def create_h_sigmoid_act() -> GraphPattern:
main_pattern = GraphPattern()
# ReLU version
pattern = GraphPattern()
input_pattern_node = pattern.add_node(
label='*INPUT_NODE*', type=GraphPattern.NON_PATTERN_NODE_TYPE)
add_node = pattern.add_node(label='ADD', type='AddV2')
relu_node = pattern.add_node(label='RELU', type='ReLU')
mul_node = pattern.add_node(label='TF_OP_MUL', type='Mul')
pattern.add_edge(input_pattern_node, add_node)
pattern.add_edge(add_node, relu_node)
pattern.add_edge(relu_node, mul_node)
main_pattern.add_pattern_alternative(pattern)
# ReLU6 version
pattern = GraphPattern()
input_pattern_node = pattern.add_node(
label='*INPUT_NODE*', type=GraphPattern.NON_PATTERN_NODE_TYPE)
add_node = pattern.add_node(label='ADD', type='AddV2')
relu6_node = pattern.add_node(label='RELU6', type='Relu6')
mul_node = pattern.add_node(label='TF_OP_MUL', type='Mul')
pattern.add_edge(input_pattern_node, add_node)
pattern.add_edge(add_node, relu6_node)
pattern.add_edge(relu6_node, mul_node)
main_pattern.add_pattern_alternative(pattern)
return main_pattern
def test_join_patterns_func_three_patterns():
pattern = (TestPattern.first_pattern + TestPattern.second_pattern +
TestPattern.third_pattern)
pattern_nodes = list(pattern.graph.nodes)
third_nodes = list(TestPattern.third_pattern.graph.nodes)
edges = list(itertools.product(pattern_nodes, third_nodes))
pattern.join_patterns(TestPattern.third_pattern, edges)
ref_pattern = GraphPattern()
_ = ref_pattern.add_node(label='first', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='second',
type=TestPattern.second_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
last_node = list(nx.topological_sort(ref_pattern.graph))[-1]
added_node = ref_pattern.add_node(label='third',
type=TestPattern.third_type)
ref_pattern.add_edge(last_node, added_node)
added_node = ref_pattern.add_node(label='third',
type=TestPattern.third_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
assert ref_pattern == pattern
def test_ops_combination_two_patterns():
pattern = TestPattern.first_pattern + TestPattern.second_pattern
ref_pattern = GraphPattern()
ref_pattern.add_node(label='first', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='second',
type=TestPattern.second_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
assert ref_pattern == pattern
pattern = TestPattern.first_pattern | TestPattern.second_pattern
ref_pattern = GraphPattern()
ref_pattern.add_node(label='first', type=TestPattern.first_type)
_ = ref_pattern.add_node(label='second', type=TestPattern.second_type)
assert ref_pattern == pattern
def create_h_swish_act() -> GraphPattern:
# TODO (vshampor): current approach with join_patterns is deficient since it does not allow to reliably
# connect nodes after a pattern has been joined. Along with the label and the type, the nodes created
# in the pattern must allow a "name" or "address" attribute, which must be a unique human readable
# string identifier of the node even if it has been joined multiple times, or perhaps each pattern
# after joining must return a list of output nodes so that these can be joined to later.
# Currently cannot specify h_swish in terms of h_sigmoid due to this.
main_pattern = GraphPattern()
# ReLU version
pattern = GraphPattern()
input_pattern_node = pattern.add_node(
label='*INPUT_NODE*', type=GraphPattern.NON_PATTERN_NODE_TYPE)
add_node = pattern.add_node(label='ADD', type='AddV2')
relu_node = pattern.add_node(label='RELU', type='ReLU')
mul_node = pattern.add_node(label='TF_OP_MUL', type='Mul')
pattern.add_edge(input_pattern_node, add_node)
pattern.add_edge(add_node, relu_node)
pattern.add_edge(relu_node, mul_node)
mul_2_node = pattern.add_node(label='MULTIPLY', type='Multiply')
pattern.add_edge(input_pattern_node, mul_2_node)
pattern.add_edge(mul_node, mul_2_node)
main_pattern.add_pattern_alternative(pattern)
# ReLU6 version
pattern = GraphPattern()
input_pattern_node = pattern.add_node(
label='*INPUT_NODE*', type=GraphPattern.NON_PATTERN_NODE_TYPE)
add_node = pattern.add_node(label='ADD', type='AddV2')
relu6_node = pattern.add_node(label='RELU6', type='Relu6')
mul_node = pattern.add_node(label='TF_OP_MUL', type='Mul')
pattern.add_edge(input_pattern_node, add_node)
pattern.add_edge(add_node, relu6_node)
pattern.add_edge(relu6_node, mul_node)
mul_2_node = pattern.add_node(label='MULTIPLY', type='Multiply')
pattern.add_edge(input_pattern_node, mul_2_node)
pattern.add_edge(mul_node, mul_2_node)
main_pattern.add_pattern_alternative(pattern)
return main_pattern
def test_join_pattern_with_special_input_node():
pattern = TestPattern.first_pattern
second_pattern = GraphPattern()
second_pattern.add_node(label='second',
type=GraphPattern.ANY_PATTERN_NODE_TYPE)
pattern.join_patterns(second_pattern)
pattern.join_patterns(TestPattern.third_pattern)
ref_pattern = GraphPattern()
ref_pattern.add_node(label='first', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='third',
type=TestPattern.third_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
assert pattern == ref_pattern
def _get_torch_hw_fused_patterns() -> HWFusedPatterns:
retval = HWFusedPatterns()
linear_ops = GraphPattern()
linear_ops.add_node(**LINEAR_OPERATIONS)
retval.register(linear_ops, LINEAR_OPERATIONS['label'], match=False)
matmul_ops = GraphPattern()
matmul_ops.add_node(**MATMUL_OPERATIONS)
retval.register(linear_ops, MATMUL_OPERATIONS['label'], match=False)
batch_norm = GraphPattern()
batch_norm.add_node(**BATCH_NORMALIZATION_OPERATIONS)
retval.register(batch_norm,
BATCH_NORMALIZATION_OPERATIONS['label'],
match=False)
atomic_activations = GraphPattern()
atomic_activations.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
swish = create_swish_act()
h_sigmoid = create_h_sigmoid_act()
h_swish = create_h_swish_act()
activations = atomic_activations | swish | h_swish | h_sigmoid
retval.register(activations, 'ACTIVATIONS', match=False)
arithmetic_ops = GraphPattern()
arithmetic_ops.add_node(**ARITHMETIC_OPERATIONS)
retval.register(arithmetic_ops,
ARITHMETIC_OPERATIONS['label'],
match=False)
batch_norm_activations_permutation = batch_norm + activations | activations + batch_norm | batch_norm | activations
retval.register(linear_ops + batch_norm_activations_permutation,
'LINEAR + BN_ACT_PERM',
match=True)
retval.register(matmul_ops + arithmetic_ops,
'MATMUL + ARITHMETIC',
match=True)
retval.register(batch_norm + activations, 'BN + ACTIVATIONS', match=True)
retval.register(activations + batch_norm, 'ACTIVATIONS + BN', match=True)
retval.register(arithmetic_ops + batch_norm_activations_permutation,
'ARITHMETIC + BN_ACT_PERM',
match=True)
group_norm = GraphPattern()
group_norm.add_node(**GROUP_NORMALIZATION_OPERATIONS)
relu = GraphPattern()
relu.add_node(**RELU_OPERATIONS)
retval.register(group_norm + relu, 'GROUP_NORM + RELU', match=True)
l2_norm = create_l2_norm()
retval.register(l2_norm, 'L2_NORM', match=True)
return retval
def test_ops_combination_three_patterns():
pattern = TestPattern.first_pattern + TestPattern.second_pattern | TestPattern.third_pattern
ref_pattern = GraphPattern()
ref_pattern.add_node(label='first', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='second',
type=TestPattern.second_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
_ = ref_pattern.add_node(label='third', type=TestPattern.third_type)
assert ref_pattern == pattern
pattern = TestPattern.first_pattern | TestPattern.second_pattern | TestPattern.third_pattern
ref_pattern = GraphPattern()
_ = ref_pattern.add_node(label='first', type=TestPattern.first_type)
_ = ref_pattern.add_node(label='second', type=TestPattern.second_type)
_ = ref_pattern.add_node(label='third', type=TestPattern.third_type)
assert ref_pattern == pattern
pattern = (TestPattern.first_pattern + TestPattern.second_pattern)
pattern_nodes = list(pattern.graph.nodes)
third_nodes = list(TestPattern.third_pattern.graph.nodes)
edges = list(itertools.product(pattern_nodes, third_nodes))
pattern.join_patterns(TestPattern.third_pattern, edges)
ref_pattern = GraphPattern()
ref_pattern.add_node(label='second', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='second',
type=TestPattern.second_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
added_node = ref_pattern.add_node(label='third',
type=TestPattern.third_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
assert ref_pattern == pattern
def test_join_patterns_func():
ref_pattern = GraphPattern()
ref_pattern.add_node(label='first', type=TestPattern.first_type)
added_node = ref_pattern.add_node(label='second',
type=TestPattern.second_type)
for node in ref_pattern.graph.nodes:
if node != added_node:
ref_pattern.add_edge(node, added_node)
first_nodes = list(TestPattern.first_pattern.graph.nodes)
second_nodes = list(TestPattern.second_pattern.graph.nodes)
edges = list(itertools.product(first_nodes, second_nodes))
pattern = copy.copy(TestPattern.first_pattern)
pattern.join_patterns(TestPattern.second_pattern, edges)
assert ref_pattern == pattern
def _get_tf_hw_fused_patterns() -> HWFusedPatterns:
retval = HWFusedPatterns()
linear_ops = GraphPattern()
linear_ops.add_node(**LINEAR_OPERATIONS)
eltwise_ops = GraphPattern()
eltwise_ops.add_node(**ELEMENTWISE_OPERATIONS)
batch_norm = GraphPattern()
batch_norm.add_node(**BATCH_NORMALIZATION_OPERATIONS)
h_sigmoid = create_h_sigmoid_act()
h_swish = create_h_swish_act()
retval.register(h_sigmoid, 'H_SIGMOID', match=True)
retval.register(h_swish, 'H_SWISH', match=True)
atomic_activations = GraphPattern()
atomic_activations.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
activations = atomic_activations | h_swish | h_sigmoid
batch_norm_activations_permutation = batch_norm + activations | activations + batch_norm
any_bn_act_combo = batch_norm | activations | batch_norm_activations_permutation
identity = GraphPattern()
identity.add_node(type=['Identity'], label='IDENTITY')
linear_ops_maybe_followed_by_identity = linear_ops | (linear_ops + identity)
agnostic_ops = GraphPattern()
agnostic_ops.add_node(**QUANTIZATION_AGNOSTIC_OPERATIONS)
any_ag_bn_act_combo = agnostic_ops + activations | any_bn_act_combo
retval.register(linear_ops_maybe_followed_by_identity, name='LINEAR', match=True)
retval.register(batch_norm_activations_permutation, name='BN_ACT_OR_ACT_BN', match=True)
retval.register(linear_ops_maybe_followed_by_identity + any_ag_bn_act_combo, 'LINEAR + ANY_AG_BN_ACT_COMBO',
match=True)
retval.register(eltwise_ops + any_ag_bn_act_combo, 'ELTWISE + ANY_AG_BN_ACT_COMBO',
match=True)
return retval
class TestPattern:
first_type = ['a', 'b']
second_type = ['c', 'd']
third_type = ['e']
forth_type = [GraphPattern.NON_PATTERN_NODE_TYPE]
fifth_type = [GraphPattern.ANY_PATTERN_NODE_TYPE]
first_pattern = GraphPattern()
first_pattern.add_node(label='first', type=first_type)
second_pattern = GraphPattern()
second_pattern.add_node(label='second', type=second_type)
third_pattern = GraphPattern()
third_pattern.add_node(label='third', type=third_type)
forth_pattern = GraphPattern()
forth_pattern.add_node(label='forth', type=forth_type)
fifth_pattern = GraphPattern()
fifth_pattern.add_node(label='fifth', type=fifth_pattern)
# pattern_with_non_pattern_nodes | pattern_with_any_pattern_nodes
# NON | ANY
# | | |
# 1 | 1
# | | |
# 2 NON | 2 ANY
# / \ / | / \ /
# 4 3 | 4 3
# | / | | /
# | / | | /
# |/ | |/
# 5 | 5
# | | |
# 6---NON | 6---ANY
pattern_with_non_pattern_nodes = GraphPattern()
pattern_with_any_pattern_nodes = GraphPattern()
common_nodes = {
'1': {
'type': 'a'
},
'2': {
'type': 'b'
},
'3': {
'type': 'c'
},
'4': {
'type': 'a'
},
'5': {
'type': 'e'
},
'6': {
'type': 'a'
}
}
non_pattern_nodes = {
'7': {
'type': GraphPattern.NON_PATTERN_NODE_TYPE
},
'8': {
'type': GraphPattern.NON_PATTERN_NODE_TYPE
},
'9': {
'type': GraphPattern.NON_PATTERN_NODE_TYPE
}
}
any_pattern_nodes = {
'7': {
'type': GraphPattern.ANY_PATTERN_NODE_TYPE
},
'8': {
'type': GraphPattern.ANY_PATTERN_NODE_TYPE
},
'9': {
'type': GraphPattern.ANY_PATTERN_NODE_TYPE
}
}
label_to_non_pattern_nodes = {}
label_to_any_pattern_nodes = {}
for label, attrs in common_nodes.items():
label_to_non_pattern_nodes[
label] = pattern_with_non_pattern_nodes.add_node(label=label,
**attrs)
label_to_any_pattern_nodes[
label] = pattern_with_any_pattern_nodes.add_node(label=label,
**attrs)
for label, attrs in non_pattern_nodes.items():
label_to_non_pattern_nodes[
label] = pattern_with_non_pattern_nodes.add_node(label=label,
**attrs)
for label, attrs in any_pattern_nodes.items():
label_to_any_pattern_nodes[
label] = pattern_with_any_pattern_nodes.add_node(label=label,
**attrs)
edges = [('1', '2'), ('2', '3'), ('2', '4'), ('4', '5'), ('5', '6'),
('3', '5'), ('7', '1'), ('8', '3'), ('9', '6')]
for edge in edges:
pattern_with_non_pattern_nodes.add_edge(
label_to_non_pattern_nodes[edge[0]],
label_to_non_pattern_nodes[edge[1]])
pattern_with_any_pattern_nodes.add_edge(
label_to_any_pattern_nodes[edge[0]],
label_to_any_pattern_nodes[edge[1]])