gpu(geometry.bvh.nodes[:split_index])
n_extra = max(1, (n_nodes - split_index)) # forbid zero size
self.extra_nodes = mapped_empty(shape=n_extra,
dtype=geometry.bvh.nodes.dtype,
write_combined=True)
if split_index < n_nodes:
logger.info('Splitting BVH between GPU and CPU memory at node %d' % split_index)
self.extra_nodes[:] = geometry.bvh.nodes[split_index:]
# See if there is enough memory to put the and/ortriangles back on the GPU
gpu_free, gpu_total = cuda.mem_get_info()
if self.triangles.nbytes < (gpu_free - min_free_gpu_mem):
self.triangles = ga.to_gpu(self.triangles)
logger.info('Optimization: Sufficient memory to move triangles onto GPU')
gpu_free, gpu_total = cuda.mem_get_info()
if self.vertices.nbytes < (gpu_free - min_free_gpu_mem):
self.vertices = ga.to_gpu(self.vertices)
logger.info('Optimization: Sufficient memory to move vertices onto GPU')
self.gpudata = make_gpu_struct(geometry_struct_size,
[Mapped(self.vertices),
Mapped(self.triangles),
self.material_codes,
self.colors, self.nodes,
Mapped(self.extra_nodes),
self.material_pointer_array,
self.surface_pointer_array,
self.world_origin,
self.world_scale,
np.int32(len(self.nodes))])
self.geometry = geometry
if print_usage:
self.print_device_usage()
logger.info(self.device_usage_str())
def __init__(self, geometry, wavelengths=None, times=None, print_usage=False, min_free_gpu_mem=300e6):
if wavelengths is None:
wavelengths = standard_wavelengths
try:
wavelength_step = np.unique(np.diff(wavelengths)).item()
except ValueError:
raise ValueError('wavelengths must be equally spaced apart.')
if times is None:
time_step = 0.05
times = np.arange(0,1000,time_step)
else:
try:
time_step = np.unique(np.diff(times)).item()
except ValueError:
raise ValueError('times must be equally spaced apart.')
geometry_source = get_cu_source('geometry_types.h')
material_struct_size = characterize.sizeof('Material', geometry_source)
surface_struct_size = characterize.sizeof('Surface', geometry_source)
dichroicprops_struct_size = characterize.sizeof('DichroicProps', geometry_source)
geometry_struct_size = characterize.sizeof('Geometry', geometry_source)
self.material_data = []
self.material_ptrs = []
def interp_material_property(wavelengths, property):
# note that it is essential that the material properties be
# interpolated linearly. this fact is used in the propagation
# code to guarantee that probabilities still sum to one.
return np.interp(wavelengths, property[:,0], property[:,1]).astype(np.float32)
for i in range(len(geometry.unique_materials)):
material = geometry.unique_materials[i]
if material is None:
raise Exception('one or more triangles is missing a material.')
refractive_index = interp_material_property(wavelengths, material.refractive_index)
refractive_index_gpu = ga.to_gpu(refractive_index)
absorption_length = interp_material_property(wavelengths, material.absorption_length)
absorption_length_gpu = ga.to_gpu(absorption_length)
scattering_length = interp_material_property(wavelengths, material.scattering_length)
scattering_length_gpu = ga.to_gpu(scattering_length)
num_comp = len(material.comp_reemission_prob)
comp_reemission_prob_gpu = [ga.to_gpu(interp_material_property(wavelengths, component)) for component in material.comp_reemission_prob]
self.material_data.append(comp_reemission_prob_gpu)
comp_reemission_prob_gpu = np.uint64(0) if len(comp_reemission_prob_gpu) == 0 else make_gpu_struct(8*len(comp_reemission_prob_gpu), comp_reemission_prob_gpu)
assert num_comp == len(material.comp_reemission_wvl_cdf), 'component arrays must be same length'
comp_reemission_wvl_cdf_gpu = [ga.to_gpu(interp_material_property(wavelengths, component)) for component in material.comp_reemission_wvl_cdf]
self.material_data.append(comp_reemission_wvl_cdf_gpu)
comp_reemission_wvl_cdf_gpu = np.uint64(0) if len(comp_reemission_wvl_cdf_gpu) == 0 else make_gpu_struct(8*len(comp_reemission_wvl_cdf_gpu), comp_reemission_wvl_cdf_gpu)
assert num_comp == len(material.comp_reemission_time_cdf), 'component arrays must be same length'
comp_reemission_time_cdf_gpu = [ga.to_gpu(interp_material_property(times, component)) for component in material.comp_reemission_time_cdf]
self.material_data.append(comp_reemission_time_cdf_gpu)
comp_reemission_time_cdf_gpu = np.uint64(0) if len(comp_reemission_time_cdf_gpu) == 0 else make_gpu_struct(8*len(comp_reemission_time_cdf_gpu), comp_reemission_time_cdf_gpu)
assert num_comp == len(material.comp_absorption_length), 'component arrays must be same length'
comp_absorption_length_gpu = [ga.to_gpu(interp_material_property(wavelengths, component)) for component in material.comp_absorption_length]
self.material_data.append(comp_absorption_length_gpu)
comp_absorption_length_gpu = np.uint64(0) if len(comp_absorption_length_gpu) == 0 else make_gpu_struct(8*len(comp_absorption_length_gpu), comp_absorption_length_gpu)
self.material_data.append(refractive_index_gpu)
self.material_data.append(absorption_length_gpu)
self.material_data.append(scattering_length_gpu)
self.material_data.append(comp_reemission_prob_gpu)
self.material_data.append(comp_reemission_wvl_cdf_gpu)
self.material_data.append(comp_reemission_time_cdf_gpu)
self.material_data.append(comp_absorption_length_gpu)
material_gpu = \
make_gpu_struct(material_struct_size,
[refractive_index_gpu, absorption_length_gpu,
scattering_length_gpu,
comp_reemission_prob_gpu,
comp_reemission_wvl_cdf_gpu,
comp_reemission_time_cdf_gpu,
comp_absorption_length_gpu,
np.uint32(num_comp),
np.uint32(len(wavelengths)),
np.float32(wavelength_step),
np.float32(wavelengths[0]),
np.uint32(len(times)),
np.float32(time_step),
np.float32(times[0])])
self.material_ptrs.append(material_gpu)
self.material_pointer_array = \
make_gpu_struct(8*len(self.material_ptrs), self.material_ptrs)
self.surface_data = []
self.surface_ptrs = []
for i in range(len(geometry.unique_surfaces)):
surface = geometry.unique_surfaces[i]
if surface is None:
# need something to copy to the surface array struct
# that is the same size as a 64-bit pointer.
# this pointer will never be used by the simulation.
self.surface_ptrs.append(np.uint64(0))
continue
detect = interp_material_property(wavelengths, surface.detect)
detect_gpu = ga.to_gpu(detect)
absorb = interp_material_property(wavelengths, surface.absorb)
absorb_gpu = ga.to_gpu(absorb)
reemit = interp_material_property(wavelengths, surface.reemit)
reemit_gpu = ga.to_gpu(reemit)
reflect_diffuse = interp_material_property(wavelengths, surface.reflect_diffuse)
reflect_diffuse_gpu = ga.to_gpu(reflect_diffuse)
reflect_specular = interp_material_property(wavelengths, surface.reflect_specular)
reflect_specular_gpu = ga.to_gpu(reflect_specular)
eta = interp_material_property(wavelengths, surface.eta)
eta_gpu = ga.to_gpu(eta)
k = interp_material_property(wavelengths, surface.k)
k_gpu = ga.to_gpu(k)
reemission_cdf = interp_material_property(wavelengths, surface.reemission_cdf)
reemission_cdf_gpu = ga.to_gpu(reemission_cdf)
if surface.dichroic_props:
props = surface.dichroic_props
transmit_pointers = []
reflect_pointers = []
angles_gpu = ga.to_gpu(np.asarray(props.angles,dtype=np.float32))
self.surface_data.append(angles_gpu)
for i,angle in enumerate(props.angles):
dichroic_reflect = interp_material_property(wavelengths, props.dichroic_reflect[i])
dichroic_reflect_gpu = ga.to_gpu(dichroic_reflect)
self.surface_data.append(dichroic_reflect_gpu)
reflect_pointers.append(dichroic_reflect_gpu)
dichroic_transmit = interp_material_property(wavelengths, props.dichroic_transmit[i])
dichroic_transmit_gpu = ga.to_gpu(dichroic_transmit)
self.surface_data.append(dichroic_transmit_gpu)
transmit_pointers.append(dichroic_transmit_gpu)
reflect_arr_gpu = make_gpu_struct(8*len(reflect_pointers),reflect_pointers)
self.surface_data.append(reflect_arr_gpu)
transmit_arr_gpu = make_gpu_struct(8*len(transmit_pointers), transmit_pointers)
self.surface_data.append(transmit_arr_gpu)
dichroic_props = make_gpu_struct(dichroicprops_struct_size,[angles_gpu,reflect_arr_gpu,transmit_arr_gpu,np.uint32(len(props.angles))])
else:
dichroic_props = np.uint64(0) #NULL
self.surface_data.append(detect_gpu)
self.surface_data.append(absorb_gpu)
self.surface_data.append(reemit_gpu)
self.surface_data.append(reflect_diffuse_gpu)
self.surface_data.append(reflect_specular_gpu)
self.surface_data.append(eta_gpu)
self.surface_data.append(k_gpu)
self.surface_data.append(dichroic_props)
surface_gpu = \
make_gpu_struct(surface_struct_size,
[detect_gpu, absorb_gpu, reemit_gpu,
reflect_diffuse_gpu,reflect_specular_gpu,
eta_gpu, k_gpu, reemission_cdf_gpu,
dichroic_props,
np.uint32(surface.model),
np.uint32(len(wavelengths)),
np.uint32(surface.transmissive),
np.float32(wavelength_step),
np.float32(wavelengths[0]),
np.float32(surface.thickness)])
self.surface_ptrs.append(surface_gpu)
self.surface_pointer_array = \
make_gpu_struct(8*len(self.surface_ptrs), self.surface_ptrs)
self.vertices = mapped_empty(shape=len(geometry.mesh.vertices),
dtype=ga.vec.float3,
write_combined=True)
self.triangles = mapped_empty(shape=len(geometry.mesh.triangles),
dtype=ga.vec.uint3,
write_combined=True)
self.vertices[:] = to_float3(geometry.mesh.vertices)
self.triangles[:] = to_uint3(geometry.mesh.triangles)
self.world_origin = ga.vec.make_float3(*geometry.bvh.world_coords.world_origin)
self.world_scale = np.float32(geometry.bvh.world_coords.world_scale)
material_codes = (((geometry.material1_index & 0xff) << 24) |
((geometry.material2_index & 0xff) << 16) |
((geometry.surface_index & 0xff) << 8)).astype(np.uint32)
self.material_codes = ga.to_gpu(material_codes)
colors = geometry.colors.astype(np.uint32)
self.colors = ga.to_gpu(colors)
self.solid_id_map = ga.to_gpu(geometry.solid_id.astype(np.uint32))
# Limit memory usage by splitting BVH into on and off-GPU parts
gpu_free, gpu_total = cuda.mem_get_info()
node_array_usage = geometry.bvh.nodes.nbytes
# Figure out how many elements we can fit on the GPU,
# but no fewer than 100 elements, and no more than the number of actual nodes
n_nodes = len(geometry.bvh.nodes)
split_index = min(
max(int((gpu_free - min_free_gpu_mem) / geometry.bvh.nodes.itemsize),100),
n_nodes
)
self.nodes = ga.to_gpu(geometry.bvh.nodes[:split_index])
n_extra = max(1, (n_nodes - split_index)) # forbid zero size
self.extra_nodes = mapped_empty(shape=n_extra,
dtype=geometry.bvh.nodes.dtype,
write_combined=True)
if split_index < n_nodes:
logger.info('Splitting BVH between GPU and CPU memory at node %d' % split_index)
self.extra_nodes[:] = geometry.bvh.nodes[split_index:]
# See if there is enough memory to put the and/ortriangles back on the GPU
gpu_free, gpu_total = cuda.mem_get_info()
if self.triangles.nbytes < (gpu_free - min_free_gpu_mem):
self.triangles = ga.to_gpu(self.triangles)
logger.info('Optimization: Sufficient memory to move triangles onto GPU')
gpu_free, gpu_total = cuda.mem_get_info()
if self.vertices.nbytes < (gpu_free - min_free_gpu_mem):
self.vertices = ga.to_gpu(self.vertices)
logger.info('Optimization: Sufficient memory to move vertices onto GPU')
self.gpudata = make_gpu_struct(geometry_struct_size,
[Mapped(self.vertices),
Mapped(self.triangles),
self.material_codes,
self.colors, self.nodes,
Mapped(self.extra_nodes),
self.material_pointer_array,
self.surface_pointer_array,
self.world_origin,
self.world_scale,
np.int32(len(self.nodes))])
self.geometry = geometry
if print_usage:
self.print_device_usage()
logger.info(self.device_usage_str())
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 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 __init__(self,
geometry,
wavelengths=None,
print_usage=False,
min_free_gpu_mem=300e6,
cl_context=None,
cl_queue=None):
log.info("GPUGeometry.__init__ min_free_gpu_mem %s ", min_free_gpu_mem)
self.geometry = geometry
self.instance_count += 1
assert self.instance_count == 1, traceback.print_stack()
self.metadata = Metadata()
self.metadata(None, 'preinfo')
self.metadata('a', "start")
self.metadata['a_min_free_gpu_mem'] = min_free_gpu_mem
if wavelengths is None:
self.wavelengths = standard_wavelengths
else:
self.wavelengths = wavelengths
try:
self.wavelength_step = np.unique(np.diff(self.wavelengths)).item()
except ValueError:
raise ValueError('wavelengths must be equally spaced apart.')
# this is where things get difficult.
# pycuda and pyopencl gives us very different methods for working with structs
#geometry_struct_size = characterize.sizeof('Geometry', geometry_source)
# Note, that unfortunately the data types returned are very different as the
if api.is_gpu_api_cuda():
self.material_data, self.material_ptrs, self.material_pointer_array = self._package_material_data_cuda(
geometry, self.wavelengths, self.wavelength_step)
self.surface_data, self.surface_ptrs, self.surface_pointer_array = self._package_surface_data_cuda(
geometry, self.wavelengths, self.wavelength_step)
elif api.is_gpu_api_opencl():
self.material_data, materials_bytes_cl = self._package_material_data_cl(
cl_context, cl_queue, geometry, self.wavelengths,
self.wavelength_step)
self.surface_data, surfaces_bytes_cl = self._package_surface_data_cl(
cl_context, cl_queue, geometry, self.wavelengths,
self.wavelength_step)
self.metadata('b', "after materials,surfaces")
if api.is_gpu_api_opencl():
self.metadata[
'b_gpu_used'] = materials_bytes_cl + surfaces_bytes_cl # opencl, we have to track this ourselves
# Load Vertices and Triangles
if api.is_gpu_api_cuda():
self.vertices = mapped_empty(shape=len(geometry.mesh.vertices),
dtype=ga.vec.float3,
write_combined=True)
self.vertices4 = np.zeros(shape=(len(self.vertices), 4),
dtype=np.float32)
self.triangles = mapped_empty(shape=len(geometry.mesh.triangles),
dtype=ga.vec.uint3,
write_combined=True)
self.triangles4 = np.zeros(shape=(len(self.triangles), 4),
dtype=np.uint32)
self.vertices[:] = to_float3(geometry.mesh.vertices)
self.vertices4[:, :-1] = self.vertices.ravel().view(
np.float32).reshape(len(self.vertices), 3) # for textures
self.triangles[:] = to_uint3(geometry.mesh.triangles)
self.triangles4[:, :-1] = self.triangles.ravel().view(
np.uint32).reshape(len(self.triangles), 3) # for textures
elif api.is_gpu_api_opencl():
self.vertices = ga.empty(cl_queue,
len(geometry.mesh.vertices),
dtype=ga.vec.float3)
self.triangles = ga.empty(cl_queue,
len(geometry.mesh.triangles),
dtype=ga.vec.uint3)
self.vertices[:] = to_float3(geometry.mesh.vertices)
self.triangles[:] = to_uint3(geometry.mesh.triangles)
if api.is_gpu_api_cuda():
self.world_origin = ga.vec.make_float3(
*geometry.bvh.world_coords.world_origin)
elif api.is_gpu_api_opencl():
self.world_origin = ga.vec.make_float3(
*geometry.bvh.world_coords.world_origin)
#self.world_origin = geometry.bvh.world_coords.world_origin
self.world_origin = ga.to_device(cl_queue, self.world_origin)
print type(self.world_origin), self.world_origin
self.world_scale = np.float32(geometry.bvh.world_coords.world_scale)
# Load material and surface indices into 8-bit codes
# check if we've reached a complexity threshold
if len(geometry.unique_materials) >= int(0xff):
raise ValueError(
'Number of materials to index has hit maximum of %d' %
(int(0xff)))
if len(geometry.unique_surfaces) >= int(0xff):
raise ValueError(
'Number of surfaces to index has hit maximum of %d' %
(int(0xff)))
# make bit code
material_codes = (((geometry.material1_index & 0xff) << 24) |
((geometry.material2_index & 0xff) << 16) |
((geometry.surface_index & 0xff) << 8)).astype(
np.uint32)
if api.is_gpu_api_cuda():
self.material_codes = ga.to_gpu(material_codes)
elif api.is_gpu_api_opencl():
self.material_codes = ga.to_device(cl_queue, material_codes)
# assign color codes
colors = geometry.colors.astype(np.uint32)
if api.is_gpu_api_cuda():
self.colors = ga.to_gpu(colors)
self.solid_id_map = ga.to_gpu(geometry.solid_id.astype(np.uint32))
elif api.is_gpu_api_opencl():
self.colors = ga.to_device(cl_queue, colors)
self.solid_id_map = ga.to_device(
cl_queue, geometry.solid_id.astype(np.uint32))
# Limit memory usage by splitting BVH into on and off-GPU parts
self.metadata('c', "after colors, idmap")
if api.is_gpu_api_cuda():
gpu_free, gpu_total = cuda.mem_get_info()
elif api.is_gpu_api_opencl():
gpu_total = self.metadata['gpu_total']
meshdef_nbytes_cl = self.vertices.nbytes + self.triangles.nbytes + self.world_origin.nbytes + self.world_scale.nbytes + self.material_codes.nbytes + self.colors.nbytes + self.solid_id_map.nbytes
self.metadata[
'c_gpu_used'] = materials_bytes_cl + surfaces_bytes_cl + meshdef_nbytes_cl
gpu_free = gpu_total - (materials_bytes_cl + surfaces_bytes_cl +
meshdef_nbytes_cl)
# Figure out how many elements we can fit on the GPU,
# but no fewer than 100 elements, and no more than the number of actual nodes
n_nodes = len(geometry.bvh.nodes)
split_index = min(
max(
int((gpu_free - min_free_gpu_mem) /
geometry.bvh.nodes.itemsize), 100), n_nodes)
print "split index=", split_index, " vs. total nodes=", n_nodes
# push nodes to GPU
if api.is_gpu_api_cuda():
self.nodes = ga.to_gpu(geometry.bvh.nodes[:split_index])
elif api.is_gpu_api_opencl():
self.nodes = ga.to_device(cl_queue,
geometry.bvh.nodes[:split_index])
n_extra = max(1, (n_nodes - split_index)) # forbid zero size
# left over nodes
if api.is_gpu_api_cuda():
self.extra_nodes = mapped_empty(shape=n_extra,
dtype=geometry.bvh.nodes.dtype,
write_combined=True)
elif api.is_gpu_api_opencl():
self.extra_nodes = ga.empty(cl_queue,
shape=n_extra,
dtype=geometry.bvh.nodes.dtype)
if split_index < n_nodes:
log.info('Splitting BVH between GPU and CPU memory at node %d' %
split_index)
self.extra_nodes[:] = geometry.bvh.nodes[split_index:]
splitting = 1
else:
splitting = 0
self.metadata('d', "after nodes")
if api.is_gpu_api_opencl():
nodes_nbytes_cl = self.nodes.nbytes
self.metadata[
'd_gpu_used'] = materials_bytes_cl + surfaces_bytes_cl + meshdef_nbytes_cl + nodes_nbytes_cl
self.metadata.array("d_nodes", geometry.bvh.nodes)
self.metadata['d_split_index'] = split_index
self.metadata['d_extra_nodes_count'] = n_extra
self.metadata['d_splitting'] = splitting
self.print_device_usage(cl_context=cl_context)
# CUDA See if there is enough memory to put the vertices and/or triangles back on the GPU
if api.is_gpu_api_cuda():
gpu_free, gpu_total = cuda.mem_get_info()
elif api.is_gpu_api_opencl():
gpu_total = self.metadata['gpu_total']
gpu_free = gpu_total - self.metadata['d_gpu_used']
self.metadata.array('e_triangles', self.triangles)
if api.is_gpu_api_cuda():
if self.triangles.nbytes < (gpu_free - min_free_gpu_mem):
self.triangles = ga.to_gpu(self.triangles)
log.info(
'Optimization: Sufficient memory to move triangles onto GPU'
)
ftriangles_gpu = 1
else:
log.warn('using host mapped memory triangles')
ftriangles_gpu = 0
elif api.is_gpu_api_opencl():
if self.triangles.nbytes < (gpu_free - min_free_gpu_mem):
#self.triangles = ga.to_device(cl_queue,self.triangles)
log.info(
'Optimization: Sufficient memory to move triangles onto GPU'
)
ftriangles_gpu = 1
else:
log.warn('using host mapped memory triangles')
ftriangles_gpu = 0
self.metadata('e', "after triangles")
self.metadata['e_triangles_gpu'] = ftriangles_gpu
if api.is_gpu_api_cuda():
gpu_free, gpu_total = cuda.mem_get_info()
elif api.is_gpu_api_opencl():
gpu_total = self.metadata['gpu_total']
gpu_free = gpu_total - self.metadata['d_gpu_used']
self.metadata.array('f_vertices', self.vertices)
if api.is_gpu_api_cuda():
if self.vertices.nbytes < (gpu_free - min_free_gpu_mem):
self.vertices = ga.to_gpu(self.vertices)
log.info(
'Optimization: Sufficient memory to move vertices onto GPU'
)
vertices_gpu = 1
else:
log.warn('using host mapped memory vertices')
vertices_gpu = 0
elif api.is_gpu_api_opencl():
if self.vertices.nbytes < (gpu_free - min_free_gpu_mem):
#self.vertices = ga.to_gpu(self.vertices)
log.info(
'Optimization: Sufficient memory to move vertices onto GPU'
)
vertices_gpu = 1
else:
log.warn('using host mapped memory vertices')
vertices_gpu = 0
self.metadata('f', "after vertices")
self.metadata['f_vertices_gpu'] = vertices_gpu
if api.is_gpu_api_cuda():
geometry_source = cutools.get_cu_source('geometry_types.h')
geometry_struct_size = characterize.sizeof('Geometry',
geometry_source)
self.gpudata = make_gpu_struct(geometry_struct_size, [
Mapped(self.vertices),
Mapped(self.triangles), self.material_codes, self.colors,
self.nodes,
Mapped(self.extra_nodes), self.material_pointer_array,
self.surface_pointer_array, self.world_origin,
self.world_scale,
np.int32(len(self.nodes))
])
elif api.is_gpu_api_opencl():
# No relevant way to pass struct into OpenCL kernel. We have to pass everything by arrays
# We then build a geometry struct later in the kernel
# provided below is example/test of passing the data
#if True: # for debuggin
if False: #
print "loading geometry_structs.cl"
geostructsmod = cltools.get_cl_module(
"geometry_structs.cl",
cl_context,
options=cltools.cl_options,
include_source_directory=True)
geostructsfunc = GPUFuncs(geostructsmod)
geostructsfunc.make_geostruct(
cl_queue, (3, ), None, self.vertices.data,
self.triangles.data, self.material_codes.data,
self.colors.data, self.nodes.data, self.extra_nodes.data,
np.int32(len(geometry.unique_materials)),
self.material_data['refractive_index'].data,
self.material_data['absorption_length'].data,
self.material_data['scattering_length'].data,
self.material_data['reemission_prob'].data,
self.material_data['reemission_cdf'].data,
np.int32(len(geometry.unique_surfaces)),
self.surface_data['detect'].data,
self.surface_data['absorb'].data,
self.surface_data['reemit'].data,
self.surface_data['reflect_diffuse'].data,
self.surface_data['reflect_specular'].data,
self.surface_data['eta'].data, self.surface_data['k'].data,
self.surface_data['reemission_cdf'].data,
self.surface_data['model'].data,
self.surface_data['transmissive'].data,
self.surface_data['thickness'].data,
self.surface_data['nplanes'].data,
self.surface_data['wire_diameter'].data,
self.surface_data['wire_pitch'].data,
self.world_origin.data, self.world_scale,
np.int32(len(self.nodes)), self.material_data['n'],
self.material_data['step'],
self.material_data["wavelength0"])
cl_queue.finish()
self.material_codes.get()
raise RuntimeError('bail')
if print_usage:
self.print_device_usage(cl_context=cl_context)
log.info(self.device_usage_str(cl_context=cl_context))
self.metadata('g', "after geometry struct")
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 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
