Skip to content

Commit

Permalink
Add chainerx.conv TensorCore test
Browse files Browse the repository at this point in the history
  • Loading branch information
@asi1024
asi1024 committed May 22, 2019
1 parent 6904ced commit b213b32
Showing 1 changed file with 132 additions and 67 deletions.
199 changes: 132 additions & 67 deletions tests/chainerx_tests/unit_tests/routines_tests/test_connection.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,39 +31,19 @@ def _create_conv_args(
return x, w, b, stride, pad, cover_all


@op_utils.op_test(['native:0', 'cuda:0'])
@chainer.testing.parameterize(*(
# without bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (5, 3), None, 1, 0),
((1, 3, 4), (5, 3, 2), None, 3, 2),
((2, 3, 4, 4), (2, 3, 3, 3), None, 2, (2, 0)),
((2, 3, 2, 6, 3), (2, 3, 1, 3, 2), None, (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_two_arrays)
]) +
# with bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (5, 3), (5,), 1, 0),
((2, 3, 4), (5, 3, 1), (5,), 1, 0),
((1, 3, 4), (5, 3, 2), (5,), 3, 2),
((2, 3, 4, 4), (2, 3, 3, 3), (2,), 1, 0),
((1, 3, 4, 4), (2, 3, 3, 3), (2,), (1, 2), 1),
((1, 3, 4, 4), (2, 3, 3, 3), (2,), 2, (2, 0)),
((1, 3, 2, 6, 3), (2, 3, 1, 3, 2), (2,), 2, (2, 0, 1)),
((1, 3, 2, 6, 3), (2, 3, 1, 3, 2), (2,), (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_three_arrays)
])
))
@chainer.testing.parameterize_pytest('cover_all', [True, False])
class TestConv(op_utils.ChainerOpTest):
def _convert_to_nhwc_layout(array):
# Converts a contiguous array to NHWC data layout.
assert isinstance(array, numpy.ndarray)
assert array.flags.c_contiguous
shape = array.shape
transposed_array = numpy.transpose(array, (0, 3, 1, 2))
transposed_array = numpy.ascontiguousarray(transposed_array)
array = numpy.transpose(transposed_array, (0, 2, 3, 1))
assert array.shape == shape
return array


class _ConvTestBase(object):

def setup(self):
if len(self.in_dtypes) == 3:
Expand Down Expand Up @@ -110,6 +90,7 @@ def generate_inputs(self):
(x_dtype, w_dtype), b_dtype = self.in_dtypes, None
x = array_utils.uniform(x_shape, x_dtype)
w = array_utils.uniform(w_shape, w_dtype)

if b_shape is None:
return x, w
else:
Expand Down Expand Up @@ -141,6 +122,63 @@ def forward_chainer(self, inputs):
return y,


@op_utils.op_test(['native:0', 'cuda:0'])
@chainer.testing.parameterize(*(
# without bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (5, 3), None, 1, 0),
((1, 3, 4), (5, 3, 2), None, 3, 2),
((2, 3, 4, 4), (2, 3, 3, 3), None, 2, (2, 0)),
((2, 3, 2, 6, 3), (2, 3, 1, 3, 2), None, (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_two_arrays)
]) +
# with bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (5, 3), (5,), 1, 0),
((2, 3, 4), (5, 3, 1), (5,), 1, 0),
((1, 3, 4), (5, 3, 2), (5,), 3, 2),
((2, 3, 4, 4), (2, 3, 3, 3), (2,), 1, 0),
((1, 3, 4, 4), (2, 3, 3, 3), (2,), (1, 2), 1),
((1, 3, 4, 4), (2, 3, 3, 3), (2,), 2, (2, 0)),
((1, 3, 2, 6, 3), (2, 3, 1, 3, 2), (2,), 2, (2, 0, 1)),
((1, 3, 2, 6, 3), (2, 3, 1, 3, 2), (2,), (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_three_arrays)
])
))
@chainer.testing.parameterize_pytest('cover_all', [True, False])
class TestConv(_ConvTestBase, op_utils.ChainerOpTest):
pass


# cudnnFindConvolutionForwardAlgorithmEx tends to choose an algorithm which
# uses TensorCore when:
# - The sizes of the input arrays are large
# - The input `x` is aligned with NHWC data layout
@op_utils.op_test(['cuda:0'])
class TestConvTensorCore(_ConvTestBase, op_utils.ChainerOpTest):

x_shape = (8, 3, 227, 227)
w_shape = (96, 3, 11, 11)
b_shape = (96,)
stride = 4
pad = 0
in_dtypes = ('float16', 'float16', 'float16')
out_dtype = 'float16'
cover_all = False

def generate_inputs(self):
x, w, b = super().generate_inputs()
return _convert_to_nhwc_layout(x), w, b


@pytest.mark.parametrize('x_shape,w_shape,b_shape,stride,pad', [
# Mismatched x and w input channels.
((1, 3, 4, 3), (5, 4, 2, 2), (5,), 3, 2),
Expand All @@ -162,40 +200,7 @@ def test_conv_invalid(
cover_all, float_dtype))


@op_utils.op_test(['native:0', 'cuda:0'])
@chainer.testing.parameterize(*(
# without bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (3, 5), None, 1, 0),
((1, 3, 4), (3, 5, 2), None, 3, 2),
((2, 3, 4, 4), (3, 2, 3, 3), None, 2, (2, 0)),
((2, 3, 5, 6, 3), (3, 2, 1, 3, 2), None, (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_two_arrays)
]) +
# with bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (3, 5), (5,), 1, 0),
((2, 3, 4), (3, 5, 1), (5,), 1, 0),
((1, 3, 4), (3, 5, 2), (5,), 3, 2),
((2, 3, 4, 4), (3, 2, 3, 3), (2,), 1, 0),
((1, 3, 4, 4), (3, 2, 3, 3), (2,), (1, 2), 1),
((1, 3, 4, 4), (3, 2, 3, 3), (2,), 2, (2, 0)),
((1, 3, 5, 6, 3), (3, 2, 1, 3, 2), (2,), 2, (2, 0, 1)),
((1, 3, 5, 6, 3), (3, 2, 1, 3, 2), (2,), (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_three_arrays)
])
))
# If None, outsize argument will be None.
@chainer.testing.parameterize_pytest('cover_all', [None, True, False])
class TestConvTranspose(op_utils.ChainerOpTest):
class _ConvTransposeTestBase(object):

def setup(self):
if len(self.in_dtypes) == 3:
Expand Down Expand Up @@ -266,6 +271,7 @@ def generate_inputs(self):
(x_dtype, w_dtype), b_dtype = self.in_dtypes, None
x = array_utils.uniform(x_shape, x_dtype)
w = array_utils.uniform(w_shape, w_dtype)

if b_shape is None:
return x, w
else:
Expand Down Expand Up @@ -298,6 +304,65 @@ def forward_chainer(self, inputs):
return y,


@op_utils.op_test(['native:0', 'cuda:0'])
@chainer.testing.parameterize(*(
# without bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (3, 5), None, 1, 0),
((1, 3, 4), (3, 5, 2), None, 3, 2),
((2, 3, 4, 4), (3, 2, 3, 3), None, 2, (2, 0)),
((2, 3, 5, 6, 3), (3, 2, 1, 3, 2), None, (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_two_arrays)
]) +
# with bias
chainer.testing.product([
chainer.testing.from_pytest_parameterize(
'x_shape,w_shape,b_shape,stride,pad', [
((1, 3), (3, 5), (5,), 1, 0),
((2, 3, 4), (3, 5, 1), (5,), 1, 0),
((1, 3, 4), (3, 5, 2), (5,), 3, 2),
((2, 3, 4, 4), (3, 2, 3, 3), (2,), 1, 0),
((1, 3, 4, 4), (3, 2, 3, 3), (2,), (1, 2), 1),
((1, 3, 4, 4), (3, 2, 3, 3), (2,), 2, (2, 0)),
((1, 3, 5, 6, 3), (3, 2, 1, 3, 2), (2,), 2, (2, 0, 1)),
((1, 3, 5, 6, 3), (3, 2, 1, 3, 2), (2,), (1, 2, 3), (2, 0, 1)),
]),
chainer.testing.from_pytest_parameterize(
'in_dtypes,out_dtype', dtype_utils.result_dtypes_three_arrays)
])
))
# If None, outsize argument will be None.
@chainer.testing.parameterize_pytest('cover_all', [None, True, False])
class TestConvTranspose(_ConvTransposeTestBase, op_utils.ChainerOpTest):
pass


# cudnnFindConvolutionForwardAlgorithmEx tends to choose an algorithm which
# uses TensorCore when:
# - The sizes of the input arrays are large
# - The input `x` is aligned with NHWC data layout
@op_utils.op_test(['cuda:0'])
@chainer.testing.parameterize_pytest('cover_all', [None, False])
class TestConvTransposeTensorCore(
_ConvTransposeTestBase, op_utils.ChainerOpTest):

x_shape = (8, 3, 227, 227)
w_shape = (3, 8, 11, 11)
b_shape = (8,)
stride = 4
pad = 0
in_dtypes = ('float16', 'float16', 'float16')
out_dtype = 'float16'

def generate_inputs(self):
x, w, b = super().generate_inputs()
return _convert_to_nhwc_layout(x), w, b


@pytest.mark.parametrize('x_shape,w_shape,b_shape,stride,pad,outsize', [
# Mismatched x and w input channels.
((1, 3, 4, 3), (5, 4, 2, 2), (5,), 3, 2, None),
Expand Down

0 comments on commit b213b32

Please sign in to comment.