randomFloat(),
randomFloat()],
rho=randomFloat(),
pre=randomFloat(),
temp=[randomFloat(),
randomFloat(),
randomFloat()])
def generateParticles(n):
return [generateTestParticle() for i in xrange(n)]
# Test 1 - increment all particle property values by 1
particles_test_data = generateParticles(32)
gpu_particles = ParticleGPUInterface(particles_test_data)
gpu_particles.run_cuda_function("increment_particle_properties")
updated_particle = gpu_particles.getResultsFromDevice()[0]
np.testing.assert_almost_equal(updated_particle.id,
particles_test_data[0].id + 1,
required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.mass,
particles_test_data[0].mass + 1,
required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.pos,
particles_test_data[0].pos + 1,
required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.vel,
particles_test_data[0].vel + 1,
required_decimal_accuray)
required_decimal_accuray = 2
def randomFloat():
return random.uniform(1, 1e3)
def generateTestParticle():
return Particle(random.randint(1, 10000), randomFloat(), randomFloat(), randomFloat(), randomFloat(), randomFloat(), randomFloat(), randomFloat(),
acc=[randomFloat(), randomFloat(), randomFloat()], rho=randomFloat(), pre=randomFloat(), temp=[randomFloat(), randomFloat(), randomFloat()])
def generateParticles(n):
return [generateTestParticle() for i in xrange(n)]
# Test 1 - increment all particle property values by 1
particles_test_data = generateParticles(32)
gpu_particles = ParticleGPUInterface(particles_test_data)
gpu_particles.run_cuda_function("increment_particle_properties")
updated_particle = gpu_particles.getResultsFromDevice()[0]
np.testing.assert_almost_equal(updated_particle.id, particles_test_data[0].id + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.mass, particles_test_data[0].mass + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.pos, particles_test_data[0].pos + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.vel, particles_test_data[0].vel + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.acc, particles_test_data[0].acc + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.pressure, particles_test_data[0].pressure + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.rho, particles_test_data[0].rho + 1, required_decimal_accuray)
np.testing.assert_almost_equal(updated_particle.temp, particles_test_data[0].temp + 1, required_decimal_accuray)
# Test 2 - Same as test 1, but apply +1 to all properties on multiple particles. Tests thread access index
def sim(particles, bound, kernel, maxiter, pnum, smooth, t_norm, x_norm, interval, savefile, timestep, mode, verbosity):
t = 0.0 # elapsed time
if(kernel == "gaussian"):
CHOOSE_KERNEL_CONST = 1
else:
CHOOSE_KERNEL_CONST = 0
if mode == "parallel":
import pycuda.driver as cuda
import pycuda.autoinit
import numpy
from pycuda.compiler import SourceModule
from gpu_interface import ParticleGPUInterface
'''
So we can do this one of two ways.
1) Keep only one copy of the system in memory.
This is what is implemented here. This does not require
data duplication, but necessitates two iterations through
the list.
2) Keep two lists. One for t = t_N, one for t = t_(N+1).
This way we can remove a for-loop, but requires more memory.
Luckily, it is likely these problems only affect the serial algorithm.
'''
# output-100.csv = prefix + interval + file extension
ary = savefile.split(".") # only split savefile once ([0]=prefix, [1]=extension)
save = 0
if mode == "parallel":
# init gpu interface, pass particles
gpu_particles = ParticleGPUInterface(particles)
print "[+] Saved @ iterations: ",
while(t < (maxiter*timestep)):
# print "t={}".format(t)
if (save*interval) == t:
# print "saving file"
fname = "%s-%d.%s" % (ary[0], int(t), ary[1])
save += 1 # bump save counter
string = "\b%d..." % int(t) # '\b' prints a backspace character to remove previous space
print string,
if mode == "parallel":
particles = gpu_particles.getResultsFromDevice()
saveParticles(particles, fname, verbosity)
# main simulation loop
if mode == "parallel":
gpu_particles.run_cuda_function('run_simulation_loops', (numpy.int32(timestep), numpy.float32(smooth), numpy.int32(CHOOSE_KERNEL_CONST)))
# Transfer the results back to CPU
# Just for testing, this should not be done here
else:
# first sim loop (could use a better name, but I have no idea what this loop is doing)
for p in particles:
# preemptively start the Velocity Verlet computation (first half of velocity update part)
p.vel += (timestep/2.0) * p.acc
p.temp = p.acc
p.acc = 0.0
p.rho = 0.0
p.pressure = 0.0
#get density
for q in particles:
p.rho += q.mass * find_kernel(CHOOSE_KERNEL_CONST, p.pos - q.pos, smooth)
# while we're iterating, add contribution from gravity
if(p.id != q.id):
p.acc += Newtonian_gravity(p,q)
# normalize density
p.rho = p.rho / len(particles)
p.pressure = pressure(p)
# second sim loop
for p in particles:
# acceleration from pressure gradient
for q in particles:
if p.id != q.id:
p.acc -= ( q.mass * ((p.pressure / (p.rho ** 2)) + (q.pressure / (q.rho ** 2))) * del_kernel(CHOOSE_KERNEL_CONST, p.pos - q.pos, smooth) ) * (1 / (np.linalg.norm(p.pos - q.pos))) * (p.pos - q.pos)
# finish velocity update
p.vel += (timestep/2.0) * p.acc
'''
Velocity Verlet integration: Works only assuming force is velocity-independent
http://en.wikipedia.org/wiki/Verlet_integration#Velocity_Verlet
'''
# iterate AGAIN to do final position updates
# save particles list to temporary holder - ensures we have consistent indexing throughout for loop
# tempp = particles
# third sim loop
for p in particles:
# perform position update
p.pos += timestep * (p.vel + (timestep/2.0)*p.temp)
# if np.linalg.norm(p.pos) > bound:
# print "Particle %d position: %f out of range at iteration %d" % (p.id, np.linalg.norm(p.pos), int(t))
# tempp.remove(p)
# particles = tempp
t += timestep # advance time
if mode == "parallel":
particles = gpu_particles.getResultsFromDevice()
# Always save the last interval
print "\b%d\n" % int(t)
fname = "%s-%d.%s" % (ary[0], int(t), ary[1])
saveParticles(particles, fname, verbosity)
# returns the last t-value, which is useful for displaying total iterations
# Also returns the final updated particles
return (t, particles)
