def time_permute(a_d, permutation, b_d=None, impl=None):
start = cuda.Event()
end = cuda.Event()
start.record()
for i in range(100):
permute(a_d, permutation, b_d, impl)
end.record()
end.synchronize()
time_taken = start.time_till(end)/100000.
return time_taken
from __future__ import print_function
import numpy as np
from pycuda import autoinit
import pycuda.gpuarray as gpuarray
from numpy.testing import assert_allclose
from cuper import permute
import itertools
for impl in ['naive', 'cuTranspose']:
for permutation in itertools.permutations((0, 1, 2)):
a = np.random.rand(32, 32, 32)
b = np.zeros((32, 32, 32), dtype=np.float64)
a_d = gpuarray.to_gpu(a)
b_d = gpuarray.to_gpu(b)
print("Testing {} implementation for permutation {} ...".format(impl, permutation), end="")
permute(a_d, permutation, b_d, impl=impl)
try:
assert_allclose(b_d.get(), a.transpose(permutation).copy())
print("Success.")
except AssertionError:
print(".................Failure.")
for permutation in itertools.permutations((0, 1, 2)):
a = np.random.rand(32, 32, 32)
b = np.zeros((32, 32, 32), dtype=np.float64)
a_d = gpuarray.to_gpu(a)
b_d = gpuarray.to_gpu(b)
print("Testing cuTranspose (inplace) implementation for permutation {} ...".format(permutation), end="")
permute(a_d, permutation, impl="cuTranspose")
try:
assert_allclose(a_d.get(), a.transpose(permutation).copy())