import numpy as np import numpy.ma as ma from pycuda.tools import make_default_context from pycuda.compiler import SourceModule from pycuda.characterize import sizeof import pycuda.driver as cuda from pycuda import gpuarray from copy import copy from itertools import izip from tools import timeit import src from geometry import standard_wavelengths from color import map_to_color import sys cuda.init() def boolean_argsort(condition): '''Returns two arrays of indicies indicating which elements of the boolean array `condition` should be exchanged so that all of the True entries occur before the false entries. Returns: tuple (a,b, ntrue) which give the indices for elements to exchange in a and b, and the number of true entries in the array in ntrue. This function in general requires fewer swaps than numpy.argsort. ''' true_indices = condition.nonzero()[0][::-1] # reverse false_indices = (~condition).nonzero()[0] length = min(len(true_indices), len(false_indices)) cut_index = np.searchsorted((true_indices[:length] - false_indices[:length]) <= 0, True) return (true_indices[:cut_index], false_indices[:cut_index], len(true_indices)) def chunk_iterator(nelements, nthreads_per_block, max_blocks): '''Iterator that yields tuples with the values requried to process a long array in multiple kernel passes on the GPU. Each yielded value is (first_index, elements_this_iteration, nblocks_this_iteration) >>> list(chunk_iterator(300, 32, 2)) [(0, 64, 2), (64, 64, 2), (128, 64, 2), (192, 64, 2), (256, 9, 1)] ''' first = 0 while first < nelements: elements_left = nelements - first blocks = int(elements_left // nthreads_per_block) if elements_left % nthreads_per_block != 0: blocks += 1 #Round up only if needed blocks = min(max_blocks, blocks) elements_this_round = min(elements_left, blocks * nthreads_per_block) yield (first, elements_this_round, blocks) first += elements_this_round def format_size(size): if size < 1e3: return '%.1f%s' % (size, ' ') elif size < 1e6: return '%.1f%s' % (size/1e3, 'K') elif size < 1e9: return '%.1f%s' % (size/1e6, 'M') else: return '%.1f%s' % (size/1e9, 'G') def format_array(name, array): return '%-15s %6s %6s' % \ (name, format_size(len(array)), format_size(array.nbytes)) class CUDAFuncs(object): '''simple container class for GPU functions as attributes''' def __init__(self, cuda_module, func_names): '''Extract __global__ functions listed in func_names from the PyCUDA module object. The functions are assigned to attributes of this object with the same name.''' for name in func_names: setattr(self, name, cuda_module.get_function(name)) class GPU(object): def __init__(self, device_id=None): '''Initialize a GPU context on the specified device. If device_id==None, the default device is used.''' if device_id is None: self.context = make_default_context() else: print 'device_id = %s' % device_id device = cuda.Device(device_id) self.context = device.make_context() print 'device %s' % self.context.get_device().name() self.context.set_cache_config(cuda.func_cache.PREFER_L1) cuda_options = ['-I' + src.dir, '--use_fast_math', '--ptxas-options=-v'] self.module = SourceModule(src.kernel, options=cuda_options, no_extern_c=True) self.geo_funcs = CUDAFuncs(self.module, ['set_wavelength_range', 'set_material', 'set_surface', 'set_global_mesh_variables', 'color_solids']) self.prop_funcs = CUDAFuncs(self.module, ['init_rng', 'propagate', 'swap']) self.nthread_per_block = 64 self.max_blocks = 1024 self.daq_module = SourceModule(src.daq, options=cuda_options, no_extern_c=True) self.daq_funcs = CUDAFuncs(self.daq_module, ['reset_earliest_time_int', 'run_daq', 'convert_sortable_int_to_float']) def print_device_usage(self): print 'device usage:' print format_array('vertices', self.vertices_gpu) print format_array('triangles', self.triangles_gpu) print format_array('lower_bounds', self.lower_bounds_gpu) print format_array('upper_bounds', self.upper_bounds_gpu) print format_array('node_map', self.node_map_gpu) print format_array('node_map_end', self.node_map_end_gpu) print '%-15s %6s %6s' % ('total', '', format_size(self.vertices_gpu.nbytes + self.triangles_gpu.nbytes + self.lower_bounds_gpu.nbytes + self.upper_bounds_gpu.nbytes + self.node_map_gpu.nbytes + self.node_map_end_gpu.nbytes)) def load_geometry(self, geometry): if not hasattr(geometry, 'mesh'): print 'WARNING: building geometry with 8-bits' geometry.build(bits=8) self.geo_funcs.set_wavelength_range(np.float32(standard_wavelengths[0]), np.float32(standard_wavelengths[-1]), np.float32(standard_wavelengths[1]-standard_wavelengths[0]), np.uint32(standard_wavelengths.size), block=(1,1,1), grid=(1,1)) self.materials = [] for i in range(len(geometry.unique_materials)): material = copy(geometry.unique_materials[i]) if material is None: raise Exception('one or more triangles is missing a material.') material.refractive_index_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, material.refractive_index[:,0], material.refractive_index[:,1]).astype(np.float32)) material.absorption_length_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, material.absorption_length[:,0], material.absorption_length[:,1]).astype(np.float32)) material.scattering_length_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, material.scattering_length[:,0], material.scattering_length[:,1]).astype(np.float32)) self.geo_funcs.set_material(np.int32(i), material.refractive_index_gpu, material.absorption_length_gpu, material.scattering_length_gpu, block=(1,1,1), grid=(1,1)) self.materials.append(material) self.surfaces = [] for i in range(len(geometry.unique_surfaces)): surface = copy(geometry.unique_surfaces[i]) if surface is None: continue surface.detect_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, surface.detect[:,0], surface.detect[:,1]).astype(np.float32)) surface.absorb_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, surface.absorb[:,0], surface.absorb[:,1]).astype(np.float32)) surface.reflect_diffuse_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, surface.reflect_diffuse[:,0], surface.reflect_diffuse[:,1]).astype(np.float32)) surface.reflect_specular_gpu = gpuarray.to_gpu(np.interp(standard_wavelengths, surface.reflect_specular[:,0], surface.reflect_specular[:,1]).astype(np.float32)) self.geo_funcs.set_surface(np.int32(i), surface.detect_gpu, surface.absorb_gpu, surface.reflect_diffuse_gpu, surface.reflect_specular_gpu, block=(1,1,1), grid=(1,1)) self.surfaces.append(surface) self.vertices_gpu = gpuarray.to_gpu(geometry.mesh.vertices.astype(np.float32).view(gpuarray.vec.float3)) triangles = \ np.empty(len(geometry.mesh.triangles), dtype=gpuarray.vec.uint4) triangles['x'] = geometry.mesh.triangles[:,0] triangles['y'] = geometry.mesh.triangles[:,1] triangles['z'] = geometry.mesh.triangles[:,2] triangles['w'] = ((geometry.material1_index & 0xff) << 24) | ((geometry.material2_index & 0xff) << 16) | ((geometry.surface_index & 0xff) << 8) self.triangles_gpu = gpuarray.to_gpu(triangles) lower_bounds_float3 = np.empty(geometry.lower_bounds.shape[0], dtype=gpuarray.vec.float3) lower_bounds_float3['x'] = geometry.lower_bounds[:,0] lower_bounds_float3['y'] = geometry.lower_bounds[:,1] lower_bounds_float3['z'] = geometry.lower_bounds[:,2] self.lower_bounds_gpu = gpuarray.to_gpu(lower_bounds_float3) upper_bounds_float3 = np.empty(geometry.upper_bounds.shape[0], dtype=gpuarray.vec.float3) upper_bounds_float3['x'] = geometry.upper_bounds[:,0] upper_bounds_float3['y'] = geometry.upper_bounds[:,1] upper_bounds_float3['z'] = geometry.upper_bounds[:,2] self.upper_bounds_gpu = gpuarray.to_gpu(upper_bounds_float3) self.colors_gpu = gpuarray.to_gpu(geometry.colors.astype(np.uint32)) self.node_map_gpu = gpuarray.to_gpu(geometry.node_map.astype(np.uint32)) self.node_map_end_gpu = gpuarray.to_gpu(geometry.node_map_end.astype(np.uint32)) self.solid_id_map_gpu = gpuarray.to_gpu(geometry.solid_id.astype(np.uint32)) self.geo_funcs.set_global_mesh_variables(self.triangles_gpu, self.vertices_gpu, self.colors_gpu, np.uint32(geometry.node_map.size-1), np.uint32(geometry.first_node), self.lower_bounds_gpu, self.upper_bounds_gpu, block=(1,1,1), grid=(1,1)) self.node_map_tex = self.module.get_texref('node_map') self.node_map_end_tex = self.module.get_texref('node_map_end') self.node_map_tex.set_address(self.node_map_gpu.gpudata, self.node_map_gpu.nbytes) self.node_map_end_tex.set_address(self.node_map_end_gpu.gpudata, self.node_map_end_gpu.nbytes) self.node_map_tex.set_format(cuda.array_format.UNSIGNED_INT32, 1) self.node_map_end_tex.set_format(cuda.array_format.UNSIGNED_INT32, 1) self.geometry = geometry print 'average of %f child nodes per node' % (np.mean(geometry.node_map_end[geometry.first_node:] - geometry.node_map[geometry.first_node:])) print '%i nodes with one child' % (np.count_nonzero((geometry.node_map_end[geometry.first_node:] - geometry.node_map[geometry.first_node:]) == 1)) print '%i leaf nodes with one child' % (np.count_nonzero((geometry.node_map_end[:geometry.first_node] - geometry.node_map[:geometry.first_node]) == 1)) self.print_device_usage() def reset_colors(self): self.colors_gpu.set_async(self.geometry.colors.astype(np.uint32)) def color_solids(self, solid_hit, colors): solid_hit_gpu = gpuarray.to_gpu(np.array(solid_hit, dtype=np.bool)) solid_colors_gpu = gpuarray.to_gpu(np.array(colors, dtype=np.uint32)) for first_triangle, triangles_this_round, blocks in chunk_iterator(self.triangles_gpu.size, self.nthread_per_block, self.max_blocks): self.geo_funcs.color_solids(np.int32(first_triangle), np.int32(triangles_this_round), self.solid_id_map_gpu, solid_hit_gpu, solid_colors_gpu, block=(self.nthread_per_block,1,1), grid=(blocks,1)) self.context.synchronize() def setup_propagate(self, seed=1): self.rng_states_gpu = cuda.mem_alloc(self.nthread_per_block * self.max_blocks * sizeof('curandStateXORWOW', '#include ')) self.prop_funcs.init_rng(np.int32(self.max_blocks * self.nthread_per_block), self.rng_states_gpu, np.uint64(seed), np.uint64(0), block=(self.nthread_per_block,1,1), grid=(self.max_blocks,1)) #self.context.synchronize() def load_photons(self, pos, dir, pol, wavelength, t0, histories=None, last_hit_triangles=None): '''Load N photons onto the GPU. pos: numpy.array(shape=(N, 3)) of photon starting positions (meters) dir: numpy.array(shape=(N, 3)) of photon starting directions (unit vectors) pol: numpy.array(shape=(N, 3)) of photon polarization directions (unit vectors) wavelength: numpy.array(shape=N) of photon wavelengths (nm) t0: numpy.array(shape=N) of photon start times (s) Optional args will be loaded with defaults on GPU if not set: histories: Bitmask of interactions that have occurred over history of photon last_hit_triangles: The triangle ID number that the photon last interacted with, if any. -1 if no triangle was hit in the last step ''' self.nphotons = len(pos) assert len(dir) == self.nphotons assert len(pol) == self.nphotons assert len(wavelength) == self.nphotons assert len(t0) == self.nphotons self.positions_gpu = gpuarray.to_gpu(pos.astype(np.float32).view(gpuarray.vec.float3)) self.directions_gpu = gpuarray.to_gpu(dir.astype(np.float32).view(gpuarray.vec.float3)) self.polarizations_gpu = gpuarray.to_gpu(pol.astype(np.float32).view(gpuarray.vec.float3)) self.wavelengths_gpu = gpuarray.to_gpu(wavelength.astype(np.float32)) self.times_gpu = gpuarray.to_gpu(t0.astype(np.float32)) if histories is not None: self.histories_gpu = gpuarray.to_gpu(histories.astype(np.uint32)) else: self.histories_gpu = gpuarray.zeros(self.nphotons, dtype=np.uint32) if last_hit_triangles is not None: self.last_hit_triangles_gpu = gpuarray.to_gpu(last_hit_triangles.astype(np.int32)) else: self.last_hit_triangles_gpu = gpuarray.empty(self.nphotons, dtype=np.int32) self.last_hit_triangles_gpu.fill(-1) def propagate(self, max_steps=10): '''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. ''' nphotons = self.nphotons step = 0 input_queue = np.zeros(shape=self.nphotons+1, dtype=np.uint32) input_queue[1:] = np.arange(self.nphotons, dtype=np.uint32) input_queue_gpu = gpuarray.to_gpu(input_queue) output_queue = np.zeros(shape=self.nphotons+1, dtype=np.uint32) output_queue[0] = 1 output_queue_gpu = gpuarray.to_gpu(output_queue) while step < max_steps: ## Just finish the rest of the steps if the # of photons is low if nphotons < self.nthread_per_block * 16 * 8: nsteps = max_steps - step else: nsteps = 1 for first_photon, photons_this_round, blocks in chunk_iterator(nphotons, self.nthread_per_block, self.max_blocks): self.prop_funcs.propagate(np.int32(first_photon), np.int32(photons_this_round), input_queue_gpu[1:], output_queue_gpu, self.rng_states_gpu, self.positions_gpu, self.directions_gpu, self.wavelengths_gpu, self.polarizations_gpu, self.times_gpu, self.histories_gpu, self.last_hit_triangles_gpu, np.int32(nsteps), block=(self.nthread_per_block,1,1), grid=(blocks, 1)) step += nsteps if step < max_steps: temp = input_queue_gpu input_queue_gpu = output_queue_gpu output_queue_gpu = temp output_queue_gpu[:1].set(np.uint32(1)) nphotons = input_queue_gpu[:1].get()[0] if gpuarray.max(self.histories_gpu).get() & (1 << 31): print >>sys.stderr, "WARNING: ABORTED PHOTONS IN THIS EVENT" if 'profile' in __builtins__: self.context.synchronize() def get_photons(self): '''Returns a dictionary of current photon state information. Contents of dictionary have the same names as the parameters to load_photons(). ''' return { 'pos' : self.positions_gpu.get().view(np.float32).reshape(self.positions_gpu.size, 3), 'dir' : self.directions_gpu.get().view(np.float32).reshape(self.directions_gpu.size, 3), 'pol' : self.polarizations_gpu.get().view(np.float32).reshape(self.polarizations_gpu.size, 3), 'wavelength' : self.wavelengths_gpu.get(), 't0' : self.times_gpu.get(), 'histories' : self.histories_gpu.get(), 'last_hit_triangles' : self.last_hit_triangles_gpu.get()} def setup_daq(self, max_pmt_id, pmt_rms=1.2e-9): self.earliest_time_gpu = gpuarray.GPUArray(shape=(max_pmt_id+1,), dtype=np.float32) self.earliest_time_int_gpu = gpuarray.GPUArray(shape=self.earliest_time_gpu.shape, dtype=np.uint32) self.channel_history_gpu = gpuarray.zeros_like(self.earliest_time_int_gpu) self.channel_q_gpu = gpuarray.zeros_like(self.earliest_time_int_gpu) self.daq_pmt_rms = pmt_rms def run_daq(self): self.daq_funcs.reset_earliest_time_int(np.float32(1e9), np.int32(len(self.earliest_time_int_gpu)), self.earliest_time_int_gpu, block=(self.nthread_per_block,1,1), grid=(len(self.earliest_time_int_gpu)//self.nthread_per_block+1,1)) self.channel_q_gpu.fill(0) self.channel_history_gpu.fill(0) #self.context.synchronize() for first_photon, photons_this_round, blocks in chunk_iterator(self.nphotons, self.nthread_per_block, self.max_blocks): self.daq_funcs.run_daq(self.rng_states_gpu, np.uint32(0x1 << 2), np.float32(self.daq_pmt_rms), np.int32(first_photon), np.int32(photons_this_round), self.times_gpu, self.histories_gpu, self.last_hit_triangles_gpu, self.solid_id_map_gpu, np.int32(len(self.earliest_time_int_gpu)), self.earliest_time_int_gpu, self.channel_q_gpu, self.channel_history_gpu, block=(self.nthread_per_block,1,1), grid=(blocks,1)) #self.context.synchronize() self.daq_funcs.convert_sortable_int_to_float(np.int32(len(self.earliest_time_int_gpu)), self.earliest_time_int_gpu, self.earliest_time_gpu, block=(self.nthread_per_block,1,1), grid=(len(self.earliest_time_int_gpu)//self.nthread_per_block+1,1)) if 'profile' in __builtins__: self.context.synchronize() def get_hits(self): return { 't': self.earliest_time_gpu.get(), 'q': self.channel_q_gpu.get().astype(np.float32), 'history': self.channel_history_gpu.get()} def __del__(self): self.context.pop()