def test_elemwise_collapse():
""" Test collapsing under many broadcast and strided pattern """
for dtype1 in ["int16", "float32", "int8"]:
for dtype2 in ["int16", "float32", "int8"]:
for shape1_, shape2_, expected in [
# 1d to test this special case
((40,), (40,), 0),
((40,), (1,), 1),
# No broadcastable dimensions
((4, 5, 6, 9), (4, 5, 6, 9), 0),
# All inputs have one(and the same) broadcastable dimension
((1, 4, 5, 9), (1, 4, 5, 9), 0),
((4, 1, 5, 9), (4, 1, 5, 9), 0),
((4, 5, 1, 9), (4, 5, 1, 9), 0),
((4, 5, 9, 1), (4, 5, 9, 1), 0),
# One inputs have one broadcastable dimension
((1, 5, 6, 9), (4, 5, 6, 9), 2),
((4, 1, 6, 9), (4, 5, 6, 9), 3),
((4, 5, 1, 9), (4, 5, 6, 9), 3),
((4, 5, 6, 1), (4, 5, 6, 9), 2),
# One inputs have two broadcastable dimension
((1, 1, 6, 9), (4, 5, 6, 9), 2),
((1, 5, 1, 9), (4, 5, 6, 9), 4),
((1, 5, 6, 1), (4, 5, 6, 9), 3),
((4, 1, 1, 9), (4, 5, 6, 9), 3),
((4, 1, 6, 1), (4, 5, 6, 9), 4),
((4, 5, 1, 1), (4, 5, 6, 9), 2),
# One inputs have tree broadcastable dimension
((1, 1, 1, 9), (4, 5, 6, 9), 2),
((1, 1, 6, 1), (4, 5, 6, 9), 3),
((1, 5, 1, 1), (4, 5, 6, 9), 3),
((4, 1, 1, 1), (4, 5, 6, 9), 2),
# One scalar
((1, 1, 1, 1), (4, 5, 6, 9), 1),
# One scalar, the other 1 broadcast dims
((1, 1, 1, 1), (4, 5, 6, 1), 1),
]:
scalar_cpu = rand((1,) * len(shape1_), dtype=dtype1)
scalar_gpu = gpu_ndarray.GpuNdArrayObject(scalar_cpu)
scalar_gpu1 = MyGpuNdArray(scalar_gpu)
for shape1, shape2 in [(shape1_, shape2_), (shape2_, shape1_)]:
a_cpu = rand(shape1, dtype=dtype1)
a = gpu_ndarray.GpuNdArrayObject(a_cpu)
a1 = MyGpuNdArray(a)
b_cpu = rand(shape2, dtype=dtype2)
b = gpu_ndarray.GpuNdArrayObject(b_cpu)
b1 = MyGpuNdArray(b)
assert len(shape1) == len(shape2)
o_shape = []
for i in range(len(shape1)):
o_shape.append(max(shape1[i], shape2[i]))
o = gpu_ndarray.empty(o_shape, dtype=(a_cpu + b_cpu).dtype)
# 1.1 Check direct collapse
nd_collaps, info = elemwise_collapses([a, b], [o])
assert nd_collaps == expected, (shape1, shape2,
nd_collaps, expected, info)
# 1.2 Check computation are still valid
f = MyGpuNdArray.gen_fct(theano.tensor.add, [a1, b1],
len(shape1))
out = f([a1, b1])
out2 = f([a1, b1], out=out)
assert out is out2
assert numpy.allclose(numpy.asarray(f([a1, b1])),
a_cpu + b_cpu)
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.adds(a1, b1)), a_cpu + b_cpu)
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.add(a1, b1)), a_cpu + b_cpu)
assert MyGpuNdArray.add(a1, b1, out=out2) is out2
# 1.3 Check work without collaping
f = MyGpuNdArray.gen_fct(theano.tensor.add, [a1, b1],
len(shape1), collapse=False)
out = f([a1, b1])
out2 = f([a1, b1], out=out)
assert out is out2
assert numpy.allclose(numpy.asarray(f([a1, b1])),
a_cpu + b_cpu)
assert numpy.allclose(numpy.asarray(MyGpuNdArray.adds(
a1, b1)), a_cpu + b_cpu)
assert numpy.allclose(numpy.asarray(MyGpuNdArray.add(
a1, b1)), a_cpu + b_cpu)
assert MyGpuNdArray.add(a1, b1, out=out2) is out2
# 2.1 What if we add a scalar?
nd_collaps, info = elemwise_collapses(
[a, b, scalar_gpu], [o])
if expected == 0:
expected2 = 1
else:
expected2 = expected
assert nd_collaps == expected2, (shape1, shape2,
nd_collaps, expected,
info)
# 2.2 Check computation
assert numpy.allclose(numpy.asarray(MyGpuNdArray.adds(
a1, b1, scalar_gpu1)),
a_cpu + b_cpu + scalar_cpu)
# 3.1 What if one of the dimensions is strided?
broadcast = any([True for i in a.shape + b.shape
if i == 1])
if expected == 0:
expected2 = 2
else:
expected2 = expected
if len(shape1_) != 4:
continue
if a.shape[0] != 1:
shape = list(shape1)
shape[0] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2,
nd_collaps, expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[1] != 1:
shape = list(shape1)
shape[1] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2,
nd_collaps, expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
pass
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[2] != 1:
shape = list(shape1)
shape[2] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::, ::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2,
nd_collaps, expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
pass
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[3] != 1:
shape = list(shape1)
shape[3] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::, ::,
::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2,
nd_collaps, expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == 1, err
pass
assert numpy.allclose(numpy.asarray(
MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
def test_elemwise_collapse():
""" Test collapsing under many broadcast and strided pattern """
for dtype1 in ["int16", "float32", "int8"]:
for dtype2 in ["int16", "float32", "int8"]:
for shape1_, shape2_, expected in [
# 1d to test this special case
((40, ), (40, ), 0),
((40, ), (1, ), 1),
# No broadcastable dimensions
((4, 5, 6, 9), (4, 5, 6, 9), 0),
# All inputs have one(and the same) broadcastable dimension
((1, 4, 5, 9), (1, 4, 5, 9), 0),
((4, 1, 5, 9), (4, 1, 5, 9), 0),
((4, 5, 1, 9), (4, 5, 1, 9), 0),
((4, 5, 9, 1), (4, 5, 9, 1), 0),
# One inputs have one broadcastable dimension
((1, 5, 6, 9), (4, 5, 6, 9), 2),
((4, 1, 6, 9), (4, 5, 6, 9), 3),
((4, 5, 1, 9), (4, 5, 6, 9), 3),
((4, 5, 6, 1), (4, 5, 6, 9), 2),
# One inputs have two broadcastable dimension
((1, 1, 6, 9), (4, 5, 6, 9), 2),
((1, 5, 1, 9), (4, 5, 6, 9), 4),
((1, 5, 6, 1), (4, 5, 6, 9), 3),
((4, 1, 1, 9), (4, 5, 6, 9), 3),
((4, 1, 6, 1), (4, 5, 6, 9), 4),
((4, 5, 1, 1), (4, 5, 6, 9), 2),
# One inputs have tree broadcastable dimension
((1, 1, 1, 9), (4, 5, 6, 9), 2),
((1, 1, 6, 1), (4, 5, 6, 9), 3),
((1, 5, 1, 1), (4, 5, 6, 9), 3),
((4, 1, 1, 1), (4, 5, 6, 9), 2),
# One scalar
((1, 1, 1, 1), (4, 5, 6, 9), 1),
# One scalar, the other 1 broadcast dims
((1, 1, 1, 1), (4, 5, 6, 1), 1),
]:
scalar_cpu = rand((1, ) * len(shape1_), dtype=dtype1)
scalar_gpu = gpu_ndarray.GpuNdArrayObject(scalar_cpu)
scalar_gpu1 = MyGpuNdArray(scalar_gpu)
for shape1, shape2 in [(shape1_, shape2_), (shape2_, shape1_)]:
a_cpu = rand(shape1, dtype=dtype1)
a = gpu_ndarray.GpuNdArrayObject(a_cpu)
a1 = MyGpuNdArray(a)
b_cpu = rand(shape2, dtype=dtype2)
b = gpu_ndarray.GpuNdArrayObject(b_cpu)
b1 = MyGpuNdArray(b)
assert len(shape1) == len(shape2)
o_shape = []
for i in range(len(shape1)):
o_shape.append(max(shape1[i], shape2[i]))
o = gpu_ndarray.empty(o_shape, dtype=(a_cpu + b_cpu).dtype)
# 1.1 Check direct collapse
nd_collaps, info = elemwise_collapses([a, b], [o])
assert nd_collaps == expected, (shape1, shape2, nd_collaps,
expected, info)
# 1.2 Check computation are still valid
f = MyGpuNdArray.gen_fct(theano.tensor.add, [a1, b1],
len(shape1))
out = f([a1, b1])
out2 = f([a1, b1], out=out)
assert out is out2
assert numpy.allclose(numpy.asarray(f([a1, b1])),
a_cpu + b_cpu)
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(a1, b1)),
a_cpu + b_cpu)
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.add(a1, b1)), a_cpu + b_cpu)
assert MyGpuNdArray.add(a1, b1, out=out2) is out2
# 1.3 Check work without collaping
f = MyGpuNdArray.gen_fct(theano.tensor.add, [a1, b1],
len(shape1),
collapse=False)
out = f([a1, b1])
out2 = f([a1, b1], out=out)
assert out is out2
assert numpy.allclose(numpy.asarray(f([a1, b1])),
a_cpu + b_cpu)
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(a1, b1)),
a_cpu + b_cpu)
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.add(a1, b1)), a_cpu + b_cpu)
assert MyGpuNdArray.add(a1, b1, out=out2) is out2
# 2.1 What if we add a scalar?
nd_collaps, info = elemwise_collapses([a, b, scalar_gpu],
[o])
if expected == 0:
expected2 = 1
else:
expected2 = expected
assert nd_collaps == expected2, (shape1, shape2,
nd_collaps, expected,
info)
# 2.2 Check computation
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(a1, b1, scalar_gpu1)),
a_cpu + b_cpu + scalar_cpu)
# 3.1 What if one of the dimensions is strided?
broadcast = any(
[True for i in a.shape + b.shape if i == 1])
if expected == 0:
expected2 = 2
else:
expected2 = expected
if len(shape1_) != 4:
continue
if a.shape[0] != 1:
shape = list(shape1)
shape[0] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2, nd_collaps,
expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[1] != 1:
shape = list(shape1)
shape[1] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2, nd_collaps,
expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
pass
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[2] != 1:
shape = list(shape1)
shape[2] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(c_cpu)[::, ::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2, nd_collaps,
expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == expected2, err
pass
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
if a.shape[3] != 1:
shape = list(shape1)
shape[3] *= 2
c_cpu = rand(shape, dtype='float32')
c = gpu_ndarray.GpuNdArrayObject(
c_cpu)[::, ::, ::, ::2]
c1 = MyGpuNdArray(c)
err = ("strided", c.shape, shape2, nd_collaps,
expected, info)
nd_collaps, info = elemwise_collapses([c, b], [o])
if broadcast:
assert nd_collaps >= expected, err
else:
assert nd_collaps == 1, err
pass
assert numpy.allclose(
numpy.asarray(MyGpuNdArray.adds(c1, b1)),
numpy.asarray(c) + b_cpu)
