Skip to content

Commit

Permalink
[CPU][ARM] Dynamic shapes support in ARM transformations (#17548)
Browse files Browse the repository at this point in the history
  • Loading branch information
@alvoron
alvoron authored Jun 12, 2023
1 parent a106eb0 commit 1588a33
Show file tree
Hide file tree
Showing 19 changed files with 912 additions and 59 deletions.
21 changes: 15 additions & 6 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_group_conv.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,11 +19,20 @@ ov::intel_cpu::ConvertGroupConvolution::ConvertGroupConvolution() {
if (!gconv) {
return false;
}

auto data_shape = gconv->get_input_shape(Inputs::Data);
const unsigned int channel_axis = 1;
const auto& input0 = gconv->input_value(0);
const auto& output_shape = gconv->get_output_partial_shape(0);
const auto& data_shape = input0.get_partial_shape();
// Weights layout GOIYX
size_t groups = gconv->get_input_shape(Inputs::Weights)[0];
if (groups == data_shape.at(1) && groups == gconv->get_output_shape(0)[1]) { // depthwise case
int64_t groups = gconv->get_input_shape(Inputs::Weights)[0];

if (data_shape[channel_axis].is_dynamic() ||
output_shape[channel_axis].is_dynamic()) {
return false;
}

if (groups == data_shape[channel_axis].get_length() &&
groups == output_shape[channel_axis].get_length()) { // depthwise case
return false;
}

Expand All @@ -33,12 +42,12 @@ ov::intel_cpu::ConvertGroupConvolution::ConvertGroupConvolution() {
groups);
replace_nodes.push_back(split_weights);

auto axis = ov::opset8::Constant::create<int64_t>(ngraph::element::i64, ngraph::Shape{}, {1});
auto axis = ov::opset8::Constant::create<int64_t>(ngraph::element::i64, ngraph::Shape{}, {channel_axis});
auto split = std::make_shared<ov::opset1::Split>(gconv->input_value(Inputs::Data), axis, groups);
replace_nodes.push_back(split);

ngraph::NodeVector concat_inputs;
for (size_t g = 0; g < groups; g++) {
for (int64_t g = 0; g < groups; g++) {
auto out = split->output(g);
auto filter = std::make_shared<ov::opset1::Squeeze>(split_weights->output(g),
ov::opset8::Constant::create<int64_t>(ngraph::element::i64, ngraph::Shape{}, {0}));
Expand Down
50 changes: 50 additions & 0 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_group_conv.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,56 @@
#include <ngraph/pattern/op/wrap_type.hpp>
#include <ngraph/pass/graph_rewrite.hpp>

/*
* Description:
* ConvertGroupConvolution detects GroupConvolution and replaces it
* with a set of Convolution operations. Number of Convolution operations
* equals to number of groups.
*
* Before:
*
* +--------------+ +---------------+
* | Input tensor | | Kernel tensor |
* +-----------+--+ +-+-------------+
* | |
* +----v---------v----+
* | Group Convolution |
* +---------+---------+
* |
* +------v------+
* | Result |
* +-------------+
*
* After:
*
* +--------------+ +--------------+ +---------------+ +--------------+
* | Input tensor | | Constant (1) | | Kernel tensor | | Constant (0) |
* +-----------+--+ +-+------------+ +-----------+---+ +-+------------+
* | | | |
* +-v---------v-+ +-v---------v-+
* | Split | | Split |
* +-+-----------+--------+ +-+---------+-+
* | | | |
* | | +-----------v--+ +-v------------+
* | | | Squeeze | | Squeeze |
* | +----------+---+--------------+ +-+------------+
* | | | |
* | | +---------------+ |
* | | | |
* +-----------v---------v------------+ +-----------v---------v------------+
* | Convolution | | Convolution |
* +-----------------------------+----+ +---+------------------------------+
* | |
* +-----v----------v------+
* | Concat |
* +----------+------------+
* |
* +----------v------------+
* | Result |
* +-----------------------+
*
*/

namespace ov {
namespace intel_cpu {

Expand Down
33 changes: 19 additions & 14 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_group_conv1d.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,24 +19,28 @@ ngraph::matcher_pass_callback ov::intel_cpu::ConvertConv1DBase::convert_conv1d_t
return false;
}

auto input_shape = conv->get_input_shape(0);
const auto& input0 = conv->input_value(0);
const auto& input_shape = input0.get_partial_shape();
// is Conv1D
if (input_shape.size() != 3) {
return false;
}

auto input = conv->input_value(0);
auto weights = conv->input_value(1);
auto input2d_shape = input_shape;
input2d_shape.push_back(1);
auto in2d_shape = std::make_shared<ov::opset8::Constant>(ngraph::element::i64, ngraph::Shape{4}, input2d_shape);

auto weights2d_shape = weights.get_shape();
weights2d_shape.push_back(1);
auto w_shape = std::make_shared<ov::opset8::Constant>(ngraph::element::i64, ngraph::Shape{weights2d_shape.size()}, weights2d_shape);

auto input2d = std::make_shared<ov::opset8::Reshape>(input, in2d_shape, true);
auto weights2d = std::make_shared<ov::opset8::Reshape>(weights, w_shape, true);
auto getUnsqueeze = [&](const ngraph::Output<ngraph::Node>& node) {
auto rank = node.get_partial_shape().rank().get_length();
return std::make_shared<ov::opset8::Unsqueeze>(node,
ov::opset1::Constant::create(ngraph::element::i64, ngraph::Shape{1}, {rank}));
};

auto input2d = getUnsqueeze(input);
auto weights2d = getUnsqueeze(weights);

auto conv2d = std::make_shared<Conv>(input2d,
weights2d,
Expand All @@ -46,8 +50,9 @@ ngraph::matcher_pass_callback ov::intel_cpu::ConvertConv1DBase::convert_conv1d_t
ngraph::Strides{conv->get_dilations()[0], 1},
conv->get_auto_pad());

auto in_shape = std::make_shared<ov::opset8::Constant>(ngraph::element::i64, ngraph::Shape{3}, conv->get_output_shape(0));
auto reshape = std::make_shared<ov::opset8::Reshape>(conv2d, in_shape, true);
auto reshape = std::make_shared<ov::opset8::Squeeze>(
conv2d,
ov::opset1::Constant::create(ngraph::element::i64, ngraph::Shape{1}, {input_shape.rank().get_length()}));

reshape->set_friendly_name(conv->get_friendly_name());
ngraph::copy_runtime_info(conv, {input2d, weights2d, conv2d, reshape});
Expand All @@ -58,16 +63,16 @@ ngraph::matcher_pass_callback ov::intel_cpu::ConvertConv1DBase::convert_conv1d_t

ov::intel_cpu::ConvertConv1D::ConvertConv1D() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::Convolution>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::any_input(ngraph::pattern::has_static_shape())},
ngraph::pattern::has_static_shape()), "ConvertConvolutionToArm");
ngraph::pattern::wrap_type<ov::opset8::Convolution>({ngraph::pattern::any_input(),
ngraph::pattern::any_input()}),
"ConvertConvolutionToArm");
register_matcher(m, convert_conv1d_to_conv2d<ov::opset8::Convolution>());
}

ov::intel_cpu::ConvertGroupConv1D::ConvertGroupConv1D() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::GroupConvolution>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::any_input(ngraph::pattern::has_static_shape())},
ngraph::pattern::has_static_shape()), "ConvertGroupConvolutionToArm");
ngraph::pattern::wrap_type<ov::opset8::GroupConvolution>({ngraph::pattern::any_input(),
ngraph::pattern::any_input()}),
"ConvertGroupConvolutionToArm");
register_matcher(m, convert_conv1d_to_conv2d<ov::opset8::GroupConvolution>());
}
45 changes: 45 additions & 0 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_group_conv1d.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,51 @@

#include <ngraph/pass/graph_rewrite.hpp>

/*
* Description:
* ConvertConv1DBase detects 1D Convolution / GroupConvolution and replaces
* it with the sequence Unsqueeze - 2D Convolution / GroupConvolution - Squeeze.
* Unsqueeze adds the additional dimension to Convolution inputs and Squeeze
* removes the additional dimention from the Convolution output.
*
* Before:
*
* +--------------+ +---------------+
* | Input tensor | | Kernel tensor |
* +-----------+--+ +-+-------------+
* | |
* +-v---------v-+
* | Convolution |
* +------+------+
* |
* +------v------+
* | Result |
* +-------------+
*
* After:
*
* +--------------+ +--------------+ +---------------+ +--------------+
* | Input tensor | | Constant (1) | | Kernel tensor | | Constant (1) |
* +-----------+--+ +-+------------+ +-----------+---+ +-+------------+
* | | | |
* +-v---------v-+ +-v---------v-+
* | Unsqueeze | | Unsqueeze |
* +------+------+ +------+------+
* | |
* +------v------------------------------------v------+ +--------------+
* | Convolution | | Constant (1) |
* +---------------------------------------------+----+ +-+------------+
* | |
* +-v---------v-+
* | Squeeze |
* +------+------+
* |
* +------v------+
* | Result |
* +-------------+
*
*/

namespace ov {
namespace intel_cpu {
class ConvertConv1DBase: public ngraph::pass::MatcherPass {
Expand Down
36 changes: 21 additions & 15 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_reduce_multi_axis.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,16 +16,26 @@ ngraph::matcher_pass_callback ov::intel_cpu::ConvertReduceMultiAxisBase::convert
if (!std::dynamic_pointer_cast<T>(reduce)) {
return false;
}
if (ngraph::shape_size(reduce->input_value(1).get_shape()) <= 1) {

const auto& input0 = reduce->input_value(0);
const auto& input1 = reduce->input_value(1);
const auto& data_shape0 = input0.get_partial_shape();
const auto& data_shape1 = input1.get_partial_shape();
if (data_shape0.is_dynamic() ||
data_shape1.is_dynamic()) {
return false;
}

if (ngraph::shape_size(input1.get_shape()) <= 1) {
return false;
}
auto reduction_axes = std::dynamic_pointer_cast<ov::opset8::Constant>(reduce->input_value(1).get_node_shared_ptr());
auto reduction_axes = std::dynamic_pointer_cast<ov::opset8::Constant>(input1.get_node_shared_ptr());
if (!reduction_axes) {
return false;
}
auto axes = reduction_axes->cast_vector<int64_t>();
ngraph::NodeVector new_ops;
std::shared_ptr<ngraph::Node> node = reduce->input_value(0).get_node_shared_ptr();
std::shared_ptr<ngraph::Node> node = input0.get_node_shared_ptr();
for (auto axis : axes) {
auto reduction_axis = ov::opset8::Constant::create<int64_t>(ngraph::element::i64, ngraph::Shape{}, {axis});
node = std::make_shared<T>(node, reduction_axis, true);
Expand All @@ -46,32 +56,28 @@ ngraph::matcher_pass_callback ov::intel_cpu::ConvertReduceMultiAxisBase::convert

ov::intel_cpu::ConvertReduceProd::ConvertReduceProd() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::ReduceProd>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::wrap_type<ov::opset8::Constant>()},
ngraph::pattern::has_static_shape()), "ConvertReduceProd");
ngraph::pattern::wrap_type<ov::opset8::ReduceProd>({ngraph::pattern::any_input(),
ngraph::pattern::wrap_type<ov::opset8::Constant>()}), "ConvertReduceProd");
register_matcher(m, convert_reduce<ov::opset8::ReduceProd>());
}

ov::intel_cpu::ConvertReduceMin::ConvertReduceMin() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::ReduceMin>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::wrap_type<ov::opset8::Constant>()},
ngraph::pattern::has_static_shape()), "ConvertReduceMin");
ngraph::pattern::wrap_type<ov::opset8::ReduceMin>({ngraph::pattern::any_input(),
ngraph::pattern::wrap_type<ov::opset8::Constant>()}), "ConvertReduceMin");
register_matcher(m, convert_reduce<ov::opset8::ReduceMin>());
}

ov::intel_cpu::ConvertReduceMax::ConvertReduceMax() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::ReduceMax>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::wrap_type<ov::opset8::Constant>()},
ngraph::pattern::has_static_shape()), "ConvertReduceMax");
ngraph::pattern::wrap_type<ov::opset8::ReduceMax>({ngraph::pattern::any_input(),
ngraph::pattern::wrap_type<ov::opset8::Constant>()}), "ConvertReduceMax");
register_matcher(m, convert_reduce<ov::opset8::ReduceMax>());
}

ov::intel_cpu::ConvertReduceSum::ConvertReduceSum() {
auto m = std::make_shared<ngraph::pattern::Matcher>(
ngraph::pattern::wrap_type<ov::opset8::ReduceSum>({ngraph::pattern::any_input(ngraph::pattern::has_static_shape()),
ngraph::pattern::wrap_type<ov::opset8::Constant>()},
ngraph::pattern::has_static_shape()), "ConvertReduceSum");
ngraph::pattern::wrap_type<ov::opset8::ReduceSum>({ngraph::pattern::any_input(),
ngraph::pattern::wrap_type<ov::opset8::Constant>()}), "ConvertReduceSum");
register_matcher(m, convert_reduce<ov::opset8::ReduceSum>());
}
44 changes: 44 additions & 0 deletions src/plugins/intel_cpu/src/transformations/cpu_opset/arm/pass/convert_reduce_multi_axis.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,50 @@
#include <ngraph/pattern/op/wrap_type.hpp>
#include <ngraph/pass/graph_rewrite.hpp>

/*
* Description:
* ConvertReduceMultiAxisBase detects Reduce operations that do not support
* multi-axis input by ACL: Min, Max, Sum, Prod. Multi-axis Reduce operation
* is replaced with a sequence of single-axe Reduce operations.
*
* Before:
*
* +--------------+ +-------------------+
* | Data | | Axes tensor [A,B] |
* +-----------+--+ +-+-----------------+
* | |
* +----v---------v----+
* | Reduce |
* +---------+---------+
* |
* +------v------+
* | Result |
* +-------------+
*
* After:
*
* +-------------+ +---------------+
* | Data | | Axes scalar A |
* +---------+---+ +----+----------+
* | |
* +-v------------v--+ +-----------------+
* | Reduce | | Axes scalar B |
* +--------------+--+ +---+-------------+
* | |
* +-v-----------v---+
* | Reduce |
* +-------+---------+
* |
* +-------v---------+
* | Reshape |
* +-------+---------+
* |
* +-------v---------+
* | Result |
* +-----------------+
*
*/

namespace ov {
namespace intel_cpu {

Expand Down
5 changes: 3 additions & 2 deletions src/plugins/intel_cpu/tests/functional/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,11 +22,11 @@ endif()
if(X86_64)
list(APPEND EXCLUDED_SOURCE_PATHS
${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests/instances/arm
${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tests/arm)
${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tests/src/arm)
else()
list(APPEND EXCLUDED_SOURCE_PATHS
${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests/instances/x64
${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tests/x64)
${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tests/src/x64)
# temporary disable all custom tests for ARM
list(APPEND EXCLUDED_SOURCE_PATHS
${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests
Expand All @@ -35,6 +35,7 @@ else()
file(GLOB_RECURSE TMP_LIST_OF_TEST_CLASSES ${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests/classes/*.cpp)
file(GLOB_RECURSE TMP_LIST_OF_COMMON_TEST_INSTANCES ${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests/instances/common/*.cpp)
file(GLOB_RECURSE TMP_LIST_OF_ARM_TEST_INSTANCES ${CMAKE_CURRENT_SOURCE_DIR}/single_layer_tests/instances/arm/*.cpp)
file(GLOB_RECURSE TMP_LIST_OF_ARM_TEST_INSTANCES ${CMAKE_CURRENT_SOURCE_DIR}/subgraph_tests/arm/*.cpp)
list(APPEND TMP_LIST_OF_EXPLICITLY_ENABLED_TESTS
${TMP_LIST_OF_TEST_CLASSES} ${TMP_LIST_OF_COMMON_TEST_INSTANCES} ${TMP_LIST_OF_ARM_TEST_INSTANCES})
set(TMP_EXPLICITLY_ENABLED_TESTS "${TMP_LIST_OF_EXPLICITLY_ENABLED_TESTS}")
Expand Down
Loading

0 comments on commit 1588a33

Please sign in to comment.