IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues

Browse Source
This commit is contained in:
Ivan
2023-03-02 21:39:57 +04:00
parent 199114a4c6 873d15ce1a
commit caea77c132
9 changed files with 63 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_binary.cpp
View File

@@ -2,6 +2,8 @@
// SPDX-License-Identifier: Apache-2.0
//
#include "transformations/common_optimizations/transpose_sinking_binary.hpp"
#include <openvino/opsets/opset10.hpp>
#include <openvino/pass/pattern/op/or.hpp>
@@ -9,7 +11,6 @@
#include "openvino/op/util/op_types.hpp"
#include "openvino/pass/pattern/op/wrap_type.hpp"
#include "transformations/common_optimizations/transpose_sinking_utils.hpp"
#include "transformations/common_optimizations/transpose_sinking_binary.hpp"
#include "transformations/rt_info/transpose_sinking_attr.hpp"
using namespace ov::pass::pattern;

7
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_data_movement.cpp
View File

@@ -23,7 +23,8 @@ ov::pass::TransposeSinkingDataMovementForward::TransposeSinkingDataMovementForwa
MATCHER_SCOPE(TransposeSinkingDataMovementForward);
auto const_label = wrap_type<Constant>();
auto transpose_label = wrap_type<Transpose>({any_input(), const_label});
auto main_node_label = wrap_type<Pad, BatchToSpace, SpaceToBatch>({transpose_label, any_input(), any_input(), any_input()});
auto main_node_label =
wrap_type<Pad, BatchToSpace, SpaceToBatch>({transpose_label, any_input(), any_input(), any_input()});
matcher_pass_callback matcher_pass_callback = [=](Matcher& m) {
const auto& pattern_to_node = m.get_pattern_map();
@@ -57,7 +58,7 @@ ov::pass::TransposeSinkingDataMovementForward::TransposeSinkingDataMovementForwa
const auto& stb = std::dynamic_pointer_cast<SpaceToBatch>(main_node);
if (bts || stb) {
main_node->input(3).replace_source_output(
ChangeValuesOrder(main_node->input_value(3), reversed_transpose_order, axis));
ChangeValuesOrder(main_node->input_value(3), reversed_transpose_order, axis));
}
main_node->validate_and_infer_types();
@@ -114,7 +115,7 @@ ov::pass::TransposeSinkingDataMovementBackward::TransposeSinkingDataMovementBack
const auto& stb = std::dynamic_pointer_cast<SpaceToBatch>(main_node);
if (bts || stb) {
main_node->input(3).replace_source_output(
ChangeValuesOrder(main_node->input_value(3), transpose_axis_order, axis));
ChangeValuesOrder(main_node->input_value(3), transpose_axis_order, axis));
}
main_node->validate_and_infer_types();
return true;

8
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_fuse.cpp
View File

@@ -8,9 +8,9 @@
#include <vector>
#include "itt.hpp"
#include "openvino/core/validation_util.hpp"
#include "openvino/opsets/opset10.hpp"
#include "openvino/pass/pattern/op/wrap_type.hpp"
#include "openvino/core/validation_util.hpp"
#include "transformations/common_optimizations/transpose_sinking_utils.hpp"
#include "transformations/utils/utils.hpp"
@@ -19,13 +19,11 @@ using namespace opset10;
ov::pass::TransposeFuse::TransposeFuse() {
MATCHER_SCOPE(TransposeFuse);
auto transpose_label =
pattern::wrap_type<Transpose>({pattern::any_input(), pattern::wrap_type<Constant>()});
auto transpose_label = pattern::wrap_type<Transpose>({pattern::any_input(), pattern::wrap_type<Constant>()});
ov::matcher_pass_callback matcher_pass_callback = [=](pattern::Matcher& m) {
const auto& pattern_to_output = m.get_pattern_map();
auto transpose_1 = pattern_to_output.at(transpose_label);
auto order_const_1 =
std::dynamic_pointer_cast<Constant>(transpose_1->input_value(1).get_node_shared_ptr());
auto order_const_1 = std::dynamic_pointer_cast<Constant>(transpose_1->input_value(1).get_node_shared_ptr());
auto consumers = transpose_1->get_output_target_inputs(0);
std::vector<int64_t> saved_order_values;

3
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_general.cpp
View File

@@ -14,6 +14,7 @@
#include "transformations/common_optimizations/transpose_sinking_binary.hpp"
#include "transformations/common_optimizations/transpose_sinking_concat.hpp"
#include "transformations/common_optimizations/transpose_sinking_data_movement.hpp"
#include "transformations/common_optimizations/transpose_sinking_interpolate.hpp"
#include "transformations/common_optimizations/transpose_sinking_split.hpp"
#include "transformations/common_optimizations/transpose_sinking_unary.hpp"
#include "transformations/utils/utils.hpp"
@@ -26,6 +27,7 @@ ov::pass::TransposeSinkingGeneralForward::TransposeSinkingGeneralForward() {
add_matcher<ov::pass::TransposeSinkingSplitForward>();
add_matcher<ov::pass::TransposeSinkingDataMovementForward>();
add_matcher<ov::pass::TransposeSinkingReductionForward>();
add_matcher<ov::pass::TransposeSinkingInterpolateForward>();
add_matcher<ov::pass::TransposeFuse>();
}
@@ -37,6 +39,7 @@ ov::pass::TransposeSinkingGeneralBackward::TransposeSinkingGeneralBackward() {
add_matcher<ov::pass::TransposeSinkingSplitBackward>();
add_matcher<ov::pass::TransposeSinkingDataMovementBackward>();
add_matcher<ov::pass::TransposeSinkingReductionBackward>();
add_matcher<ov::pass::TransposeSinkingInterpolateBackward>();
add_matcher<ov::pass::TransposeFuse>();
}

28
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_interpolate.cpp
View File

@@ -44,12 +44,11 @@ ov::pass::TransposeSinkingInterpolateForward::TransposeSinkingInterpolateForward
main_node->input(0).replace_source_output(transpose_parent);
const auto transpose_axis_order = transpose_const->get_axis_vector_val();
const auto reversed_transpose_order = ReverseTransposeOrder(transpose_axis_order);
auto axis = std::make_shared<Constant>(element::i32, Shape{}, std::vector<int32_t>{0});
const auto& interpolate = std::dynamic_pointer_cast<Interpolate>(main_node);
auto data = std::make_shared<Constant>(element::i32, Shape{reversed_transpose_order.size()}, reversed_transpose_order);
auto data = std::make_shared<Constant>(element::i32, Shape{transpose_axis_order.size()}, transpose_axis_order);
const auto& indices = main_node->input_value(3);
auto new_axis = std::make_shared<Gather>(data, indices, axis);
@@ -57,13 +56,13 @@ ov::pass::TransposeSinkingInterpolateForward::TransposeSinkingInterpolateForward
if (interpolate) {
op::v4::Interpolate::InterpolateAttrs attrs = interpolate->get_attrs();
if (!attrs.pads_begin.empty() || !attrs.pads_end.empty()) {
const auto &order_size = reversed_transpose_order.size();
const auto& order_size = transpose_axis_order.size();
attrs.pads_begin.resize(order_size);
attrs.pads_end.resize(order_size);
std::vector<size_t> new_pads_begin(order_size), new_pads_end(order_size);
for (size_t i = 0; i < order_size; ++i) {
new_pads_begin[i] = attrs.pads_begin[reversed_transpose_order[i]];
new_pads_end[i] = attrs.pads_end[reversed_transpose_order[i]];
new_pads_begin[i] = attrs.pads_begin[transpose_axis_order[i]];
new_pads_end[i] = attrs.pads_end[transpose_axis_order[i]];
}
std::swap(attrs.pads_begin, new_pads_begin);
std::swap(attrs.pads_end, new_pads_end);
@@ -71,7 +70,6 @@ ov::pass::TransposeSinkingInterpolateForward::TransposeSinkingInterpolateForward
}
}
main_node->validate_and_infer_types();
TransposeInputsInfo transpose_input_info = {transpose, transpose_const, 0};
for (auto& new_node : sink_forward::InsertOutputTransposes(main_node, transpose_input_info)) {
@@ -95,9 +93,9 @@ ov::pass::TransposeSinkingInterpolateBackward::TransposeSinkingInterpolateBackwa
auto transpose_const_label = wrap_type<Constant>();
auto transpose_label =
wrap_type<Transpose>({main_node_label, transpose_const_label}, [](const Output<Node>& output) -> bool {
return has_static_rank()(output) && is_sinking_node(output);
});
wrap_type<Transpose>({main_node_label, transpose_const_label}, [](const Output<Node>& output) -> bool {
return has_static_rank()(output) && is_sinking_node(output);
});
matcher_pass_callback matcher_pass_callback = [=](Matcher& m) {
const auto& pattern_to_output = m.get_pattern_value_map();
@@ -107,7 +105,7 @@ ov::pass::TransposeSinkingInterpolateBackward::TransposeSinkingInterpolateBackwa
for (auto& new_node : sink_backward::InsertTransposeBeforeNode(main_node,
transpose_const,
/* input_indexes= */ {0})) {
/* input_indexes= */ {0})) {
register_new_node(new_node);
}
@@ -115,21 +113,23 @@ ov::pass::TransposeSinkingInterpolateBackward::TransposeSinkingInterpolateBackwa
RemoveSingleOutputConsumers(main_node);
const auto transpose_axis_order = transpose_const->get_axis_vector_val();
const auto reversed_transpose_order = ReverseTransposeOrder(transpose_axis_order);
auto axis = std::make_shared<Constant>(element::i32, Shape{}, std::vector<int32_t>{0});
auto data = std::make_shared<Constant>(element::i32, Shape{transpose_axis_order.size()}, transpose_axis_order);
auto data =
std::make_shared<Constant>(element::i32, Shape{reversed_transpose_order.size()}, reversed_transpose_order);
const auto& indices = main_node->input_value(3);
auto new_axis = std::make_shared<Gather>(data, indices, axis);
const auto& interpolate = std::dynamic_pointer_cast<Interpolate>(main_node);
if (interpolate) {
op::v4::Interpolate::InterpolateAttrs attrs = interpolate->get_attrs();
if (!attrs.pads_begin.empty() || !attrs.pads_end.empty()) {
const auto &order_size = transpose_axis_order.size();
const auto& order_size = reversed_transpose_order.size();
attrs.pads_begin.resize(order_size);
attrs.pads_end.resize(order_size);
std::vector<size_t> new_pads_begin(order_size), new_pads_end(order_size);
for (size_t i = 0; i < order_size; ++i) {
new_pads_begin[i] = attrs.pads_begin[transpose_axis_order[i]];
new_pads_end[i] = attrs.pads_end[transpose_axis_order[i]];
new_pads_begin[i] = attrs.pads_begin[reversed_transpose_order[i]];
new_pads_end[i] = attrs.pads_end[reversed_transpose_order[i]];
}
std::swap(attrs.pads_begin, new_pads_begin);
std::swap(attrs.pads_end, new_pads_end);

52
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_reduction.cpp
View File

@@ -8,11 +8,11 @@
#include <vector>
#include "itt.hpp"
#include "openvino/core/validation_util.hpp"
#include "openvino/opsets/opset10.hpp"
#include "openvino/op/util/arithmetic_reductions_keep_dims.hpp"
#include "openvino/op/util/logical_reduction_keep_dims.hpp"
#include "openvino/pass/pattern/op/wrap_type.hpp"
#include "openvino/core/validation_util.hpp"
#include "transformations/common_optimizations/transpose_sinking_utils.hpp"
#include "transformations/utils/utils.hpp"
@@ -79,19 +79,16 @@ bool get_keep_dims(const std::shared_ptr<Node>& reduction) {
keep_dims = arithmetic_reduce->get_keep_dims();
return keep_dims;
}
}
} // namespace
ov::pass::TransposeSinkingReductionForward::TransposeSinkingReductionForward() {
MATCHER_SCOPE(TransposeSinkingReductionForward);
auto transpose_label =
pattern::wrap_type<Transpose>({pattern::any_input(), pattern::wrap_type<Constant>()},
pattern::consumers_count(1));
auto reduce_or_squeeze_label =
pattern::wrap_type<op::util::ArithmeticReductionKeepDims,
op::util::LogicalReductionKeepDims,
Squeeze,
Unsqueeze>({transpose_label, pattern::wrap_type<Constant>()});
auto transpose_label = pattern::wrap_type<Transpose>({pattern::any_input(), pattern::wrap_type<Constant>()},
pattern::consumers_count(1));
auto reduce_or_squeeze_label = pattern::
wrap_type<op::util::ArithmeticReductionKeepDims, op::util::LogicalReductionKeepDims, Squeeze, Unsqueeze>(
{transpose_label, pattern::wrap_type<Constant>()});
ov::matcher_pass_callback matcher_pass_callback = [=](pattern::Matcher& m) {
const auto& pattern_to_output = m.get_pattern_value_map();
@@ -106,9 +103,9 @@ ov::pass::TransposeSinkingReductionForward::TransposeSinkingReductionForward() {
return false;
auto unsqueeze = std::dynamic_pointer_cast<Unsqueeze>(reduction);
auto rank =
unsqueeze ? reduction->get_output_partial_shape(0).rank() : reduction->get_input_partial_shape(0).rank();
unsqueeze ? reduction->get_output_partial_shape(0).rank() : reduction->get_input_partial_shape(0).rank();
auto non_negative_axes =
normalize_axes(reduction->get_friendly_name(), reduction_axes->cast_vector<int64_t>(), rank);
normalize_axes(reduction->get_friendly_name(), reduction_axes->cast_vector<int64_t>(), rank);
auto transpose_order_values = transpose_order->cast_vector<size_t>();
std::vector<size_t> new_values;
@@ -138,14 +135,13 @@ ov::pass::TransposeSinkingReductionForward::TransposeSinkingReductionForward() {
}
}
auto new_transpose_order = std::make_shared<Constant>(transpose_order->get_element_type(),
Shape{transpose_order_values.size()},
transpose_order_values);
Shape{transpose_order_values.size()},
transpose_order_values);
std::shared_ptr<Node> new_reduction;
if (!unsqueeze) {
auto new_const = std::make_shared<Constant>(reduction_axes->get_element_type(),
reduction_axes->get_shape(),
new_values);
auto new_const =
std::make_shared<Constant>(reduction_axes->get_element_type(), reduction_axes->get_shape(), new_values);
new_reduction = reduction->clone_with_new_inputs({transpose->input_value(0), new_const});
} else {
new_reduction = reduction->clone_with_new_inputs({transpose->input_value(0), reduction->input_value(1)});
@@ -167,14 +163,11 @@ ov::pass::TransposeSinkingReductionForward::TransposeSinkingReductionForward() {
ov::pass::TransposeSinkingReductionBackward::TransposeSinkingReductionBackward() {
MATCHER_SCOPE(TransposeSinkingReductionBackward);
auto reduce_or_squeeze_label =
pattern::wrap_type<op::util::ArithmeticReductionKeepDims,
op::util::LogicalReductionKeepDims,
Squeeze,
Unsqueeze>({pattern::any_input(), pattern::wrap_type<Constant>()},
transpose_sinking::HasSameOutputTransposeNodes);
auto transpose_label =
pattern::wrap_type<Transpose>({reduce_or_squeeze_label, pattern::wrap_type<Constant>()});
auto reduce_or_squeeze_label = pattern::
wrap_type<op::util::ArithmeticReductionKeepDims, op::util::LogicalReductionKeepDims, Squeeze, Unsqueeze>(
{pattern::any_input(), pattern::wrap_type<Constant>()},
transpose_sinking::HasSameOutputTransposeNodes);
auto transpose_label = pattern::wrap_type<Transpose>({reduce_or_squeeze_label, pattern::wrap_type<Constant>()});
ov::matcher_pass_callback matcher_pass_callback = [=](pattern::Matcher& m) {
const auto& pattern_to_output = m.get_pattern_value_map();
@@ -188,9 +181,9 @@ ov::pass::TransposeSinkingReductionBackward::TransposeSinkingReductionBackward()
auto unsqueeze = std::dynamic_pointer_cast<Unsqueeze>(reduction);
auto rank =
unsqueeze ? reduction->get_output_partial_shape(0).rank() : reduction->get_input_partial_shape(0).rank();
unsqueeze ? reduction->get_output_partial_shape(0).rank() : reduction->get_input_partial_shape(0).rank();
auto non_negative_axes =
normalize_axes(reduction->get_friendly_name(), reduction_axes->cast_vector<int64_t>(), rank);
normalize_axes(reduction->get_friendly_name(), reduction_axes->cast_vector<int64_t>(), rank);
auto transpose_order_values = transpose_order->cast_vector<size_t>();
auto old_transpose_order_values = transpose_order_values;
@@ -251,9 +244,8 @@ ov::pass::TransposeSinkingReductionBackward::TransposeSinkingReductionBackward()
copy_runtime_info({transpose, reduction}, {new_transpose, new_reduction});
register_new_node(new_transpose);
} else {
auto new_const = std::make_shared<Constant>(reduction_axes->get_element_type(),
reduction_axes->get_shape(),
new_values);
auto new_const =
std::make_shared<Constant>(reduction_axes->get_element_type(), reduction_axes->get_shape(), new_values);
auto new_transpose = transpose->clone_with_new_inputs({reduction->input_value(0), new_transpose_order});
auto new_reduction = reduction->clone_with_new_inputs({new_transpose, new_const});
replace_node(transpose, new_reduction);

7
src/common/transformations/src/transformations/common_optimizations/transpose_sinking_unary.cpp
View File

@@ -55,7 +55,9 @@ ov::pass::TransposeSinkingUnaryForward::TransposeSinkingUnaryForward() {
MATCHER_SCOPE(TransposeSinkingUnaryForward);
auto transpose_label = wrap_type<Transpose>({any_input(), any_input()});
auto unary_label = wrap_type<UnaryElementwiseArithmetic, Clamp, Elu, SoftPlus, LogicalNot, Convert, IsInf, IsNaN, IsFinite>({transpose_label});
auto unary_label =
wrap_type<UnaryElementwiseArithmetic, Clamp, Elu, SoftPlus, LogicalNot, Convert, IsInf, IsNaN, IsFinite>(
{transpose_label});
ov::matcher_pass_callback matcher_pass_callback = [=](Matcher& m) {
const auto& pattern_to_output = m.get_pattern_value_map();
@@ -90,7 +92,8 @@ ov::pass::TransposeSinkingUnaryBackward::TransposeSinkingUnaryBackward() {
auto unary_label =
wrap_type<UnaryElementwiseArithmetic, Clamp, Elu, SoftPlus, LogicalNot, Convert, IsInf, IsNaN, IsFinite>(
{any_input()}, unary_restrictions);
{any_input()},
unary_restrictions);
auto transpose_const_label = wrap_type<Constant>();

17
src/common/transformations/tests/common_optimizations/transpose_sinking_pad_test.cpp
View File

@@ -292,14 +292,15 @@ INSTANTIATE_TEST_SUITE_P(TransposeSinkingPadForwardSingleConsumerTestSuite,
::testing::Values(element::f32)),
TransposeSinkingPadTestFixture::get_test_name);
INSTANTIATE_TEST_SUITE_P(TransposeSinkingPadBackwardSingleConsumerTestSuite,
TransposeSinkingPadTestFixture,
::testing::Combine(::testing::Values(CREATE_PASS_FACTORY(TransposeSinkingDataMovementBackward)),
::testing::ValuesIn(pad_operations_numbers),
::testing::Values(backward::single_consumer::CreateFunction),
::testing::Values(backward::single_consumer::CreateReferenceFunction),
::testing::Values(element::f32)),
TransposeSinkingPadTestFixture::get_test_name);
INSTANTIATE_TEST_SUITE_P(
TransposeSinkingPadBackwardSingleConsumerTestSuite,
TransposeSinkingPadTestFixture,
::testing::Combine(::testing::Values(CREATE_PASS_FACTORY(TransposeSinkingDataMovementBackward)),
::testing::ValuesIn(pad_operations_numbers),
::testing::Values(backward::single_consumer::CreateFunction),
::testing::Values(backward::single_consumer::CreateReferenceFunction),
::testing::Values(element::f32)),
TransposeSinkingPadTestFixture::get_test_name);
INSTANTIATE_TEST_SUITE_P(
TransposeSinkingPadBackwardSingleConsumerMultiTransposesTestSuite,

2
src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
View File

@@ -166,7 +166,7 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model) {
pass_config->disable<ov::pass::AddFakeQuantizeFusion>();
// TransposeReduction can be enabled when Transpose-Conv-Transpose patterns will be handled in ngraph
// transformations
pass_config->disable<ov::pass::TransposeReduction>();
pass_config->disable<ov::pass::TransposeSinkingReductionForward>();
// Operations Max and Min aren't supported
pass_config->disable<ov::pass::ConcatReduceFusion>();