def local_search(points, weights, distances, upper_bound, lower_bound,
random_states):
thread_id = cuda.threadIdx.x
if thread_id < len(points):
pass
# TODO rozmiar tej tablicy musi byc stalą czasu kompilacji , wtf
tmp_point = cuda.local.array(20, float64)
for i in range(20):
tmp_point[i] = points[thread_id][i]
for index in range(20):
direction = xoroshiro128p_uniform_float32(random_states,
thread_id) > 0.5
step = xoroshiro128p_uniform_float32(random_states, thread_id)
if direction:
length = upper_bound - tmp_point[index]
tmp_point[index] = tmp_point[index] + step * length
else:
length = tmp_point[index] - lower_bound
tmp_point[index] = tmp_point[index] - step * length
val1 = qap_device(tmp_point, weights, distances)
val2 = qap_device(points[thread_id], weights, distances)
if val1 < val2:
for i in range(20):
points[thread_id][i] = tmp_point[i]
break
def loop(_, r, V, rngs, w, d, tavg, bold_state, bold_out, I, Delta, eta,
tau, J, cr, cv, r_sigma, V_sigma):
it = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.x
nt = cuda.blockDim.x * cuda.gridDim.x
itx = cuda.threadIdx.x
# if it==0: print('hello from ', cuda.blockIdx.x, cuda.threadIdx.x)
# if it==0: print("NT =", NT)
o_tau = nb.float32(1 / tau)
# if it==0: print("o_tau = ", o_tau)
assert r.shape[0] == V.shape[
0] == nh # shape asserts help numba optimizer
assert r.shape[1] == V.shape[1] == nn
# if it==0: print("creating nrV shared..")
nrV = cuda.shared.array((2, blockDim_x), nb.float32)
# if it==0: print("zeroing tavg..")
for j in range(nto):
for i in range(nn):
tavg[j, 0, i, it] = nb.float32(0.0)
tavg[j, 1, i, it] = nb.float32(0.0)
# if it==0: print('tavg zero\'d', -1, nh - 1)
for t0 in range(-1, nh - 1):
# if it==0: print('t0=', t0)
t = nh - 1 if t0 < 0 else t0
# if it==0: print('t=', t)
t1 = t0 + 1
# if it==0: print('t1=', t1)
# if it==0: print('nh//nto', nh // nto)
# if it==0: print('t1=', t1)
t0_nto = t0 // (nh // nto)
# if it==0: print(t, t1, t0_nto)
for i in range(nn):
rc = nb.float32(0) # using array here costs 50%+
Vc = nb.float32(0)
for j in range(nn):
dij = (t - d[i, j] + nh) & (nh - 1)
rc += w[i, j] * cfpre(r[dij, j, it], r[t, i, it])
Vc += w[i, j] * cfpre(V[dij, j, it], V[t, i, it])
rc = cfpost(rc)
Vc = cfpost(Vc)
# RNG + Box Muller
pi_2 = nb.float32(np.pi * 2)
u1 = xoroshiro128p_uniform_float32(
rngs, t1 * nt * nn * 2 + i * nt * 2 + it)
u2 = xoroshiro128p_uniform_float32(
rngs, t1 * nt * nn * 2 + i * nt * 2 + it + nt)
z0 = math.sqrt(-nb.float32(2.0) * math.log(u1)) * math.cos(
pi_2 * u2)
z1 = math.sqrt(-nb.float32(2.0) * math.log(u1)) * math.sin(
pi_2 * u2)
# RK4
rk4_rV(nrV, r[t, i, it], V[t, i, it], o_tau, pi, tau, Delta,
eta, J, I, cr, rc, cv, Vc, r_sigma, V_sigma, z0, z1)
r[t1, i, it] = nrV[0, itx]
V[t1, i, it] = nrV[1, itx]
# if it==0: print(nrV[0, it], nrV[1, it], o_nh)
tavg[t0_nto, 0, i, it] += nrV[0, itx] * o_nh
tavg[t0_nto, 1, i, it] += nrV[1, itx] * o_nh
# if it==0: print(t1, o_nh, tavg[t0_nto, 0, i, it], tavg[t0_nto, 1, i, it])
bold_out[i, it] = fmri_gpu(it, bold_state[i], nrV[0, itx], dt)
def mc_pi(states, iterations, out):
tid = cuda.grid(1)
inside = 0
for i in range(iterations):
x = xoroshiro128p_uniform_float32(states, tid)
y = xoroshiro128p_uniform_float32(states, tid)
if x**2 + y**2 <= 1.0:
inside += 1
out[tid] = 4.0 * inside / iterations
def simulate_pi(rng_states, iterations, out):
thread_id = cuda.grid(1)
inside = 0
for i in range(iterations):
x = xoroshiro128p_uniform_float32(rng_states, thread_id)
y = xoroshiro128p_uniform_float32(rng_states, thread_id)
if x**2 + y**2 <= 1.0:
inside += 1
out[thread_id] = 4.0 * inside / iterations
def compute_pi(rng_states, iterations, out):
"""Find the maximum value in values and store in result[0]"""
thread_id = cuda.grid(1)
# Compute pi by drawing random (x, y) points and finding what
# fraction lie inside a unit circle
inside = 0
for i in range(iterations):
x = xoroshiro128p_uniform_float32(rng_states, thread_id)
y = xoroshiro128p_uniform_float32(rng_states, thread_id)
if x**2 + y**2 <= 1.0:
inside += 1
out[thread_id] = 4.0 * inside / iterations
def compute_pi(rng_states, iterations, out):
"""Find the maximum value in values and store in result[0]"""
thread_id = cuda.grid(1)
# Compute pi by drawing random (x, y) points and finding what
# fraction lie inside a unit circle
inside = 0
for i in range(iterations):
x = xoroshiro128p_uniform_float32(rng_states, thread_id)
y = xoroshiro128p_uniform_float32(rng_states, thread_id)
if x**2 + y**2 <= 1.0:
inside += 1
out[thread_id] = 4.0 * inside / iterations
def mutate_kernel(d_next_gen, d_is_elite, rng_states, mutate_prob):
r"""Perform mutation by randomly swapping two CPs.
"""
i = cuda.grid(1)
if i < d_next_gen.shape[0]:
if d_is_elite[i] == False:
rnd = xoroshiro128p_uniform_float32(rng_states, i)
if rnd < mutate_prob:
rnd = xoroshiro128p_uniform_float32(rng_states, i)
idx1 = int(math.floor(rnd * d_next_gen.shape[1]))
rnd = xoroshiro128p_uniform_float32(rng_states, i)
idx2 = int(math.floor(rnd * d_next_gen.shape[1]))
tmp = d_next_gen[i, idx1]
d_next_gen[i, idx1] = d_next_gen[i, idx2]
d_next_gen[i, idx2] = tmp
def Busqueda_MetropolisCUDA(M,individuo,probabilidades,AristMono,numColores,numNodos,rng_states,id):
if AristMono != 0:
for i in prange(busqueda_vecindario):
nodo = 0
vacia = 0
#print(i)
while vacia == 0:
bolsaProbabilistica = int(bolsaProbabilidad_gpu(probabilidades,numColores,rng_states,id)) # Elige una bolsa con probabilidad a su número de nodos
#print(bolsaProbabilistica)
vacia = esVacia_gpu(individuo[bolsaProbabilistica],numNodos) #verifica que la bolsa no esté vacía
#print(id, vacia)
while nodo == 0: #Selecciona un nodo al azar
r = int(xoroshiro128p_uniform_float32(rng_states, id)*numNodos) #selecciona un número al azar del cero al número de nodos
nodo = individuo[bolsaProbabilistica][r]
monoAct = NAMBolsa_gpu(individuo[bolsaProbabilistica], r, M,numNodos) # calcula el número de aristas monocromáticas del nodo en la bolsa elegida
#print(id, r, nodo, bolsaProbabilistica,monoAct)
BolsaNueva = bolsaAleatoria_gpu(bolsaProbabilistica, numColores,rng_states,id)
#print(bolsaProbabilistica, BolsaNueva, r)
monopost = NAMBolsa_gpu(individuo[BolsaNueva], r, M,numNodos)
delta = monopost - monoAct
#print(delta, AristMono)
if probAcepta_gpu(delta,rng_states,id):
#print("acepta")
AristMono = AristMono + delta
#print(AristMono)
individuo[bolsaProbabilistica][nodo] = 0 # Elimina el nodo de la bolsa
individuo[BolsaNueva][nodo] = 1 # Inserta el nodo en otra bolsa al azar
if AristMono == 0:
break
def Busqueda_EscalandoCUDA(M, individuo, probabilidades, AristMono, numColores,
numNodos, rng_states, id):
if AristMono != 0:
for i in prange(busqueda_vecindario):
r = 0
vacia = 0
while vacia == 0:
bolsaProbabilistica = int(
bolsaProbabilidad_gpu(probabilidades, numColores,
rng_states, id)
) # Elige una bolsa con probabilidad a su número de nodos
vacia = esVacia_gpu(
individuo[bolsaProbabilistica],
numNodos) #verifica que la bolsa no esté vacía
while r == 0: #Selecciona un nodo al azar
nodo = int(
xoroshiro128p_uniform_float32(rng_states, id) * numNodos
) #selecciona un número al azar del cero al número de nodos
r = individuo[bolsaProbabilistica][nodo]
monoAct = NAMBolsa_gpu(
individuo[bolsaProbabilistica], nodo, M, numNodos
) # calcula el número de aristas monocromáticas del nodo en la bolsa elegida
BolsaNueva = bolsaAleatoria_gpu(bolsaProbabilistica, numColores,
rng_states, id)
monopost = NAMBolsa_gpu(individuo[BolsaNueva], nodo, M, numNodos)
delta = monopost - monoAct
if monopost > monoAct:
AristMono = AristMono + delta
individuo[bolsaProbabilistica][
nodo] = 0 # Elimina el nodo de la bolsa
individuo[BolsaNueva][
nodo] = 1 # Inserta el nodo en otra bolsa al azar
if AristMono == 0:
break
def packet(rng_states, thread_id, q):
x = 0
y = False
while y == False:
y = (xoroshiro128p_uniform_float32(rng_states, thread_id) > q)
x += 1
return x
def cross_over_1p(parent1, parent2, rng_states, d_pop, d_next_gen, i):
r"""Perform 1-point crossover. Copy a portion of parent1, then fill the
rest with parent2.
"""
m = d_pop.shape[1]
rnd = xoroshiro128p_uniform_float32(rng_states, i)
split = int(math.floor(rnd * m))
# copy from parent1
for j in range(split):
d_next_gen[i, j] = d_pop[parent1, j]
# copy from parent2
idx = split
for j in range(m):
cp2 = d_pop[parent2, j]
repeat = False
for k in range(split):
cp1 = d_next_gen[i, k]
if cp1 == cp2:
repeat = True
break
if repeat == False:
d_next_gen[i, idx] = cp2
idx += 1
if idx == m:
break
def scatter2(threadindex, rng_states, neutron):
"isotropic scattering kernel with a uniform sampling of the polar angle and azimuthal angle"
# randomly pick direction
theta = xoroshiro128p_uniform_float32(rng_states, threadindex) * pi
phi = xoroshiro128p_uniform_float32(rng_states, threadindex) * (2 * pi)
cos_t, sin_t = cos(theta), sin(theta)
sin_p, cos_p = sin(phi), cos(phi)
# compute velocity
vx, vy, vz = neutron[3:6]
vi = sqrt(vx * vx + vy * vy + vz * vz)
vx = vi * sin_t * cos_p
vy = vi * sin_t * sin_p
vz = vi * cos_t
neutron[3:6] = vx, vy, vz
neutron[-1] *= sin_t * (pi / 2)
return
def tournament(rng_states, i, pop_size, d_nonelite, num_elites, d_fitness_all,
tournament_size):
r"""Randomly choose candidates from nonelite individuals then choose the
best as one parent.
"""
rnd = xoroshiro128p_uniform_float32(rng_states, i)
num_nonelite = d_nonelite.size
parent = d_nonelite[int(math.floor(rnd * num_nonelite))]
min_fitness = d_fitness_all[parent]
for j in range(tournament_size - 1):
rnd = xoroshiro128p_uniform_float32(rng_states, i)
new_parent = parent = d_nonelite[int(math.floor(rnd * num_nonelite))]
if min_fitness < d_fitness_all[new_parent]:
parent = new_parent
min_fitness = d_fitness_all[new_parent]
return parent
def scatter(threadindex, rng_states, neutron):
"isotropic scattering kernel with a uniform sampling of 4pi solid angle"
# randomly pick direction
cos_t = xoroshiro128p_uniform_float32(rng_states, threadindex) * 2 - 1
phi = xoroshiro128p_uniform_float32(rng_states, threadindex) * (2 * pi)
if cos_t > 1: cos_t = 1
sin_t = sqrt(1 - cos_t * cos_t)
sin_p, cos_p = sin(phi), cos(phi)
# compute velocity
vx, vy, vz = neutron[3:6]
vi = sqrt(vx * vx + vy * vy + vz * vz)
vx = vi * sin_t * cos_p
vy = vi * sin_t * sin_p
vz = vi * cos_t
neutron[3:6] = vx, vy, vz
return
def find_spread_gpu(graph, active, new_active, new_ones, mc, p, rng_states):
# Get abosolute position of current thread
thread_id = cuda.grid(1)
# Because of fixed block sizes, some of the threads won't be needed
if thread_id >= mc:
return
done = False
while not done:
done = True
for j in range(new_active[thread_id].shape[0]):
if new_active[thread_id][j]:
for k in range(graph.shape[0]):
if graph[j][k] and p > xoroshiro128p_uniform_float32(
rng_states, thread_id):
new_ones[thread_id][k] = True
for j in range(new_active[thread_id].shape[0]):
if new_ones[thread_id][j] and (not active[thread_id][j]):
active[thread_id][j] = True
new_active[thread_id][j] = True
done = False
else:
new_active[thread_id][j] = False
new_ones[thread_id][j] = False
def move(rng_states, start_x, start_y, out_x, out_y, doms, rs, domhits, domhitstimes):
thread_id = cuda.grid(1)
def rng():
return xoroshiro128p_uniform_float32(rng_states, thread_id)
x = start_x
y = start_y
d = rng()*math.pi*2
vx = math.cos(d)
vy = math.sin(d)
absorbed = False
time = 0
while not absorbed:
if rng() < 0.02:#1:
d = xoroshiro128p_uniform_float32(rng_states, thread_id)*math.pi*2
vx = math.cos(d)
vy = math.sin(d)
if rng() < 0.02:#05:
absorbed = True
x += vx
y += vy
for i in range(len(doms)):
domx = doms[i,0]
domy = doms[i,1]
r = rs[i]
if r >= (math.sqrt((domx-x)**2 + (domy-y)**2)):
domhits[thread_id, i] += 1
domhitstimes[thread_id, i] = time
absorbed = True
time += 1
out_x[thread_id] = x
out_y[thread_id] = y
def rng_kernel_float32(states, out, count, distribution):
thread_id = cuda.grid(1)
for i in range(count):
if distribution == UNIFORM:
out[thread_id * count + i] = xoroshiro128p_uniform_float32(states, thread_id)
elif distribution == NORMAL:
out[thread_id * count + i] = xoroshiro128p_normal_float32(states, thread_id)
def _detect_gpu(matrix, vec, rng_states):
thread_id = cuda.grid(1)
if thread_id < vec.shape[0]:
l = matrix.shape[0]
x = int(xoroshiro128p_uniform_float32(rng_states, thread_id) * l)
y = int(xoroshiro128p_uniform_float32(rng_states, thread_id) * l)
ret = 0
for m in [x, y]:
m_inv = x + y - m
for n in range(l):
if matrix[m, n] > 0 or matrix[m_inv, n] > 0:
if m != n and m_inv != n:
ret += (abs(m - n) - abs(m_inv - n)) * (
matrix[m, n] - matrix[m_inv, n])
if ret > 0:
vec[thread_id, 0] = x
vec[thread_id, 1] = y
def recombination(
inp_weights, out_weights, n_inp_ia, tot_ia, n_weights, rng_states
): #n_inp_ia: number of ia in input, tot_ia: total number of ia to be generated, n_weights: total number of weighta for ia
pos = cuda.grid(1)
if pos < tot_ia:
ia_1 = int(
xoroshiro128p_uniform_float32(rng_states, cuda.grid(1)) * n_inp_ia)
ia_2 = int(
xoroshiro128p_uniform_float32(rng_states, cuda.grid(1)) * n_inp_ia)
cut = int(
xoroshiro128p_uniform_float32(rng_states, cuda.grid(1)) *
n_weights)
for i in range(n_weights):
if i < cut:
out_weights[pos][i] = inp_weights[ia_1][i]
else:
out_weights[pos][i] = inp_weights[ia_2][i]
def sample_kernel(rng_states,weight,old_particle_pos,particle_pos):
tx = int(cuda.threadIdx.x) # this is the unique thread ID within a 1D block
ty = int(cuda.blockIdx.x) # Similarly, this is t
thread_id = cuda.grid(1)
tt=xoroshiro128p_uniform_float32(rng_states,thread_id)
if tt<0.01:
particle_pos[ty][0]=xoroshiro128p_uniform_float32(rng_states,thread_id)*max_x
particle_pos[ty][1]=xoroshiro128p_uniform_float32(rng_states,thread_id)*max_y
else:
t=xoroshiro128p_uniform_float32(rng_states,thread_id)
for i in range(len(weight)):
if t-weight[i]<0:
particle_pos[ty][0]=old_particle_pos[i][0]
particle_pos[ty][1]=old_particle_pos[i][1]
break
else:
t-=weight[i]
def bolsaAleatoriaProbabilidadCUDA (probabilidades, numColores,rng_states,id):
r = xoroshiro128p_uniform_float32(rng_states, id) #selecciona un número al azar del cero al uno
l = 0
for i in range(numColores): #recorre hasta el número de bolsas
if (r >= l and r < l + probabilidades[i]): #si cae entre l y la probabilidad de la bolsa i
return i #retorna el indice i
else:
l = l + probabilidades[i] #si no a la variable l le suma probabilidad de la bolsa i
return i
def mutation(inp_weights, n_ia, n_weights, rng_states, prob=0.1):
# Thread id in a 1D block
tx = cuda.threadIdx.x
# Block id in a 1D grid
ty = cuda.blockIdx.x
if ty < n_ia and tx < n_weights:
a = xoroshiro128p_uniform_float32(rng_states, cuda.grid(1))
if a < prob:
inp_weights[ty][tx] += (xoroshiro128p_normal_float32(
rng_states, cuda.grid(1))) / 5.
def recombination_2(inp_weights, out_weights, n_inp_ia, tot_ia, n_weights,
rng_states):
# Thread id in a 1D block
tx = cuda.threadIdx.x
# Block id in a 1D grid
ty = cuda.blockIdx.x
if ty < tot_ia and tx < n_weights:
ia_rng = int(
xoroshiro128p_uniform_float32(rng_states, cuda.grid(1)) * n_inp_ia)
out_weights[ty][tx] = inp_weights[ia_rng][tx]
def probabilidadAceptarCUDA(delta, rng_states, id):
if delta < 0:
return True
else:
P = math.exp(-delta/k*T)
r = xoroshiro128p_uniform_float32(rng_states, id) # selecciona un número al azar del cero al uno
if r < P:
return True
else:
return False
def crossoverUniform1(popvec_in, mother, father, ii, popvec_out, config_i,
rng_states, tid, tmp):
chr_sz = config_i[cfg.CHROMO_SIZE]
# uniform crossover leading to 1 child
for j in range(0, chr_sz):
if (xoroshiro128p_uniform_float32(rng_states, tid) < 0.50):
popvec_out[ii * chr_sz + j] = popvec_in[mother * chr_sz + j]
else:
popvec_out[ii * chr_sz + j] = popvec_in[father * chr_sz + j]
def initialize_kernel(d_pop_init, rng_states):
r"""Generate random numbers.
"""
i = cuda.grid(1)
pop_size = d_pop_init.shape[0]
m = d_pop_init.shape[1]
if i < pop_size:
for j in range(m):
rnd = xoroshiro128p_uniform_float32(rng_states, i)
d_pop_init[i, j] = rnd
def sample(q_array, input, sigma, rng):
pos = cuda.grid(1)
if pos < q_array.shape[1]:
euler = cuda.local.array(shape=(3), dtype=float32)
for i in range(euler.shape[0]):
euler[i] = input[i] + (xoroshiro128p_uniform_float32(rng, pos) -
0.5)
quaternion_add_euler(q_array[:, pos], euler)
def propagate(
threadindex, rng_states,
in_neutron,
square, width, height, radius,
wl_distr, Lambda0, dLambda, E0, dE,
xw, yh, dist, pmul
):
r1 = xoroshiro128p_uniform_float32(rng_states, threadindex)
r2 = xoroshiro128p_uniform_float32(rng_states, threadindex)
r3 = xoroshiro128p_uniform_float32(rng_states, threadindex)
r4 = xoroshiro128p_uniform_float32(rng_states, threadindex)
r5 = xoroshiro128p_uniform_float32(rng_states, threadindex)
if square:
x = width * (r1 - 0.5)
y = height * (r2 - 0.5)
else:
chi=2*math.pi*r1
r=math.sqrt(r2)*radius
x=r*math.cos(chi)
y=r*math.sin(chi)
in_neutron[:3] = x, y, 0.
# choose final vector
target = cuda.local.array(shape=3, dtype=NB_FLOAT)
target[0] = target[1] = 0.0
target[2] = dist
vec_f = cuda.local.array(shape=3, dtype=NB_FLOAT)
solidangle = randvec_target_rect(target, xw, yh, r3, r4, vec_f)
# vector from moderator to final position is
# (vec_f[0]-x, vec_f[1]-y, dist)
dx = vec_f[0]-x; dy = vec_f[1]-y
dist1 = math.sqrt(dx*dx+dy*dy+dist*dist)
# velocity scalar
if wl_distr:
L = Lambda0+dLambda*(r5*2-1)
v = K2V*(2*math.pi/L)
else:
E = E0+dE*(r5*2-1)
v = SE2V*math.sqrt(E)
in_neutron[3:6] = v*dx/dist1, v*dy/dist1, v*dist/dist1
in_neutron[-2] = 0
in_neutron[-1] = pmul*solidangle
return
def move(points, best_point_index, forces, random_states):
thread_id = cuda.threadIdx.x
if thread_id == best_point_index:
return
step = xoroshiro128p_uniform_float32(random_states, thread_id)
for k in range(20):
points[thread_id][
k] = points[thread_id][k] + step * forces[thread_id][k]
def initial_state(state_array: cuda.devicearry,
initial_state: cuda.devicearray,
state_prob: cuda.devicearray,
rng_states):
tx = cuda.threadIdx.x
ty = cuda.blockIdx.y
bw = cuda.blockDim.x
pos = tx + ty * bw
if pos < state_array.shape[0]:
# for i in range(state_num.shape[0])
for i in range(6):
state_array[i, pos] =initial_state[i] + xoroshiro128p_uniform_float32(rng_states, tx)
def random_3d(arr, rng_states):
# Per-dimension thread indices and strides
startx, starty, startz = cuda.grid(3)
stridex, stridey, stridez = cuda.gridsize(3)
# Linearized thread index
tid = (startz * stridey * stridex) + (starty * stridex) + startx
# Use strided loops over the array to assign a random value to each entry
for i in range(startz, arr.shape[0], stridez):
for j in range(starty, arr.shape[1], stridey):
for k in range(startx, arr.shape[2], stridex):
arr[i, j, k] = xoroshiro128p_uniform_float32(
rng_states, tid)
