def dirichlet(a, times=1):
a = np.array(a)
normalize_scale = [
a.shape[-1], a.shape[-1] * times,
int(a.size / a.shape[-1]) * times
]
normal_block = int(500)
normal_grid = int(normalize_scale[2] / normal_block + 1)
normalize_scale = gpuarray.to_gpu(np.array(normalize_scale,
dtype=np.int32))
a, b, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.float32, np.float32, a, 1.0)
func = Sampler.get_function('rand_Gamma')
func(randomseed,
output,
matrix_scale,
a,
b,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
func = Sampler.get_function('normalize')
func(output,
matrix_scale,
normalize_scale,
grid=(normal_grid, 1, 1),
block=(normal_block, 1, 1))
output = output.get()
return output
def chisquare(n, times=1, device='cpu'):
n, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.int32, np.float32, n)
func = Sampler.get_function('rand_Chisquare')
func(randomseed,
output,
matrix_scale,
n,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
if device == 'cpu':
output = output.get()
if single_number:
return output[0]
return output
def exponential(Lambda, times=1, device='cpu'):
Lambda, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.float32, np.float32, Lambda)
func = Sampler.get_function('rand_Exponential')
func(randomseed,
output,
matrix_scale,
Lambda,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
if device == 'cpu':
output = output.get()
if single_number:
return output[0]
return output
def hypergeometric(ng, nb, ns, times=1, device='cpu'):
ng, nb, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.float32, np.int32, ng, nb)
ns = gpuarray.to_gpu(np.array(ns, dtype=np.int32, order='C'))
prob = ng / (ng + nb)
func = Sampler.get_function('rand_Hypergeometric')
func(randomseed,
output,
matrix_scale,
prob,
ns,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
if device == 'cpu':
output = output.get()
if single_number:
return output[0]
return output
def noncentral_chisquare(n, delta, times=1, var=1, device='cpu'):
# n,delta,var must be same shape
n = np.array(n, dtype=np.int32, order='C')
delta, var, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.float32, np.float32, delta, var)
func = Sampler.get_function('rand_Noncentral_Chisquare')
func(randomseed,
output,
matrix_scale,
n,
delta,
var,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
if device == 'cpu':
output = output.get()
if single_number:
return output[0]
return output
def triangular(left=0, mid=1, right=1, times=1, device='cpu'):
left, output, matrix_scale, randomseed, partition, single_number = para_preprocess(
times, np.float32, np.float32, left)
mid = gpuarray.to_gpu(np.array(mid, dtype=np.float32, order='C'))
right = gpuarray.to_gpu(np.array(right, dtype=np.float32, order='C'))
func = Sampler.get_function('rand_Triangular')
func(randomseed,
output,
matrix_scale,
left,
mid,
right,
grid=(partition[0], 1, 1),
block=(partition[1], 1, 1))
if device == 'cpu':
output = output.get()
if single_number:
return output[0]
return output