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

Merge remote-tracking branch 'upstream/master' into add_mxnet_operations

Browse Source
This commit is contained in:
yekruglov 2021-12-16 18:28:14 +03:00
commit c6ab942ada
138 changed files with 6286 additions and 2092 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
docs/template_plugin/backend/CMakeLists.txt
View File

@ -38,17 +38,10 @@ target_include_directories(interpreter_backend PUBLIC $<BUILD_INTERFACE:${CMAKE_
file(GLOB_RECURSE all_backends_src "${CMAKE_CURRENT_SOURCE_DIR}/*.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/*.hpp")
add_clang_format_target(interpreter_backend_clang FOR_SOURCES ${all_backends_src})
# developer package
openvino_developer_export_targets(COMPONENT core TARGETS interpreter_backend)
install(TARGETS interpreter_backend
RUNTIME DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL
ARCHIVE DESTINATION ${IE_CPACK_RUNTIME_PATH} COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL
LIBRARY DESTINATION ${IE_CPACK_LIBRARY_PATH} COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL)
if(NOT BUILD_SHARED_LIBS)
install(TARGETS interpreter_backend
RUNTIME DESTINATION tests COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL
ARCHIVE DESTINATION tests COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL
LIBRARY DESTINATION tests COMPONENT tests OPTIONAL EXCLUDE_FROM_ALL)
endif()
# install
ov_install_static_lib(interpreter_backend template)

43
docs/template_plugin/backend/evaluates_map.cpp
View File

@ -1707,7 +1707,24 @@ bool evaluate(const shared_ptr<op::v0::Log>& op, const HostTensorVector& outputs
}
namespace ctc_loss_v4 {
template <element::Type_t t1, element::Type_t t2>
template <element::Type_t t1,
element::Type_t t2,
typename std::enable_if<!std::is_floating_point<typename element_type_traits<t1>::value_type>::value &&
!std::is_same<typename element_type_traits<t1>::value_type, bfloat16>::value &&
!std::is_same<typename element_type_traits<t1>::value_type, float16>::value,
bool>::type = true>
inline void evaluate(const shared_ptr<op::v4::CTCLoss>& op,
const HostTensorVector& outputs,
const HostTensorVector& inputs) {
OPENVINO_ASSERT(false, "The data type for logits is expected to be a floating point type. Got:", element::Type(t1));
}
template <element::Type_t t1,
element::Type_t t2,
typename std::enable_if<std::is_floating_point<typename element_type_traits<t1>::value_type>::value ||
std::is_same<typename element_type_traits<t1>::value_type, bfloat16>::value ||
std::is_same<typename element_type_traits<t1>::value_type, float16>::value,
bool>::type = true>
inline void evaluate(const shared_ptr<op::v4::CTCLoss>& op,
const HostTensorVector& outputs,
const HostTensorVector& inputs) {
@ -1944,6 +1961,30 @@ bool evaluate(const shared_ptr<op::v0::RNNCell>& op, const HostTensorVector& out
return true;
}
template <element::Type_t ET>
bool evaluate(const shared_ptr<op::v0::LSTMCell>& op, const HostTensorVector& outputs, const HostTensorVector& inputs) {
using T = typename element_type_traits<ET>::value_type;
runtime::reference::lstm_cell<T>(inputs[0]->get_data_ptr<ET>(),
inputs[0]->get_shape(),
inputs[1]->get_data_ptr<ET>(),
inputs[1]->get_shape(),
inputs[2]->get_data_ptr<ET>(),
inputs[2]->get_shape(),
inputs[3]->get_data_ptr<ET>(),
inputs[3]->get_shape(),
inputs[4]->get_data_ptr<ET>(),
inputs[4]->get_shape(),
inputs[5]->get_data_ptr<ET>(),
inputs[5]->get_shape(),
outputs[0]->get_data_ptr<ET>(),
outputs[1]->get_data_ptr<ET>(),
op->get_activations()[0],
op->get_activations()[1],
op->get_activations()[2],
op->get_clip());
return true;
}
template <element::Type_t ET>
bool evaluate(const shared_ptr<op::v4::LSTMCell>& op, const HostTensorVector& outputs, const HostTensorVector& inputs) {
using T = typename element_type_traits<ET>::value_type;

1
docs/template_plugin/backend/opset_int_tbl.hpp
View File

@ -20,6 +20,7 @@ NGRAPH_OP(Gelu, op::v0)
NGRAPH_OP(GRN, op::v0)
NGRAPH_OP(HardSigmoid, op::v0)
NGRAPH_OP(LRN, ngraph::op::v0)
NGRAPH_OP(LSTMCell, op::v0)
NGRAPH_OP(MVN, ngraph::op::v0)
NGRAPH_OP(NormalizeL2, op::v0)
NGRAPH_OP(PriorBox, ngraph::op::v0)

1
docs/template_plugin/src/CMakeLists.txt
View File

@ -37,4 +37,3 @@ set_target_properties(${TARGET_NAME} PROPERTIES INTERPROCEDURAL_OPTIMIZATION_REL
# ie_register_plugins(MAIN_TARGET ${TARGET_NAME}
# POSSIBLE_PLUGINS ${TARGET_NAME})
# [cmake:plugin]
ov_install_static_lib(interpreter_backend tests)

182
docs/template_plugin/tests/functional/op_reference/einsum.cpp Normal file
View File

@ -0,0 +1,182 @@
// Copyright (C) 2021 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <gtest/gtest.h>
#include "openvino/opsets/opset7.hpp"
#include "openvino/opsets/opset1.hpp"
#include "base_reference_test.hpp"
using namespace reference_tests;
using namespace ov;
namespace {
struct EinsumParams {
std::vector<Tensor> inputs;
std::string equation;
Tensor expectedResult;
std::string testcaseName;
};
struct Builder : ParamsBuilder<EinsumParams> {
REFERENCE_TESTS_ADD_SET_PARAM(Builder, inputs);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, equation);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, expectedResult);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, testcaseName);
};
class ReferenceEinsumTest : public testing::TestWithParam<EinsumParams>, public CommonReferenceTest {
public:
void SetUp() override {
auto params = GetParam();
function = CreateModel(params);
for (const auto& input_tensor : params.inputs) {
inputData.push_back(input_tensor.data);
}
refOutData = {params.expectedResult.data};
}
static std::string getTestCaseName(const testing::TestParamInfo<EinsumParams>& obj) {
auto param = obj.param;
std::ostringstream result;
result << "iType=" << param.inputs[0].type;
result << "_iShape=" << param.inputs[0].shape;
result << "_equation=" << param.equation;
result << "_eType=" << param.expectedResult.type;
result << "_eShape=" << param.expectedResult.shape;
if (param.testcaseName != "") {
result << "_=" << param.testcaseName;
}
return result.str();
}
private:
static std::shared_ptr<Model> CreateModel(const EinsumParams& params) {
OutputVector output_vector;
ParameterVector param_vector;
for (const auto& input_tensor : params.inputs) {
auto param = std::make_shared<opset1::Parameter>(input_tensor.type, input_tensor.shape);
output_vector.push_back(param);
param_vector.push_back(param);
}
const auto einsum = std::make_shared<opset7::Einsum>(output_vector, params.equation);
const auto f = std::make_shared<Model>(OutputVector{einsum}, param_vector);
return f;
}
};
TEST_P(ReferenceEinsumTest, CompareWithRefs) {
Exec();
}
template <element::Type_t ET>
std::vector<EinsumParams> generateParams() {
using T = typename element_type_traits<ET>::value_type;
std::vector<EinsumParams> params {
Builder {}
.inputs({{ET, {1, 2}, std::vector<T>{1, 2}},
{ET, {3, 4}, std::vector<T>{3, 4, 5, 6,
7, 8, 9, 10,
11, 12, 13, 14}}})
.equation("ab,cd->abcd")
.expectedResult({ET, {1, 2, 3, 4}, std::vector<T>{3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 6, 8, 10, 12,
14, 16, 18, 20, 22, 24, 26, 28}})
.testcaseName("einsum_no_reduction"),
Builder {}
.inputs({{ET, {1, 2, 3}, std::vector<T>{1, 2, 3, 4, 5, 6}}})
.equation("ijk->kij")
.expectedResult({ET, {3, 1, 2}, std::vector<T>{1, 4, 2, 5, 3, 6}})
.testcaseName("einsum_transpose"),
Builder {}
.inputs({{ET, {2, 3}, std::vector<T>{1, 2, 3, 4, 5, 6}}})
.equation("ab->a")
.expectedResult({ET, {2}, std::vector<T>{6, 15}})
.testcaseName("einsum_reduce"),
Builder {}
.inputs({{ET, {2, 3}, std::vector<T>{1, 2, 3, 4, 5, 6}},
{ET, {3, 2}, std::vector<T>{1, 2, 3, 4, 5, 6}}})
.equation("ab,bc->ac")
.expectedResult({ET, {2, 2}, std::vector<T>{22, 28, 49, 64}})
.testcaseName("einsum_matrix_multiplication"),
Builder {}
.inputs({{ET, {2, 4}, std::vector<T>{1, 3, 2, 7, 5, 6, 0, 1}},
{ET, {4, 3, 1}, std::vector<T>{1, 2, 3, 4, 5, 6, 5, 7, 3, 7, 9, 1}},
{ET, {4, 3}, std::vector<T>{4, 3, 1, 6, 4, 2, 2, 5, 3, 1, 9, 4}}})
.equation("ab,bcd,bc->ca")
.expectedResult({ET, {3, 2}, std::vector<T>{145, 171, 703, 231, 85, 91}})
.testcaseName("einsum_multiple_multiplication"),
Builder {}
.inputs({{ET, {2, 2, 3}, std::vector<T>{1, 3, 2, 7, 5, 6, 3, 5, 2, 1, 0, 7}}})
.equation("a...->...")
.expectedResult({ET, {2, 3}, std::vector<T>{4, 8, 4, 8, 5, 13}})
.testcaseName("einsum_ellipsis_one_input_reduction"),
Builder {}
.inputs({{ET, {2, 2, 3}, std::vector<T>{1, 3, 2, 7, 5, 6, 3, 5, 2, 1, 0, 7}}})
.equation("a...->...a")
.expectedResult({ET, {2, 3, 2}, std::vector<T>{1, 3, 3, 5, 2, 2, 7, 1, 5, 0, 6, 7}})
.testcaseName("einsum_ellipsis_one_input_transpose"),
Builder {}
.inputs({{ET, {2, 2, 3}, std::vector<T>{1, 3, 2, 7, 5, 6, 3, 5, 2, 1, 0, 7}},
{ET, {1}, std::vector<T>{2}}})
.equation("ab...,...->ab...")
.expectedResult({ET, {2, 2, 3}, std::vector<T>{2, 6, 4, 14, 10, 12, 6, 10, 4, 2, 0, 14}})
.testcaseName("einsum_ellipsis_mul_by_1dscalar"),
Builder {}
.inputs({{ET, {1, 1, 4, 3}, std::vector<T>{1, 3, 2, 7, 5, 6, 3, 5, 2, 1, 0, 7}},
{ET, {3, 4, 2, 1}, std::vector<T>{3, 1, 6, 2, 3, 10, 9, 8, 2, 9, 3, 2,
4, 2, 3, 1, 9, 1, 11, 4, 7, 2, 3, 1}}})
.equation("a...j,j...->a...")
.expectedResult({ET, {1, 4, 2, 4}, std::vector<T>{27, 85, 37, 66, 30, 58, 50, 8,
37, 123, 55, 83, 16, 48, 24, 30,
29, 83, 43, 52, 20, 92, 44, 24,
24, 96, 48, 30, 13, 67, 31, 15}})
.testcaseName("einsum_ellipsis_complex_mul"),
Builder {}
.inputs({{ET, {1, 3, 3}, std::vector<T>{1, 2, 3, 4, 5, 6, 7, 8, 9}}})
.equation("kii->ki")
.expectedResult({ET, {1, 3}, std::vector<T>{1, 5, 9}})
.testcaseName("einsum_diagonal"),
Builder {}
.inputs({{ET, {2, 3, 3, 2, 4}, std::vector<T>{4, 2, 5, 4, 5, 5, 1, 1, 3, 3, 1, 1, 2, 2, 4, 1, 3, 4,
4, 5, 1, 3, 1, 3, 1, 4, 3, 5, 4, 4, 5, 4, 4, 5, 4, 2,
2, 2, 3, 3, 1, 1, 4, 3, 4, 2, 2, 1, 1, 2, 3, 1, 1, 4,
2, 3, 1, 3, 4, 2, 5, 5, 3, 4, 3, 4, 5, 4, 4, 5, 1, 3,
4, 4, 5, 3, 1, 3, 2, 5, 3, 2, 5, 4, 4, 2, 4, 4, 1, 4,
4, 5, 4, 4, 4, 2, 3, 3, 4, 2, 4, 2, 5, 1, 3, 2, 4, 3,
5, 1, 2, 3, 1, 1, 2, 5, 1, 1, 2, 1, 4, 5, 3, 4, 1, 3,
3, 1, 3, 2, 4, 5, 1, 1, 5, 4, 5, 2, 2, 3, 3, 1, 2, 4}},
{ET, {3, 2, 1}, std::vector<T>{1, 4, 4, 5, 3, 3}}})
.equation("abbac,bad->ad")
.expectedResult({ET, {2, 1}, std::vector<T>{123, 129}})
.testcaseName("einsum_diagonal_with_matmul"),
};
return params;
}
std::vector<EinsumParams> generateCombinedParams() {
const std::vector<std::vector<EinsumParams>> generatedParams {
generateParams<element::Type_t::i32>(),
generateParams<element::Type_t::f32>(),
};
std::vector<EinsumParams> combinedParams;
for (const auto& params : generatedParams) {
combinedParams.insert(combinedParams.end(), params.begin(), params.end());
}
return combinedParams;
}
INSTANTIATE_TEST_SUITE_P(smoke_Einsum_With_Hardcoded_Refs, ReferenceEinsumTest,
testing::ValuesIn(generateCombinedParams()), ReferenceEinsumTest::getTestCaseName);
} // namespace

246
docs/template_plugin/tests/functional/op_reference/extract_image_patches.cpp Normal file
View File

@ -0,0 +1,246 @@
// Copyright (C) 2021 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <gtest/gtest.h>
#include "openvino/opsets/opset3.hpp"
#include "openvino/opsets/opset1.hpp"
#include "base_reference_test.hpp"
using namespace reference_tests;
using namespace ov;
namespace {
struct ExtractImagePatchesParams {
Tensor data;
Shape sizes;
Strides strides;
Shape rates;
op::PadType autoPad;
Tensor expectedResult;
std::string testcaseName;
};
struct Builder : ParamsBuilder<ExtractImagePatchesParams> {
REFERENCE_TESTS_ADD_SET_PARAM(Builder, data);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, sizes);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, strides);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, rates);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, autoPad);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, expectedResult);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, testcaseName);
};
class ReferenceExtractImagePatchesTest : public testing::TestWithParam<ExtractImagePatchesParams>, public CommonReferenceTest {
public:
void SetUp() override {
auto params = GetParam();
function = CreateModel(params);
inputData = {params.data.data};
refOutData = {params.expectedResult.data};
}
static std::string getTestCaseName(const testing::TestParamInfo<ExtractImagePatchesParams>& obj) {
auto param = obj.param;
std::ostringstream result;
result << "dType=" << param.data.type;
result << "_dShape=" << param.data.shape;
result << "_sizes=" << param.sizes;
result << "_strides=" << param.strides;
result << "_rates=" << param.rates;
result << "_autoPad=" << param.autoPad;
result << "_eType=" << param.expectedResult.type;
result << "_eShape=" << param.expectedResult.shape;
if (param.testcaseName != "") {
result << "_=" << param.testcaseName;
}
return result.str();
}
private:
static std::shared_ptr<Model> CreateModel(const ExtractImagePatchesParams& params) {
const auto data = std::make_shared<opset1::Parameter>(params.data.type, params.data.shape);
const auto extrace_image_patches = std::make_shared<opset3::ExtractImagePatches>(data,
params.sizes,
params.strides,
params.rates,
params.autoPad);
const auto f = std::make_shared<Model>(extrace_image_patches, ParameterVector{data});
return f;
}
};
TEST_P(ReferenceExtractImagePatchesTest, CompareWithRefs) {
Exec();
}
template <element::Type_t ET>
std::vector<ExtractImagePatchesParams> generateParams() {
using T = typename element_type_traits<ET>::value_type;
std::vector<ExtractImagePatchesParams> params {
Builder {}
.data({ET, {1, 1, 10, 10}, std::vector<T>{
1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47, 48, 49, 50,
51, 52, 53, 54, 55, 56, 57, 58, 59, 60,
61, 62, 63, 64, 65, 66, 67, 68, 69, 70,
71, 72, 73, 74, 75, 76, 77, 78, 79, 80,
81, 82, 83, 84, 85, 86, 87, 88, 89, 90,
91, 92, 93, 94, 95, 96, 97, 98, 99, 100}})
.sizes({3, 3})
.strides({5, 5})
.rates({1, 1})
.autoPad(op::PadType::VALID)
.expectedResult({ET, {1, 9, 2, 2}, std::vector<T>{
1, 6, 51, 56,
2, 7, 52, 57,
3, 8, 53, 58,
11, 16, 61, 66,
12, 17, 62, 67,
13, 18, 63, 68,
21, 26, 71, 76,
22, 27, 72, 77,
23, 28, 73, 78}}),
Builder {}
.data({ET, {1, 1, 10, 10}, std::vector<T>{
1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47, 48, 49, 50,
51, 52, 53, 54, 55, 56, 57, 58, 59, 60,
61, 62, 63, 64, 65, 66, 67, 68, 69, 70,
71, 72, 73, 74, 75, 76, 77, 78, 79, 80,
81, 82, 83, 84, 85, 86, 87, 88, 89, 90,
91, 92, 93, 94, 95, 96, 97, 98, 99, 100}})
.sizes({4, 4})
.strides({8, 8})
.rates({1, 1})
.autoPad(op::PadType::VALID)
.expectedResult({ET, {1, 16, 1, 1}, std::vector<T>{
1, 2, 3, 4,
11, 12, 13, 14,
21, 22, 23, 24,
31, 32, 33, 34}}),
Builder {}
.data({ET, {1, 1, 10, 10}, std::vector<T>{
1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47, 48, 49, 50,
51, 52, 53, 54, 55, 56, 57, 58, 59, 60,
61, 62, 63, 64, 65, 66, 67, 68, 69, 70,
71, 72, 73, 74, 75, 76, 77, 78, 79, 80,
81, 82, 83, 84, 85, 86, 87, 88, 89, 90,
91, 92, 93, 94, 95, 96, 97, 98, 99, 100}})
.sizes({4, 4})
.strides({9, 9})
.rates({1, 1})
.autoPad(op::PadType::SAME_UPPER)
.expectedResult({ET, {1, 16, 2, 2}, std::vector<T>{
0, 0, 0, 89,
0, 0, 81, 90,
0, 0, 82, 0,
0, 0, 83, 0,
0, 9, 0, 99,
1, 10, 91, 100,
2, 0, 92, 0,
3, 0, 93, 0,
0, 19, 0, 0,
11, 20, 0, 0,
12, 0, 0, 0,
13, 0, 0, 0,
0, 29, 0, 0,
21, 30, 0, 0,
22, 0, 0, 0,
23, 0, 0, 0}}),
Builder {}
.data({ET, {1, 1, 10, 10}, std::vector<T>{
1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47, 48, 49, 50,
51, 52, 53, 54, 55, 56, 57, 58, 59, 60,
61, 62, 63, 64, 65, 66, 67, 68, 69, 70,
71, 72, 73, 74, 75, 76, 77, 78, 79, 80,
81, 82, 83, 84, 85, 86, 87, 88, 89, 90,
91, 92, 93, 94, 95, 96, 97, 98, 99, 100}})
.sizes({3, 3})
.strides({5, 5})
.rates({2, 2})
.autoPad(op::PadType::VALID)
.expectedResult({ET, {1, 9, 2, 2}, std::vector<T>{
1, 6, 51, 56,
3, 8, 53, 58,
5, 10, 55, 60,
21, 26, 71, 76,
23, 28, 73, 78,
25, 30, 75, 80,
41, 46, 91, 96,
43, 48, 93, 98,
45, 50, 95, 100}}),
Builder {}
.data({ET, {1, 2, 5, 5}, std::vector<T>{
1, 2, 3, 4, 5,
6, 7, 8, 9, 10,
11, 12, 13, 14, 15,
16, 17, 18, 19, 20,
21, 22, 23, 24, 25,
26, 27, 28, 29, 30,
31, 32, 33, 34, 35,
36, 37, 38, 39, 40,
41, 42, 43, 44, 45,
46, 47, 48, 49, 50}})
.sizes({2, 2})
.strides({3, 3})
.rates({1, 1})
.autoPad(op::PadType::VALID)
.expectedResult({ET, {1, 8, 2, 2}, std::vector<T>{
1, 4, 16, 19,
26, 29, 41, 44,
2, 5, 17, 20,
27, 30, 42, 45,
6, 9, 21, 24,
31, 34, 46, 49,
7, 10, 22, 25,
32, 35, 47, 50}}),
};
return params;
}
std::vector<ExtractImagePatchesParams> generateCombinedParams() {
const std::vector<std::vector<ExtractImagePatchesParams>> generatedParams {
generateParams<element::Type_t::i8>(),
generateParams<element::Type_t::i16>(),
generateParams<element::Type_t::i32>(),
generateParams<element::Type_t::i64>(),
generateParams<element::Type_t::u8>(),
generateParams<element::Type_t::u16>(),
generateParams<element::Type_t::u32>(),
generateParams<element::Type_t::u64>(),
generateParams<element::Type_t::bf16>(),
generateParams<element::Type_t::f16>(),
generateParams<element::Type_t::f32>(),
generateParams<element::Type_t::f64>(),
};
std::vector<ExtractImagePatchesParams> combinedParams;
for (const auto& params : generatedParams) {
combinedParams.insert(combinedParams.end(), params.begin(), params.end());
}
return combinedParams;
}
INSTANTIATE_TEST_SUITE_P(smoke_ExtractImagePatches_With_Hardcoded_Refs, ReferenceExtractImagePatchesTest,
testing::ValuesIn(generateCombinedParams()), ReferenceExtractImagePatchesTest::getTestCaseName);
} // namespace

589
docs/template_plugin/tests/functional/op_reference/lstm_cell.cpp
View File

@ -4,7 +4,8 @@
#include <gtest/gtest.h>
#include "openvino/op/lstm_cell.hpp"
#include "openvino/opsets/opset4.hpp"
#include "openvino/opsets/opset1.hpp"
#include "base_reference_test.hpp"
using namespace reference_tests;
@ -12,13 +13,6 @@ using namespace ov;
namespace {
struct LSTMCellParams {
LSTMCellParams(
int32_t batchSize, int32_t inputSize, int32_t hiddenSize, int32_t gatesCount,
const Tensor& X, const Tensor& W, const Tensor& R, const Tensor& H_t, const Tensor& C_t, const Tensor& B,
const Tensor& Ho, const Tensor& Co, const std::string& testcaseName = "") :
batchSize(batchSize), inputSize(inputSize), hiddenSize(hiddenSize), gatesCount(gatesCount),
X(X), W(W), R(R), H_t(H_t), C_t(C_t), B(B), Ho(Ho), Co(Co), testcaseName(testcaseName) {}
int32_t batchSize;
int32_t inputSize;
int32_t hiddenSize;
@ -34,6 +28,22 @@ struct LSTMCellParams {
std::string testcaseName;
};
struct Builder : ParamsBuilder<LSTMCellParams> {
REFERENCE_TESTS_ADD_SET_PARAM(Builder, batchSize);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, inputSize);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, hiddenSize);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, gatesCount);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, X);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, W);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, R);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, H_t);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, C_t);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, B);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, Ho);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, Co);
REFERENCE_TESTS_ADD_SET_PARAM(Builder, testcaseName);
};
class ReferenceLSTMCellTest : public testing::TestWithParam<LSTMCellParams>, public CommonReferenceTest {
public:
void SetUp() override {
@ -63,26 +73,24 @@ public:
result << "_hoType=" << param.Ho.type;
result << "_hoShape=" << param.Ho.shape;
result << "_coType=" << param.Co.type;
result << "_coShape=" << param.Co.shape;
if (param.testcaseName != "") {
result << "_coShape=" << param.Co.shape;
result << "_=" << param.testcaseName;
} else {
result << "_coShape=" << param.Co.shape;
}
return result.str();
}
private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const auto X = std::make_shared<op::v0::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<op::v0::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<op::v0::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<op::v0::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<op::v0::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<op::v0::Parameter>(params.B.type, params.B.shape);
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<op::v4::LSTMCell>(X,
std::make_shared<opset4::LSTMCell>(X,
H_t,
C_t,
op::util::convert_lstm_node_format(W, op::util::LSTMWeightsFormat::IOFC),
@ -107,15 +115,15 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const auto X = std::make_shared<op::v0::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<op::v0::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<op::v0::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<op::v0::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<op::v0::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<op::v0::Parameter>(params.B.type, params.B.shape);
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<op::v4::LSTMCell>(X,
std::make_shared<opset4::LSTMCell>(X,
H_t,
C_t,
op::util::convert_lstm_node_format(W, op::util::LSTMWeightsFormat::IOFC),
@ -142,15 +150,15 @@ private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const float clip_threshold = 3.5f;
const auto X = std::make_shared<op::v0::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<op::v0::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<op::v0::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<op::v0::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<op::v0::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<op::v0::Parameter>(params.B.type, params.B.shape);
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<op::v4::LSTMCell>(X,
std::make_shared<opset4::LSTMCell>(X,
H_t,
C_t,
W,
@ -179,36 +187,130 @@ TEST_P(ReferenceLSTMCellTestBiasClip, CompareWithRefs) {
Exec();
}
class ReferenceLSTMCellV1Test : public ReferenceLSTMCellTest {
private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<opset1::LSTMCell>(X,
H_t,
C_t,
op::util::convert_lstm_node_format(W, op::util::LSTMWeightsFormat::IOFC),
op::util::convert_lstm_node_format(R, op::util::LSTMWeightsFormat::IOFC),
op::util::convert_lstm_node_format(B, op::util::LSTMWeightsFormat::IOFC),
params.hiddenSize);
auto function = std::make_shared<Model>(lstm_cell->outputs(), ParameterVector{X, H_t, C_t, W, R, B});
return function;
}
};
class ReferenceLSTMCellV1TestBiasDefaultAttrs : public ReferenceLSTMCellTestBiasDefaultAttrs {
private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<opset1::LSTMCell>(X,
H_t,
C_t,
op::util::convert_lstm_node_format(W, op::util::LSTMWeightsFormat::IOFC),
op::util::convert_lstm_node_format(R, op::util::LSTMWeightsFormat::IOFC),
op::util::convert_lstm_node_format(B, op::util::LSTMWeightsFormat::IOFC),
params.hiddenSize);
auto function = std::make_shared<Model>(lstm_cell->outputs(), ParameterVector{X, H_t, C_t, W, R, B});
return function;
}
};
class ReferenceLSTMCellV1TestBiasClip : public ReferenceLSTMCellTestBiasClip {
private:
static std::shared_ptr<Model> CreateFunction(const LSTMCellParams& params) {
const float clip_threshold = 3.5f;
const auto X = std::make_shared<opset1::Parameter>(params.X.type, params.X.shape);
const auto W = std::make_shared<opset1::Parameter>(params.W.type, params.W.shape);
const auto R = std::make_shared<opset1::Parameter>(params.R.type, params.R.shape);
const auto H_t = std::make_shared<opset1::Parameter>(params.H_t.type, params.H_t.shape);
const auto C_t = std::make_shared<opset1::Parameter>(params.C_t.type, params.C_t.shape);
const auto B = std::make_shared<opset1::Parameter>(params.B.type, params.B.shape);
const auto lstm_cell =
std::make_shared<opset1::LSTMCell>(X,
H_t,
C_t,
W,
R,
B,
params.hiddenSize,
op::LSTMWeightsFormat::IFCO,
std::vector<std::string>{"sigmoid", "tanh", "tanh"},
std::vector<float>{},
std::vector<float>{},
clip_threshold);
auto function = std::make_shared<Model>(lstm_cell->outputs(), ParameterVector{X, H_t, C_t, W, R, B});
return function;
}
};
TEST_P(ReferenceLSTMCellV1Test, CompareWithRefs) {
Exec();
}
TEST_P(ReferenceLSTMCellV1TestBiasDefaultAttrs, CompareWithRefs) {
Exec();
}
TEST_P(ReferenceLSTMCellV1TestBiasClip, CompareWithRefs) {
Exec();
}
template <element::Type_t ET>
std::vector<LSTMCellParams> generateParams() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
LSTMCellParams(
2, 3, 3, 4,
Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}),
Tensor(ET, {4 * 3}, std::vector<T>(4 * 3, 0.f)),
Tensor(ET, {2, 3}, std::vector<T>{0.81457126f, 0.61109227f, 0.769522f, 0.52239674f, 0.4324641f, 0.63183f}),
Tensor(ET, {2, 3}, std::vector<T>{1.4444952f, 0.9635685f, 1.2875274f, 0.8053419f, 0.7184521f, 0.95803297f}),
"lstm_cell_zero_bias_default_attrs"),
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>(4 * 3, 0.f)))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81457126f, 0.61109227f, 0.769522f, 0.52239674f, 0.4324641f, 0.63183f}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.4444952f, 0.9635685f, 1.2875274f, 0.8053419f, 0.7184521f, 0.95803297f}))
.testcaseName("lstm_cell_zero_bias_default_attrs")
};
return params;
}
@ -232,53 +334,56 @@ template <element::Type_t ET>
std::vector<LSTMCellParams> generateParamsBiasDefaultAttrs() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
LSTMCellParams(
2, 3, 3, 4,
Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}),
Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}),
Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
0.76665538549423218,
0.82509011030197144,
0.6479143500328064,
0.66586339473724365,
0.74838578701019287}),
Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
0.74838578701019287}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
1.1150213479995728,
1.4578367471694946,
1.0649888515472412,
0.93761754035949707,
1.3659683465957642}),
"lstm_cell_bias_default_attrs"),
1.3659683465957642}))
.testcaseName("lstm_cell_bias_default_attrs"),
};
return params;
}
@ -302,53 +407,56 @@ template <element::Type_t ET>
std::vector<LSTMCellParams> generateParamsBiasClip() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
LSTMCellParams(
2, 3, 3, 4,
Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}),
Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}),
Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}),
Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}),
Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
0.76665538549423218,
0.82387429475784302,
0.6479143500328064,
0.66586339473724365,
0.74838578701019287}),
Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
0.74838578701019287}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
1.1150213479995728,
1.4510968923568726,
1.0649888515472412,
0.93761754035949707,
1.3659683465957642}),
"lstm_cell_bias_clip"),
1.3659683465957642}))
.testcaseName("lstm_cell_bias_clip"),
};
return params;
}
@ -376,4 +484,211 @@ INSTANTIATE_TEST_SUITE_P(smoke_LSTMCell_With_Hardcoded_Refs, ReferenceLSTMCellTe
INSTANTIATE_TEST_SUITE_P(smoke_LSTMCell_With_Hardcoded_Refs, ReferenceLSTMCellTestBiasClip,
testing::ValuesIn(generateCombinedParamsBiasClip()), ReferenceLSTMCellTest::getTestCaseName);
} // namespace
template <element::Type_t ET>
std::vector<LSTMCellParams> generateParamsV1() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>(4 * 3, 0.f)))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81457126f, 0.61109227f, 0.769522f, 0.52239674f, 0.4324641f, 0.63183f}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.4444952f, 0.9635685f, 1.2875274f, 0.8053419f, 0.7184521f, 0.95803297f}))
.testcaseName("lstm_cell_v1_zero_bias_default_attrs")
};
return params;
}
std::vector<LSTMCellParams> generateCombinedParamsV1() {
const std::vector<std::vector<LSTMCellParams>> generatedParams {
generateParamsV1<element::Type_t::bf16>(),
generateParamsV1<element::Type_t::f16>(),
generateParamsV1<element::Type_t::f32>(),
generateParamsV1<element::Type_t::f64>(),
};
std::vector<LSTMCellParams> combinedParams;
for (const auto& params : generatedParams) {
combinedParams.insert(combinedParams.end(), params.begin(), params.end());
}
return combinedParams;
}
template <element::Type_t ET>
std::vector<LSTMCellParams> generateParamsBiasDefaultAttrsV1() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
0.76665538549423218,
0.82509011030197144,
0.6479143500328064,
0.66586339473724365,
0.74838578701019287}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
1.1150213479995728,
1.4578367471694946,
1.0649888515472412,
0.93761754035949707,
1.3659683465957642}))
.testcaseName("lstm_cell_v1_bias_default_attrs"),
};
return params;
}
std::vector<LSTMCellParams> generateCombinedParamsBiasDefaultAttrsV1() {
const std::vector<std::vector<LSTMCellParams>> generatedParams {
generateParamsBiasDefaultAttrsV1<element::Type_t::bf16>(),
generateParamsBiasDefaultAttrsV1<element::Type_t::f16>(),
generateParamsBiasDefaultAttrsV1<element::Type_t::f32>(),
generateParamsBiasDefaultAttrsV1<element::Type_t::f64>(),
};
std::vector<LSTMCellParams> combinedParams;
for (const auto& params : generatedParams) {
combinedParams.insert(combinedParams.end(), params.begin(), params.end());
}
return combinedParams;
}
template <element::Type_t ET>
std::vector<LSTMCellParams> generateParamsBiasClipV1() {
using T = typename element_type_traits<ET>::value_type;
std::vector<LSTMCellParams> params {
Builder {}
.batchSize(2)
.inputSize(3)
.hiddenSize(3)
.gatesCount(4)
.X(Tensor(ET, {2, 3}, std::vector<T>{
0.81342685f, 0.84108883f, 0.8152282f, 0.46893653f, 0.0901856f, 0.37088776f}))
.W(Tensor(ET, {4 * 3, 3}, std::vector<T>{
3.3330739e-01f, 3.6229487e-04f, 4.6773660e-01f, 4.3046016e-01f, 7.3950343e-02f, 3.8063636e-01f,
9.6921772e-01f, 9.6897459e-01f, 6.2964785e-01f, 3.1134409e-01f, 8.4709978e-01f, 9.4928098e-01f,
6.1676943e-01f, 6.6020679e-01f, 1.9072217e-01f, 8.8032126e-02f, 4.0472135e-01f, 6.8342745e-01f,
8.3432144e-01f, 4.4928190e-01f, 7.9524308e-01f, 5.3966165e-01f, 8.5936421e-01f, 8.3136767e-01f,
5.5125546e-02f, 4.7791195e-01f, 3.5788772e-01f, 6.7507404e-01f, 2.1716513e-01f, 2.7473119e-01f,
3.3999152e-02f, 9.6835363e-01f, 3.7581277e-01f, 2.4026000e-01f, 6.7418844e-01f, 3.4199652e-01f}))
.R(Tensor(ET, {4 * 3, 3}, std::vector<T>{
0.0987983f, 0.52032113f, 0.5848073f, 0.5356095f, 0.74497133f, 0.73260087f,
0.1700787f, 0.45684233f, 0.1495722f, 0.42734373f, 0.4433832f, 0.25906256f,
0.03854987f, 0.47480518f, 0.37215272f, 0.99890584f, 0.74019486f, 0.3518967f,
0.6881257f, 0.8170279f, 0.54088944f, 0.81225616f, 0.14619833f, 0.42941234f,
0.86843914f, 0.45967972f, 0.6237719f, 0.11074839f, 0.6029616f, 0.3149305f,
0.46504205f, 0.5843412f, 0.8733427f, 0.7687243f, 0.07074859f, 0.39188156f}))
.H_t(Tensor(ET, {2, 3}, std::vector<T>{
0.77956f, 0.5331557f, 0.04297554f, 0.7962175f, 0.7635707f, 0.11989366f}))
.C_t(Tensor(ET, {2, 3}, std::vector<T>{
0.8488452f, 0.18851636f, 0.5020695f, 0.29716516f, 0.06740791f, 0.45384037f}))
.B(Tensor(ET, {4 * 3}, std::vector<T>{1.07393714f,
1.15248052f,
1.16671345f,
0.21450312f,
1.2380678f,
1.51688835f,
0.46718366f,
0.91810346f,
1.1274234f,
0.51022074f,
1.11389844f,
0.74174305f}))
.Ho(Tensor(ET, {2, 3}, std::vector<T>{0.81014400720596313,
0.76665538549423218,
0.82387429475784302,
0.6479143500328064,
0.66586339473724365,
0.74838578701019287}))
.Co(Tensor(ET, {2, 3}, std::vector<T>{1.6800162792205811,
1.1150213479995728,
1.4510968923568726,
1.0649888515472412,
0.93761754035949707,
1.3659683465957642}))
.testcaseName("lstm_cell_v1_bias_clip"),
};
return params;
}
std::vector<LSTMCellParams> generateCombinedParamsBiasClipV1() {
const std::vector<std::vector<LSTMCellParams>> generatedParams {
generateParamsBiasClipV1<element::Type_t::bf16>(),
generateParamsBiasClipV1<element::Type_t::f16>(),
generateParamsBiasClipV1<element::Type_t::f32>(),
generateParamsBiasClipV1<element::Type_t::f64>(),
};
std::vector<LSTMCellParams> combinedParams;
for (const auto& params : generatedParams) {
combinedParams.insert(combinedParams.end(), params.begin(), params.end());
}
return combinedParams;
}
INSTANTIATE_TEST_SUITE_P(smoke_LSTMCellV1_With_Hardcoded_Refs, ReferenceLSTMCellV1Test,
testing::ValuesIn(generateCombinedParamsV1()), ReferenceLSTMCellV1Test::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_LSTMCellV1_With_Hardcoded_Refs, ReferenceLSTMCellV1TestBiasDefaultAttrs,
testing::ValuesIn(generateCombinedParamsBiasDefaultAttrsV1()), ReferenceLSTMCellV1Test::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_LSTMCellV1_With_Hardcoded_Refs, ReferenceLSTMCellV1TestBiasClip,
testing::ValuesIn(generateCombinedParamsBiasClipV1()), ReferenceLSTMCellV1Test::getTestCaseName);
} // namespace

481
docs/template_plugin/tests/functional/op_reference/topk.cpp
View File

@ -4,8 +4,8 @@
#include <gtest/gtest.h>
#include "openvino/op/topk.hpp"
#include "openvino/op/constant.hpp"
#include "openvino/opsets/opset3.hpp"
#include "openvino/opsets/opset1.hpp"
#include "base_reference_test.hpp"
using namespace reference_tests;
@ -15,7 +15,7 @@ namespace {
struct TopKParams {
TopKParams(
const Tensor& A, const Tensor& k, const int64_t axis,
const op::v1::TopK::Mode mode, const op::v1::TopK::SortType sort,
const opset1::TopK::Mode mode, const opset1::TopK::SortType sort,
const Tensor& result0, const Tensor& result1, const size_t outIdx,
const std::string& testcaseName = "") :
A(A), k(k), axis(axis), mode(mode), sort(sort),
@ -25,8 +25,8 @@ struct TopKParams {
Tensor A;
Tensor k;
int64_t axis;
op::v1::TopK::Mode mode;
op::v1::TopK::SortType sort;
opset1::TopK::Mode mode;
opset1::TopK::SortType sort;
Tensor result0;
Tensor result1;
size_t outIdx;
@ -71,7 +71,6 @@ struct TopKParamsResnet50 {
std::string testcaseName;
};
class ReferenceTopKTestResnet50 : public testing::TestWithParam<TopKParamsResnet50>, public CommonReferenceTest {
public:
void SetUp() override {
@ -101,18 +100,18 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const TopKParamsResnet50& params) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto B = std::make_shared<op::v1::TopK>(A,
op::v0::Constant::create(element::i64, {}, {5}),
const auto B = std::make_shared<opset1::TopK>(A,
opset1::Constant::create(element::i64, {}, {5}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES);
const auto C = std::make_shared<op::v1::TopK>(A,
op::v0::Constant::create(element::i64, {}, {1}),
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES);
const auto C = std::make_shared<opset1::TopK>(A,
opset1::Constant::create(element::i64, {}, {1}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES);
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES);
const auto out5_value = B->output(0);
const auto out5_index = B->output(1);
@ -220,12 +219,12 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = op::v0::Constant::create(params.k.type,
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<op::v1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto B = std::make_shared<opset1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(B->outputs(), ParameterVector{A});
return f;
}
@ -253,8 +252,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::NONE,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::NONE,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape, std::vector<size_t> shape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -292,8 +291,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::NONE,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::NONE,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -331,8 +330,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape, std::vector<size_t> shape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -366,8 +365,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -401,8 +400,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape, std::vector<size_t> shape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -440,8 +439,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
}({128, 1000})),
Tensor(ET2, {}, std::vector<T2>{5}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {128, 5}, [](std::vector<size_t> rshape) -> std::vector<T>{
std::vector<T> expected_value;
for (size_t i = 0; i < rshape[0]; i++) {
@ -467,8 +466,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{5, 4, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{3, 4, 0}),
0,
@ -478,8 +477,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {3}, std::vector<T>{3, 5, 4}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{0, 3, 4}),
0,
@ -489,8 +488,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{1, 2, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{1, 2, 0}),
0,
@ -500,8 +499,8 @@ std::vector<TopKParams> generateParamsMaxMinSort() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {3}, std::vector<T>{3, 1, 2}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{0, 1, 2}),
0,
@ -536,7 +535,7 @@ std::vector<TopKParams> generateCombinedParamsMaxMinSort() {
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestMaxMinSort,
testing::ValuesIn(generateCombinedParamsMaxMinSort()), ReferenceTopKTest::getTestCaseName);
class ReferenceTopKTestV3 : public ReferenceTopKTest {
class ReferenceTopKTestBackend : public ReferenceTopKTest {
public:
void SetUp() override {
auto params = GetParam();
@ -547,18 +546,18 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = op::v0::Constant::create(params.k.type,
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<op::v3::TopK>(A, k, params.axis, params.mode, params.sort);
const auto B = std::make_shared<opset1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(B->outputs(), ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestV3, CompareWithRefs) {
TEST_P(ReferenceTopKTestBackend, CompareWithRefs) {
Exec();
}
@ -572,8 +571,8 @@ std::vector<TopKParams> generateParamsV3() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{5, 4, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{3, 4, 0}),
0,
@ -583,8 +582,8 @@ std::vector<TopKParams> generateParamsV3() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {3}, std::vector<T>{3, 5, 4}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{0, 3, 4}),
0,
@ -594,8 +593,8 @@ std::vector<TopKParams> generateParamsV3() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{1, 2, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{1, 2, 0}),
0,
@ -605,8 +604,8 @@ std::vector<TopKParams> generateParamsV3() {
Tensor(ET, {5}, std::vector<T>{3, 1, 2, 5, 4}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_INDICES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_INDICES,
Tensor(ET, {3}, std::vector<T>{3, 1, 2}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{0, 1, 2}),
0,
@ -615,7 +614,7 @@ std::vector<TopKParams> generateParamsV3() {
return params;
}
std::vector<TopKParams> generateCombinedParamsV3() {
std::vector<TopKParams> generateCombinedParamsBackend() {
const std::vector<std::vector<TopKParams>> generatedParams {
generateParamsMaxMinSort<element::Type_t::i8, element::Type_t::i64, element::Type_t::i32>(),
generateParamsMaxMinSort<element::Type_t::i16, element::Type_t::i64, element::Type_t::i32>(),
@ -638,8 +637,8 @@ std::vector<TopKParams> generateCombinedParamsV3() {
return combinedParams;
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestV3,
testing::ValuesIn(generateCombinedParamsV3()), ReferenceTopKTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestBackend,
testing::ValuesIn(generateCombinedParamsBackend()), ReferenceTopKTest::getTestCaseName);
class ReferenceTopKTest1dMaxMin : public ReferenceTopKTest {
public:
@ -673,12 +672,12 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params, size_t out_idx) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = op::v0::Constant::create(params.k.type,
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<op::v1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto B = std::make_shared<opset1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(OutputVector{B->output(out_idx)}, ParameterVector{A});
return f;
}
@ -698,8 +697,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{6}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET_OUT, {6}, std::vector<T_OUT>{5, 4, 3, 2, 1, 0}),
0,
@ -709,8 +708,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{6}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET_OUT, {6}, std::vector<T_OUT>{5, 4, 3, 2, 1, 0}),
1,
@ -720,8 +719,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{6, 5, 4}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{5, 4, 3}),
0,
@ -731,8 +730,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{6, 5, 4}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{5, 4, 3}),
1,
@ -742,8 +741,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1}, std::vector<T>{6}),
Tensor(ET_OUT, {1}, std::vector<T_OUT>{5}),
0,
@ -753,8 +752,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1}, std::vector<T>{6}),
Tensor(ET_OUT, {1}, std::vector<T_OUT>{5}),
1,
@ -764,8 +763,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{6}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET_OUT, {6}, std::vector<T_OUT>{5, 4, 3, 2, 1, 0}),
0,
@ -775,8 +774,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{6}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {6}, std::vector<T>{1, 2, 3, 4, 5, 6}),
Tensor(ET_OUT, {6}, std::vector<T_OUT>{5, 4, 3, 2, 1, 0}),
1,
@ -786,8 +785,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{1, 2, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{5, 4, 3}),
0,
@ -797,8 +796,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{3}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {3}, std::vector<T>{1, 2, 3}),
Tensor(ET_OUT, {3}, std::vector<T_OUT>{5, 4, 3}),
1,
@ -808,8 +807,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1}, std::vector<T>{1}),
Tensor(ET_OUT, {1}, std::vector<T_OUT>{5}),
0,
@ -819,8 +818,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
Tensor(ET, {6}, std::vector<T>{6, 5, 4, 3, 2, 1}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1}, std::vector<T>{1}),
Tensor(ET_OUT, {1}, std::vector<T_OUT>{5}),
1,
@ -832,8 +831,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
10, 12, 9, 4, 8, 2, 11, 7, 6, 3, 5, 1
}),
@ -849,8 +848,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
10, 12, 9, 4, 8, 2, 11, 7, 6, 3, 5, 1
}),
@ -882,8 +881,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 3, 2, 4}, std::vector<T>{
169, 241, 177, 249, 185, 233, 170, 242, 178, 250, 186, 258, 171, 243, 179, 251,
187, 259, 172, 224, 180, 252, 188, 260, 149, 221, 157, 229, 165, 113, 150, 222,
@ -923,8 +922,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 3, 2, 4}, std::vector<T>{
169, 241, 177, 249, 185, 233, 170, 242, 178, 250, 186, 258, 171, 243, 179, 251,
187, 259, 172, 224, 180, 252, 188, 260, 149, 221, 157, 229, 165, 113, 150, 222,
@ -948,8 +947,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 2}, std::vector<T>{
10, 12, 9, 4, 11, 7, 6, 3
}),
@ -965,8 +964,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 2}, std::vector<T>{
10, 12, 9, 4, 11, 7, 6, 3
}),
@ -982,8 +981,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1, 2}, std::vector<T>{
10, 12, 11, 7
}),
@ -999,8 +998,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1, 2}, std::vector<T>{
10, 12, 11, 7
}),
@ -1016,8 +1015,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
8, 2, 10, 4, 12, 9, 5, 1, 6, 3, 11, 7
}),
@ -1033,8 +1032,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
8, 2, 10, 4, 12, 9, 5, 1, 6, 3, 11, 7
}),
@ -1050,8 +1049,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 2}, std::vector<T>{
8, 2, 10, 4, 5, 1, 6, 3
}),
@ -1067,8 +1066,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 2}, std::vector<T>{
8, 2, 10, 4, 5, 1, 6, 3
}),
@ -1084,8 +1083,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1, 2}, std::vector<T>{
8, 2, 5, 1
}),
@ -1101,8 +1100,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1, 2}, std::vector<T>{
8, 2, 5, 1
}),
@ -1118,8 +1117,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{4}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {4, 3}, std::vector<T>{
12, 11, 10, 9, 8, 7, 6, 2, 5, 3, 1, 4
}),
@ -1135,8 +1134,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{4}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {4, 3}, std::vector<T>{
12, 11, 10, 9, 8, 7, 6, 2, 5, 3, 1, 4
}),
@ -1152,8 +1151,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3}, std::vector<T>{
12, 11, 10, 9, 8, 7
}),
@ -1169,8 +1168,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3}, std::vector<T>{
12, 11, 10, 9, 8, 7
}),
@ -1186,8 +1185,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1, 3}, std::vector<T>{
12, 11, 10
}),
@ -1203,8 +1202,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {1, 3}, std::vector<T>{
12, 11, 10
}),
@ -1220,8 +1219,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1}, std::vector<T>{
4, 3
}),
@ -1237,8 +1236,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 1}, std::vector<T>{
4, 3
}),
@ -1254,8 +1253,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{4}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {4, 3}, std::vector<T>{
3, 1, 4, 6, 2, 5, 9, 8, 7, 12, 11, 10
}),
@ -1271,8 +1270,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{4}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {4, 3}, std::vector<T>{
3, 1, 4, 6, 2, 5, 9, 8, 7, 12, 11, 10
}),
@ -1288,8 +1287,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3}, std::vector<T>{
3, 1, 4, 6, 2, 5
}),
@ -1305,8 +1304,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{2}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3}, std::vector<T>{
3, 1, 4, 6, 2, 5
}),
@ -1322,8 +1321,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::NONE,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::NONE,
Tensor(ET, {1, 3}, std::vector<T>{
3, 1, 4
}),
@ -1339,8 +1338,8 @@ std::vector<TopKParams> generateParams1dMaxMin() {
}),
Tensor(ET2, {}, std::vector<T2>{1}),
0,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::NONE,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::NONE,
Tensor(ET, {1, 3}, std::vector<T>{
3, 1, 4
}),
@ -1380,12 +1379,12 @@ INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTest1dMaxM
class ReferenceTopKTestInt64 : public ReferenceTopKTest1dMaxMin {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params, size_t out_idx) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = op::v0::Constant::create(params.k.type,
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<op::v1::TopK>(A,
const auto B = std::make_shared<opset1::TopK>(A,
k,
params.axis,
params.mode,
@ -1412,8 +1411,8 @@ std::vector<TopKParams> generateParamsInt64() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
10, 12, 9, 4, 8, 2, 11, 7, 6, 3, 5, 1
}),
@ -1428,8 +1427,8 @@ std::vector<TopKParams> generateParamsInt64() {
}),
Tensor(ET2, {}, std::vector<T2>{3}),
1,
op::v1::TopK::Mode::MAX,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 3, 2}, std::vector<T>{
10, 12, 9, 4, 8, 2, 11, 7, 6, 3, 5, 1
}),
@ -1468,12 +1467,12 @@ public:
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<op::v0::Parameter>(params.A.type,
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = op::v0::Constant::create(params.k.type,
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<op::v1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto B = std::make_shared<opset1::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(OutputVector{B->output(1)}, ParameterVector{A});
return f;
}
@ -1493,8 +1492,8 @@ std::vector<TopKParams> generateParamsSingleOutput() {
Tensor(ET, {2, 3, 2}, std::vector<T>{12, 2, 10, 9, 8, 4, 6, 1, 5, 3, 11, 7}),
Tensor(ET2, {}, std::vector<T2>{2}),
1,
op::v1::TopK::Mode::MIN,
op::v1::TopK::SortType::SORT_VALUES,
opset1::TopK::Mode::MIN,
opset1::TopK::SortType::SORT_VALUES,
Tensor(ET, {2, 2, 2}, std::vector<T>{}),
Tensor(ET_OUT, {2, 2, 2}, std::vector<T_OUT>{2, 0, 1, 2, 1, 0, 0, 1}),
0,
@ -1530,19 +1529,181 @@ INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestSingle
testing::ValuesIn(generateCombinedParamsSingleOutput()), ReferenceTopKTest::getTestCaseName);
TEST(ReferenceTopKTestInvalid, topk_v1_invalid_strings) {
const auto data = std::make_shared<op::v0::Parameter>(element::f32, Shape{1, 2, 3});
const auto k = op::v0::Constant::create(element::i64, Shape{}, {1});
EXPECT_THROW(op::v1::TopK(data, k, 0, "max", "invalid_mode"), ngraph::CheckFailure);
EXPECT_THROW(op::v1::TopK(data, k, 0, "invalid_sort", "index"), ngraph::CheckFailure);
const auto data = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 2, 3});
const auto k = opset1::Constant::create(element::i64, Shape{}, {1});
EXPECT_THROW(opset1::TopK(data, k, 0, "max", "invalid_mode"), ngraph::CheckFailure);
EXPECT_THROW(opset1::TopK(data, k, 0, "invalid_sort", "index"), ngraph::CheckFailure);
}
TEST(ReferenceTopKTestInvalid, topk_v1_invalid_k) {
const auto data = std::make_shared<op::v0::Parameter>(element::f32, Shape{1, 2, 3});
const auto k_non_scalar = op::v0::Constant::create(element::i64, Shape{2}, {1, 2});
EXPECT_THROW(op::v1::TopK(data, k_non_scalar, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_float = op::v0::Constant::create(element::f32, Shape{}, {1.0f});
EXPECT_THROW(op::v1::TopK(data, k_float, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_negative = op::v0::Constant::create(element::i8, Shape{}, {-1});
EXPECT_THROW(op::v1::TopK(data, k_negative, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto data = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 2, 3});
const auto k_non_scalar = opset1::Constant::create(element::i64, Shape{2}, {1, 2});
EXPECT_THROW(opset1::TopK(data, k_non_scalar, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_float = opset1::Constant::create(element::f32, Shape{}, {1.0f});
EXPECT_THROW(opset1::TopK(data, k_float, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_negative = opset1::Constant::create(element::i8, Shape{}, {-1});
EXPECT_THROW(opset1::TopK(data, k_negative, 0, "max", "index"), ngraph::NodeValidationFailure);
}
class ReferenceTopKTestResnet50V3 : public ReferenceTopKTestResnet50 {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParamsResnet50& params) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto B = std::make_shared<opset3::TopK>(A,
opset1::Constant::create(element::i64, {}, {5}),
1,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES);
const auto C = std::make_shared<opset3::TopK>(A,
opset1::Constant::create(element::i64, {}, {1}),
1,
opset1::TopK::Mode::MAX,
opset1::TopK::SortType::SORT_VALUES);
const auto out5_value = B->output(0);
const auto out5_index = B->output(1);
const auto out1_value = C->output(0);
const auto out1_index = C->output(1);
const auto f = std::make_shared<Model>(OutputVector{out5_value, out5_index, out1_value, out1_index}, ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestResnet50V3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestResnet50V3,
testing::ValuesIn(generateCombinedParamsResnet50()), ReferenceTopKTestResnet50V3::getTestCaseName);
class ReferenceTopKTestMaxMinSortV3 : public ReferenceTopKTestMaxMinSort {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<opset3::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(B->outputs(), ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestMaxMinSortV3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestMaxMinSortV3,
testing::ValuesIn(generateCombinedParamsMaxMinSort()), ReferenceTopKTestMaxMinSortV3::getTestCaseName);
class ReferenceTopKTestBackendV3 : public ReferenceTopKTestBackend {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<opset3::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(B->outputs(), ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestBackendV3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestBackendV3,
testing::ValuesIn(generateCombinedParamsBackend()), ReferenceTopKTestBackendV3::getTestCaseName);
class ReferenceTopKTest1dMaxMinV3 : public ReferenceTopKTest1dMaxMin {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params, size_t out_idx) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<opset3::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(OutputVector{B->output(out_idx)}, ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTest1dMaxMinV3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTest1dMaxMinV3,
testing::ValuesIn(generateCombinedParams1dMaxMin()), ReferenceTopKTest1dMaxMinV3::getTestCaseName);
class ReferenceTopKTestInt64V3 : public ReferenceTopKTestInt64 {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params, size_t out_idx) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<opset3::TopK>(A,
k,
params.axis,
params.mode,
params.sort,
element::i64);
const auto f = std::make_shared<Model>(OutputVector{B->output(out_idx)}, ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestInt64V3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestInt64V3,
testing::ValuesIn(generateCombinedParamsInt64()), ReferenceTopKTestInt64V3::getTestCaseName);
class ReferenceTopKTestSingleOutputV3 : public ReferenceTopKTestSingleOutput {
private:
static std::shared_ptr<Model> CreateFunction(const TopKParams& params) {
const auto A = std::make_shared<opset1::Parameter>(params.A.type,
params.A.shape);
const auto k = opset1::Constant::create(params.k.type,
params.k.shape,
params.k.data.data());
const auto B = std::make_shared<opset3::TopK>(A, k, params.axis, params.mode, params.sort);
const auto f = std::make_shared<Model>(OutputVector{B->output(1)}, ParameterVector{A});
return f;
}
};
TEST_P(ReferenceTopKTestSingleOutputV3, CompareWithRefs) {
Exec();
}
INSTANTIATE_TEST_SUITE_P(smoke_TopK_With_Hardcoded_Refs, ReferenceTopKTestSingleOutputV3,
testing::ValuesIn(generateCombinedParamsSingleOutput()), ReferenceTopKTestSingleOutputV3::getTestCaseName);
TEST(ReferenceTopKTestInvalidV3, topk_v3_invalid_strings) {
const auto data = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 2, 3});
const auto k = opset1::Constant::create(element::i64, Shape{}, {1});
EXPECT_THROW(opset3::TopK(data, k, 0, "max", "invalid_mode"), ngraph::CheckFailure);
EXPECT_THROW(opset3::TopK(data, k, 0, "invalid_sort", "index"), ngraph::CheckFailure);
}
TEST(ReferenceTopKTestInvalidV3, topk_v3_invalid_k) {
const auto data = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 2, 3});
const auto k_non_scalar = opset1::Constant::create(element::i64, Shape{2}, {1, 2});
EXPECT_THROW(opset3::TopK(data, k_non_scalar, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_float = opset1::Constant::create(element::f32, Shape{}, {1.0f});
EXPECT_THROW(opset3::TopK(data, k_float, 0, "max", "index"), ngraph::NodeValidationFailure);
const auto k_negative = opset1::Constant::create(element::i8, Shape{}, {-1});
EXPECT_THROW(opset3::TopK(data, k_negative, 0, "max", "index"), ngraph::NodeValidationFailure);
}
} // namespace

1
inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp
View File

@ -319,7 +319,6 @@ void MKLDNNGraph::InitGraph() {
SortTopologically();
InitDescriptors();
RemoveDroppedEdges();
InitOptimalPrimitiveDescriptors();

6
inference-engine/src/mkldnn_plugin/mkldnn_node.cpp
View File

@ -717,7 +717,7 @@ void MKLDNNNode::initDescriptor(const NodeConfig& config) {
selectedPD->setConfig(rightConfig);
}
void MKLDNNNode::prepareMemory(const NodeDesc *selected_pd, mkldnn::primitive_desc_iterator& itpd) {
void MKLDNNNode::prepareMemory(mkldnn::primitive_desc_iterator& itpd) {
for (size_t i = 0; i < getChildEdges().size(); i++) {
auto &dstMemPtr = getChildEdgeAt(i)->getMemoryPtr();
if (!dstMemPtr || !dstMemPtr->GetPrimitivePtr())
@ -1049,7 +1049,9 @@ void MKLDNNNode::setDynamicBatchLim(int lim) {
}
}
void MKLDNNNode::appendPostOpArgs(const mkldnn::primitive_attr& attr) {
void MKLDNNNode::appendPostOpArgs(const mkldnn::primitive_attr& attr,
std::unordered_map<int, mkldnn::memory>& primArgs,
const std::vector<MKLDNNMemoryPtr>& binaryPostOpsArgs) {
auto post_ops = attr.get_post_ops();
int idx = 0;
for (int i = 0; i < post_ops.len(); i++) {

10
inference-engine/src/mkldnn_plugin/mkldnn_node.h
View File

@ -208,7 +208,9 @@ public:
return 1;
}
void appendPostOpArgs(const mkldnn::primitive_attr& attr);
static void appendPostOpArgs(const mkldnn::primitive_attr& attr,
std::unordered_map<int, mkldnn::memory>& primArgs,
const std::vector<MKLDNNMemoryPtr>& binaryPostOpsArgs);
bool isFusedWith(Type type) const;
@ -425,7 +427,7 @@ public:
if (impl_type == selected_pd->getImplementationType() &&
descsCompatible(srcDescs, selected_pd->getConfig().inConfs) &&
descsCompatible(dstDescs, selected_pd->getConfig().outConfs)) {
prepareMemory(selected_pd, itpd);
prepareMemory(itpd);
PD prim_desc = createPd<PD, D, FPD>(desc);
return {itpd.get()};
}
@ -722,6 +724,8 @@ protected:
supportedPrimitiveDescriptors.push_back({config, implType});
}
void prepareMemory(mkldnn::primitive_desc_iterator& itpd);
bool isDynamic = false;
bool inputShapesDefined() const;
@ -746,6 +750,7 @@ protected:
}
std::vector<VectorDims> lastInputDims = {};
std::shared_ptr<ngraph::Node> opToShapeInfer;
private:
@ -788,7 +793,6 @@ private:
return PD(*selected_desc_ptr, engine);
}
void prepareMemory(const NodeDesc *selected_pd, mkldnn::primitive_desc_iterator& itpd);
enum LOOK { LOOK_UP = 1, LOOK_DOWN = 2 };
ConstantType checkConstant(LOOK look, std::vector<MKLDNNNodePtr>& checkNodes);

31
inference-engine/src/mkldnn_plugin/mkldnn_plugin.cpp
View File

@ -504,23 +504,24 @@ Engine::LoadExeNetworkImpl(const InferenceEngine::CNNNetwork &network, const std
OV_ITT_SCOPED_TASK(itt::domains::MKLDNNPlugin, "Engine::LoadExeNetworkImpl");
// verification of supported input
InferenceEngine::InputsDataMap _networkInputs = network.getInputsInfo();
for (const auto &ii : _networkInputs) {
for (const auto &ii : network.getInputsInfo()) {
auto input_precision = ii.second->getPrecision();
if (input_precision != InferenceEngine::Precision::FP64 &&
input_precision != InferenceEngine::Precision::FP32 &&
input_precision != InferenceEngine::Precision::I32 &&
input_precision != InferenceEngine::Precision::U32 &&
input_precision != InferenceEngine::Precision::U16 &&
input_precision != InferenceEngine::Precision::I16 &&
input_precision != InferenceEngine::Precision::I8 &&
input_precision != InferenceEngine::Precision::U8 &&
input_precision != InferenceEngine::Precision::BF16 &&
input_precision != InferenceEngine::Precision::BOOL &&
input_precision != InferenceEngine::Precision::I64 &&
input_precision != InferenceEngine::Precision::U64) {
using hash_t = std::hash<typename std::underlying_type<Precision::ePrecision>::type>;
static const std::unordered_set<Precision::ePrecision, hash_t> supported_precisions = {
Precision::U8, Precision::I8,
Precision::U16, Precision::I16,
Precision::U32, Precision::I32,
Precision::U64, Precision::I64,
Precision::BF16, Precision::FP16,
Precision::FP32, Precision::FP64,
Precision::BOOL
};
if (!supported_precisions.count(input_precision)) {
IE_THROW(NotImplemented)
<< "Input image format " << input_precision << " is not supported yet...";
<< "Input image format " << input_precision << " is not supported yet...";
}
}

1
inference-engine/src/mkldnn_plugin/mkldnn_primitive.h
View File

@ -18,7 +18,6 @@ public:
operator bool() const;
MKLDNNPrimitive& operator=(const std::shared_ptr<mkldnn::primitive>& primitive);
mkldnn::primitive operator*();
void reset(mkldnn::primitive* primitive);
private:

616
inference-engine/src/mkldnn_plugin/nodes/common/cpu_convert.cpp
View File

@ -4,27 +4,208 @@
#include "cpu_convert.h"
#include "cpu_memcpy.h"
#include "utils/bfloat16.hpp"
#include <utils/bfloat16.hpp>
#include <utils/general_utils.h>
#include <mkldnn_selective_build.h>
#include <ie_parallel.hpp>
#include <openvino/core/type/float16.hpp>
#include <cpu/x64/jit_generator.hpp>
#include <algorithm>
#include <type_traits>
#include <tuple>
#include <ie_parallel.hpp>
#include <cmath>
using namespace MKLDNNPlugin;
using namespace InferenceEngine;
using namespace dnnl::impl::cpu::x64;
using namespace dnnl::impl::utils;
using namespace Xbyak;
namespace {
template<typename srcType, typename dstType>
void convert(const void *srcPtr, void *dstPtr, const size_t size) {
if (std::is_same<srcType, dstType>::value) {
cpu_memcpy(dstPtr, srcPtr, size*sizeof(dstType));
} else {
const srcType *srcData = reinterpret_cast<const srcType *>(srcPtr);
dstType *dstData = reinterpret_cast<dstType *>(dstPtr);
template <typename src_t, typename dst_t>
void convert_vec(jit_generator & gen,
const RegExp & src,
const RegExp & dst);
parallel_for(size, [&](size_t i) {
dstData[i] = static_cast<dstType>(srcData[i]);
template <>
void convert_vec<ov::float16, float>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.xmm3;
auto const & f32vec = gen.ymm4;
gen.movdqu(f16vec, gen.xword[src]);
gen.vcvtph2ps(f32vec, f16vec);
gen.vmovups(gen.yword[dst], f32vec);
}
template <>
void convert_vec<float, ov::float16>(jit_generator & gen,
const RegExp & src,
const RegExp & dst) {
auto const & f16vec = gen.xmm3;
auto const & f32vec = gen.ymm4;
gen.vmovups(f32vec, gen.yword[src]);
gen.vcvtps2ph(f16vec, f32vec, 0);
gen.movdqu(gen.xword[dst], f16vec);
}
class jit_convert_array : public jit_generator {
DECLARE_CPU_JIT_AUX_FUNCTIONS(jit_convert_array)
void generate() override {
const size_t vlen = 8u;
const size_t vlen_log2 = 3;
auto reg_src = rax;
auto reg_dst = rbx;
auto reg_sz = rdx;
Label tail, exit;
preamble();
mov(reg_src, ptr[param1 + offsetof(args_t, src)]);
mov(reg_dst, ptr[param1 + offsetof(args_t, out)]);
mov(reg_sz, ptr[param1 + offsetof(args_t, count)]);
xor_(rsi, rsi);
mov(r8, reg_sz);
shr(r8, vlen_log2);
foreach(rsi, 1, r8, [&, this](const Xbyak::Reg64& idx) {
_convert_vec(*this, reg_src, reg_dst);
add(reg_src, _src_size * vlen);
add(reg_dst, _dst_size * vlen);
});
L(tail);
shl(rsi, vlen_log2);
sub(reg_sz, rsi);
test(reg_sz, reg_sz);
jz(exit);
// allocate array for 8 floats on stack
sub(rsp, vlen * sizeof(float));
mov(r8, rsp);
vpxor(ymm4, ymm4, ymm4);
vmovups(yword[r8], ymm4);
// Tail conversion
copy(r8, reg_src, reg_sz, _src_size);
_convert_vec(*this, r8, r8);
copy(reg_dst, r8, reg_sz, _dst_size);
// Free the array on stack
add(rsp, vlen * sizeof(float));
L(exit);
postamble();
}
void foreach(const Xbyak::Reg64& idx,
size_t step,
const Xbyak::Reg64& end,
std::function<void(const Xbyak::Reg64&)> && fn) {
Label loop, exit;
L(loop);
cmp(idx, end);
jge(exit);
fn(idx);
add(idx, step);
jmp(loop);
L(exit);
}
void copy(const Xbyak::Reg64& dst,
const Xbyak::Reg64& src,
const Xbyak::Reg64& size,
size_t item_size) {
push(rsi);
push(r15);
xor_(rsi, rsi);
auto address_frame = [this](size_t size) -> const AddressFrame& {
switch (size) {
case 1: return byte;
case 2: return word;
case 4: return dword;
case 8: return qword;
default:
break;
}
return ptr;
};
const auto & addr_frame = address_frame(item_size);
foreach(rsi, 1, size, [&, this](const Xbyak::Reg64& idx) {
mov(r15, addr_frame[src + idx * item_size]);
mov(addr_frame[dst + idx * item_size], r15);
});
pop(r15);
pop(rsi);
}
public:
typedef struct {
const void* src;
void* out;
const size_t count;
} args_t;
typedef void (*fn_t)(const args_t*);
typedef void (*convert_vec_t)(jit_generator &,
const RegExp &,
const RegExp &);
jit_convert_array(convert_vec_t convert_vec,
size_t src_size,
size_t dst_size)
: _convert_vec(convert_vec)
, _src_size(src_size)
, _dst_size(dst_size) {}
template<typename src_t, typename dst_t>
static fn_t get() {
if (mayiuse(avx2) && cpu().has(util::Cpu::tF16C)) {
static jit_convert_array converter(convert_vec<src_t, dst_t>, sizeof(src_t), sizeof(dst_t));
auto & generator = static_cast<jit_generator&>(converter);
generator.create_kernel();
return (fn_t)generator.jit_ker();
}
return nullptr;
}
private:
convert_vec_t _convert_vec;
size_t _src_size;
size_t _dst_size;
};
template <typename TI, typename TO>
void jit_convert(const TI* arg, TO* out, size_t count) {
using jit_impl = jit_convert_array;
static auto converter = jit_impl::get<TI, TO>();
if (converter) {
typename jit_impl::args_t args = { arg, out, count };
converter(&args);
} else {
for (size_t i = 0; i < count; ++i) {
out[i] = static_cast<TO>(arg[i]);
}
}
}
@ -35,84 +216,391 @@ struct PrecisionInfo {
template <>
struct PrecisionInfo<Precision::BF16> {
using value_type = MKLDNNPlugin::bfloat16_t;
using value_type = bfloat16_t;
};
template <>
struct PrecisionInfo<Precision::FP16> {
using value_type = ov::float16;
};
template <>
struct PrecisionInfo<Precision::BOOL> {
using value_type = uint8_t;
};
template<typename T,
typename U = typename std::conditional<
std::is_same<ov::float16, T>::value
|| std::is_same<bfloat16_t, T>::value,
float, T>::type>
struct Range {
const std::tuple<U, U> & fit(const Precision & prec);
private:
std::tuple<U, U> _range {
std::numeric_limits<T>::lowest(),
std::numeric_limits<T>::max()
};
};
template<typename T, typename U>
const std::tuple<U, U> & Range<T, U>::fit(const Precision & prec) {
if (prec.is_float()) {
double lbound, ubound;
switch (prec) {
case Precision::BF16:
lbound = static_cast<double>(std::numeric_limits<bfloat16_t>::lowest());
ubound = static_cast<double>(std::numeric_limits<bfloat16_t>::max());
break;
case Precision::FP16:
lbound = static_cast<double>(std::numeric_limits<ov::float16>::lowest());
ubound = static_cast<double>(std::numeric_limits<ov::float16>::max());
break;
case Precision::FP32:
lbound = static_cast<double>(std::numeric_limits<float>::lowest());
ubound = static_cast<double>(std::numeric_limits<float>::max());
break;
case Precision::FP64:
lbound = std::numeric_limits<double>::lowest();
ubound = std::numeric_limits<double>::max();
break;
default:
IE_THROW() << "Unsupported precision";
}
std::get<0>(_range) = static_cast<U>(std::max(static_cast<double>(std::get<0>(_range)), lbound));
std::get<1>(_range) = static_cast<U>(std::min(static_cast<double>(std::get<1>(_range)), ubound));
} else {
int64_t lbound;
uint64_t ubound;
switch (prec) {
case Precision::BOOL:
case Precision::U8:
lbound = static_cast<int64_t>(std::numeric_limits<uint8_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<uint8_t>::max());
break;
case Precision::I8:
lbound = static_cast<int64_t>(std::numeric_limits<int8_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<int8_t>::max());
break;
case Precision::U16:
lbound = static_cast<int64_t>(std::numeric_limits<uint16_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<uint16_t>::max());
break;
case Precision::I16:
lbound = static_cast<int64_t>(std::numeric_limits<int16_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<int16_t>::max());
break;
case Precision::U32:
lbound = static_cast<int64_t>(std::numeric_limits<uint32_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<uint32_t>::max());
break;
case Precision::I32:
lbound = static_cast<int64_t>(std::numeric_limits<int32_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<int32_t>::max());
break;
case Precision::U64:
lbound = static_cast<int64_t>(std::numeric_limits<uint64_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<uint64_t>::max());
break;
case Precision::I64:
lbound = static_cast<int64_t>(std::numeric_limits<int64_t>::lowest());
ubound = static_cast<uint64_t>(std::numeric_limits<int64_t>::max());
break;
default:
IE_THROW() << "Unsupported precision";
}
using ltype = typename std::conditional<
std::is_floating_point<U>::value,
double, int64_t>::type;
using utype = typename std::conditional<
std::is_floating_point<U>::value,
double, uint64_t>::type;
std::get<0>(_range) = static_cast<U>(std::max(static_cast<ltype>(std::get<0>(_range)), static_cast<ltype>(lbound)));
std::get<1>(_range) = static_cast<U>(std::min(static_cast<utype>(std::get<1>(_range)), static_cast<utype>(ubound)));
}
return _range;
}
struct ConvertContext {
const void *srcPtr;
void *dstPtr;
size_t size;
Precision interimPrc;
Precision dstPrc;
bool converted;
template<typename T>
std::tuple<T, T> range() const {
Range<T> r;
r.fit(interimPrc);
return r.fit(dstPrc);
}
};
template<typename T>
struct ConvertPrecision {
using src_t = typename std::tuple_element<0, T>::type;
using dst_t = typename std::tuple_element<1, T>::type;
struct ConvertPrecision;
template<typename src_t, typename dst_t>
struct ConvertPrecision<std::tuple<src_t, dst_t>> {
void operator()(ConvertContext & ctx) {
convert<src_t, dst_t>(ctx.srcPtr, ctx.dstPtr, ctx.size);
auto src = static_cast<const src_t *>(ctx.srcPtr);
auto dst = static_cast<dst_t *>(ctx.dstPtr);
src_t lbound, ubound;
std::tie(lbound, ubound) = ctx.range<src_t>();
if (std::is_integral<src_t>::value
|| ctx.interimPrc.is_float()
|| std::is_integral<dst_t>::value) {
parallel_for(ctx.size, [&](size_t i) {
dst[i] = static_cast<dst_t>(std::max(std::min(src[i], ubound), lbound));
});
} else {
parallel_for(ctx.size, [&](size_t i) {
dst[i] = static_cast<dst_t>(std::trunc(std::max(std::min(src[i], ubound), lbound)));
});
}
ctx.converted = true;
}
};
template<>
struct ConvertPrecision<std::tuple<float, bfloat16_t>> {
void operator()(ConvertContext & ctx) {
auto src = static_cast<const float *>(ctx.srcPtr);
auto dst = static_cast<bfloat16_t *>(ctx.dstPtr);
if (ctx.interimPrc.is_float()) {
parallel_for(ctx.size, [&](size_t i) {
dst[i] = static_cast<bfloat16_t>(src[i]);
});
} else {
float lbound, ubound;
std::tie(lbound, ubound) = ctx.range<float>();
parallel_for(ctx.size, [&](size_t i) {
dst[i] = static_cast<bfloat16_t>(std::trunc(std::max(std::min(src[i], ubound), lbound)));
});
}
ctx.converted = true;
}
};
template<>
struct ConvertPrecision<std::tuple<bfloat16_t, float>> {
void operator()(ConvertContext & ctx) {
auto src = static_cast<const bfloat16_t *>(ctx.srcPtr);
auto dst = static_cast<float *>(ctx.dstPtr);
if (ctx.interimPrc.is_float()) {
parallel_for(ctx.size, [&](size_t i) {
dst[i] = static_cast<float>(src[i]);
});
} else {
float lbound, ubound;
std::tie(lbound, ubound) = ctx.range<bfloat16_t>();
parallel_for(ctx.size, [&](size_t i) {
dst[i] = std::trunc(std::max(std::min(static_cast<float>(src[i]), ubound), lbound));
});
}
ctx.converted = true;
}
};
template<typename src_t>
struct ConvertPrecision<std::tuple<src_t, ov::float16>> {
void operator()(ConvertContext & ctx) {
auto src = static_cast<const src_t *>(ctx.srcPtr);
auto dst = static_cast<ov::float16 *>(ctx.dstPtr);
constexpr size_t batch = 64;
const size_t iterations = MKLDNNPlugin::div_up(ctx.size, batch);
typedef float batch_type[batch];
src_t lbound, ubound;
std::tie(lbound, ubound) = ctx.range<src_t>();
if (std::is_integral<src_t>::value
|| ctx.interimPrc.is_float()) {
parallel_for(iterations, [&](size_t i) {
batch_type tmp;
const size_t offset = i * batch;
const size_t current_batch_size = std::min(ctx.size - offset, batch);
for (size_t j = 0; j < current_batch_size; ++j) // src_t -> fp32
tmp[j] = static_cast<float>(std::max(std::min(src[offset + j], ubound), lbound));
jit_convert(tmp, dst + offset, current_batch_size); // fp32 -> fp16
});
} else {
parallel_for(iterations, [&](size_t i) {
batch_type tmp;
const size_t offset = i * batch;
const size_t current_batch_size = std::min(ctx.size - offset, batch);
for (size_t j = 0; j < current_batch_size; ++j) // src_t -> fp32
tmp[j] = static_cast<float>(std::trunc(std::max(std::min(src[offset + j], ubound), lbound)));
jit_convert(tmp, dst + offset, current_batch_size); // fp32 -> fp16
});
}
ctx.converted = true;
}
};
template<typename dst_t>
struct ConvertPrecision<std::tuple<ov::float16, dst_t>> {
void operator()(ConvertContext & ctx) {
auto src = static_cast<const ov::float16 *>(ctx.srcPtr);
auto dst = static_cast<dst_t *>(ctx.dstPtr);
constexpr size_t batch = 64;
const size_t iterations = MKLDNNPlugin::div_up(ctx.size, batch);
typedef float batch_type[batch];
float lbound, ubound;
std::tie(lbound, ubound) = ctx.range<ov::float16>();
if (ctx.interimPrc.is_float()
|| std::is_integral<dst_t>::value) {
parallel_for(iterations, [&](size_t i) {
batch_type tmp;
const size_t offset = i * batch;
const size_t current_batch_size = std::min(ctx.size - offset, batch);
jit_convert(src + offset, tmp, current_batch_size); // fp16 -> fp32
for (size_t j = 0; j < current_batch_size; ++j) // fp32 -> dst_t
dst[offset + j] = static_cast<dst_t>(std::max(std::min(tmp[j], ubound), lbound));
});
} else {
parallel_for(iterations, [&](size_t i) {
batch_type tmp;
const size_t offset = i * batch;
const size_t current_batch_size = std::min(ctx.size - offset, batch);
jit_convert(src + offset, tmp, current_batch_size); // fp16 -> fp32
for (size_t j = 0; j < current_batch_size; ++j) // fp32 -> dst_t
dst[offset + j] = static_cast<dst_t>(std::trunc(std::max(std::min(tmp[j], ubound), lbound)));
});
}
ctx.converted = true;
}
};
template<>
struct ConvertPrecision<std::tuple<ov::float16, ov::float16>> {
void operator()(ConvertContext & ctx) {
auto src = static_cast<const ov::float16 *>(ctx.srcPtr);
auto dst = static_cast<ov::float16 *>(ctx.dstPtr);
constexpr size_t batch = 64;
const size_t iterations = MKLDNNPlugin::div_up(ctx.size, batch);
typedef float batch_type[batch];
float lbound, ubound;
std::tie(lbound, ubound) = ctx.range<ov::float16>();
if (ctx.interimPrc.is_float()) {
cpu_memcpy(dst, src, ctx.size * sizeof(ov::float16));
} else {
parallel_for(iterations, [&](size_t i) {
batch_type tmp;
const size_t offset = i * batch;
const size_t current_batch_size = std::min(ctx.size - offset, batch);
jit_convert(src + offset, tmp, current_batch_size); // fp16 -> fp32
for (size_t j = 0; j < current_batch_size; ++j) // truncate fp32
tmp[j] = std::trunc(std::max(std::min(tmp[j], ubound), lbound));
jit_convert(tmp, dst + offset, current_batch_size); // fp32 -> fp16
});
}
ctx.converted = true;
}
};
bool isConversionTruncatesRange(const Precision & from, const Precision & to) {
return to.bitsSize() < from.bitsSize()
|| (from.is_float() && !to.is_float()) // float -> integral
|| (from.isSigned() != to.isSigned()) // signed <-> unsigned
|| (to == Precision::BOOL && from != to); // T -> bool
}
} // namespace
#define MKLDNN_CVT(ST, DT) OV_CASE2(Precision::ST, Precision::DT, PrecisionInfo<Precision::ST>::value_type, PrecisionInfo<Precision::DT>::value_type)
void cpu_convert(const void *srcPtr, void *dstPtr, Precision srcPrc, Precision dstPrc, const size_t size) {
using namespace MKLDNNPlugin;
#define MKLDNN_CVT_LIST \
MKLDNN_CVT(U8, I8), MKLDNN_CVT(U8, U16), MKLDNN_CVT(U8, I16), MKLDNN_CVT(U8, U32), \
MKLDNN_CVT(U8, I32), MKLDNN_CVT(U8, U64), MKLDNN_CVT(U8, I64), MKLDNN_CVT(U8, FP32), \
MKLDNN_CVT(U8, FP16), MKLDNN_CVT(U8, BF16), MKLDNN_CVT(U8, FP64), MKLDNN_CVT(U8, BOOL), \
MKLDNN_CVT(I8, U8), MKLDNN_CVT(I8, U16), MKLDNN_CVT(I8, I16), MKLDNN_CVT(I8, U32), \
MKLDNN_CVT(I8, I32), MKLDNN_CVT(I8, U64), MKLDNN_CVT(I8, I64), MKLDNN_CVT(I8, FP32), \
MKLDNN_CVT(I8, FP16), MKLDNN_CVT(I8, BF16), MKLDNN_CVT(I8, FP64), MKLDNN_CVT(I8, BOOL), \
MKLDNN_CVT(U16, U8), MKLDNN_CVT(U16, I8), MKLDNN_CVT(U16, I16), MKLDNN_CVT(U16, U32), \
MKLDNN_CVT(U16, I32), MKLDNN_CVT(U16, U64), MKLDNN_CVT(U16, I64), MKLDNN_CVT(U16, FP32), \
MKLDNN_CVT(U16, FP16), MKLDNN_CVT(U16, BF16), MKLDNN_CVT(U16, FP64), MKLDNN_CVT(U16, BOOL), \
MKLDNN_CVT(I16, U8), MKLDNN_CVT(I16, I8), MKLDNN_CVT(I16, U16), MKLDNN_CVT(I16, U32), \
MKLDNN_CVT(I16, I32), MKLDNN_CVT(I16, U64), MKLDNN_CVT(I16, I64), MKLDNN_CVT(I16, FP32), \
MKLDNN_CVT(I16, FP16), MKLDNN_CVT(I16, BF16), MKLDNN_CVT(I16, FP64), MKLDNN_CVT(I16, BOOL), \
MKLDNN_CVT(U32, U8), MKLDNN_CVT(U32, I8), MKLDNN_CVT(U32, U16), MKLDNN_CVT(U32, I16), \
MKLDNN_CVT(U32, I32), MKLDNN_CVT(U32, U64), MKLDNN_CVT(U32, I64), MKLDNN_CVT(U32, FP32), \
MKLDNN_CVT(U32, FP16), MKLDNN_CVT(U32, BF16), MKLDNN_CVT(U32, FP64), MKLDNN_CVT(U32, BOOL), \
MKLDNN_CVT(I32, U8), MKLDNN_CVT(I32, I8), MKLDNN_CVT(I32, U16), MKLDNN_CVT(I32, I16), \
MKLDNN_CVT(I32, U32), MKLDNN_CVT(I32, U64), MKLDNN_CVT(I32, I64), MKLDNN_CVT(I32, FP32), \
MKLDNN_CVT(I32, FP16), MKLDNN_CVT(I32, BF16), MKLDNN_CVT(I32, FP64), MKLDNN_CVT(I32, BOOL), \
MKLDNN_CVT(U64, U8), MKLDNN_CVT(U64, I8), MKLDNN_CVT(U64, U16), MKLDNN_CVT(U64, I16), \
MKLDNN_CVT(U64, U32), MKLDNN_CVT(U64, I32), MKLDNN_CVT(U64, I64), MKLDNN_CVT(U64, FP32), \
MKLDNN_CVT(U64, FP16), MKLDNN_CVT(U64, BF16), MKLDNN_CVT(U64, FP64), MKLDNN_CVT(U64, BOOL), \
MKLDNN_CVT(I64, U8), MKLDNN_CVT(I64, I8), MKLDNN_CVT(I64, U16), MKLDNN_CVT(I64, I16), \
MKLDNN_CVT(I64, U32), MKLDNN_CVT(I64, I32), MKLDNN_CVT(I64, U64), MKLDNN_CVT(I64, FP32), \
MKLDNN_CVT(I64, FP16), MKLDNN_CVT(I64, BF16), MKLDNN_CVT(I64, FP64), MKLDNN_CVT(I64, BOOL), \
MKLDNN_CVT(FP32, U8), MKLDNN_CVT(FP32, I8), MKLDNN_CVT(FP32, U16), MKLDNN_CVT(FP32, I16), \
MKLDNN_CVT(FP32, U32), MKLDNN_CVT(FP32, I32), MKLDNN_CVT(FP32, U64), MKLDNN_CVT(FP32, I64), \
MKLDNN_CVT(FP32, FP16), MKLDNN_CVT(FP32, BF16), MKLDNN_CVT(FP32, FP64), MKLDNN_CVT(FP32, BOOL), \
MKLDNN_CVT(FP16, U8), MKLDNN_CVT(FP16, I8), MKLDNN_CVT(FP16, U16), MKLDNN_CVT(FP16, I16), \
MKLDNN_CVT(FP16, U32), MKLDNN_CVT(FP16, I32), MKLDNN_CVT(FP16, U64), MKLDNN_CVT(FP16, I64), \
MKLDNN_CVT(FP16, FP32), MKLDNN_CVT(FP16, BF16), MKLDNN_CVT(FP16, FP64), MKLDNN_CVT(FP16, BOOL), \
MKLDNN_CVT(BF16, U8), MKLDNN_CVT(BF16, I8), MKLDNN_CVT(BF16, U16), MKLDNN_CVT(BF16, I16), \
MKLDNN_CVT(BF16, U32), MKLDNN_CVT(BF16, I32), MKLDNN_CVT(BF16, U64), MKLDNN_CVT(BF16, I64), \
MKLDNN_CVT(BF16, FP32), MKLDNN_CVT(BF16, FP16), MKLDNN_CVT(BF16, FP64), MKLDNN_CVT(BF16, BOOL), \
MKLDNN_CVT(FP64, U8), MKLDNN_CVT(FP64, I8), MKLDNN_CVT(FP64, U16), MKLDNN_CVT(FP64, I16), \
MKLDNN_CVT(FP64, U32), MKLDNN_CVT(FP64, I32), MKLDNN_CVT(FP64, U64), MKLDNN_CVT(FP64, I64), \
MKLDNN_CVT(FP64, FP32), MKLDNN_CVT(FP64, FP16), MKLDNN_CVT(FP64, BF16), MKLDNN_CVT(FP64, BOOL), \
MKLDNN_CVT(BOOL, U8), MKLDNN_CVT(BOOL, I8), MKLDNN_CVT(BOOL, U16), MKLDNN_CVT(BOOL, I16), \
MKLDNN_CVT(BOOL, U32), MKLDNN_CVT(BOOL, I32), MKLDNN_CVT(BOOL, U64), MKLDNN_CVT(BOOL, I64), \
MKLDNN_CVT(BOOL, FP32), MKLDNN_CVT(BOOL, FP16), MKLDNN_CVT(BOOL, BF16), MKLDNN_CVT(BOOL, FP64), \
MKLDNN_CVT(U8, U8), MKLDNN_CVT(I8, I8), MKLDNN_CVT(U16, U16), MKLDNN_CVT(I16, I16), \
MKLDNN_CVT(U32, U32), MKLDNN_CVT(I32, I32), MKLDNN_CVT(U64, U64), MKLDNN_CVT(I64, I64), \
MKLDNN_CVT(FP32, FP32), MKLDNN_CVT(FP16, FP16), MKLDNN_CVT(BF16, BF16), MKLDNN_CVT(FP64, FP64), \
MKLDNN_CVT(BOOL, BOOL)
void cpu_convert(const void *srcPtr, void *dstPtr, Precision srcPrc, Precision dstPrc, const size_t size) {
cpu_convert(srcPtr, dstPtr, srcPrc, dstPrc, dstPrc, size);
}
void cpu_convert(const void *srcPtr,
void *dstPtr,
InferenceEngine::Precision srcPrc,
InferenceEngine::Precision interimPrc,
InferenceEngine::Precision dstPrc,
const size_t size) {
if (srcPtr == nullptr || dstPtr == nullptr)
IE_THROW() << "cpu_convert has null data pointer";
if (srcPrc == dstPrc) {
cpu_memcpy(dstPtr, srcPtr, size*dstPrc.size());
return;
if (srcPrc == dstPrc && srcPrc == interimPrc) {
cpu_memcpy(dstPtr, srcPtr, size * dstPrc.size());
} else {
ConvertContext ctx = {
srcPtr,
dstPtr,
size,
interimPrc,
dstPrc,
false
};
OV_SWITCH(MKLDNNPlugin, ConvertPrecision, ctx, std::tie(srcPrc, dstPrc), MKLDNN_CVT_LIST);
if (!ctx.converted)
IE_THROW() << "cpu_convert can't convert from: " << srcPrc << " precision to: " << dstPrc;
}
ConvertContext ctx = { srcPtr, dstPtr, size, false };
OV_SWITCH(MKLDNNPlugin, ConvertPrecision, ctx, std::tie(srcPrc, dstPrc),
MKLDNN_CVT(U8, I8), MKLDNN_CVT(U8, U16), MKLDNN_CVT(U8, I16),
MKLDNN_CVT(U8, I32), MKLDNN_CVT(U8, U64), MKLDNN_CVT(U8, I64),
MKLDNN_CVT(U8, FP32), MKLDNN_CVT(U8, BF16), MKLDNN_CVT(U8, BOOL),
MKLDNN_CVT(I8, U8), MKLDNN_CVT(I8, U16), MKLDNN_CVT(I8, I16),
MKLDNN_CVT(I8, I32), MKLDNN_CVT(I8, U64), MKLDNN_CVT(I8, I64),
MKLDNN_CVT(I8, FP32), MKLDNN_CVT(I8, BF16), MKLDNN_CVT(I8, BOOL),
MKLDNN_CVT(U16, U8), MKLDNN_CVT(U16, I8), MKLDNN_CVT(U16, I16),
MKLDNN_CVT(U16, I32), MKLDNN_CVT(U16, U64), MKLDNN_CVT(U16, I64),
MKLDNN_CVT(U16, FP32), MKLDNN_CVT(U16, BF16), MKLDNN_CVT(U16, BOOL),
MKLDNN_CVT(I16, U8), MKLDNN_CVT(I16, I8), MKLDNN_CVT(I16, U16),
MKLDNN_CVT(I16, I32), MKLDNN_CVT(I16, U64), MKLDNN_CVT(I16, I64),
MKLDNN_CVT(I16, FP32), MKLDNN_CVT(I16, BF16), MKLDNN_CVT(I16, BOOL),
MKLDNN_CVT(I32, U8), MKLDNN_CVT(I32, I8), MKLDNN_CVT(I32, U16),
MKLDNN_CVT(I32, I16), MKLDNN_CVT(I32, U64), MKLDNN_CVT(I32, I64),
MKLDNN_CVT(I32, FP32), MKLDNN_CVT(I32, BF16), MKLDNN_CVT(I32, BOOL),
MKLDNN_CVT(U64, U8), MKLDNN_CVT(U64, I8), MKLDNN_CVT(U64, U16),
MKLDNN_CVT(U64, I16), MKLDNN_CVT(U64, I32), MKLDNN_CVT(U64, I64),
MKLDNN_CVT(U64, FP32), MKLDNN_CVT(U64, BF16), MKLDNN_CVT(U64, BOOL),
MKLDNN_CVT(I64, U8), MKLDNN_CVT(I64, I8), MKLDNN_CVT(I64, U16),
MKLDNN_CVT(I64, I16), MKLDNN_CVT(I64, I32), MKLDNN_CVT(I64, U64),
MKLDNN_CVT(I64, FP32), MKLDNN_CVT(I64, BF16), MKLDNN_CVT(I64, BOOL),
MKLDNN_CVT(FP32, U8), MKLDNN_CVT(FP32, I8), MKLDNN_CVT(FP32, U16),
MKLDNN_CVT(FP32, I16), MKLDNN_CVT(FP32, I32), MKLDNN_CVT(FP32, U64),
MKLDNN_CVT(FP32, I64), MKLDNN_CVT(FP32, BF16), MKLDNN_CVT(FP32, BOOL),
MKLDNN_CVT(BF16, U8), MKLDNN_CVT(BF16, I8), MKLDNN_CVT(BF16, U16),
MKLDNN_CVT(BF16, I16), MKLDNN_CVT(BF16, I32), MKLDNN_CVT(BF16, U64),
MKLDNN_CVT(BF16, I64), MKLDNN_CVT(BF16, FP32), MKLDNN_CVT(BF16, BOOL),
MKLDNN_CVT(BOOL, U8), MKLDNN_CVT(BOOL, I8), MKLDNN_CVT(BOOL, U16),
MKLDNN_CVT(BOOL, I16), MKLDNN_CVT(BOOL, I32), MKLDNN_CVT(BOOL, U64),
MKLDNN_CVT(BOOL, I64), MKLDNN_CVT(BOOL, FP32), MKLDNN_CVT(BOOL, BF16),
MKLDNN_CVT(FP64, U8), MKLDNN_CVT(FP64, I8), MKLDNN_CVT(FP64, U16),
MKLDNN_CVT(FP64, I16), MKLDNN_CVT(FP64, I32), MKLDNN_CVT(FP64, U64),
MKLDNN_CVT(FP64, I64), MKLDNN_CVT(FP64, FP32), MKLDNN_CVT(FP64, BF16), MKLDNN_CVT(FP64, BOOL),
MKLDNN_CVT(U32, U8), MKLDNN_CVT(U32, I8), MKLDNN_CVT(U32, U16),
MKLDNN_CVT(U32, I16), MKLDNN_CVT(U32, I32), MKLDNN_CVT(U32, U64),
MKLDNN_CVT(U32, I64), MKLDNN_CVT(U32, FP32), MKLDNN_CVT(U32, BF16), MKLDNN_CVT(U32, BOOL));
if (!ctx.converted)
IE_THROW() << "cpu_convert can't convert from: " << srcPrc << " precision to: " << dstPrc;
}
#undef MKLDNN_CVT
#undef MKLDNN_CVT_LIST

29
inference-engine/src/mkldnn_plugin/nodes/common/cpu_convert.h
View File

@ -19,5 +19,32 @@
* number of elements in buffers to be converted
* @return none.
*/
void cpu_convert(const void *srcPtr,
void *dstPtr,
InferenceEngine::Precision srcPrc,
InferenceEngine::Precision dstPrc,
const size_t size);
void cpu_convert(const void *srcPtr, void *dstPtr, InferenceEngine::Precision srcPrc, InferenceEngine::Precision dstPrc, const size_t size);
/**
* @brief Copy size elements from buffer specified srcPtr pointer to buffer specified dstPtr.
* If the precisions srcPrc and dstPrc are different, a conversion from srcPrc to dstPrc is performed.
* @param srcPtr
* pointer to the buffer to convert from
* @param dstPtr
* pointer to the buffer to convert to
* @param srcPrc
* precision the buffer from which convert
* @param interimPrc
* intermediate precision used for type truncation
* @param dstPrc
* precision the buffer to which convert
* @param size
* number of elements in buffers to be converted
* @return none.
*/
void cpu_convert(const void *srcPtr,
void *dstPtr,
InferenceEngine::Precision srcPrc,
InferenceEngine::Precision interimPrc,
InferenceEngine::Precision dstPrc,
const size_t size);

2
inference-engine/src/mkldnn_plugin/nodes/mkldnn_conv_node.cpp
View File

@ -979,7 +979,7 @@ void MKLDNNConvolutionNode::prepareParams() {
primArgs[DNNL_ARG_BIAS] = getBias();
}
appendPostOpArgs(*pAttrLocal);
appendPostOpArgs(*pAttrLocal, primArgs, binaryPostOpsArgs);
}
void MKLDNNConvolutionNode::executeDynamicImpl(dnnl::stream strm) {

21
inference-engine/src/mkldnn_plugin/nodes/mkldnn_convert_node.cpp
View File

@ -7,7 +7,8 @@
#include "common/cpu_convert.h"
#include "common/blocked_desc_creator.h"
#include <ngraph/opsets/opset1.hpp>
#include "utils/ngraph_utils.hpp"
#include <ie_ngraph_utils.hpp>
#include <utils/ngraph_utils.hpp>
using namespace mkldnn;
using namespace MKLDNNPlugin;
@ -26,14 +27,17 @@ bool MKLDNNConvertNode::isSupportedOperation(const std::shared_ptr<const ngraph:
return true;
}
MKLDNNConvertNode::MKLDNNConvertNode(const std::shared_ptr<ngraph::Node>& op, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache) :
MKLDNNNode(op, eng, cache) {
MKLDNNConvertNode::MKLDNNConvertNode(const std::shared_ptr<ngraph::Node>& op, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache)
: MKLDNNNode(op, eng, cache) {
std::string errorMessage;
if (isSupportedOperation(op, errorMessage)) {
errorPrefix = "Convert node with name '" + getName() + "'";
} else {
IE_THROW(NotImplemented) << errorMessage;
}
auto convert = ov::as_type_ptr<const ngraph::opset1::Convert>(op);
origPrc = details::convertPrecision(convert->get_destination_type());
}
std::vector<VectorDims> MKLDNNConvertNode::shapeInfer() const {
@ -42,7 +46,8 @@ std::vector<VectorDims> MKLDNNConvertNode::shapeInfer() const {
MKLDNNConvertNode::MKLDNNConvertNode(const Shape &shape, const InferenceEngine::Precision &inPrc, const InferenceEngine::Precision &outPrc,
const std::string &nodeName, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache)
: MKLDNNNode("Convert", nodeName, eng, cache) {
: MKLDNNNode("Convert", nodeName, eng, cache)
, origPrc(outPrc) {
inputShapes.push_back(shape);
addOriginalInputPrecision(inPrc);
outputShapes.push_back(shape);
@ -147,7 +152,13 @@ void MKLDNNConvertNode::execute(mkldnn::stream strm) {
void* srcPtr = parentMem.GetPtr();
void* dstPtr = childMem.GetPtr();
cpu_convert(srcPtr, dstPtr, parentMem.getDesc().getPrecision(), childMem.getDesc().getPrecision(), parentPaddElemCount);
cpu_convert(srcPtr,
dstPtr,
parentMem.getDesc().getPrecision(),
origPrc,
childMem.getDesc().getPrecision(),
parentPaddElemCount);
}
bool MKLDNNConvertNode::created() const {

1
inference-engine/src/mkldnn_plugin/nodes/mkldnn_convert_node.h
View File

@ -49,6 +49,7 @@ public:
private:
MemoryDescPtr input;
MemoryDescPtr output;
InferenceEngine::Precision origPrc;
std::string errorPrefix;
};

48
inference-engine/src/mkldnn_plugin/nodes/mkldnn_cum_sum_node.cpp
View File

@ -12,6 +12,7 @@
#include "ie_precision.hpp"
#include <ie_ngraph_utils.hpp>
#include "mkldnn_cum_sum_node.h"
#include "utils/bfloat16.hpp"
using namespace MKLDNNPlugin;
using namespace InferenceEngine;
@ -70,8 +71,7 @@ void MKLDNNCumSumNode::initSupportedPrimitiveDescriptors() {
return;
dataPrecision = getOriginalInputPrecisionAtPort(CUM_SUM_DATA);
if (dataPrecision != Precision::I8 && dataPrecision != Precision::U8 && dataPrecision != Precision::I16 && dataPrecision != Precision::I32 &&
dataPrecision != Precision::FP32 && dataPrecision != Precision::I64 && dataPrecision != Precision::U64 && dataPrecision != Precision::BF16)
if (!one_of(dataPrecision, Precision::I8, Precision::U8, Precision::I16, Precision::BF16, Precision::I32, Precision::FP32, Precision::I64, Precision::U64))
IE_THROW() << errorPrefix << " has unsupported 'data' input precision: " << dataPrecision.name();
if (inputShapes.size() == numOfInputs) {
@ -95,43 +95,17 @@ void MKLDNNCumSumNode::execute(mkldnn::stream strm) {
if (inputShapes.size() == numOfInputs)
axis = getAxis(getParentEdgeAt(AXIS)->getMemory(), getParentEdgeAt(CUM_SUM_DATA)->getMemory());
switch (dataPrecision) {
case Precision::I8 : {
exec<int8_t>();
break;
}
case Precision::U8 : {
exec<uint8_t>();
break;
}
case Precision::I16 : {
exec<int16_t>();
break;
}
case Precision::I32 : {
exec<int32_t>();
break;
}
case Precision::FP32 : {
exec<float>();
break;
}
case Precision::I64 : {
exec<int64_t>();
break;
}
case Precision::U64 : {
exec<uint64_t>();
break;
}
default : {
std::string errorMsg = errorPrefix + " has unsupported 'data' input precision: " + dataPrecision.name();
IE_THROW() << errorMsg;
}
}
OV_SWITCH(MKLDNNPlugin, CumSumExecute, this, dataPrecision,
OV_CASE(Precision::I8, int8_t),
OV_CASE(Precision::U8, uint8_t),
OV_CASE(Precision::I16, int16_t),
OV_CASE(Precision::BF16, bfloat16_t),
OV_CASE(Precision::I32, int32_t),
OV_CASE(Precision::FP32, float),
OV_CASE(Precision::I64, int64_t),
OV_CASE(Precision::U64, uint64_t))
}
template <typename dataType>
void MKLDNNCumSumNode::exec() {
const auto *input = reinterpret_cast<const dataType *>(getParentEdgeAt(CUM_SUM_DATA)->getMemoryPtr()->GetPtr());

7
inference-engine/src/mkldnn_plugin/nodes/mkldnn_cum_sum_node.h
View File

@ -47,6 +47,13 @@ private:
InferenceEngine::Precision dataPrecision;
std::string errorPrefix;
template<typename T>
struct CumSumExecute {
void operator()(MKLDNNCumSumNode* node) {
node->exec<T>();
}
};
};
} // namespace MKLDNNPlugin

611
inference-engine/src/mkldnn_plugin/nodes/mkldnn_deconv_node.cpp
View File

@ -13,34 +13,38 @@
#include <mkldnn_extension_utils.h>
#include "ie_parallel.hpp"
#include "utils/general_utils.h"
#include <ngraph/opsets/opset1.hpp>
#include <cpu/x64/cpu_isa_traits.hpp>
#include <nodes/common/cpu_memcpy.h>
#include <memory_desc/cpu_memory_desc_utils.h>
#include "memory_desc/dnnl_blocked_memory_desc.h"
#include "utils/cpu_utils.hpp"
#include <ngraph/opsets/opset1.hpp>
#include <utils/shape_inference/static_shape.hpp>
#include <utils/shape_inference/shape_inference.hpp>
#include <ie_ngraph_utils.hpp>
#include "convolution_shape_inference.hpp"
using namespace mkldnn;
using namespace MKLDNNPlugin;
using namespace InferenceEngine;
bool MKLDNNDeconvolutionNode::isSupportedOperation(const std::shared_ptr<const ngraph::Node>& op, std::string& errorMessage) noexcept {
try {
if (isDynamicNgraphNode(op)) {
errorMessage = "Doesn't support op with dynamic shapes";
return false;
}
if (std::dynamic_pointer_cast<const ngraph::opset1::ConvolutionBackpropData>(op) == nullptr &&
std::dynamic_pointer_cast<const ngraph::opset1::GroupConvolutionBackpropData>(op) == nullptr) {
errorMessage = "Only opset1 ConvolutionBackpropData and GroupConvolutionBackpropData operations are supported";
return false;
}
size_t ndims = op->get_input_shape(0).size();
size_t ndims = op->get_input_partial_shape(0).rank().get_length();
if ((ndims < 3) || (ndims > 5)) {
errorMessage = "Only 3D, 4D and 5D blobs are supported as input";
return false;
}
if (op->get_input_partial_shape(1).is_dynamic() || (op->get_input_size() > 2 && op->get_input_partial_shape(2).is_dynamic())) {
errorMessage = "Doesn't support dynamic shapes for 'weights' and 'output_shape' inputs";
return false;
}
} catch (...) {
return false;
}
@ -58,15 +62,14 @@ MKLDNNDeconvolutionNode::MKLDNNDeconvolutionNode(const std::shared_ptr<ngraph::N
auto convBackprop = std::dynamic_pointer_cast<const ngraph::opset1::ConvolutionBackpropData>(op);
auto groupConvBackprop = std::dynamic_pointer_cast<const ngraph::opset1::GroupConvolutionBackpropData>(op);
const auto dataShape = op->get_input_shape(0);
weightDims = op->get_input_shape(1);
const auto outShape = op->get_shape();
OC = outShape[1];
IC = dataShape[1];
const auto& weightDims = getWeightDims();
if (convBackprop) {
algorithm = DeconvolutionCommon;
IC = weightDims[0];
OC = weightDims[1];
groupNum = 1;
withGroups = false;
@ -78,10 +81,17 @@ MKLDNNDeconvolutionNode::MKLDNNDeconvolutionNode(const std::shared_ptr<ngraph::N
}
paddingL = convBackprop->get_pads_begin();
paddingR = convBackprop->get_pads_end();
outputPadding = convBackprop->get_output_padding();
autoPad = one_of(convBackprop->get_auto_pad(), ov::op::PadType::SAME_LOWER, ov::op::PadType::SAME_UPPER);
} else if (groupConvBackprop) {
algorithm = DeconvolutionGrouped;
groupNum = weightDims[0];
IC = groupNum * weightDims[1];
OC = groupNum * weightDims[2];
withGroups = groupNum > 1;
isDW = withGroups && groupNum == OC && groupNum == IC;
@ -93,10 +103,26 @@ MKLDNNDeconvolutionNode::MKLDNNDeconvolutionNode(const std::shared_ptr<ngraph::N
}
paddingL = groupConvBackprop->get_pads_begin();
paddingR = groupConvBackprop->get_pads_end();
outputPadding = groupConvBackprop->get_output_padding();
autoPad = one_of(groupConvBackprop->get_auto_pad(), ov::op::PadType::SAME_LOWER, ov::op::PadType::SAME_UPPER);
}
for (int i = 0; i < dilation.size(); i++) {
kernel.push_back(weightDims[withGroups + 2 + i]);
}
externOutShape = inputShapes.size() == 3;
if (externOutShape && isDynamicNode()) {
bool isConstOutShape = ngraph::is_type<ov::op::v0::Constant>(op->get_input_node_shared_ptr(2));
if (isConstOutShape) {
lastOutputSpatialDims = ov::as_type<ov::op::v0::Constant>(op->get_input_node_ptr(2))->cast_vector<int32_t>();
}
const auto spDimsNum = getInputShapeAtPort(0).getRank() - 2;
if (getInputShapeAtPort(2).getStaticDims()[0] != spDimsNum || (isConstOutShape && lastOutputSpatialDims.size() != spDimsNum)) {
IE_THROW() << "'output_shape' input has incorrect number of elements. Expected = " << spDimsNum;
}
}
} else {
IE_THROW(NotImplemented) << errorMessage;
}
@ -113,14 +139,6 @@ InferenceEngine::Blob::Ptr MKLDNNDeconvolutionNode::createWeiBlobAsIO(InferenceE
auto const blbSize = blb->GetSize();
// WA: In int8 case, we are processing weights using internal blob.
// So we disconnect constant node containing weights from the graph and then don't use it.
if (getParentEdges().size() == 3) {
removeEdge(getParentEdgeAt(2));
inputShapes.erase(inputShapes.begin() + 2);
}
removeEdge(getParentEdgeAt(1));
inputShapes.erase(inputShapes.begin() + 1);
InferenceEngine::SizeVector dimsForBlockedDesc{dims};
std::swap(dimsForBlockedDesc[withGroups + 0], dimsForBlockedDesc[withGroups + 1]);
@ -160,13 +178,16 @@ bool MKLDNNDeconvolutionNode::canBeExecutedInInt8() const {
if (!withGroups && stride.back() > 3)
return false;
if (!impl::cpu::x64::mayiuse(impl::cpu::x64::avx512_common)) {
auto inDims = getOutputShapeAtPort(0).getStaticDims();
const auto& inMaxDims = getOutputShapeAtPort(0).getMaxDims();
if (std::any_of(inMaxDims.begin(), inMaxDims.end(), [](Dim dim) { return dim == Shape::UNDEFINED_DIM; })) {
return false;
}
// heuristicConst = 2^26
// heuristicParam = IC^2 * SP
auto heuristicConst = 67108864;
auto heuristicParam = IC * IC;
for (int i = 2; i < inDims.size(); i++)
heuristicParam *= inDims[i];
for (int i = 2; i < inMaxDims.size(); i++)
heuristicParam *= inMaxDims[i];
if (heuristicParam > heuristicConst)
return false;
}
@ -203,10 +224,65 @@ bool MKLDNNDeconvolutionNode::canFuse(const MKLDNNNodePtr& node) const {
return (fusedWith.empty() && node->canBePerformedAsScaleShift(this));
}
void MKLDNNDeconvolutionNode::getSupportedDescriptors() {
if (!descs_fwd.empty() && !descs_bwd.empty())
return;
void MKLDNNDeconvolutionNode::initPadding(std::shared_ptr<ngraph::Node> op, const Shape &inDims, const std::vector<int32_t>& outSpDims) {
std::vector<ov::StaticShape> input_shapes{inDims.getStaticDims(), getWeightDims()};
ov::StaticShape output_shape_input;
if (externOutShape) {
IE_ASSERT(outSpDims.size() == getInputShapeAtPort(2).getStaticDims()[0]);
input_shapes.push_back({outSpDims.size()});
for (size_t i = 0; i < outSpDims.size(); i++) {
output_shape_input.push_back(outSpDims[i]);
}
}
if (getAlgorithm() == DeconvolutionCommon) {
auto deconv = ngraph::as_type_ptr<ngraph::op::v1::ConvolutionBackpropData>(op);
IE_ASSERT(ov::op::v1::resolve_auto_pad_for_shape_back_prop(deconv.get(), paddingL, paddingR, input_shapes, output_shape_input, 2, 2));
} else if (getAlgorithm() == DeconvolutionGrouped) {
auto deconv = ngraph::as_type_ptr<ngraph::op::v1::GroupConvolutionBackpropData>(op);
IE_ASSERT(ov::op::v1::resolve_auto_pad_for_shape_back_prop(deconv.get(), paddingL, paddingR, input_shapes, output_shape_input, 2, 3));
}
}
std::pair<VectorDims, VectorDims> MKLDNNDeconvolutionNode::makeDummyInOutShape() {
auto inShape = MemoryDescUtils::makeDummyShape(getInputShapeAtPort(0));
auto outShape = getOutputShapeAtPort(0);
if (isDynamicNode()) {
if (externOutShape) {
if (lastOutputSpatialDims.empty()) {
const auto& shape = getOutputShapeAtPort(0);
lastOutputSpatialDims.resize(shape.getRank() - 2);
const auto& minDims = shape.getMinDims();
const auto& maxDims = shape.getMaxDims();
const auto& dims = shape.getDims();
for (size_t i = 0; i < dims.size() - 2; ++i) {
lastOutputSpatialDims[i] = dims[i + 2] == Shape::UNDEFINED_DIM ? std::min(maxDims[i + 2],
std::max(minDims[i + 2], static_cast<Dim>(64))) : dims[i + 2];
}
}
ov::CoordinateDiff pb = autoPad ? ov::CoordinateDiff(paddingL.size(), 0) : paddingL;
ov::CoordinateDiff pe = autoPad ? ov::CoordinateDiff(paddingR.size(), 0) : paddingR;
auto inputDims = inShape.getStaticDims();
const auto& weightDims = getWeightDims();
const size_t wghOffset = getAlgorithm() == DeconvolutionGrouped ? 1 : 0;
for (size_t i = 0; i < inputDims.size() - 2; i++) {
inputDims[2 + i] = ((lastOutputSpatialDims[i] - (dilation[i] + 1) *
(weightDims[wghOffset + 2 + i] - 1) - 1 + pb[i] + pe[i] - outputPadding[i])) /
stride[i] + 1;
}
inShape = Shape(inputDims);
}
initPadding(opToShapeInfer, inShape, lastOutputSpatialDims);
outShape = Shape(shapeInferInternal(inShape.getStaticDims(), lastOutputSpatialDims));
}
return {inShape.getStaticDims(), outShape.getStaticDims()};
}
void MKLDNNDeconvolutionNode::getSupportedDescriptors() {
isInt8 = canBeExecutedInInt8();
InferenceEngine::Precision inPrecision = getOriginalInputPrecisionAtPort(0);
@ -236,21 +312,17 @@ void MKLDNNDeconvolutionNode::getSupportedDescriptors() {
if (getChildEdges().empty())
IE_THROW() << errorPrefix << " has incorrect number of output edges";
for (int i = 0; i < paddingR.size(); i++) {
int with_group = getAlgorithm() == DeconvolutionGrouped ? 1 : 0;
int krn = weightDims[with_group + 2 + i];
int src = getOutputShapeAtPort(0).getStaticDims()[2 + i];
int dst = getInputShapeAtPort(0).getStaticDims()[2 + i];
krn = (krn - 1)*(dilation[i] + 1) + 1;
int calc_dst = (src - krn + paddingL[i]) / stride[i] + 1;
paddingR[i] = (dst - calc_dst) * stride[i];
}
VectorDims inDims, outDims;
std::tie(inDims, outDims) = makeDummyInOutShape();
inShape = Shape(inDims);
Shape outShape(outDims);
initPaddingR(inShape, outShape);
if (isInt8) {
int8WeightDims = getWeightDims();
// WA: if int8 deconvolution is supported, we create internal weights blob in IO format
std::swap(weightDims[withGroups + 0], weightDims[withGroups + 1]);
internalBlobs.push_back(createWeiBlobAsIO(weightDims));
std::swap(int8WeightDims[withGroups + 0], int8WeightDims[withGroups + 1]);
internalBlobs.push_back(createWeiBlobAsIO(int8WeightDims));
auto format = getInputShapeAtPort(0).getRank() == 5 ? dnnl::memory::format_tag::ndhwc : dnnl::memory::format_tag::nhwc;
MemoryDescPtr in_candidate = std::make_shared<DnnlBlockedMemoryDesc>(getInputShapeAtPort(0), inputDataType, format);
MemoryDescPtr out_candidate = std::make_shared<DnnlBlockedMemoryDesc>(getOutputShapeAtPort(0), outputDataType, format);
@ -262,18 +334,31 @@ void MKLDNNDeconvolutionNode::getSupportedDescriptors() {
createDescriptor({in_candidate}, {out_candidate});
}
}
setPostOps(attr);
setPostOps(attr, outShape.getStaticDims());
}
void MKLDNNDeconvolutionNode::setPostOps(mkldnn::primitive_attr &attr) {
void MKLDNNDeconvolutionNode::initPaddingR(const Shape &inShape, const Shape &outShape) {
for (int i = 0; i < paddingR.size(); i++) {
int with_group = getAlgorithm() == DeconvolutionGrouped ? 1 : 0;
const auto& weightDims = getWeightDims();
int krn = weightDims[with_group + 2 + i];
int src = outShape.getStaticDims()[2 + i];
int dst = inShape.getStaticDims()[2 + i];
krn = (krn - 1)*(dilation[i] + 1) + 1;
int calc_dst = (src - krn + paddingL[i]) / stride[i] + 1;
paddingR[i] = (dst - calc_dst) * stride[i];
}
}
void MKLDNNDeconvolutionNode::setPostOps(mkldnn::primitive_attr &attr, const VectorDims &dims) {
mkldnn::post_ops ops;
auto getBinPostOpShape = [&](){
const auto outShape = getOutputShapeAtPort(0).getStaticDims();
const auto outShapeRank = getOutputShapeAtPort(0).getRank();
const auto chIdx = getFusingAxis();
std::vector<size_t> binaryShape(outShapeRank, 1);
binaryShape[chIdx] = outShape[chIdx];
binaryShape[chIdx] = dims[chIdx];
return binaryShape;
};
@ -282,7 +367,7 @@ void MKLDNNDeconvolutionNode::setPostOps(mkldnn::primitive_attr &attr) {
// TODO [DS]: change to shape from memory
constexpr int align = 16;
// use legacy depthwise since backprop convolution does not support binary post ops
eltwiseNode->appendPostOps(ops, getOutputShapeAtPort(0).getStaticDims(), align);
eltwiseNode->appendPostOps(ops, dims, align);
continue;
}
if (auto* fakeQuantizeNode = dynamic_cast<MKLDNNFakeQuantizeNode *>(node.get())) {
@ -339,80 +424,277 @@ bool MKLDNNDeconvolutionNode::created() const {
return getType() == Deconvolution;
}
void MKLDNNDeconvolutionNode::createPrimitive() {
if (prim)
return;
if (isInt8) {
auto prim_desc = createPrimitiveDescriptor<deconvolution_forward::primitive_desc,
deconvolution_forward::desc>(attr);
prim.reset(new deconvolution_forward(prim_desc));
auto src = getParentEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
auto dst = getChildEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
primArgs = {{DNNL_ARG_SRC, src}, {DNNL_ARG_WEIGHTS, internalBlobMemory[0]->GetPrimitive()}, {DNNL_ARG_DST, dst}};
} else {
auto prim_desc = createPrimitiveDescriptor<convolution_backward_data::primitive_desc,
convolution_backward_data::desc, convolution_forward::primitive_desc>(attr);
prim.reset(new convolution_backward_data(prim_desc));
auto src = getParentEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
auto weights = getParentEdgeAt(1)->getMemory().GetPrimitive();
auto dst = getChildEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
primArgs = {{DNNL_ARG_DIFF_DST, src}, {DNNL_ARG_WEIGHTS, weights}, {DNNL_ARG_DIFF_SRC, dst}};
bool MKLDNNDeconvolutionNode::needShapeInfer() const {
if (inputShapesModified()) {
return true;
}
if (externOutShape) {
if (lastOutputSpatialDims != readOutputSpatialDims()) {
return true;
}
}
appendPostOpArgs(attr);
return false;
}
void MKLDNNDeconvolutionNode::createDescriptor(const std::vector<MemoryDescPtr> &inputDesc,
const std::vector<MemoryDescPtr> &outputDesc) {
const auto in_candidate = MemoryDescUtils::convertToDnnlBlockedMemoryDesc(*inputDesc[0]);
const auto out_candidate = MemoryDescUtils::convertToDnnlBlockedMemoryDesc(*outputDesc[0]);
std::vector<VectorDims> MKLDNNDeconvolutionNode::shapeInfer() const {
const auto &dataMemPtr = getParentEdgesAtPort(0)[0]->getMemoryPtr();
std::vector<int32_t> outSpDims;
if (externOutShape) {
outSpDims = readOutputSpatialDims();
}
return {shapeInferInternal(dataMemPtr->getStaticDims(), outSpDims)};
}
// grouping and autoblicking is not compatible
if ((withGroups && !isDW) && (in_candidate.blocksExtended() || out_candidate.blocksExtended()))
return;
VectorDims MKLDNNDeconvolutionNode::shapeInferInternal(const VectorDims &inDims, std::vector<int32_t> outSpDims) const {
std::vector<ov::StaticShape> inputShapes = {
inDims,
getWeightDims()
};
std::map<size_t, std::shared_ptr<ngraph::runtime::HostTensor>> inputValues;
if (externOutShape) {
if (outSpDims.size() != getInputShapeAtPort(2).getStaticDims()[0]) {
IE_THROW() << "Can't compute output shape for node with name: " << getName()
<< ", because the node has 'output_shape' input, but provided output spatial dims number is incorrect";
}
inputShapes.push_back({outSpDims.size()});
inputValues.insert({2, std::make_shared<ngraph::runtime::HostTensor>(ngraph::element::Type_t::i32,
inputShapes.back().to_shape(),
outSpDims.data())});
}
std::vector<ov::StaticShape> outputShapes(1);
shape_inference(opToShapeInfer.get(), inputShapes, outputShapes, inputValues);
return outputShapes.back().to_shape();
}
void MKLDNNDeconvolutionNode::execute(mkldnn::stream strm) {
if (!execPtr) {
IE_THROW() << "Can't execute Deconvolution node with name: " << getName() << ", because executor is not compiled";
}
execPtr->exec(strm);
if (externOutShape) {
lastOutputSpatialDims = readOutputSpatialDims();
}
}
std::shared_ptr<MKLDNNDescriptor> MKLDNNDeconvolutionNode::createDefaultMkldnnDeconvDesc(const mkldnn::memory::desc& srcDesc,
const mkldnn::memory::desc& wghDesc,
const mkldnn::memory::desc& dstDesc,
bool isWinograd) const {
mkldnn::algorithm alg = isWinograd ? mkldnn::algorithm::convolution_winograd : mkldnn::algorithm::convolution_direct;
std::shared_ptr<convolution_backward_data::desc> deconv_desc;
std::shared_ptr<convolution_forward::primitive_desc> fwd_conv_pd;
std::tie(deconv_desc, fwd_conv_pd) = createDescriptorInternalDefault(srcDesc, wghDesc, dstDesc, alg);
if (fwd_conv_pd->get(true) == nullptr) {
IE_THROW() << "Forward convolution primitive descriptor is nullable for node with name: " << getName();
}
return std::make_shared<MKLDNNDescriptor>(deconv_desc, fwd_conv_pd);
}
std::shared_ptr<MKLDNNDescriptor> MKLDNNDeconvolutionNode::createInt8MkldnnDeconvDesc(const mkldnn::memory::desc& srcDesc,
const mkldnn::memory::desc& wghDesc,
const mkldnn::memory::desc& dstDesc) const {
return std::make_shared<MKLDNNDescriptor>(createDescriptorInternalInt8(srcDesc, wghDesc, dstDesc));
}
void MKLDNNDeconvolutionNode::createDeconvPrim(std::shared_ptr<MKLDNNDescriptor> desc,
MKLDNNMemoryPtr srcMemPtr,
MKLDNNMemoryPtr wghMemPtr,
MKLDNNMemoryPtr dstMemPtr,
AttrPtr attr,
impl_desc_type selectedImpl) {
auto itpd = desc->createPrimitiveDescriptorIterator(getEngine(), *attr);
while (static_cast<bool>(itpd)) {
impl_desc_type impl_type = parse_impl_name(itpd.impl_info_str());
if (impl_type == selectedImpl) {
if (isInt8) {
if (internalBlobMemory.empty()) {
prepareMemory(itpd);
}
auto prim_desc = deconvolution_forward::primitive_desc(itpd.get());
execPtr = std::make_shared<DeconvExecutorInt8>(prim_desc, srcMemPtr, internalBlobMemory.front(), dstMemPtr, *attr,
binaryPostOpsArgs, getEngine());
} else {
auto prim_desc = convolution_backward_data::primitive_desc(itpd.get());
execPtr = std::make_shared<DeconvExecutorDefault>(prim_desc, srcMemPtr, wghMemPtr, dstMemPtr, *attr,
binaryPostOpsArgs, getEngine());
}
return;
}
if (!itpd.next_impl()) {
auto inDesc = mkldnn::memory::desc(MKLDNNExtensionUtils::convertToDnnlDims(srcMemPtr->getStaticDims()),
memory::data_type::f32,
memory::format_tag::any);
auto wghDesc = mkldnn::memory::desc(MKLDNNExtensionUtils::convertToDnnlDims(wghMemPtr->getStaticDims()),
memory::data_type::f32,
memory::format_tag::any);
auto outDesc = mkldnn::memory::desc(MKLDNNExtensionUtils::convertToDnnlDims(dstMemPtr->getStaticDims()),
memory::data_type::f32,
memory::format_tag::any);
std::shared_ptr<MKLDNNDescriptor> anyDeconvDesc = createDefaultMkldnnDeconvDesc(inDesc, wghDesc, outDesc, false);
auto anyDeconvItpd = anyDeconvDesc->createPrimitiveDescriptorIterator(getEngine(), *attr);
if (static_cast<bool>(anyDeconvItpd)) {
auto prim_desc = convolution_backward_data::primitive_desc(anyDeconvItpd.get());
execPtr = std::make_shared<DeconvExecutorDefault>(prim_desc, srcMemPtr, wghMemPtr, dstMemPtr, *attr,
binaryPostOpsArgs, getEngine());
return;
}
}
}
IE_THROW() << "Primitive descriptor was not found for node " << getName() << ".";
}
void MKLDNNDeconvolutionNode::prepareParams() {
auto srcMemPtr = getParentEdgesAtPort(0)[0]->getMemoryPtr();
auto dstMemPtr = getChildEdgesAtPort(0)[0]->getMemoryPtr();
if (!dstMemPtr || !dstMemPtr->GetPrimitivePtr())
IE_THROW() << "Destination memory didn't allocate.";
if (!srcMemPtr || !srcMemPtr->GetPrimitivePtr())
IE_THROW() << "Input memory didn't allocate.";
const NodeDesc *selected_pd = getSelectedPrimitiveDescriptor();
if (selected_pd == nullptr)
IE_THROW() << "Preferable primitive descriptor is not set for node " << getName() << ".";
auto inMemoryDesc = getParentEdgesAtPort(0).front()->getMemory().GetDescWithType<DnnlMemoryDesc>();
auto outMemoryDesc = getChildEdgesAtPort(0).front()->getMemory().GetDescWithType<DnnlMemoryDesc>();
auto initPrimitiveAttr = [&]() {
mkldnn::primitive_attr attr;
setPostOps(attr, dstMemPtr->getStaticDims());
return std::make_shared<mkldnn::primitive_attr>(std::move(attr));
};
AttrPtr pAttrLocal;
if (isDynamicNode()) {
if (!pAttr) {
pAttr = initPrimitiveAttr();
}
pAttrLocal = pAttr;
if (autoPad || externOutShape) {
initPadding(opToShapeInfer, inMemoryDesc->getShape(), externOutShape ? readOutputSpatialDims() : std::vector<int32_t>{});
}
initPaddingR(inMemoryDesc->getShape(), outMemoryDesc->getShape());
} else {
pAttrLocal = initPrimitiveAttr();
}
const auto in_candidate = inMemoryDesc->getDnnlDesc();
const auto out_candidate = outMemoryDesc->getDnnlDesc();
mkldnn::memory::desc wgh_candidate;
if (isInt8) {
if (internalBlobMemory.empty()) {
wgh_candidate = mkldnn::memory::desc(MKLDNNExtensionUtils::convertToDnnlDims(int8WeightDims), memory::data_type::s8, memory::format_tag::any);
} else {
wgh_candidate = internalBlobMemory.front()->GetDescWithType<DnnlMemoryDesc>()->getDnnlDesc();
}
} else {
wgh_candidate = getParentEdgesAtPort(1).front()->getMemory().GetDescWithType<DnnlMemoryDesc>()->getDnnlDesc();
}
std::shared_ptr<MKLDNNDescriptor> desc;
if (isInt8) {
desc = createInt8MkldnnDeconvDesc(in_candidate, wgh_candidate, out_candidate);
} else {
desc = createDefaultMkldnnDeconvDesc(in_candidate, wgh_candidate, out_candidate,
selected_pd->getImplementationType() == MKLDNNPlugin::impl_desc_type::jit_avx512_winograd);
}
createDeconvPrim(desc, srcMemPtr, getParentEdgesAtPort(1)[0]->getMemoryPtr(), dstMemPtr, pAttrLocal, selected_pd->getImplementationType());
}
void MKLDNNDeconvolutionNode::createPrimitive() {
if (inputShapesDefined()) {
if (needPrepareParams())
prepareParams();
updateLastInputDims();
}
}
MKLDNNDeconvolutionNode::DefaultDeconvDescs MKLDNNDeconvolutionNode::createDescriptorInternalDefault(const mkldnn::memory::desc& in_candidate,
const mkldnn::memory::desc& wgh_candidate,
const mkldnn::memory::desc& out_candidate,
mkldnn::algorithm alg) const {
auto convertDims = [] (const std::vector<ptrdiff_t>& orig_dims) {
return memory::dims(orig_dims.begin(), orig_dims.end());
};
std::shared_ptr<mkldnn::convolution_forward::desc> conv_desc;
conv_desc = std::make_shared<convolution_forward::desc>(prop_kind::forward_inference, alg,
out_candidate, wgh_candidate, in_candidate,
convertDims(stride),
convertDims(dilation),
convertDims(paddingL),
convertDims(paddingR));
std::shared_ptr<mkldnn::convolution_backward_data::desc> deconv_desc;
deconv_desc = std::make_shared<convolution_backward_data::desc>(alg, out_candidate, wgh_candidate,
in_candidate,
convertDims(stride),
convertDims(dilation),
convertDims(paddingL),
convertDims(paddingR));
auto fwd_conv_pd = std::make_shared<convolution_forward::primitive_desc>(*conv_desc, getEngine(), true);
return {deconv_desc, fwd_conv_pd};
}
MKLDNNDeconvolutionNode::Int8DeconvDesc MKLDNNDeconvolutionNode::createDescriptorInternalInt8(const mkldnn::memory::desc& in_candidate,
const mkldnn::memory::desc& wgh_candidate,
const mkldnn::memory::desc& out_candidate) const {
auto convertDims = [] (const std::vector<ptrdiff_t>& orig_dims) {
return memory::dims(orig_dims.begin(), orig_dims.end());
};
MKLDNNDeconvolutionNode::Int8DeconvDesc deconv_desc;
deconv_desc = std::make_shared<mkldnn::deconvolution_forward::desc>(prop_kind::forward_inference, mkldnn::algorithm::deconvolution_direct,
in_candidate, wgh_candidate, out_candidate,
convertDims(stride), convertDims(dilation),
convertDims(paddingL), convertDims(paddingR));
return deconv_desc;
}
void MKLDNNDeconvolutionNode::createDescriptor(const std::vector<MemoryDescPtr> &inputDesc,
const std::vector<MemoryDescPtr> &outputDesc) {
auto inDesc = inputDesc[0]->isDefined() ? inputDesc[0] : inputDesc[0]->cloneWithNewDims(inShape.getStaticDims());
auto dnnlInDesc = MemoryDescUtils::convertToDnnlBlockedMemoryDesc(*inDesc);
auto in_candidate = dnnlInDesc.getDnnlDesc();
auto outDesc = outputDesc[0];
if (!outDesc->isDefined()) {
const auto outShape = shapeInferInternal(inDesc->getShape().getStaticDims(), lastOutputSpatialDims);
outDesc = outDesc->cloneWithNewDims(outShape);
}
auto dnnlOutDesc = MemoryDescUtils::convertToDnnlBlockedMemoryDesc(*outDesc);
auto out_candidate = dnnlOutDesc.getDnnlDesc();
// grouping and autoblocking is not compatible
if ((withGroups && !isDW) && (dnnlInDesc.blocksExtended() || dnnlOutDesc.blocksExtended()))
return;
if (isInt8) {
mkldnn::memory::desc wgh_candidate(MKLDNNExtensionUtils::convertToDnnlDims(weightDims), memory::data_type::s8, memory::format_tag::any);
std::shared_ptr<mkldnn::deconvolution_forward::desc> deconv_desc;
deconv_desc.reset(new deconvolution_forward::desc(prop_kind::forward_inference, mkldnn::algorithm::deconvolution_direct,
in_candidate.getDnnlDesc(), wgh_candidate, out_candidate.getDnnlDesc(),
convertDims(stride), convertDims(dilation),
convertDims(paddingL), convertDims(paddingR)));
descs.emplace_back(deconv_desc);
mkldnn::memory::desc wgh_candidate(MKLDNNExtensionUtils::convertToDnnlDims(int8WeightDims), memory::data_type::s8, memory::format_tag::any);
descs.emplace_back(createDescriptorInternalInt8(in_candidate, wgh_candidate, out_candidate));
} else {
mkldnn::memory::desc wgh_candidate(MKLDNNExtensionUtils::convertToDnnlDims(weightDims), in_candidate.getDataType(), memory::format_tag::any);
mkldnn::memory::desc wgh_candidate(MKLDNNExtensionUtils::convertToDnnlDims(getWeightDims()),
dnnlInDesc.getDataType(), memory::format_tag::any);
for (auto alg : {mkldnn::algorithm::convolution_winograd, mkldnn::algorithm::convolution_direct}) {
std::shared_ptr<mkldnn::convolution_forward::desc> conv_desc;
conv_desc.reset(new convolution_forward::desc(prop_kind::forward_inference, alg,
out_candidate.getDnnlDesc(), wgh_candidate, in_candidate.getDnnlDesc(),
convertDims(stride),
convertDims(dilation),
convertDims(paddingL),
convertDims(paddingR)));
std::shared_ptr<mkldnn::convolution_backward_data::desc> deconv_desc;
deconv_desc.reset(new convolution_backward_data::desc(alg, out_candidate.getDnnlDesc(), wgh_candidate,
in_candidate.getDnnlDesc(),
convertDims(stride),
convertDims(dilation),
convertDims(paddingL),
convertDims(paddingR)));
descs_fwd.push_back(conv_desc);
descs_bwd.push_back(deconv_desc);
auto fwd_conv_pd = std::make_shared<convolution_forward::primitive_desc>(*conv_desc, getEngine(), true);
std::shared_ptr<convolution_backward_data::desc> deconv_desc;
std::shared_ptr<convolution_forward::primitive_desc> fwd_conv_pd;
std::tie(deconv_desc, fwd_conv_pd) = createDescriptorInternalDefault(in_candidate, wgh_candidate, out_candidate, alg);
if (fwd_conv_pd->get(true) == nullptr)
continue;
descs.emplace_back(deconv_desc, fwd_conv_pd);
}
}
@ -420,15 +702,25 @@ void MKLDNNDeconvolutionNode::createDescriptor(const std::vector<MemoryDescPtr>
std::shared_ptr<MemoryDesc> MKLDNNDeconvolutionNode::getSrcMemDesc(mkldnn::primitive_desc_iterator &primitive_desc_it, size_t idx) {
if (idx == 2) {
return std::make_shared<CpuBlockedMemoryDesc>(getOriginalInputPrecisionAtPort(2), Shape(getInputShapeAtPort(2).getStaticDims()));
return std::make_shared<CpuBlockedMemoryDesc>(InferenceEngine::Precision::I32, Shape(getInputShapeAtPort(2).getStaticDims()));
} else if (idx > 0 && isInt8) {
// we need to store 'weight' input as edge,
// because at this moment we can't simple replace internal blob with input, since we need to save weight data as is, but with different order
return std::make_shared<CpuBlockedMemoryDesc>(getOriginalInputPrecisionAtPort(idx), Shape(getInputShapeAtPort(idx).getStaticDims()));
}
auto desc = idx > 0 ? primitive_desc_it.weights_desc(idx - 1) : isInt8 ? primitive_desc_it.src_desc(idx) : primitive_desc_it.diff_dst_desc(idx);
if (getInputShapeAtPort(idx).isDynamic()) {
return MKLDNNExtensionUtils::makeUndefinedDesc(desc, getInputShapeAtPort(idx));
}
return MKLDNNExtensionUtils::makeDescriptor(desc);
}
std::shared_ptr<MemoryDesc> MKLDNNDeconvolutionNode::getDstMemDesc(mkldnn::primitive_desc_iterator &primitive_desc_it, size_t idx) {
auto desc = isInt8 ? primitive_desc_it.dst_desc(idx) : primitive_desc_it.diff_src_desc(idx);
if (getOutputShapeAtPort(idx).isDynamic()) {
return MKLDNNExtensionUtils::makeUndefinedDesc(desc, getOutputShapeAtPort(idx));
}
return MKLDNNExtensionUtils::makeDescriptor(desc);
}
@ -446,4 +738,117 @@ InferenceEngine::Precision MKLDNNDeconvolutionNode::getRuntimePrecision() const
return getMaxPrecision(inputPrecisions);
}
MKLDNNDeconvolutionNode::DeconvExecutor::IntermReorder::IntermReorder(MKLDNNMemoryPtr memFrom,
const mkldnn::memory::desc& descTo,
const mkldnn::engine& engine) : m_memFrom(memFrom) {
m_memTo = std::make_shared<MKLDNNMemory>(engine);
m_memTo->Create(MKLDNNExtensionUtils::makeDescriptor(descTo));
m_reorder = mkldnn::reorder(m_memFrom->GetPrimitive(), m_memTo->GetPrimitive());
}
MKLDNNDeconvolutionNode::DeconvExecutor::IntermReorder::IntermReorder(const mkldnn::memory::desc& descFrom,
MKLDNNMemoryPtr memTo,
const mkldnn::engine& engine) : m_memTo(memTo) {
m_memFrom = std::make_shared<MKLDNNMemory>(engine);
m_memFrom->Create(MKLDNNExtensionUtils::makeDescriptor(descFrom));
m_reorder = mkldnn::reorder(m_memFrom->GetPrimitive(), m_memTo->GetPrimitive());
}
void MKLDNNDeconvolutionNode::DeconvExecutor::IntermReorder::exec(mkldnn::stream strm) {
auto src = m_memFrom->GetPrimitive();
auto dst = m_memTo->GetPrimitive();
m_reorder.execute(strm, src, dst);
}
void MKLDNNDeconvolutionNode::DeconvExecutor::exec(mkldnn::stream strm) {
for (auto &inReorder : inputReorders) {
inReorder.exec(strm);
}
(*execPrim).execute(strm, primArgs);
for (auto &outReorder : outputReorders) {
outReorder.exec(strm);
}
}
MKLDNNDeconvolutionNode::DeconvExecutorDefault::DeconvExecutorDefault(const mkldnn::convolution_backward_data::primitive_desc& pd,
MKLDNNMemoryPtr inMem,
MKLDNNMemoryPtr weightMem,
MKLDNNMemoryPtr outMem,
const mkldnn::primitive_attr &attr,
const std::vector<MKLDNNMemoryPtr>& binPostOpsArgs,
const mkldnn::engine& engine) {
execPrim.reset(new mkldnn::convolution_backward_data(pd));
if (inMem->GetPrimitive().get_desc() != pd.diff_dst_desc()) {
inputReorders.push_back(IntermReorder(inMem, pd.diff_dst_desc(), engine));
primArgs[DNNL_ARG_DIFF_DST] = inputReorders.back().getToMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_DIFF_DST] = inMem->GetPrimitive();
}
if (weightMem->GetPrimitive().get_desc() != pd.weights_desc()) {
inputReorders.push_back(IntermReorder(weightMem, pd.weights_desc(), engine));
primArgs[DNNL_ARG_WEIGHTS] = inputReorders.back().getToMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_WEIGHTS] = weightMem->GetPrimitive();
}
if (outMem->GetPrimitive().get_desc() != pd.diff_src_desc()) {
outputReorders.push_back(IntermReorder(pd.diff_src_desc(), outMem, engine));
primArgs[DNNL_ARG_DIFF_SRC] = outputReorders.back().getFromMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_DIFF_SRC] = outMem->GetPrimitive();
}
MKLDNNNode::appendPostOpArgs(attr, primArgs, binPostOpsArgs);
}
MKLDNNDeconvolutionNode::DeconvExecutorInt8::DeconvExecutorInt8(const mkldnn::deconvolution_forward::primitive_desc& pd,
MKLDNNMemoryPtr inMem,
MKLDNNMemoryPtr weightMem,
MKLDNNMemoryPtr outMem,
const mkldnn::primitive_attr &attr,
const std::vector<MKLDNNMemoryPtr>& binPostOpsArgs,
const mkldnn::engine& engine) {
execPrim.reset(new mkldnn::deconvolution_forward(pd));
if (inMem->GetPrimitive().get_desc() != pd.src_desc()) {
inputReorders.push_back(IntermReorder(inMem, pd.src_desc(), engine));
primArgs[DNNL_ARG_SRC] = inputReorders.back().getToMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_SRC] = inMem->GetPrimitive();
}
if (weightMem->GetPrimitive().get_desc() != pd.weights_desc()) {
inputReorders.push_back(IntermReorder(weightMem, pd.weights_desc(), engine));
primArgs[DNNL_ARG_WEIGHTS] = inputReorders.back().getToMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_WEIGHTS] = weightMem->GetPrimitive();
}
if (outMem->GetPrimitive().get_desc() != pd.dst_desc()) {
outputReorders.push_back(IntermReorder(pd.dst_desc(), outMem, engine));
primArgs[DNNL_ARG_DST] = outputReorders.back().getFromMem()->GetPrimitive();
} else {
primArgs[DNNL_ARG_DST] = outMem->GetPrimitive();
}
MKLDNNNode::appendPostOpArgs(attr, primArgs, binPostOpsArgs);
}
std::vector<int32_t> MKLDNNDeconvolutionNode::readOutputSpatialDims() const {
if (getParentEdges().size() < 3) {
IE_THROW() << "Can't get output spatial dims. Inputs number = " << getParentEdges().size();
}
const auto &shapeMemPtr = getParentEdgesAtPort(2)[0]->getMemoryPtr();
if (!shapeMemPtr || !shapeMemPtr->GetPrimitivePtr()) {
IE_THROW() << "'output_shape' input memory is not allocated.";
}
const auto spDimsNum = getInputShapeAtPort(0).getRank() - 2;
if (shapeMemPtr->getStaticDims()[0] != spDimsNum) {
IE_THROW() << "Can't read output spatial dims, beause 'output_shape' input has incorrect number of elements";
}
const int32_t *outShapePtr = reinterpret_cast<const int32_t *>(shapeMemPtr->GetPtr());
std::vector<int32_t> outSpDims(outShapePtr, outShapePtr + shapeMemPtr->getStaticDims()[0]);
return outSpDims;
}
REG_MKLDNN_PRIM_FOR(MKLDNNDeconvolutionNode, Deconvolution);

113
inference-engine/src/mkldnn_plugin/nodes/mkldnn_deconv_node.h
View File

@ -13,6 +13,10 @@
namespace MKLDNNPlugin {
class MKLDNNDeconvolutionNode : public MKLDNNNode {
using DefaultDeconvDescs = std::pair<std::shared_ptr<mkldnn::convolution_backward_data::desc>,
std::shared_ptr<mkldnn::convolution_forward::primitive_desc>>;
using Int8DeconvDesc = std::shared_ptr<mkldnn::deconvolution_forward::desc>;
public:
MKLDNNDeconvolutionNode(const std::shared_ptr<ngraph::Node>& op, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache);
@ -39,27 +43,120 @@ public:
static bool isSupportedOperation(const std::shared_ptr<const ngraph::Node>& op, std::string& errorMessage) noexcept;
bool canFuse(const MKLDNNNodePtr& node) const override;
const InferenceEngine::SizeVector& getWeightDims() { return weightDims; }
const std::vector<ptrdiff_t>& getStride() { return stride; }
const VectorDims& getWeightDims() const { return getInputShapeAtPort(1).getStaticDims(); }
const std::vector<ptrdiff_t>& getStride() const { return stride; }
void prepareParams() override;
void execute(mkldnn::stream strm) override;
void executeDynamicImpl(mkldnn::stream strm) override { execute(strm); }
bool needShapeInfer() const override;
std::vector<VectorDims> shapeInfer() const override;
private:
class DeconvExecutor {
protected:
class IntermReorder {
public:
IntermReorder(MKLDNNMemoryPtr memFrom, const mkldnn::memory::desc& descTo, const mkldnn::engine& engine);
IntermReorder(const mkldnn::memory::desc& descFrom, MKLDNNMemoryPtr memTo, const mkldnn::engine& engine);
MKLDNNMemoryPtr getFromMem() const { return m_memFrom; }
MKLDNNMemoryPtr getToMem() const { return m_memTo; }
void exec(mkldnn::stream strm);
private:
MKLDNNMemoryPtr m_memFrom;
MKLDNNMemoryPtr m_memTo;
mkldnn::reorder m_reorder;
};
public:
void exec(mkldnn::stream strm);
virtual ~DeconvExecutor() = default;
protected:
DeconvExecutor() = default;
std::vector<IntermReorder> inputReorders;
MKLDNNPrimitive execPrim;
std::vector<IntermReorder> outputReorders;
std::unordered_map<int, mkldnn::memory> primArgs;
};
using executorPtr = std::shared_ptr<DeconvExecutor>;
executorPtr execPtr = nullptr;
class DeconvExecutorDefault : public DeconvExecutor {
public:
DeconvExecutorDefault(const mkldnn::convolution_backward_data::primitive_desc& pd,
MKLDNNMemoryPtr inMem,
MKLDNNMemoryPtr weightMem,
MKLDNNMemoryPtr outMem,
const mkldnn::primitive_attr &attr,
const std::vector<MKLDNNMemoryPtr>& binPostOpsArgs,
const mkldnn::engine& engine);
};
class DeconvExecutorInt8 : public DeconvExecutor {
public:
DeconvExecutorInt8(const mkldnn::deconvolution_forward::primitive_desc& pd,
MKLDNNMemoryPtr inMem,
MKLDNNMemoryPtr weightMem,
MKLDNNMemoryPtr outMem,
const mkldnn::primitive_attr &attr,
const std::vector<MKLDNNMemoryPtr>& binPostOpsArgs,
const mkldnn::engine& engine);
};
bool withGroups = false;
bool isDW = false;
bool isInt8 = false;
bool autoPad = false;
bool externOutShape = false;
size_t groupNum = 1;
size_t IC;
size_t OC;
std::vector<ptrdiff_t> kernel;
std::vector<ptrdiff_t> stride;
std::vector<ptrdiff_t> dilation;
std::vector<ptrdiff_t> paddingL;
std::vector<ptrdiff_t> paddingR;
InferenceEngine::SizeVector weightDims;
std::vector<std::shared_ptr<mkldnn::convolution_forward::desc>> descs_fwd;
std::vector<std::shared_ptr<mkldnn::convolution_backward_data::desc>> descs_bwd;
ov::CoordinateDiff paddingL;
ov::CoordinateDiff paddingR;
ov::CoordinateDiff outputPadding;
std::vector<int32_t> lastOutputSpatialDims;
VectorDims int8WeightDims;
Shape inShape;
AttrPtr pAttr;
mkldnn::primitive_attr attr;
void setPostOps(mkldnn::primitive_attr &attr);
void setPostOps(mkldnn::primitive_attr &attr, const VectorDims &dims);
VectorDims shapeInferInternal(const VectorDims &inDims, std::vector<int32_t> outSpDims) const;
void initPadding(std::shared_ptr<ngraph::Node> op, const Shape &inShape, const std::vector<int32_t>& outSpDims);
void initPaddingR(const Shape &inShape, const Shape &outShape);
std::vector<int32_t> readOutputSpatialDims() const;
std::pair<VectorDims, VectorDims> makeDummyInOutShape();
DefaultDeconvDescs createDescriptorInternalDefault(const mkldnn::memory::desc& in_candidate,
const mkldnn::memory::desc& wgh_candidate,
const mkldnn::memory::desc& out_candidate,
mkldnn::algorithm alg) const;
Int8DeconvDesc createDescriptorInternalInt8(const mkldnn::memory::desc& in_candidate,
const mkldnn::memory::desc& wgh_candidate,
const mkldnn::memory::desc& out_candidate) const;
std::shared_ptr<MKLDNNDescriptor> createDefaultMkldnnDeconvDesc(const mkldnn::memory::desc& srcDesc,
const mkldnn::memory::desc& wghDesc,
const mkldnn::memory::desc& dstDesc,
bool isWinograd) const;
std::shared_ptr<MKLDNNDescriptor> createInt8MkldnnDeconvDesc(const mkldnn::memory::desc& srcDesc,
const mkldnn::memory::desc& wghDesc,
const mkldnn::memory::desc& dstDesc) const;
void createDeconvPrim(std::shared_ptr<MKLDNNDescriptor> desc,
MKLDNNMemoryPtr srcMemPtr,
MKLDNNMemoryPtr wghMemPtr,
MKLDNNMemoryPtr dstMemPtr,
AttrPtr attr,
impl_desc_type selectedImpl);
std::string errorPrefix;

2
inference-engine/src/mkldnn_plugin/nodes/mkldnn_fullyconnected_node.cpp
View File

@ -147,7 +147,7 @@ void MKLDNNFullyConnectedNode::createPrimitive() {
else
primArgs = {{DNNL_ARG_SRC, src}, {DNNL_ARG_WEIGHTS, getParentEdgeAt(WEIGHTS_ID)->getMemory().GetPrimitive()}, {DNNL_ARG_DST, dst}};
appendPostOpArgs(*attr);
appendPostOpArgs(*attr, primArgs, binaryPostOpsArgs);
}
void MKLDNNFullyConnectedNode::execute(mkldnn::stream strm) {

2
inference-engine/src/mkldnn_plugin/nodes/mkldnn_matmul_node.cpp
View File

@ -421,7 +421,7 @@ void MKLDNNMatMulNode::prepareParams() {
if (withBiases)
primArgs[DNNL_ARG_BIAS] = getParentEdgeAt(2)->getMemoryPtr()->GetPrimitive();
appendPostOpArgs(*attr);
appendPostOpArgs(*attr, primArgs, binaryPostOpsArgs);
}
void MKLDNNMatMulNode::executeDynamicImpl(dnnl::stream strm) {

2
inference-engine/thirdparty/clDNN/api/intel_gpu/runtime/tensor.hpp vendored
View File

@ -85,6 +85,7 @@ struct format {
bs_fs_zyx_bsv16_fsv16, ///< format used for 3D blocked convolution (batch and features blocked by 16)
bs_fs_yx_bsv16_fsv16, ///< format used for 2D blocked convolution (batch and features blocked by 16)
bs_fs_yx_bsv4_fsv4, ///< format used for 2D blocked convolution (batch and features blocked by 4)
bs_fs_yx_bsv8_fsv4, ///< format used for 2D blocked convolution (batch and features blocked by 8 and 4)
bs_fs_yx_bsv4_fsv2, ///< format used for 2D blocked convolution (batch blocked by 4, features blocked by 2)
bs_fs_zyx_bsv4_fsv4, ///< format used for 3D blocked convolution (batch and features blocked by 4)
bs_fs_zyx_bsv4_fsv2, ///< format used for 3D blocked convolution (batch blocked by 4, features blocked by 2)
@ -255,6 +256,7 @@ struct format {
{ bs_fs_zyx_bsv16_fsv16, { 1, 1, 3, 0, "bfzyx", "bfxyz", {{0, 16 }, {1, 16}}}},
{ bs_fs_yx_bsv16_fsv16, { 1, 1, 2, 0, "bfyx", "bfxy?", {{0, 16 }, {1, 16}}}},
{ bs_fs_yx_bsv4_fsv4, { 1, 1, 2, 0, "bfyx", "bfxy?", {{0, 4 }, {1, 4}}}},
{ bs_fs_yx_bsv8_fsv4, { 1, 1, 2, 0, "bfyx", "bfxy?", {{0, 8 }, {1, 4}}}},
{ bs_fs_yx_bsv4_fsv2, { 1, 1, 2, 0, "bfyx", "bfxy?", {{0, 4 }, {1, 2}}}},
{ bs_fs_zyx_bsv4_fsv4, { 1, 1, 3, 0, "bfzyx", "bfxyz", {{0, 4 }, {1, 4}}}},
{ bs_fs_zyx_bsv4_fsv2, { 1, 1, 3, 0, "bfzyx", "bfxyz", {{0, 4 }, {1, 2}}}},

6
inference-engine/thirdparty/clDNN/kernel_selector/common/tensor_type.cpp vendored
View File

@ -29,6 +29,7 @@ DataTensor::DataChannelArray DataTensor::dataChannelArray {{
{ DataLayout::bs_fs_zyx_bsv16_fsv16, { 0, 1, 2, -1, 3, 4 } },
{ DataLayout::bs_fs_yx_bsv16_fsv16, { 0, 1, -1, -1, 2, 3 } },
{ DataLayout::bs_fs_yx_bsv4_fsv4, { 0, 1, -1, -1, 2, 3 } },
{ DataLayout::bs_fs_yx_bsv8_fsv4, { 0, 1, -1, -1, 2, 3 } },
{ DataLayout::bs_fs_yx_bsv4_fsv2, { 0, 1, -1, -1, 2, 3 } },
{ DataLayout::bs_fs_yx_bsv32_fsv32, { 0, 1, -1, -1, 2, 3 } },
{ DataLayout::bs_fs_yx_bsv32_fsv16, { 0, 1, -1, -1, 2, 3 } },
@ -206,6 +207,11 @@ NDims DataTensor::GetSimpleDims(const std::vector<size_t>& d, DataLayout l) {
newDims[2] = RoundUp(newDims[2], 4);
newDims[3] = RoundUp(newDims[3], 4);
break;
case bs_fs_yx_bsv8_fsv4:
assert(newDims.size() == 4);
newDims[2] = RoundUp(newDims[2], 4);
newDims[3] = RoundUp(newDims[3], 8);
break;
case bs_fs_yx_bsv4_fsv2:
assert(newDims.size() == 4);
newDims[2] = RoundUp(newDims[2], 2);

1
inference-engine/thirdparty/clDNN/kernel_selector/common/tensor_type.h vendored
View File

@ -39,6 +39,7 @@ enum DataLayout {
bs_fs_yx_bsv16_fsv16, // batch, feature, 2D spatial. Blocks of 16 batch and channels
bs_fs_zyx_bsv16_fsv16, // batch, feature, 3D spatial. Blocks of 16 batch and channels
bs_fs_yx_bsv4_fsv4, // batch, feature, 2D spatial. Blocks of 4 batch and 4 channels
bs_fs_yx_bsv8_fsv4, // batch, feature, 2D spatial. Blocks of 8 batch and 4 channels
bs_fs_yx_bsv4_fsv2, // batch, feature, 2D spatial. Blocks of 4 batch and 2 channels
bs_fs_yx_bsv32_fsv32, // batch, feature, 2D spatial. Blocks of 32 batch and 32 channels
bs_fs_yx_bsv32_fsv16, // batch, feature, 2D spatial. Blocks of 32 batch and 16 channels

3
inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convert_color/convert_color_kernel_base.cpp vendored
View File

@ -43,6 +43,9 @@ JitConstants ConvertColorKernelBase::GetJitConstants(const convert_color_params&
case color_format::NV12:
jit.AddConstant(MakeJitConstant("CONVERT_FROM_NV12", ""));
break;
case color_format::I420:
jit.AddConstant(MakeJitConstant("CONVERT_FROM_I420", ""));
break;
default:
IE_THROW() << "Not supported input color format";
}

5
inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_base.cpp vendored
View File

@ -196,15 +196,16 @@ KernelsData ConvolutionKernelBase::GetCommonKernelsData(const Params& params,
return {};
}
auto preferredWeightsLayout = GetPreferredWeightsLayout(newParams);
bool succeed = UpdateWeightsParams(newParams,
options,
GetPreferredWeightsLayout(newParams),
preferredWeightsLayout,
kd.weightsReorderParams,
GetSupportedKey(),
newParams.groups,
newParams.transposed);
bool bSupportedWeightsLayout = newParams.weights.GetLayout() == GetPreferredWeightsLayout(newParams);
bool bSupportedWeightsLayout = newParams.weights.GetLayout() == preferredWeightsLayout;
const bool bWeightsOK = bSupportedWeightsLayout || options.allowStaticInputReordering;
if (!succeed || !bWeightsOK) {

21
inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/gemm/gemm_kernel_tiled_opt.cpp vendored
View File

@ -13,6 +13,8 @@ ParamsKey GemmKernelTiledOpt::GetSupportedKey() const {
k.EnableInputDataType(Datatype::F32);
k.EnableOutputDataType(Datatype::F16);
k.EnableOutputDataType(Datatype::F32);
k.EnableOutputDataType(Datatype::INT8);
k.EnableOutputDataType(Datatype::UINT8);
k.EnableInputLayout(DataLayout::bfyx);
k.EnableOutputLayout(DataLayout::bfyx);
k.EnableInputLayout(DataLayout::bfzyx);
@ -21,6 +23,7 @@ ParamsKey GemmKernelTiledOpt::GetSupportedKey() const {
k.EnableOutputLayout(DataLayout::bfwzyx);
k.EnableBatching();
k.EnableDifferentTypes();
return k;
}
@ -117,25 +120,29 @@ JitConstants GemmKernelTiledOpt::GetJitConstants(const gemm_params& params) cons
if (tuning_data.tile_k_size > tuning_data.simd_size) {
jit.AddConstants({
MakeJitConstant("A_VEC_SIZE", tuning_data.tile_k_size / tuning_data.simd_size),
MakeJitConstant("A_FLOATN", std::string("UNIT_TYPE") + toCodeString(tuning_data.tile_k_size / tuning_data.simd_size)),
MakeJitConstant("A_FLOATN", std::string("CAT(INPUT0_TYPE, ") + toCodeString(tuning_data.tile_k_size / tuning_data.simd_size) + ")"),
});
} else {
jit.AddConstants({
MakeJitConstant("A_VEC_SIZE", 1),
MakeJitConstant("A_FLOATN", std::string("UNIT_TYPE")),
MakeJitConstant("A_FLOATN", std::string("INPUT0_TYPE")),
});
}
if (tuning_data.tile_n_size > tuning_data.simd_size) {
jit.AddConstants({
MakeJitConstant("B_VEC_SIZE", b_vec_size),
MakeJitConstant("B_FLOATN", std::string("UNIT_TYPE") + toCodeString(b_vec_size)),
MakeJitConstant("B_FLOATN", std::string("CAT(INPUT1_TYPE, ") + toCodeString(b_vec_size) + ")"),
MakeJitConstant("OUTPUT_TYPE_VEC", std::string("CAT(OUTPUT_TYPE, ") + toCodeString(b_vec_size) + ")"),
MakeJitConstant("ACCUMULATOR_TYPE_VEC", std::string("CAT(ACCUMULATOR_TYPE, ") + toCodeString(b_vec_size) + ")"),
});
} else {
b_vec_size = 1;
jit.AddConstants({
MakeJitConstant("B_VEC_SIZE", 1),
MakeJitConstant("B_FLOATN", std::string("UNIT_TYPE")),
MakeJitConstant("B_VEC_SIZE", b_vec_size),
MakeJitConstant("B_FLOATN", std::string("INPUT1_TYPE")),
MakeJitConstant("OUTPUT_TYPE_VEC", std::string("OUTPUT_TYPE")),
MakeJitConstant("ACCUMULATOR_TYPE_VEC", std::string("ACCUMULATOR_TYPE")),
});
}
@ -183,6 +190,10 @@ bool GemmKernelTiledOpt::Validate(const Params& params, const optional_params& o
if ((gmm_params.transpose_input0 || gmm_params.transpose_input1) && gemm_leftovers)
return false;
for (size_t i = 1; i < gmm_params.inputs.size(); i++)
if (gmm_params.inputs[0].GetDType() != gmm_params.inputs[i].GetDType())
return false;
return true;
}
} // namespace kernel_selector

50
inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/convert_color_ref.cl vendored
View File

@ -5,11 +5,14 @@
#include "include/batch_headers/fetch_data.cl"
#include "include/batch_headers/data_types.cl"
#ifdef CONVERT_FROM_NV12
#if defined(CONVERT_FROM_NV12) || defined(CONVERT_FROM_I420)
#ifdef BUFFER_MEM
KERNEL(convert_color_ref)(const __global INPUT0_TYPE* input_y,
#if INPUTS_COUNT == 2
const __global INPUT1_TYPE* input_uv,
KERNEL(convert_color_ref)(const __global INPUT0_TYPE* input1,
#if INPUTS_COUNT > 1
const __global INPUT1_TYPE* input2,
#if INPUTS_COUNT == 3
const __global INPUT2_TYPE* input3,
#endif
#endif
__global OUTPUT_TYPE* output) {
@ -17,16 +20,19 @@ KERNEL(convert_color_ref)(const __global INPUT0_TYPE* input_y,
const uint y = get_global_id(1);
const uint x = get_global_id(2);
float Y = input_y[GET_DATA_INDEX(INPUT0, b, 0, y, x)];
float Y = input1[GET_DATA_INDEX(INPUT0, b, 0, y, x)];
#if INPUTS_COUNT == 2
float U = input_uv[GET_DATA_INDEX(INPUT1, b, 0, y / 2, x / 2)];
float V = input_uv[GET_DATA_INDEX(INPUT1, b, 1, y / 2, x / 2)];
#if INPUTS_COUNT == 3
float U = input2[GET_DATA_INDEX(INPUT1, b, 0, y / 2, x / 2)];
float V = input3[GET_DATA_INDEX(INPUT2, b, 0, y / 2, x / 2)];
#elif INPUTS_COUNT == 2
float U = input2[GET_DATA_INDEX(INPUT1, b, 0, y / 2, x / 2)];
float V = input2[GET_DATA_INDEX(INPUT1, b, 1, y / 2, x / 2)];
#else // Single plane
uint input_uv_offset = INPUT0_SIZE_X * INPUT0_SIZE_Y / 3 * 2;
float U = input_y[GET_DATA_INDEX(INPUT0, b, 0, y / 2, (x / 2) * 2) + input_uv_offset];
float V = input_y[GET_DATA_INDEX(INPUT0, b, 1, y / 2, (x / 2) * 2) + input_uv_offset];
float U = input1[GET_DATA_INDEX(INPUT0, b, 0, y / 2, (x / 2) * 2) + input_uv_offset];
float V = input1[GET_DATA_INDEX(INPUT0, b, 1, y / 2, (x / 2) * 2) + input_uv_offset];
#endif
float Ycomponent = mad(Y, 1.164f, -18.624f);
@ -57,9 +63,12 @@ KERNEL(convert_color_ref)(const __global INPUT0_TYPE* input_y,
#ifdef SURFACE_MEM
KERNEL(convert_color_ref)(read_only image2d_t input_y,
#if INPUTS_COUNT == 2
read_only image2d_t input_uv,
KERNEL(convert_color_ref)(read_only image2d_t input1,
#if INPUTS_COUNT > 1
read_only image2d_t input2,
#if INPUTS_COUNT == 3
read_only image2d_t input3,
#endif
#endif
__global OUTPUT_TYPE* output) {
@ -67,17 +76,22 @@ KERNEL(convert_color_ref)(read_only image2d_t input_y,
const uint y = get_global_id(1);
const uint x = get_global_id(2);
float4 Y = read_imagef(input_y, (int2)(x, y));
float4 Y = read_imagef(input1, (int2)(x, y));
float Ycomponent = mad(Y.x, 296.82f, -18.624f);
#if INPUTS_COUNT == 2
float4 UV = read_imagef(input_uv, (int2)(x / 2, y / 2));
#if INPUTS_COUNT == 3
float4 U = read_imagef(input2, (int2)(x / 2, y / 2));
float4 V = read_imagef(input3, (int2)(x / 2, y / 2));
float Ucomponent = mad(U.x, 255.0f, -128.f);
float Vcomponent = mad(V.x, 255.0f, -128.f);
#elif INPUTS_COUNT == 2
float4 UV = read_imagef(input2, (int2)(x / 2, y / 2));
float Ucomponent = mad(UV.x, 255.0f, -128.f);
float Vcomponent = mad(UV.y, 255.0f, -128.f);
#else // Single plane
uint input_y_offset = INPUT0_SIZE_Y / 3 * 2;
float4 U = read_imagef(input_y, (int2)((x / 2) * 2, y / 2 + input_y_offset));
float4 V = read_imagef(input_y, (int2)((x / 2) * 2 + 1, y / 2 + input_y_offset));
float4 U = read_imagef(input1, (int2)((x / 2) * 2, y / 2 + input_y_offset));
float4 V = read_imagef(input1, (int2)((x / 2) * 2 + 1, y / 2 + input_y_offset));
float Ucomponent = mad(U.x, 255.0f, -128.f);
float Vcomponent = mad(V.x, 255.0f, -128.f);
#endif

24
inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/gemm_tiled_opt.cl vendored
View File

@ -3,7 +3,7 @@
//
#include "include/batch_headers/fetch_data.cl"
#include "include/unit_type.cl"
#include "include/batch_headers/data_types.cl"
#define unroll_for __attribute__((opencl_unroll_hint)) for
@ -14,17 +14,17 @@
#endif // INPUT0_TYPE_SIZE == 4
#if TILE_K > SIMD_WIDTH
#define BLOCK_READ_A(ptr, offset) CAT(UNIT_BLOCK_READ, A_VEC_SIZE)(ptr, offset)
#define BLOCK_READ_A(ptr, offset) BLOCK_READN(INPUT0_TYPE, A_VEC_SIZE, ptr, offset)
#else // TILE_K > SIMD_WIDTH
#define BLOCK_READ_A(ptr, offset) UNIT_BLOCK_READ(ptr, offset)
#define BLOCK_READ_A(ptr, offset) BLOCK_READN(INPUT0_TYPE, 1, ptr, offset)
#endif // TILE_K > SIMD_WIDTH
#if TILE_N > SIMD_WIDTH
#define BLOCK_READ_B(ptr, offset) CAT(UNIT_BLOCK_READ, B_VEC_SIZE)(ptr, offset)
#define BLOCK_WRITE_C(ptr, offset, data) CAT(UNIT_BLOCK_WRITE, B_VEC_SIZE)(ptr, offset, data)
#define BLOCK_READ_B(ptr, offset) BLOCK_READN(INPUT1_TYPE, B_VEC_SIZE, ptr, offset)
#define BLOCK_WRITE_C(ptr, offset, data) BLOCK_WRITEN(OUTPUT_TYPE, B_VEC_SIZE, ptr, offset, data)
#else // TILE_N > SIMD_WIDTH
#define BLOCK_READ_B(ptr, offset) UNIT_BLOCK_READ(ptr, offset)
#define BLOCK_WRITE_C(ptr, offset, data) UNIT_BLOCK_WRITE(ptr, offset, data)
#define BLOCK_READ_B(ptr, offset) BLOCK_READN(INPUT1_TYPE, 1, ptr, offset)
#define BLOCK_WRITE_C(ptr, offset, data) BLOCK_WRITEN(OUTPUT_TYPE, 1, ptr, offset, data)
#endif // TILE_N > SIMD_WIDTH
inline uint FUNC(get_input0_batch_offset)(uint b, uint f, uint w, uint z) {
@ -294,9 +294,9 @@ KERNEL(gemm_tiled_opt)(
#if TILE_N_NOT_DIVISIBLE
if (b_raw_global_id < N) {
#ifdef INPUT2_TYPE
OUTPUT_TYPE dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id] + TO_ACCUMULATOR_TYPE(BETA) * c_ptr[sglid];
ACCUMULATOR_TYPE dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id] + TO_ACCUMULATOR_TYPE(BETA) * c_ptr[sglid];
#else // INPUT2_TYPE
OUTPUT_TYPE dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id];
ACCUMULATOR_TYPE dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id];
#endif // INPUT2_TYPE
#if HAS_FUSED_OPS
@ -316,9 +316,9 @@ KERNEL(gemm_tiled_opt)(
#ifdef INPUT2_TYPE
B_FLOATN c_val = BLOCK_READ_B(c_ptr, 0);
B_FLOATN dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id] + TO_ACCUMULATOR_TYPE(BETA) * c_val;
ACCUMULATOR_TYPE_VEC dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id] + TO_ACCUMULATOR_TYPE(BETA) * c_val;
#else // INPUT2_TYPE
B_FLOATN dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id];
ACCUMULATOR_TYPE_VEC dequantized = TO_ACCUMULATOR_TYPE(ALPHA) * c_tile[write_id];
#endif // INPUT2_TYPE
#if HAS_FUSED_OPS
@ -327,7 +327,7 @@ KERNEL(gemm_tiled_opt)(
#else // FUSED_OPS_CAN_USE_PRELOAD
FUSED_OPS_VEC;
#endif // FUSED_OPS_CAN_USE_PRELOAD
B_FLOATN res = FUSED_OPS_RESULT_VEC;
OUTPUT_TYPE_VEC res = FUSED_OPS_RESULT_VEC;
BLOCK_WRITE_C(d_ptr, 0, res);
#else // HAS_FUSED_OPS
BLOCK_WRITE_C(d_ptr, 0, dequantized);

33
inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/include/batch_headers/fetch_data.cl vendored
View File

@ -506,6 +506,22 @@ inline uint get_bs_fs_zyx_bsv_fsv_index(uint b, uint f, uint z, uint y, uint x,
CAT(prefix, _PAD_BEFORE_SIZE_X), \
CAT(prefix, _PAD_AFTER_SIZE_X), 4, 4)
#define GET_DATA_BS_FS_YX_BSV8_FSV4_INDEX(prefix, b, f, y, x) \
get_bs_fs_zyx_bsv_fsv_index( \
b, f, 0, y, x, \
CAT(prefix, _SIZE_X), \
CAT(prefix, _SIZE_Y), \
CAT(prefix, _SIZE_Z), \
CAT(prefix, _FEATURE_NUM), \
CAT(prefix, _PAD_BEFORE_FEATURE_NUM), \
CAT(prefix, _PAD_AFTER_FEATURE_NUM), \
CAT(prefix, _PAD_BEFORE_SIZE_Z), \
CAT(prefix, _PAD_AFTER_SIZE_Z), \
CAT(prefix, _PAD_BEFORE_SIZE_Y), \
CAT(prefix, _PAD_AFTER_SIZE_Y), \
CAT(prefix, _PAD_BEFORE_SIZE_X), \
CAT(prefix, _PAD_AFTER_SIZE_X), 8, 4)
#define GET_DATA_BS_FS_YX_BSV4_FSV2_INDEX(prefix, b, f, y, x) \
get_bs_fs_zyx_bsv_fsv_index( \
b, f, 0, y, x, \
@ -605,6 +621,23 @@ inline uint get_bs_fs_zyx_bsv_fsv_index(uint b, uint f, uint z, uint y, uint x,
CAT(prefix, _PAD_BEFORE_SIZE_X), \
CAT(prefix, _PAD_AFTER_SIZE_X), 4, 4)
#define GET_DATA_BS_FS_YX_BSV8_FSV4_INDEX_SAFE(prefix, b, f, y, x) \
get_bs_fs_zyx_bsv_fsv_index_safe( \
b, f, 0, y, x, \
CAT(prefix, _SIZE_X), \
CAT(prefix, _SIZE_Y), \
CAT(prefix, _SIZE_Z), \
CAT(prefix, _FEATURE_NUM), \
CAT(prefix, _BATCH_NUM), \
CAT(prefix, _PAD_BEFORE_FEATURE_NUM), \
CAT(prefix, _PAD_AFTER_FEATURE_NUM), \
CAT(prefix, _PAD_BEFORE_SIZE_Z), \
CAT(prefix, _PAD_AFTER_SIZE_Z), \
CAT(prefix, _PAD_BEFORE_SIZE_Y), \
CAT(prefix, _PAD_AFTER_SIZE_Y), \
CAT(prefix, _PAD_BEFORE_SIZE_X), \
CAT(prefix, _PAD_AFTER_SIZE_X), 8, 4)
#define GET_DATA_BS_FS_YX_BSV4_FSV2_INDEX_SAFE(prefix, b, f, y, x) \
get_bs_fs_zyx_bsv_fsv_index_safe( \
b, f, 0, y, x, \

2
inference-engine/thirdparty/clDNN/kernel_selector/core/common/jitter.cpp vendored
View File

@ -334,6 +334,7 @@ JitDefinitions DataTensorJitConstant::GetDefinitions() const {
layout == DataLayout::fs_b_yx_fsv32 ||
layout == DataLayout::bs_fs_yx_bsv16_fsv16 ||
layout == DataLayout::bs_fs_yx_bsv4_fsv4 ||
layout == DataLayout::bs_fs_yx_bsv8_fsv4 ||
layout == DataLayout::bs_fs_yx_bsv4_fsv2 ||
layout == DataLayout::bs_fs_yx_bsv32_fsv16 ||
layout == DataLayout::bs_fs_yx_bsv32_fsv32) {
@ -346,6 +347,7 @@ JitDefinitions DataTensorJitConstant::GetDefinitions() const {
layout == DataLayout::bs_fs_yx_bsv32_fsv32 ||
layout == DataLayout::bs_fs_yx_bsv32_fsv16 ||
layout == DataLayout::bs_fs_yx_bsv4_fsv4 ||
layout == DataLayout::bs_fs_yx_bsv8_fsv4 ||
layout == DataLayout::bs_fs_yx_bsv4_fsv2 ||
layout == DataLayout::bs_fs_yx_bsv16_fsv16)
safe_index_func_val = "GET_DATA_" + layout_str + "_INDEX_SAFE(" + _name + ", b, f, y, x)";

1
inference-engine/thirdparty/clDNN/kernel_selector/core/kernel_selector_common.cpp vendored
View File

@ -105,6 +105,7 @@ std::string toString(DataLayout l) {
case kernel_selector::DataLayout::bs_fs_yx_bsv16_fsv16: return "BS_FS_YX_BSV16_FSV16";
case kernel_selector::DataLayout::bs_fs_zyx_bsv16_fsv16: return "BS_FS_ZYX_BSV16_FSV16";
case kernel_selector::DataLayout::bs_fs_yx_bsv4_fsv4: return "BS_FS_YX_BSV4_FSV4";
case kernel_selector::DataLayout::bs_fs_yx_bsv8_fsv4: return "BS_FS_YX_BSV8_FSV4";
case kernel_selector::DataLayout::bs_fs_yx_bsv4_fsv2: return "BS_FS_YX_BSV4_FSV2";
case kernel_selector::DataLayout::bs_fs_yx_bsv32_fsv32: return "BS_FS_YX_BSV32_FSV32";
case kernel_selector::DataLayout::bs_fs_yx_bsv32_fsv16: return "BS_FS_YX_BSV32_FSV16";

2
inference-engine/thirdparty/clDNN/src/binary_convolution.cpp vendored
View File

@ -125,7 +125,7 @@ binary_convolution_inst::typed_primitive_inst(network& network, binary_convoluti
"Only one-dimensional batch size are supported");
CLDNN_ERROR_LESS_THAN(node.id(),
"Weights feature maps number",
(input_inst.size.feature[0] + pad.feature[0]) / split,
input_inst.size.feature[0],
"input feature maps number",
filter_inst.size.feature[0],
"Weights/ifm mismatch");

13
inference-engine/thirdparty/clDNN/src/convolution.cpp vendored
View File

@ -97,7 +97,7 @@ layout convolution_inst::calc_output_layout(convolution_node const& node) {
input_layout.format == format::image_2d_weights_winograd_6x3_s1_xfbyb)
CLDNN_ERROR_MESSAGE(
node.id(),
"Input for convolution should not be in windograd weights format - it is reserved for weights only");
"Input for convolution should not be in winograd weights format - it is reserved for weights only");
if (input_layout.format == format::winograd_2x3_s1_data) {
CLDNN_ERROR_NOT_EQUAL(node.id(),
@ -369,10 +369,19 @@ convolution_inst::typed_primitive_inst(network& network, convolution_node const&
"Only one-dimensional batch size are supported");
CLDNN_ERROR_LESS_THAN(node.id(),
"Weights feature maps number",
(input_inst.size.feature[0] + pad.feature[0]) / split,
input_inst.size.feature[0],
"input feature maps number",
weights_ifm,
"Weights/ifm mismatch");
if (!argument.grouped_weights_shape && !format::is_grouped(filter_inst.format)) {
CLDNN_ERROR_NOT_EQUAL(node.id(),
"Weights feature maps number",
input_inst.size.feature[0],
"input feature maps number",
weights_ifm,
"Weights/ifm mismatch");
}
}
}
} // namespace cldnn

24
inference-engine/thirdparty/clDNN/src/deconvolution.cpp vendored
View File

@ -82,11 +82,11 @@ layout deconvolution_inst::calc_output_layout(deconvolution_node const& node) {
int32_t off_factor = -2;
size_t spatial_dims = cldnn::format::traits(input_layout.format).spatial_num;
CLDNN_ERROR_GREATER_THAN(node.id(),
"number of spatial dimensions",
spatial_dims,
"expected number of dimensions",
3,
"As for now, deconvolutions with more than 3 dimensions are not supported");
"number of spatial dimensions",
spatial_dims,
"expected number of dimensions",
3,
"As for now, deconvolutions with more than 3 dimensions are not supported");
int32_t x = off_factor * pad.spatial[0] + (input_layout.size.spatial[0] - 1) * strd.spatial[0] + filter_size.spatial[0];
int32_t y = 1;
@ -208,6 +208,7 @@ deconvolution_inst::typed_primitive_inst(network& network, deconvolution_node co
1,
"Spatial[0] of bias should be 1. Bias isn't 1D vector.");
}
CLDNN_ERROR_NOT_EQUAL(node.id(),
"deconvolution padding filling value",
node.get_output_layout().data_padding.filling_value(),
@ -240,10 +241,19 @@ deconvolution_inst::typed_primitive_inst(network& network, deconvolution_node co
"Only one-dimensional features are supported");
CLDNN_ERROR_LESS_THAN(node.id(),
"Weights feature maps number",
(input_inst.size.feature[0] + pad.feature[0]) / split,
input_inst.size.feature[0],
"input feature maps number",
weights_ifm,
"Weights/ifm mimsmatch");
"Weights/ifm mismatch");
if (!argument.grouped_weights_shape && !format::is_grouped(filter_inst.format)) {
CLDNN_ERROR_NOT_EQUAL(node.id(),
"Weights feature maps number",
input_inst.size.feature[0],
"input feature maps number",
weights_ifm,
"Weights/ifm mismatch");
}
}
}
} // namespace cldnn

33
inference-engine/thirdparty/clDNN/src/graph_optimizer/reorder_inputs.cpp vendored
View File

@ -536,7 +536,7 @@ void reorder_inputs::run(program& p, layout_optimizer& lo, reorder_factory& rf)
}
};
const auto reorder_input_deconvolution = [&p, &lo, &rf](typed_program_node<deconvolution>& deconv_node) {
const auto reorder_input_and_weights_deconvolution = [&p, &lo, &rf](typed_program_node<deconvolution>& deconv_node) {
auto& input = deconv_node.input();
auto input_layout = input.get_output_layout();
auto new_format = lo.get_preferred_format(deconv_node);
@ -547,14 +547,41 @@ void reorder_inputs::run(program& p, layout_optimizer& lo, reorder_factory& rf)
p.add_intermediate(reorder.first, deconv_node, 0, !reorder.second);
}
}
auto& weights = deconv_node.weights();
auto weights_layout = weights.get_output_layout();
if (!format::is_simple_data_format(weights_layout.format) && !weights.is_type<data>() && !weights.is_constant()) {
auto dims = weights_layout.format.dimension();
auto preferred_format = dims <= 4 ? format::bfyx : dims == 5 ? format::bfzyx : format::bfwzyx;
auto reorder = rf.get_reorder(weights.id(), weights_layout,
layout{ weights_layout.data_type, preferred_format, weights_layout.size });
if (reorder.first) {
p.add_intermediate(reorder.first, deconv_node, 1, !reorder.second);
}
}
};
const auto reorder_weights_convolution = [&p, &lo, &rf](typed_program_node<convolution>& conv_node) {
auto& weights = conv_node.weights();
auto weights_layout = weights.get_output_layout();
if (!format::is_simple_data_format(weights_layout.format) && !weights.is_type<data>() && !weights.is_constant()) {
auto dims = weights_layout.format.dimension();
auto preferred_format = dims <= 4 ? format::bfyx : dims == 5 ? format::bfzyx : format::bfwzyx;
auto reorder = rf.get_reorder(weights.id(), weights_layout,
layout{ weights_layout.data_type, preferred_format, weights_layout.size });
if (reorder.first) {
p.add_intermediate(reorder.first, conv_node, 1, !reorder.second);
}
}
};
for (auto& prim : p.get_processing_order()) {
program_helpers::do_for_types<detection_output, binary_convolution, deconvolution>(
program_helpers::do_for_types<detection_output, binary_convolution, deconvolution, convolution>(
*prim,
reorder_input_detection_output,
reorder_input_binary_convolution,
reorder_input_deconvolution);
reorder_input_and_weights_deconvolution,
reorder_weights_convolution);
}
for (auto n : p.get_processing_order()) {

5
inference-engine/thirdparty/clDNN/src/impls/ocl/convolution.cpp vendored
View File

@ -225,6 +225,11 @@ attach_convolution_impl::attach_convolution_impl() {
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv2),

7
inference-engine/thirdparty/clDNN/src/impls/ocl/eltwise.cpp vendored
View File

@ -214,6 +214,13 @@ attach_eltwise_impl::attach_eltwise_impl() {
std::make_tuple(data_types::i32, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::i64, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i64, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv4_fsv2),

5
inference-engine/thirdparty/clDNN/src/impls/onednn/concatenation_onednn.cpp vendored
View File

@ -119,6 +119,11 @@ attach_concatenation_onednn::attach_concatenation_onednn() {
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv8_fsv4),
});
}

5
inference-engine/thirdparty/clDNN/src/impls/onednn/convolution_onednn.cpp vendored
View File

@ -256,6 +256,11 @@ attach_convolution_onednn::attach_convolution_onednn() {
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv2),

5
inference-engine/thirdparty/clDNN/src/impls/onednn/deconvolution_onednn.cpp vendored
View File

@ -199,6 +199,11 @@ attach_deconvolution_onednn::attach_deconvolution_onednn() {
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv4_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::i8, format::bs_fs_yx_bsv8_fsv4),
std::make_tuple(data_types::f32, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::f16, format::bs_fs_yx_bsv4_fsv2),
std::make_tuple(data_types::u8, format::bs_fs_yx_bsv4_fsv2),

1
inference-engine/thirdparty/clDNN/src/impls/onednn/utils.cpp vendored
View File

@ -91,6 +91,7 @@ dnnl::memory::format_tag convert_data_format(cldnn::format fmt) {
case cldnn::format::bs_fs_yx_bsv16_fsv16: return dnnl::memory::format_tag::NChw16n16c;
case cldnn::format::bs_fs_yx_bsv32_fsv32: return dnnl::memory::format_tag::NChw32n32c;
case cldnn::format::bs_fs_yx_bsv4_fsv4: return dnnl::memory::format_tag::ABcd4a4b;
case cldnn::format::bs_fs_yx_bsv8_fsv4: return dnnl::memory::format_tag::ABcd8a4b;
case cldnn::format::bs_fs_yx_bsv4_fsv2: return dnnl::memory::format_tag::ABcd4a2b;
case cldnn::format::bs_fs_yx_bsv32_fsv16: return dnnl::memory::format_tag::NChw32n16c;
case cldnn::format::bs_fs_zyx_bsv16_fsv16: return dnnl::memory::format_tag::NCdhw16n16c;

2
inference-engine/thirdparty/clDNN/src/include/to_string_utils.h vendored
View File

@ -97,6 +97,8 @@ inline std::string fmt_to_str(format fmt) {
return "bs_fs_yx_bsv4_fsv2";
case format::bs_fs_yx_bsv4_fsv4:
return "bs_fs_yx_bsv4_fsv4";
case format::bs_fs_yx_bsv8_fsv4:
return "bs_fs_yx_bsv8_fsv4";
case format::bs_fs_yx_bsv32_fsv32:
return "bs_fs_yx_bsv32_fsv32";
case format::b_fs_zyx_fsv16:

4
inference-engine/thirdparty/clDNN/src/kernel_selector_helper.cpp vendored
View File

@ -136,6 +136,8 @@ kernel_selector::data_layout to_data_layout(format f) {
return kernel_selector::data_layout::bs_fs_yx_bsv32_fsv16;
case format::bs_fs_yx_bsv4_fsv4:
return kernel_selector::data_layout::bs_fs_yx_bsv4_fsv4;
case format::bs_fs_yx_bsv8_fsv4:
return kernel_selector::data_layout::bs_fs_yx_bsv8_fsv4;
case format::bs_fs_yx_bsv4_fsv2:
return kernel_selector::data_layout::bs_fs_yx_bsv4_fsv2;
case format::bs_fs_yx_bsv32_fsv32:
@ -193,6 +195,8 @@ cldnn::format from_data_layout(kernel_selector::data_layout l) {
return cldnn::format::bs_fs_yx_bsv4_fsv2;
case kernel_selector::data_layout::bs_fs_yx_bsv4_fsv4:
return cldnn::format::bs_fs_yx_bsv4_fsv4;
case kernel_selector::data_layout::bs_fs_yx_bsv8_fsv4:
return cldnn::format::bs_fs_yx_bsv8_fsv4;
case kernel_selector::data_layout::bs_fs_yx_bsv32_fsv32:
return cldnn::format::bs_fs_yx_bsv32_fsv32;
case kernel_selector::data_layout::nv12:

26
inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp vendored
View File

@ -284,10 +284,11 @@ bool layout_optimizer::can_fuse_reorder(program_node& prev, program_node& next,
return true;
if (next.is_type<convolution>() &&
(fmt_prev == format::b_fs_yx_fsv4 || fmt_prev == format::bs_fs_yx_bsv4_fsv4) &&
(fmt_prev == format::b_fs_yx_fsv4 || fmt_prev == format::bs_fs_yx_bsv4_fsv4 || fmt_prev == format::bs_fs_yx_bsv8_fsv4) &&
((fmt_next == format::b_fs_yx_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
(fmt_next == format::bs_fs_yx_bsv32_fsv32 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
(fmt_next == format::bs_fs_yx_bsv4_fsv4 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
(fmt_next == format::bs_fs_yx_bsv8_fsv4 && (prev_output_layout.size.feature[0] == 3 || prev_output_layout.size.feature[0] == 4)) ||
(fmt_next == format::b_fs_yx_fsv16 && next_output_layout.size.feature[0] >= 16 &&
(prev_output_layout.size.feature[0] == 3 || (prev_output_layout.size.feature[0] == 4 && (prev_dt == data_types::u8 || prev_dt == data_types::i8))))))
return true;
@ -1269,6 +1270,7 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
format::bs_fs_yx_bsv32_fsv16,
format::bs_fs_yx_bsv32_fsv32,
format::bs_fs_yx_bsv4_fsv4,
format::bs_fs_yx_bsv8_fsv4,
format::bs_fs_yx_bsv4_fsv2,
format::bs_fs_zyx_bsv4_fsv4,
format::bs_fs_zyx_bsv4_fsv2,
@ -1320,17 +1322,27 @@ impl_types layout_optimizer::get_preferred_impl_type(program_node& node, format
impl_candidate = impl_types::ocl;
}
size_t eltw_dep = 0;
for (auto& fo : node.get_fused_primitives()) {
if (fo.node->is_type<eltwise>()) {
auto in_layout = node.get_dependency(fo.dep_start_idx).get_output_layout();
auto out_layout = node.get_output_layout();
auto in_dt = in_layout.data_type;
auto out_dt = out_layout.data_type;
if ((out_layout.count() == in_layout.count()) &&
(data_type_traits::is_floating_point(in_dt) || data_type_traits::is_floating_point(out_dt)) && in_dt != out_dt &&
fo.node->as<eltwise>().get_primitive()->needs_onednn_sum_post_op(in_layout)) {
impl_candidate = impl_types::ocl;
break;
if (fo.node->as<eltwise>().get_primitive()->needs_onednn_sum_post_op(in_layout)) {
if ((out_layout.count() == in_layout.count()) &&
(data_type_traits::is_floating_point(in_dt) || data_type_traits::is_floating_point(out_dt)) && in_dt != out_dt) {
impl_candidate = impl_types::ocl;
break;
}
if (in_layout.size == out_layout.size && in_layout.format == out_layout.format && in_layout.data_padding == out_layout.data_padding &&
data_type_traits::size_of(in_dt) == data_type_traits::size_of(out_dt)) {
if (eltw_dep > 0) {
impl_candidate = impl_types::ocl;
break;
}
eltw_dep = fo.dep_start_idx;
}
}
} else if (fo.node->is_type<activation>()) {
// Some activations aren't implemented in oneDNN
@ -1453,7 +1465,7 @@ format layout_optimizer::get_preferred_format(program_node& node) {
if (data_type_traits::is_floating_point(conv.get_output_layout().data_type) || ws.spatial[0] != 7 || conv.get_primitive()->groups > 1)
expected = format::bfyx;
else
expected = format::bs_fs_yx_bsv4_fsv4;
expected = format::bs_fs_yx_bsv8_fsv4;
auto conv_output_layout = conv.get_output_layout();
auto weights_layout = conv.weights(0).get_output_layout();

18
inference-engine/thirdparty/clDNN/src/network.cpp vendored
View File

@ -514,15 +514,17 @@ void network::allocate_primitives() {
can_reuse_eltwise_mem = true;
}
if (_primitives.find(eltw_in.id()) != _primitives.end() && _primitives.find(node->id()) != _primitives.end()) {
auto& eltw_inst = _primitives.at(eltw_in.id());
auto& prim_inst = _primitives.at(node->id());
auto eltw_mem_type = eltw_inst->output_memory().get_allocation_type();
auto prim_mem_type = prim_inst->output_memory().get_allocation_type();
if (!can_reuse_eltwise_mem) {
if (_primitives.find(eltw_in.id()) != _primitives.end() && _primitives.find(node->id()) != _primitives.end()) {
auto& eltw_inst = _primitives.at(eltw_in.id());
auto& prim_inst = _primitives.at(node->id());
auto eltw_mem_type = eltw_inst->output_memory().get_allocation_type();
auto prim_mem_type = prim_inst->output_memory().get_allocation_type();
// Keep lockable memory type for `prim_inst` output if needed
if (eltw_mem_type != prim_mem_type && eltw_mem_type != allocation_type::cl_mem && eltw_mem_type != allocation_type::usm_host)
can_reuse_eltwise_mem = false;
// Keep lockable memory type for `prim_inst` output if needed
if (eltw_mem_type != prim_mem_type && eltw_mem_type != allocation_type::cl_mem && eltw_mem_type != allocation_type::usm_host)
can_reuse_eltwise_mem = false;
}
}
if (fused_op.node->as<eltwise>().get_primitive()->needs_onednn_sum_post_op(eltw_in_layout) && !can_reuse_eltwise_mem) {

14
inference-engine/thirdparty/clDNN/src/program.cpp vendored
View File

@ -741,10 +741,10 @@ program_node& program::get_or_create(std::shared_ptr<primitive> prim) {
}
void program::add_intermediate(program_node& node,
program_node& next,
size_t prev_idx,
bool connect_int_node_with_old_dep,
bool move_usrs_of_prev_to_node) {
program_node& next,
size_t prev_idx,
bool connect_int_node_with_old_dep,
bool move_usrs_of_prev_to_node) {
if (connect_int_node_with_old_dep && !node.dependencies.empty())
throw std::invalid_argument(
"Node which is about to be added in between two other nodes should not have any existing dependencies");
@ -1112,8 +1112,8 @@ void program::remove_nodes(std::vector<program_node*>& to_remove) {
// TODO: break this function into number of smaller ones + add per-primitive fields (possibly use
// primitive_inst::to_string?)
void program::dump_program(const char* stage,
bool with_full_info,
std::function<bool(program_node const&)> const& filter) const {
bool with_full_info,
std::function<bool(program_node const&)> const& filter) const {
std::string path = get_dir_path(options);
if (path.empty() || !with_full_info) {
return;
@ -1230,7 +1230,7 @@ void program::save_pass_info(std::string pass_name) {
}
void program::add_optimized_primitive_info(primitive_id optimized_primitive_id,
std::vector<primitive_id> replaced_with_ids) {
std::vector<primitive_id> replaced_with_ids) {
for (auto& e : optimized) {
auto it = std::find_if(e.second.begin(), e.second.end(), [&optimized_primitive_id](const primitive_id& id) {
return optimized_primitive_id == id;

43
inference-engine/thirdparty/clDNN/src/program_helpers.cpp vendored
View File

@ -139,30 +139,25 @@ std::pair<bool, bool> program_helpers::are_layouts_identical(layout const& l1, l
return {false, false};
if (l1.get_linear_size() != l2.get_linear_size())
return {false, false};
if ((l1.format == format::b_fs_yx_fsv4 && l2.format != format::b_fs_yx_fsv4) ||
(l2.format == format::b_fs_yx_fsv4 && l1.format != format::b_fs_yx_fsv4) ||
(l1.format == format::fs_b_yx_fsv32 && l2.format != format::fs_b_yx_fsv32) ||
(l2.format == format::fs_b_yx_fsv32 && l1.format != format::fs_b_yx_fsv32) ||
(l1.format == format::b_fs_yx_fsv16 && l2.format != format::b_fs_yx_fsv16) ||
(l2.format == format::b_fs_yx_fsv16 && l1.format != format::b_fs_yx_fsv16) ||
(l1.format == format::b_fs_yx_fsv32 && l2.format != format::b_fs_yx_fsv32) ||
(l2.format == format::b_fs_yx_fsv32 && l1.format != format::b_fs_yx_fsv32) ||
(l1.format == format::b_fs_zyx_fsv32 && l2.format != format::b_fs_zyx_fsv32) ||
(l2.format == format::b_fs_zyx_fsv32 && l1.format != format::b_fs_zyx_fsv32) ||
(l1.format == format::b_fs_zyx_fsv16 && l2.format != format::b_fs_zyx_fsv16) ||
(l2.format == format::b_fs_zyx_fsv16 && l1.format != format::b_fs_zyx_fsv16) ||
(l1.format == format::bs_fs_yx_bsv4_fsv4 && l2.format != format::bs_fs_yx_bsv4_fsv4) ||
(l2.format == format::bs_fs_yx_bsv4_fsv4 && l1.format != format::bs_fs_yx_bsv4_fsv4) ||
(l1.format == format::bs_fs_yx_bsv4_fsv2 && l2.format != format::bs_fs_yx_bsv4_fsv2) ||
(l2.format == format::bs_fs_yx_bsv4_fsv2 && l1.format != format::bs_fs_yx_bsv4_fsv2) ||
(l1.format == format::bs_fs_yx_bsv32_fsv16 && l2.format != format::bs_fs_yx_bsv32_fsv16) ||
(l2.format == format::bs_fs_yx_bsv32_fsv16 && l1.format != format::bs_fs_yx_bsv32_fsv16) ||
(l1.format == format::bs_fs_yx_bsv32_fsv32 && l2.format != format::bs_fs_yx_bsv32_fsv32) ||
(l2.format == format::bs_fs_yx_bsv32_fsv32 && l1.format != format::bs_fs_yx_bsv32_fsv32) ||
(l1.format == format::bs_fs_yx_bsv16_fsv16 && l2.format != format::bs_fs_yx_bsv16_fsv16) ||
(l2.format == format::bs_fs_yx_bsv16_fsv16 && l1.format != format::bs_fs_yx_bsv16_fsv16) ||
(l1.format == format::bs_fs_zyx_bsv16_fsv16 && l2.format != format::bs_fs_zyx_bsv16_fsv16) ||
(l2.format == format::bs_fs_zyx_bsv16_fsv16 && l1.format != format::bs_fs_zyx_bsv16_fsv16))
auto check_format = [&l1, &l2](cldnn::format format) {
return (l1.format == format && l2.format != format) ||
(l2.format == format && l1.format != format);
};
if (check_format(format::b_fs_yx_fsv4) ||
check_format(format::fs_b_yx_fsv32) ||
check_format(format::b_fs_yx_fsv16) ||
check_format(format::b_fs_yx_fsv32) ||
check_format(format::b_fs_zyx_fsv32) ||
check_format(format::b_fs_zyx_fsv16) ||
check_format(format::bs_fs_yx_bsv4_fsv4) ||
check_format(format::bs_fs_yx_bsv8_fsv4) ||
check_format(format::bs_fs_yx_bsv4_fsv2) ||
check_format(format::bs_fs_yx_bsv32_fsv16) ||
check_format(format::bs_fs_yx_bsv32_fsv32) ||
check_format(format::bs_fs_yx_bsv16_fsv16) ||
check_format(format::bs_fs_zyx_bsv16_fsv16))
return {false, false};
auto l1_pitch = l1.get_pitches();

3
inference-engine/thirdparty/clDNN/src/program_node.cpp vendored
View File

@ -428,7 +428,8 @@ dnnl::post_ops program_node::try_optimize_post_ops(dnnl::post_ops& p_ops, const
// Ignore optimized operations for "previous" operation in our operation pair
while (type_is_any_optimized(prev_type) && cur_post_op_idx < post_ops_size - 1) {
prev_post_op_idx++;
cur_post_op_idx++;
if (prev_post_op_idx == cur_post_op_idx)
cur_post_op_idx++;
prev_type = cur_post_ops[prev_post_op_idx].op_type;
cur_type = cur_post_ops[cur_post_op_idx].op_type;
}

226
inference-engine/thirdparty/clDNN/tests/test_cases/convert_color_gpu_test.cpp vendored
View File

@ -16,15 +16,15 @@ using namespace cldnn;
using namespace ::tests;
template <typename T, typename U>
void createReferenceData(const T* arg_y, const T* arg_uv, U* out_ptr,
size_t batch_size, size_t image_h, size_t image_w,
size_t stride_y, size_t stride_uv, bool to_rgb) {
for (int batch = 0; batch < batch_size; batch++) {
void createReferenceDataNV12(const T* arg_y, const T* arg_uv, U* out_ptr,
size_t batch_size, size_t image_h, size_t image_w,
size_t stride_y, size_t stride_uv, bool to_rgb) {
for (size_t batch = 0; batch < batch_size; ++batch) {
U* out = out_ptr + batch * image_w * image_h * 3;
auto y_ptr = arg_y + batch * stride_y;
auto uv_ptr = arg_uv + batch * stride_uv;
for (int h = 0; h < image_h; h++) {
for (int w = 0; w < image_w; w++) {
for (size_t h = 0; h < image_h; ++h) {
for (size_t w = 0; w < image_w; ++w) {
auto y_index = h * image_w + w;
auto y_val = static_cast<float>(y_ptr[y_index]);
auto uv_index = (h / 2) * image_w + (w / 2) * 2;
@ -87,8 +87,8 @@ TEST(convert_color, nv12_to_rgb_two_planes_buffer_fp32) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<float, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
createReferenceDataNV12<float, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<float> output_ptr(output, get_test_stream());
@ -126,8 +126,8 @@ TEST(convert_color, nv12_to_bgr_two_planes_buffer_fp32) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, false);
createReferenceDataNV12<float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, false);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<float> output_ptr(output, get_test_stream());
@ -166,8 +166,8 @@ TEST(convert_color, nv12_to_rgb_two_planes_buffer_u8) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<uint8_t, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
createReferenceDataNV12<uint8_t, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<uint8_t> output_ptr(output, get_test_stream());
@ -206,8 +206,8 @@ TEST(convert_color, nv12_to_rgb_two_planes_buffer_fp16) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<FLOAT16, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
createReferenceDataNV12<FLOAT16, float>(input_y_data.data(), input_uv_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<uint16_t> output_ptr(output, get_test_stream());
@ -243,8 +243,8 @@ TEST(convert_color, nv12_to_rgb_single_plane_buffer_fp32) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<float, float>(input_data.data(), input_data.data() + height * width, ref_res.data(),
1, height, width, input_height * width, input_height * width, true);
createReferenceDataNV12<float, float>(input_data.data(), input_data.data() + height * width, ref_res.data(),
1, height, width, input_height * width, input_height * width, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<float> output_ptr(output, get_test_stream());
@ -279,8 +279,8 @@ TEST(convert_color, nv12_to_rgb_single_plane_buffer_u8) {
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceData<uint8_t, float>(input_data.data(), input_data.data() + height * width, ref_res.data(),
1, height, width, input_height * width, input_height * width, true);
createReferenceDataNV12<uint8_t, float>(input_data.data(), input_data.data() + height * width, ref_res.data(),
1, height, width, input_height * width, input_height * width, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<uint8_t> output_ptr(output, get_test_stream());
@ -356,8 +356,8 @@ TEST(convert_color, nv12_to_rgb_two_planes_surface_u8) {
auto outputs = network.execute();
std::vector<float> reference_results(width * height * 3);
createReferenceData<uint8_t, float>(data.data(), data.data() + height * width, reference_results.data(),
1, height, width, height * width, height * width / 2, true);
createReferenceDataNV12<uint8_t, float>(data.data(), data.data() + height * width, reference_results.data(),
1, height, width, height * width, height * width / 2, true);
auto output_prim = outputs.begin()->second.get_memory();
cldnn::mem_lock<float> output_ptr(output_prim, get_test_stream());
@ -419,8 +419,8 @@ TEST(convert_color, nv12_to_rgb_single_plane_surface_u8) {
auto outputs = network.execute();
std::vector<float> reference_results(width * height * 3);
createReferenceData<uint8_t, float>(input_data.data(), input_data.data() + height * width, reference_results.data(),
1, height, width, input_height * width, input_height * width, true);
createReferenceDataNV12<uint8_t, float>(input_data.data(), input_data.data() + height * width, reference_results.data(),
1, height, width, input_height * width, input_height * width, true);
auto output_prim = outputs.begin()->second.get_memory();
cldnn::mem_lock<float> output_ptr(output_prim, get_test_stream());
@ -429,3 +429,185 @@ TEST(convert_color, nv12_to_rgb_single_plane_surface_u8) {
}
checkStatus(clReleaseMemObject(nv12_image), "clReleaseMemObject");
}
template <typename T>
std::tuple<T, T, T> yuv_pixel_to_rgb(float y_val, float u_val, float v_val) {
auto c = y_val - 16.f;
auto d = u_val - 128.f;
auto e = v_val - 128.f;
auto clip = [](float a) -> T {
if (std::is_integral<T>()) {
return static_cast<T>(std::min(std::max(std::round(a), 0.f), 255.f));
} else {
return static_cast<T>(std::min(std::max(a, 0.f), 255.f));
}
};
auto b = clip(1.164f * c + 2.018f * d);
auto g = clip(1.164f * c - 0.391f * d - 0.813f * e);
auto r = clip(1.164f * c + 1.596f * e);
return std::tuple<T, T, T>{r, g, b};
}
template <typename T, typename U>
void createReferenceDataI420(const T* arg_y, const T* arg_u, const T* arg_v, U* out_ptr,
size_t batch_size, size_t image_h, size_t image_w,
size_t stride_y, size_t stride_uv, bool rgb_color_format) {
for (size_t batch = 0; batch < batch_size; ++batch) {
U* out = out_ptr + batch * image_w * image_h * 3;
auto y_ptr = arg_y + batch * stride_y;
auto u_ptr = arg_u + batch * stride_uv;
auto v_ptr = arg_v + batch * stride_uv;
for (size_t h = 0; h < image_h; ++h) {
for (size_t w = 0; w < image_w; ++w) {
auto y_index = h * image_w + w;
auto y_val = static_cast<float>(y_ptr[y_index]);
auto uv_index = (h / 2) * (image_w / 2) + (w / 2);
auto u_val = static_cast<float>(u_ptr[uv_index]);
auto v_val = static_cast<float>(v_ptr[uv_index]);
T r, g, b;
std::tie(r, g, b) = yuv_pixel_to_rgb<U>(y_val, u_val, v_val);
if (rgb_color_format) {
out[y_index * 3] = r;
out[y_index * 3 + 1] = g;
out[y_index * 3 + 2] = b;
} else {
out[y_index * 3] = b;
out[y_index * 3 + 1] = g;
out[y_index * 3 + 2] = r;
}
}
}
}
}
TEST(convert_color, i420_to_rgb_three_planes_buffer_fp32) {
auto& engine = get_test_engine();
int width = 224;
int height = 448;
auto input_y = engine.allocate_memory({ data_types::f32, format::byxf, { 1, 1, width, height } });
auto input_u = engine.allocate_memory({ data_types::f32, format::byxf, { 1, 1, width / 2 , height / 2 } });
auto input_v = engine.allocate_memory({ data_types::f32, format::byxf, { 1, 1, width / 2 , height / 2 } });
std::vector<float> input_y_data = generate_random_1d<float>(width * height, 0, 255);
std::vector<float> input_u_data = generate_random_1d<float>(width * height / 4, 0, 255);
std::vector<float> input_v_data = generate_random_1d<float>(width * height / 4, 0, 255);
set_values(input_y, input_y_data);
set_values(input_u, input_u_data);
set_values(input_v, input_v_data);
layout output_layout(data_types::f32, cldnn::format::byxf, { 1, 3, width, height });
topology topology;
topology.add(input_layout("input_y", input_y->get_layout()));
topology.add(input_layout("input_u", input_u->get_layout()));
topology.add(input_layout("input_v", input_v->get_layout()));
topology.add(convert_color("convert_color", { "input_y", "input_u", "input_v" }, cldnn::convert_color::color_format::I420, cldnn::convert_color::color_format::RGB,
cldnn::convert_color::memory_type::buffer, output_layout));
network network(engine, topology);
network.set_input_data("input_y", input_y);
network.set_input_data("input_u", input_u);
network.set_input_data("input_v", input_v);
auto outputs = network.execute();
std::vector<float> ref_res(width * height * 3);
createReferenceDataI420<float, float>(input_y_data.data(), input_u_data.data(), input_v_data.data(), ref_res.data(),
1, height, width, height * width, height * width / 2, true);
auto output = outputs.at("convert_color").get_memory();
cldnn::mem_lock<float> output_ptr(output, get_test_stream());
for (size_t i = 0; i < ref_res.size(); ++i) {
EXPECT_NEAR(ref_res[i], output_ptr[i], 1.001f);
}
}
TEST(convert_color, i420_to_rgb_three_planes_surface_u8) {
int width = 224;
int height = 448;
auto ocl_instance = std::make_shared<OpenCL>();
device_query query(engine_types::ocl, runtime_types::ocl, static_cast<void*>(ocl_instance->_context.get()));
auto devices = query.get_available_devices();
auto engine_config = cldnn::engine_configuration();
auto engine = engine::create(engine_types::ocl, runtime_types::ocl, devices.begin()->second, engine_config);
if (!engine->get_device_info().supports_image) {
GTEST_SKIP() << "Device doesn't support images";
}
int data_size = width * (height + height / 2);
std::vector<uint8_t> data = generate_random_1d<uint8_t>(data_size, 0, 255);
cl_int err;
cl_image_format image_format;
image_format.image_channel_order = CL_R;
image_format.image_channel_data_type = CL_UNORM_INT8;
cl_image_desc image_desc = { CL_MEM_OBJECT_IMAGE2D, (size_t)width, (size_t)height, 0,
0, 0, 0, 0, 0, { nullptr } };
cl_mem i420_image_plane_y = clCreateImage(ocl_instance->_context.get(), CL_MEM_READ_WRITE, &image_format, &image_desc, nullptr, &err);
checkStatus(err, "Creating i420 image plane_y failed");
image_desc.image_width = width / 2;
image_desc.image_height = height / 2;
cl_mem i420_image_plane_u = clCreateImage(ocl_instance->_context.get(), CL_MEM_READ_WRITE, &image_format, &image_desc, nullptr, &err);
checkStatus(err, "Creating i420 image plane_u failed");
cl_mem i420_image_plane_v = clCreateImage(ocl_instance->_context.get(), CL_MEM_READ_WRITE, &image_format, &image_desc, nullptr, &err);
checkStatus(err, "Creating i420 image plane_v failed");
size_t origin[3] = { 0, 0, 0 };
size_t y_region[3] = { (size_t)width, (size_t)height, 1 };
size_t uv_region[3] = { (size_t)width / 2, (size_t)height / 2, 1 };
err = clEnqueueWriteImage(ocl_instance->_queue.get(), i420_image_plane_y, true, origin, y_region, 0, 0, &data[0], 0, nullptr, nullptr);
checkStatus(err, "Writing i420 image plane_y failed");
err = clEnqueueWriteImage(ocl_instance->_queue.get(), i420_image_plane_u, true, origin, uv_region, 0, 0, &data[width * height], 0, nullptr, nullptr);
checkStatus(err, "Writing i420 image plane_u failed");
err = clEnqueueWriteImage(ocl_instance->_queue.get(), i420_image_plane_v, true, origin, uv_region, 0, 0, &data[width * (height + height / 4)], 0, nullptr, nullptr);
checkStatus(err, "Writing i420 image plane_v failed");
auto input = input_layout("input", { data_types::u8, format::nv12, { 1, 1, width, height } });
auto input2 = input_layout("input2", { data_types::u8, format::nv12, { 1, 1, width / 2, height / 2 } });
auto input3 = input_layout("input3", { data_types::u8, format::nv12, { 1, 1, width / 2, height / 2 } });
auto output_format = cldnn::format::byxf;
layout output_layout(data_types::f32, output_format, { 1, 3, width, height });
auto input_memory = engine->share_image(input.layout, i420_image_plane_y);
auto input_memory2 = engine->share_image(input2.layout, i420_image_plane_u);
auto input_memory3 = engine->share_image(input3.layout, i420_image_plane_v);
topology topology;
topology.add(input);
topology.add(input2);
topology.add(input3);
topology.add(convert_color("convert_color", { "input", "input2", "input3" }, cldnn::convert_color::color_format::I420, cldnn::convert_color::color_format::RGB,
cldnn::convert_color::memory_type::image, output_layout));
network network(*engine, topology);
network.set_input_data("input", input_memory);
network.set_input_data("input2", input_memory2);
network.set_input_data("input3", input_memory3);
auto outputs = network.execute();
std::vector<float> reference_results(width * height * 3);
createReferenceDataI420<uint8_t, float>(data.data(), data.data() + height * width, data.data() + width * (height + height / 4), reference_results.data(),
1, height, width, height * width, height * width / 2, true);
auto output_prim = outputs.begin()->second.get_memory();
cldnn::mem_lock<float> output_ptr(output_prim, get_test_stream());
for (auto i = 0; i < reference_results.size(); i++) {
EXPECT_NEAR(reference_results[i], output_ptr[i], 1.001f);
}
checkStatus(clReleaseMemObject(i420_image_plane_y), "clReleaseMemObject");
checkStatus(clReleaseMemObject(i420_image_plane_u), "clReleaseMemObject");
checkStatus(clReleaseMemObject(i420_image_plane_v), "clReleaseMemObject");
}

816
inference-engine/thirdparty/clDNN/tests/test_cases/convolution_gpu_test.cpp vendored
View File

File diff suppressed because it is too large Load Diff

37
inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp vendored
View File

@ -681,7 +681,7 @@ TEST_P(conv_fp32_reorder_fsv16_to_bfyx_conv, basic) {
reorder("reorder_fsv16", "input", format::b_fs_yx_fsv16, data_types::f32),
convolution("conv_prim", "reorder_fsv16", { "weights" }, p.groups, p.stride, p.pad, p.dilation),
reorder("reorder_bfyx", "conv_prim", format::bfyx, data_types::f32),
convolution("conv_output", "reorder_bfyx", { "weights_dw" }, 1, dw_stride, p.pad, p.dilation),
convolution("conv_output", "reorder_bfyx", { "weights_dw" }, p.out_shape.feature[0], dw_stride, p.pad, p.dilation),
activation("activation", "conv_output", activation_func::abs),
reorder("reorder_output", "activation", p.default_format, data_types::f32)
);
@ -3264,6 +3264,35 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, gemm_2in_quantize_u8,
//gemm_test_params{ CASE_GEMM_2IN_FP32_1, 3, 4 },
}));
class gemm_2in_quantize_float_in : public GemmFusingTest {};
TEST_P(gemm_2in_quantize_float_in, basic) {
auto p = GetParam();
create_topologies(input_layout("input0", get_input_layout(p, 0)),
input_layout("input1", get_input_layout(p, 1)),
data("in_lo", get_mem(get_per_channel_layout(p), 0)),
data("in_hi", get_mem(get_per_channel_layout(p), 1, max_random)),
data("out_lo", get_mem(get_single_element_layout(p), 0)),
data("out_hi", get_mem(get_single_element_layout(p), 255)),
gemm("gemm_prim", { "input0", "input1" }, data_types::f32),
quantize("quantize", "gemm_prim", "in_lo", "in_hi", "out_lo", "out_hi", 256, data_types::u8),
reorder("reorder_bfyx", "quantize", p.default_format, data_types::f32)
);
implementation_desc gemm_impl = { format::bfyx, "gemm_tiled_opt" };
bo_fused.set_option(build_option::force_implementations({ {"gemm_prim", gemm_impl} }));
tolerance = 1.0f;
execute(p);
}
INSTANTIATE_TEST_SUITE_P(fusings_gpu, gemm_2in_quantize_float_in,
::testing::ValuesIn(std::vector<gemm_test_params>{
gemm_test_params{ CASE_GEMM_2IN_FP16_1, 3, 4 },
gemm_test_params{ CASE_GEMM_2IN_FP32_1, 3, 4 },
gemm_test_params{ CASE_GEMM_ELTWISE_2IN_FP16_1, 3, 4 },
gemm_test_params{ CASE_GEMM_ELTWISE_2IN_FP32_1, 3, 4 },
}));
class gemm_2in_scale : public GemmFusingTest {};
TEST_P(gemm_2in_scale, basic) {
auto p = GetParam();
@ -10059,7 +10088,7 @@ TEST_P(conv_fp32_reorder_bfyx_to_fsv32_conv_subtract, have_subtract_per_feature)
data("weights_dw", get_mem(dw_weights_layout, -127, 127)),
convolution("conv_prim", "input", { "weights" }, p.groups, p.stride, p.pad, p.dilation),
reorder("reorder_fsv32", "conv_prim", format::fs_b_yx_fsv32, data_types::f32, values_to_subtract),
convolution("conv_output", "reorder_fsv32", { "weights_dw" }, 1, dw_stride, p.pad, p.dilation),
convolution("conv_output", "reorder_fsv32", { "weights_dw" }, p.out_shape.feature[0], dw_stride, p.pad, p.dilation),
activation("activation", "conv_output", activation_func::abs)
);
@ -10088,7 +10117,7 @@ TEST_P(conv_fp32_reorder_bfyx_to_fsv32_conv_fused_activation, have_fused_activat
convolution("conv_prim", "input", { "weights" }, p.groups, p.stride, p.pad, p.dilation),
reorder("reorder_fsv32", "conv_prim", format::fs_b_yx_fsv32, data_types::f32),
activation("activation_quantize", "reorder_fsv32", activation_func::relu),
convolution("conv_output", "activation_quantize", { "weights_dw" }, 1, dw_stride, p.pad, p.dilation),
convolution("conv_output", "activation_quantize", { "weights_dw" }, p.out_shape.feature[0], dw_stride, p.pad, p.dilation),
activation("activation", "conv_output", activation_func::abs)
);
@ -10116,7 +10145,7 @@ TEST_P(conv_fp32_reorder_bfyx_to_fsv32_conv_data_padding, have_data_padding) {
data("weights_dw", get_mem(dw_weights_layout, -127, 127)),
convolution("conv_prim", "input", { "weights" }, p.groups, p.stride, p.pad, p.dilation),
reorder("reorder_fsv32", "conv_prim", layout(data_types::f32, format::fs_b_yx_fsv32, dw_tensor, padding{ {0, 0, 1, 1}, 0 })),
convolution("conv_output", "reorder_fsv32", { "weights_dw" }, 1, dw_stride, p.pad, p.dilation),
convolution("conv_output", "reorder_fsv32", { "weights_dw" }, p.out_shape.feature[0], dw_stride, p.pad, p.dilation),
activation("activation", "conv_output", activation_func::abs),
activation("activation2", "conv_prim", activation_func::abs),
eltwise("add_bias", { "activation", "activation2" }, eltwise_mode::sum)

50
inference-engine/thirdparty/clDNN/tests/test_cases/memory_test.cpp vendored
View File

@ -43,7 +43,7 @@ TEST(memory_tests, DISABLED_network_creation_loop)
{
engine eng;
memory in = memory::allocate(eng, layout{ data_types::f32, format::bfyx,{ 1, 1, 1000, 1000 } });
memory in = memory::allocate(eng, layout{ data_types::f32, format::bfyx, { 1, 1, 1000, 1000 } });
topology tpl{
input_layout("in", in->get_layout()),
@ -66,7 +66,7 @@ TEST(memory_pool, basic_non_padded_relu_pipe) {
auto x_size = 1;
auto y_size = 1;
auto input = engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
auto input = engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
topology topology;
topology.add(input_layout("input", input->get_layout()));
@ -86,7 +86,7 @@ TEST(memory_pool, basic_non_padded_relu_pipe) {
network.set_input_data("input", input);
auto outputs = network.execute();
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t) 64);
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t)64);
}
TEST(memory_pool, basic_non_padded_relu_and_pooling_pipe) {
@ -99,13 +99,13 @@ TEST(memory_pool, basic_non_padded_relu_and_pooling_pipe) {
auto x_size = 4;
auto y_size = 4;
auto input = engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
auto input = engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
topology topology;
topology.add(input_layout("input", input->get_layout()));
topology.add(activation("relu", "input", activation_func::relu));
topology.add(activation("relu1", "relu", activation_func::relu));
topology.add(pooling("pool1", "relu1",pooling_mode::max, { 1,1,3,3 }, { 1,1,2,2 }));
topology.add(pooling("pool1", "relu1", pooling_mode::max, { 1, 1, 3, 3 }, { 1, 1, 2, 2 }));
topology.add(activation("relu2", "pool1", activation_func::relu));
topology.add(activation("relu3", "relu2", activation_func::relu));
topology.add(activation("relu4", "relu3", activation_func::relu));
@ -133,7 +133,7 @@ TEST(memory_pool, multi_outputs_network) {
auto x_size = 4;
auto y_size = 4;
auto input = engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
auto input = engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
topology topology;
topology.add(input_layout("input", input->get_layout()));
@ -153,7 +153,7 @@ TEST(memory_pool, multi_outputs_network) {
network.set_input_data("input", input);
auto outputs = network.execute();
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t)1536);
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t) 1536);
}
TEST(memory_pool, oooq) {
@ -171,14 +171,14 @@ TEST(memory_pool, oooq) {
auto x_size = 4;
auto y_size = 4;
auto input = engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
auto input = engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(x_size, y_size), feature(feature_num), batch(batch_num)) } });
topology topology;
topology.add(input_layout("input", input->get_layout()));
topology.add(activation("relu1", "input", activation_func::relu));
topology.add(activation("relu2", "input", activation_func::relu));
topology.add(activation("relu3", "input", activation_func::relu));
topology.add(concatenation("concat1", { "relu1", "relu2"},concatenation::along_f));
topology.add(concatenation("concat1", { "relu1", "relu2" },concatenation::along_f));
topology.add(activation("relu4", "concat1", activation_func::relu));
topology.add(activation("relu5", "relu3", activation_func::relu));
topology.add(concatenation("concat2", { "relu4", "relu5" }, concatenation::along_f));
@ -209,7 +209,7 @@ TEST(memory_pool, DISABLED_shared_mem_pool_same_topology_twice) {
auto inp_x_size = 4;
auto inp_y_size = 4;
auto input = engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(inp_x_size, inp_y_size), feature(feature_num), batch(batch_num)) } });
auto input = engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(inp_x_size, inp_y_size), feature(feature_num), batch(batch_num)) } });
set_values(input,
{ 1.0f, 2.5f, 3.0f, 4.0f, 5.0f, 2.0f, 2.0f, 3.0f, 6.1f, 4.7f, 1.0f, 1.0f, 8.2f, 1.0f, 2.0f, 1.0f,
@ -227,7 +227,7 @@ TEST(memory_pool, DISABLED_shared_mem_pool_same_topology_twice) {
topology.add(activation("relu4", "concat1", activation_func::relu));
topology.add(activation("relu5", "relu3", activation_func::relu));
topology.add(concatenation("concat2", { "relu4", "relu5" }, concatenation::along_f));
topology.add(activation("relu6", "concat2", activation_func::linear, {1.0f, 0.5f}));
topology.add(activation("relu6", "concat2", activation_func::linear, { 1.0f, 0.5f }));
build_options bo;
bo.set_option(build_option::optimize_data(true));
@ -286,8 +286,8 @@ TEST(memory_pool, DISABLED_shared_mem_pool_same_topology_twice_weights) {
auto inp_x_size = 4;
auto inp_y_size = 4;
auto input= engine->allocate_memory({ data_types::f32, format::bfyx,{ tensor(spatial(inp_x_size, inp_y_size), feature(feature_num), batch(batch_num)) } });
auto weights = engine->allocate_memory({ data_types::f32,format::bfyx,{ 1, 1, 3, 2 } });
auto input= engine->allocate_memory({ data_types::f32, format::bfyx, { tensor(spatial(inp_x_size, inp_y_size), feature(feature_num), batch(batch_num)) } });
auto weights = engine->allocate_memory({ data_types::f32, format::bfyx, { 1, 1, 3, 2 } });
std::vector<float> dummy_input_data_1 = {
/*f0 xy*/ 0.8f, 0.65f, 0.1f, 1.0f, 1.0f, 0.5f, 0.11f, 0.33f, 0.66f, 0.11f, 0.22f, 0.33f, 0.99f, 0.8f, 0.7f, 0.5f,
@ -373,10 +373,10 @@ TEST(memory_pool, shared_mem_pool_diff_batches) {
layout lay_batch_8 = { dt, fmt, { tensor(spatial(inp_x_size, inp_y_size), feature(feature_num), batch(batch_8)) }};
auto input_1 = engine->allocate_memory(lay_batch_1);
auto input_8 = engine->allocate_memory(lay_batch_8);
auto weights = engine->allocate_memory({ dt, fmt, { 1, 1, 3, 2 } });
auto weights = engine->allocate_memory({ dt, fmt, { 1, 3, 3, 2 } });
std::vector<float> dummy_input_data_1 = generate_random_1d<float>(batch_1*feature_num*inp_x_size*inp_y_size, 0, 1);
std::vector<float> dummy_input_data_8 = generate_random_1d<float>(batch_8*feature_num*inp_x_size*inp_y_size, 0, 1);
std::vector<float> dummy_input_data_1 = generate_random_1d<float>(batch_1 * feature_num * inp_x_size * inp_y_size, 0, 1);
std::vector<float> dummy_input_data_8 = generate_random_1d<float>(batch_8 * feature_num * inp_x_size * inp_y_size, 0, 1);
set_values(input_1, dummy_input_data_1);
set_values(input_8, dummy_input_data_8);
@ -396,14 +396,14 @@ TEST(memory_pool, shared_mem_pool_diff_batches) {
auto outputs = network_first.execute();
auto dev_info = engine->get_device_info();
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t)3928);
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t) 4744);
topo.change_input_layout("input", input_1->get_layout());//change input layout to batch=1
network network_second(*engine, topo, bo);
network_second.set_input_data("input", input_1);
auto outputs_second = network_second.execute();
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t)4328);
EXPECT_EQ(engine->get_max_used_device_memory(), (uint64_t) 5912);
}
TEST(memory_pool, shared_dep_two_output) {
@ -459,20 +459,20 @@ TEST(memory_pool, non_opt_intermidate_opt_after) {
auto input_memory1 = engine.allocate_memory(input_layout1);
auto input_memory2 = engine.allocate_memory(input_layout2);
auto scale_memory = engine.allocate_memory(layout(cldnn::data_types::f32, cldnn::format::bfyx, { 1,1,1,1 }));
auto scale_memory = engine.allocate_memory(layout(cldnn::data_types::f32, cldnn::format::bfyx, { 1, 1, 1, 1 }));
auto data_memory = cldnn::data("scale_mem", scale_memory);
set_values(input_memory1, { 1.0f, 2.0f, 3.0f, 4.0f });
set_values(input_memory2, { 5.0f, 6.0f, 7.0f, 8.0f });
set_values(scale_memory, { 1.0f});
set_values(scale_memory, { 1.0f });
auto reshape_tensor = cldnn::tensor(8, 1, 1, 1);
auto input = cldnn::input_layout("input1", input_layout1);
auto input2 = cldnn::input_layout("input2", input_layout2);
auto concat = cldnn::concatenation("concat", { "input1", "input2" }, cldnn::concatenation::along_b);
auto reshape = cldnn::reshape("reshape", "concat", reshape_tensor);
auto crop1 = cldnn::crop("crop1", "reshape", { 1,1,1,1 }, { 0, 0, 0, 0 });
auto crop2 = cldnn::crop("crop2", "reshape", { 1,1,1,1 }, { 1, 0, 0, 0 });
auto crop1 = cldnn::crop("crop1", "reshape", { 1, 1, 1, 1 }, { 0, 0, 0, 0 });
auto crop2 = cldnn::crop("crop2", "reshape", { 1, 1, 1, 1 }, { 1, 0, 0, 0 });
auto eltwise1 = cldnn::scale("elt1", "crop1", "scale_mem");
auto eltwise2 = cldnn::scale("elt2", "crop2", "scale_mem");
@ -508,7 +508,7 @@ TEST(memory_pool, add_mem_dep_test) {
auto input_layout1 = layout(cldnn::data_types::f32, cldnn::format::bfyx, { 1, 2, 2, 2 });
auto input_memory1 = engine.allocate_memory(input_layout1);
auto scale_memory = engine.allocate_memory(layout(cldnn::data_types::f32, cldnn::format::bfyx, { 1,1,1,1 }));
auto scale_memory = engine.allocate_memory(layout(cldnn::data_types::f32, cldnn::format::bfyx, { 1, 1, 1, 1 }));
auto data_memory = cldnn::data("scale_mem", scale_memory);
set_values(input_memory1, { 1.0f, 2.0f, 3.0f, 4.0f,
@ -518,8 +518,8 @@ TEST(memory_pool, add_mem_dep_test) {
auto input = cldnn::input_layout("input1", input_layout1);
auto actv1 = cldnn::activation("input_activ1", "input1", activation_func::abs);
auto actv2 = cldnn::activation("input_activ2", "input1", activation_func::abs);
auto crop1 = cldnn::crop("crop1", "input_activ1", { 1,1,2,2 }, { 0, 0, 0, 0 });
auto crop2 = cldnn::crop("crop2", "input_activ2", { 1,1,2,2 }, { 0, 1, 0, 0 });
auto crop1 = cldnn::crop("crop1", "input_activ1", { 1, 1, 2, 2 }, { 0, 0, 0, 0 });
auto crop2 = cldnn::crop("crop2", "input_activ2", { 1, 1, 2, 2 }, { 0, 1, 0, 0 });
auto eltwise1 = cldnn::scale("elt1", "crop1", "scale_mem");
auto eltwise2 = cldnn::scale("elt2", "crop2", "scale_mem");
auto actv3 = cldnn::activation("out3", "elt1", activation_func::abs);

2
samples/c/common/opencv_c_wrapper/CMakeLists.txt
View File

@ -9,7 +9,7 @@ file(GLOB SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp ${CMAKE_CURRENT_SOURCE_DIR}/
file(GLOB HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/*.h)
# create library
add_library(${TARGET_NAME} SHARED EXCLUDE_FROM_ALL ${HEADERS} ${SOURCES})
add_library(${TARGET_NAME} SHARED ${HEADERS} ${SOURCES})
# Find OpenCV components if exist
find_package(OpenCV COMPONENTS core imgproc imgcodecs QUIET)

103
samples/c/common/opencv_c_wrapper/bmp_reader.c
View File

@ -1,31 +1,72 @@
#include "bmp_reader.h"
#include <memory.h>
#include <stdio.h>
#include <stdlib.h>
int readBmpImage(const char* fileName, BitMap* image) {
FILE* input = fopen(fileName, "rb");
size_t cnt;
int status = 0;
FILE* input = 0;
if (input == NULL) {
printf("[BMP] file %s is not opened\n", fileName);
return 1;
if (NULL == fileName || NULL == image) {
printf("[BMP] bad arguments\n");
status = -1;
goto Exit;
}
fread(&image->header.type, 2, 1, input);
memset(image, 0, sizeof(BitMap));
input = fopen(fileName, "rb");
if (input == NULL) {
printf("[BMP] file %s is not opened\n", fileName);
status = 1;
goto Exit;
}
cnt = fread(&image->header.type, sizeof(image->header.type), sizeof(unsigned char), input);
if (cnt != sizeof(image->header.type)) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
if (image->header.type != 'M' * 256 + 'B') {
printf("[BMP] file is not bmp type\n");
return 2;
status = 2;
goto Exit;
}
fread(&image->header.size, 4, 1, input);
fread(&image->header.reserved, 4, 1, input);
fread(&image->header.offset, 4, 1, input);
cnt = fread(&image->header.size, sizeof(image->header.size), sizeof(unsigned char), input);
if (cnt != sizeof(image->header.size)) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
fread(&image->infoHeader, sizeof(BmpInfoHeader), 1, input);
cnt = fread(&image->header.reserved, sizeof(image->header.reserved), sizeof(unsigned char), input);
if (cnt != sizeof(image->header.reserved)) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
cnt = fread(&image->header.offset, sizeof(image->header.offset), sizeof(unsigned char), input);
if (cnt != sizeof(image->header.offset)) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
cnt = fread(&image->infoHeader, sizeof(BmpInfoHeader), sizeof(unsigned char), input);
if (cnt != sizeof(image->header.offset)) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
image->width = image->infoHeader.width;
image->height = image->infoHeader.height;
image->height = abs(image->infoHeader.height);
if (image->infoHeader.bits != 24) {
printf("[BMP] 24bpp only supported. But input has: %d\n", image->infoHeader.bits);
@ -38,21 +79,49 @@ int readBmpImage(const char* fileName, BitMap* image) {
}
int padSize = image->width & 3;
size_t row_size = (size_t)image->width * 3;
char pad[3];
size_t size = image->width * image->height * 3;
size_t size = row_size * image->height;
image->data = malloc(sizeof(char) * size);
if (NULL == image->data) {
printf("[BMP] memory allocation failed\n");
return 5;
}
fseek(input, image->header.offset, 0);
if (0 != fseek(input, image->header.offset, SEEK_SET)) {
printf("[BMP] file seek error\n");
status = 2;
goto Exit;
}
// reading by rows in invert vertically
int i;
for (i = 0; i < image->height; i++) {
unsigned int storeAt = image->infoHeader.height < 0 ? i : (unsigned int)image->height - 1 - i;
fread(image->data + image->width * 3 * storeAt, image->width * 3, 1, input);
fread(pad, padSize, 1, input);
cnt = fread(image->data + row_size * storeAt, row_size, sizeof(unsigned char), input);
if (cnt != row_size) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
cnt = fread(pad, padSize, sizeof(unsigned char), input);
if (cnt != padSize) {
printf("[BMP] file read error\n");
status = 2;
goto Exit;
}
}
fclose(input);
return 0;
Exit:
if (0 != status && NULL != image && NULL != image->data) {
free(image->data);
}
if (NULL != input) {
fclose(input);
}
return status;
}

2
samples/cpp/common/format_reader/CMakeLists.txt
View File

@ -13,7 +13,7 @@ source_group("src" FILES ${LIBRARY_SRC})
source_group("include" FILES ${LIBRARY_HEADERS})
# Create library file from sources.
add_library(${TARGET_NAME} SHARED EXCLUDE_FROM_ALL ${MAIN_SRC} ${LIBRARY_HEADERS})
add_library(${TARGET_NAME} SHARED ${MAIN_SRC} ${LIBRARY_HEADERS})
# Find OpenCV components if exist
find_package(OpenCV COMPONENTS core imgproc imgcodecs QUIET)

50
src/bindings/python/tests/test_frontend/test_frontend_onnx.py
View File

@ -26,6 +26,32 @@ def create_onnx_model():
return make_model(graph, producer_name="ngraph ONNX Importer")
def create_onnx_model_with_subgraphs():
A = onnx.helper.make_tensor_value_info("A", onnx.TensorProto.FLOAT, [3])
B = onnx.helper.make_tensor_value_info("B", onnx.TensorProto.FLOAT, [3])
add_out = onnx.helper.make_tensor_value_info("add_out", onnx.TensorProto.FLOAT, [3])
sub_out = onnx.helper.make_tensor_value_info("sub_out", onnx.TensorProto.FLOAT, [3])
add = onnx.helper.make_node("Add", inputs=["A", "B"], outputs=["add_out"])
sub = onnx.helper.make_node("Sub", inputs=["A", "B"], outputs=["sub_out"])
then_body = make_graph([add], "then_body", [], [add_out])
else_body = make_graph([sub], "else_body", [], [sub_out])
if_node = onnx.helper.make_node(
"If",
inputs=["cond"],
outputs=["res"],
then_branch=then_body,
else_branch=else_body
)
cond = onnx.helper.make_tensor_value_info("cond", onnx.TensorProto.BOOL, [])
res = onnx.helper.make_tensor_value_info("res", onnx.TensorProto.FLOAT, [3])
graph = make_graph([if_node], "graph", [cond, A, B], [res])
return make_model(graph, producer_name="ngraph ONNX Importer")
def run_function(function, *inputs, expected):
runtime = get_runtime()
computation = runtime.computation(function)
@ -37,15 +63,18 @@ def run_function(function, *inputs, expected):
fem = FrontEndManager()
onnx_model_filename = "model.onnx"
onnx_model_with_subgraphs_filename = "model_subgraphs.onnx"
ONNX_FRONTEND_NAME = "onnx"
def setup_module():
onnx.save_model(create_onnx_model(), onnx_model_filename)
onnx.save_model(create_onnx_model_with_subgraphs(), onnx_model_with_subgraphs_filename)
def teardown_module():
os.remove(onnx_model_filename)
os.remove(onnx_model_with_subgraphs_filename)
def skip_if_onnx_frontend_is_disabled():
@ -72,17 +101,29 @@ def test_convert():
run_function(function, a, b, expected=[expected])
def test_decode_and_convert():
@pytest.mark.parametrize("model_filename, inputs, expected", [
[onnx_model_filename,
[np.array([[1, 2], [3, 4]], dtype=np.float32),
np.array([[2, 3], [4, 5]], dtype=np.float32)],
np.array([[1.5, 5], [10.5, 18]], dtype=np.float32)],
[onnx_model_with_subgraphs_filename,
[np.array(False, dtype=bool),
np.array([1, 2, 3], dtype=np.float32),
np.array([2, 3, 5], dtype=np.float32)],
np.array([-1, -1, -2], dtype=np.float32)],
])
def test_decode_and_convert(model_filename, inputs, expected):
skip_if_onnx_frontend_is_disabled()
fe = fem.load_by_framework(framework=ONNX_FRONTEND_NAME)
assert fe
model = fe.load(onnx_model_filename)
model = fe.load(model_filename)
assert model
decoded_function = fe.decode(model)
assert decoded_function
for op in decoded_function.get_ordered_ops():
assert op.get_type_name() in ["Parameter", "Constant", "ONNXFrameworkNode",
"ONNXSubgraphFrameworkNode", "Result"]
@ -92,10 +133,7 @@ def test_decode_and_convert():
for op in decoded_function.get_ordered_ops():
assert op.get_type_name() not in ["ONNXFrameworkNode", "ONNXSubgraphFrameworkNode"]
a = np.array([[1, 2], [3, 4]], dtype=np.float32)
b = np.array([[2, 3], [4, 5]], dtype=np.float32)
expected = np.array([[1.5, 5], [10.5, 18]], dtype=np.float32)
run_function(decoded_function, a, b, expected=[expected])
run_function(decoded_function, *inputs, expected=[expected])
def test_load_by_model():

7
src/bindings/python/tests/test_onnx/test_backend.py
View File

@ -101,16 +101,9 @@ tests_expected_to_fail = [
(
xfail_issue_FLOAT_LIKE,
"OnnxBackendNodeModelTest.test_cast_BFLOAT16_to_FLOAT_cpu",
"OnnxBackendNodeModelTest.test_cast_FLOAT16_to_DOUBLE_cpu",
"OnnxBackendNodeModelTest.test_cast_FLOAT16_to_FLOAT_cpu",
"OnnxBackendNodeModelTest.test_cast_FLOAT_to_BFLOAT16_cpu",
"OnnxBackendNodeModelTest.test_castlike_BFLOAT16_to_FLOAT_expanded_cpu",
"OnnxBackendNodeModelTest.test_castlike_FLOAT16_to_DOUBLE_expanded_cpu",
"OnnxBackendNodeModelTest.test_castlike_FLOAT16_to_FLOAT_expanded_cpu",
"OnnxBackendNodeModelTest.test_castlike_FLOAT_to_BFLOAT16_expanded_cpu",
"OnnxBackendNodeModelTest.test_max_float16_cpu",
"OnnxBackendNodeModelTest.test_min_float16_cpu",
"OnnxBackendNodeModelTest.test_mod_mixed_sign_float16_cpu",
),
(
xfail_issue_49207,

6
src/bindings/python/tests/test_onnx/test_zoo_models.py
View File

@ -171,12 +171,6 @@ if len(zoo_models) > 0:
test_cases = backend_test.test_cases["OnnxBackendModelExecutionTest"]
if tests.MODEL_ZOO_XFAIL:
execution_xfail_list = [
# New Python API - fp16 blob
(xfail_issue_67415, "test_MSFT_opset7_fp16_inception_v1_onnxzoo_lotus_inception_v1_cpu"),
(xfail_issue_67415, "test_MSFT_opset7_fp16_shufflenet_onnxzoo_lotus_shufflenet_cpu"),
(xfail_issue_67415, "test_MSFT_opset8_fp16_inception_v1_onnxzoo_lotus_inception_v1_cpu"),
(xfail_issue_67415, "test_MSFT_opset8_fp16_shufflenet_onnxzoo_lotus_shufflenet_cpu"),
# ONNX Model Zoo
(xfail_issue_39669, "test_onnx_model_zoo_text_machine_comprehension_t5_model_t5_encoder_12_t5_encoder_cpu"),
(xfail_issue_39669, "test_onnx_model_zoo_text_machine_comprehension_t5_model_t5_decoder_with_lm_head_12_t5_decoder_with_lm_head_cpu"),

8
src/bindings/python/tests/test_utils/test_utils.py
View File

@ -1,9 +1,9 @@
# Copyright (C) 2021 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
from openvino.runtime import Function
from openvino.runtime.impl import Shape, Type
from openvino.runtime.impl.op import Parameter
from openvino.runtime import Model
from openvino.runtime import Shape, Type
from openvino.runtime.op import Parameter
import openvino.runtime.opset8 as ops
@ -11,7 +11,7 @@ def get_test_function():
element_type = Type.f32
param = Parameter(element_type, Shape([1, 3, 22, 22]))
relu = ops.relu(param)
func = Function([relu], [param], "test")
func = Model([relu], [param], "test")
assert func is not None
return func

2
src/core/include/openvino/op/i420_to_bgr.hpp
View File

@ -34,6 +34,8 @@ class OPENVINO_API I420toBGR : public util::ConvertColorI420Base {
public:
OPENVINO_OP("I420toBGR", "opset8", util::ConvertColorI420Base);
BWDCMP_RTTI_DECLARATION;
I420toBGR() = default;
/// \brief Constructs a conversion operation from input image in I420 format

2
src/core/include/openvino/op/i420_to_rgb.hpp
View File

@ -34,6 +34,8 @@ class OPENVINO_API I420toRGB : public util::ConvertColorI420Base {
public:
OPENVINO_OP("I420toRGB", "opset8", util::ConvertColorI420Base);
BWDCMP_RTTI_DECLARATION;
I420toRGB() = default;
/// \brief Constructs a conversion operation from input image in I420 format

4
src/core/reference/CMakeLists.txt
View File

@ -22,10 +22,6 @@ ie_faster_build(${TARGET_NAME}
UNITY
PCH PRIVATE "src/precomp.hpp")
if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
target_compile_options(${TARGET_NAME} PUBLIC /wd4146)
endif()
target_compile_definitions(${TARGET_NAME} PRIVATE XBYAK_NO_OP_NAMES XBYAK64)
if(NOT BUILD_SHARED_LIBS)

8
src/core/reference/include/ngraph/runtime/reference/abs.hpp
View File

@ -5,11 +5,17 @@
#pragma once
#include <cstddef>
#include <type_traits>
namespace ngraph {
namespace runtime {
namespace reference {
template <typename T>
template <typename T, typename std::enable_if<std::is_unsigned<T>::value, bool>::type = true>
void abs(const T* arg, T* out, size_t count) {
std::copy(arg, arg + count, out);
}
template <typename T, typename std::enable_if<!std::is_unsigned<T>::value, bool>::type = true>
void abs(const T* arg, T* out, size_t count) {
for (size_t i = 0; i < count; i++) {
// TODO: generic "abs" doesn't work here for some reason.

3
src/core/reference/include/ngraph/runtime/reference/max.hpp
View File

@ -16,8 +16,7 @@ namespace runtime {
namespace reference {
template <typename T>
void max(const T* arg, T* out, const Shape& in_shape, const AxisSet& reduction_axes) {
T minval =
std::numeric_limits<T>::has_infinity ? T(-std::numeric_limits<T>::infinity()) : std::numeric_limits<T>::min();
T minval = std::numeric_limits<T>::lowest();
constexpr bool dont_keep_dims_in_output = false;
const auto out_shape = reduce(in_shape, reduction_axes, dont_keep_dims_in_output);

12
src/core/reference/include/ngraph/runtime/reference/sigmoid.hpp
View File

@ -6,11 +6,21 @@
#include <cmath>
#include <cstddef>
#include <type_traits>
namespace ngraph {
namespace runtime {
namespace reference {
template <typename T>
template <typename T, typename std::enable_if<std::is_integral<T>::value, bool>::type = true>
void sigmoid(const T* arg, T* out, size_t count) {
T exp_value;
for (size_t i = 0; i < count; i++) {
exp_value = std::exp(-static_cast<typename std::make_signed<T>::type>(arg[i]));
out[i] = 1 / (1 + exp_value);
}
}
template <typename T, typename std::enable_if<!std::is_integral<T>::value, bool>::type = true>
void sigmoid(const T* arg, T* out, size_t count) {
T exp_value;
for (size_t i = 0; i < count; i++) {

2
src/core/src/op/i420_to_bgr.cpp
View File

@ -6,6 +6,8 @@
#include "itt.hpp"
BWDCMP_RTTI_DEFINITION(ov::op::v8::I420toBGR);
ov::op::v8::I420toBGR::I420toBGR(const Output<Node>& arg)
: util::ConvertColorI420Base(arg, util::ConvertColorI420Base::ColorConversion::I420_TO_BGR) {
constructor_validate_and_infer_types();

2
src/core/src/op/i420_to_rgb.cpp
View File

@ -6,6 +6,8 @@
#include "itt.hpp"
BWDCMP_RTTI_DEFINITION(ov::op::v8::I420toRGB);
ov::op::v8::I420toRGB::I420toRGB(const Output<Node>& arg)
: util::ConvertColorI420Base(arg, util::ConvertColorI420Base::ColorConversion::I420_TO_RGB) {
constructor_validate_and_infer_types();

3
src/core/src/op/range.cpp
View File

@ -314,7 +314,8 @@ void static check_step(const op::v0::Range* node, T step) {
template <typename T>
static typename std::enable_if<std::is_integral<T>::value, T>::type adjust_for_step_and_sign(T span, T step) {
return ceil_div(span < 0 ? -span : span, step < 0 ? -step : step);
return ceil_div(span < 0 ? -static_cast<typename std::make_signed<T>::type>(span) : span,
step < 0 ? -static_cast<typename std::make_signed<T>::type>(step) : step);
}
template <typename T>

2
src/core/tests/frontend/shared/CMakeLists.txt
View File

@ -11,7 +11,7 @@ add_library(${TARGET_NAME} STATIC EXCLUDE_FROM_ALL ${LIBRARY_SRC} ${LIBRARY_HEAD
target_include_directories(${TARGET_NAME} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
target_include_directories(${TARGET_NAME} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/../..)
target_link_libraries(${TARGET_NAME} PUBLIC frontend_common interpreter_backend engines_test_util
target_link_libraries(${TARGET_NAME} PUBLIC frontend_common engines_test_util
ngraph cnpy commonTestUtils ngraph_test_util openvino::util)
target_compile_definitions(${TARGET_NAME}

56
src/core/tests/models/onnx/softmax_axis_1_opset11.prototxt Normal file
View File

@ -0,0 +1,56 @@
ir_version: 3
producer_name: "nGraph ONNX Importer"
graph {
node {
input: "x"
output: "y"
op_type: "Softmax"
attribute {
name: "axis"
i: 1
type: INT
}
}
name: "test_softmax_axis_1"
input {
name: "x"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
output {
name: "y"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
}
opset_import {
version: 11
}

56
src/core/tests/models/onnx/softmax_axis_negative_1_opset11.prototxt Normal file
View File

@ -0,0 +1,56 @@
ir_version: 3
producer_name: "nGraph ONNX Importer"
graph {
node {
input: "x"
output: "y"
op_type: "Softmax"
attribute {
name: "axis"
i: -1
type: INT
}
}
name: "test_softmax_axis_0"
input {
name: "x"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
output {
name: "y"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
}
opset_import {
version: 11
}

56
src/core/tests/models/onnx/softmax_axis_negative_1_opset13.prototxt Normal file
View File

@ -0,0 +1,56 @@
ir_version: 3
producer_name: "nGraph ONNX Importer"
graph {
node {
input: "x"
output: "y"
op_type: "Softmax"
attribute {
name: "axis"
i: -1
type: INT
}
}
name: "test_softmax_axis_0"
input {
name: "x"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
output {
name: "y"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 3
}
dim {
dim_value: 4
}
dim {
dim_value: 5
}
}
}
}
}
}
opset_import {
version: 13
}

168
src/core/tests/onnx/onnx_import.in.cpp
View File

@ -380,7 +380,7 @@ NGRAPH_TEST(${BACKEND_NAME}, onnx_model_initializer_wo_input) {
test_case.run();
}
NGRAPH_TEST(onnx_${BACKEND_NAME}, onnx_expand_function) {
NGRAPH_TEST(${BACKEND_NAME}, onnx_expand_function) {
const auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/quantization/dynamicquantizelinear.onnx"));
@ -392,7 +392,7 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, onnx_expand_function) {
test_case.run();
}
NGRAPH_TEST(onnx_${BACKEND_NAME}, onnx_expand_function_dependency_to_created_subgraph) {
NGRAPH_TEST(${BACKEND_NAME}, onnx_expand_function_dependency_to_created_subgraph) {
const auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/transformations/greater_or_equal.onnx"));
@ -403,7 +403,7 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, onnx_expand_function_dependency_to_created_sub
test_case.run();
}
NGRAPH_TEST(onnx_${BACKEND_NAME}, onnx_expand_context_dependent_function) {
NGRAPH_TEST(${BACKEND_NAME}, onnx_expand_context_dependent_function) {
auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/transformations/softmax_crossentropy_consumed.onnx"));
@ -690,19 +690,24 @@ NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_1D) {
}
namespace {
// common input for all Softmax 3D test cases (Shape = {3,4,5})
// clang-format off
const std::vector<float> SOFTMAX_INPUT = {
2.75793882, -0.50841322, 0.82013929, -0.62409912, -0.96136118, 0.21004745, 1.38337255,
1.19030397, 2.0940445, -0.03551657, -0.78686039, 1.992782, 0.04300319, -0.29230777,
-0.56797112, -1.26732165, -0.61935399, 0.57670432, 0.92844898, 2.82469233,
2.75793882, -0.50841322, 0.82013929, -0.62409912, -0.96136118,
0.21004745, 1.38337255, 1.19030397, 2.0940445, -0.03551657,
-0.78686039, 1.992782, 0.04300319, -0.29230777, -0.56797112,
-1.26732165, -0.61935399, 0.57670432, 0.92844898, 2.82469233,
0.98721677, -0.05100663, -1.21178917, -0.17530157, 1.40051805, -0.13259761, -1.14313018,
0.2673723, -0.87996154, 1.29053106, 1.55, 0.8396538, 1.20729817, 0.23727845,
-0.89113606, -1.70909842, 0.26460363, -0.70566808, 2.383518, 1.07024615,
0.98721677, -0.05100663, -1.21178917, -0.17530157, 1.40051805,
-0.13259761, -1.14313018, 0.2673723, -0.87996154, 1.29053106,
1.55, 0.8396538, 1.20729817, 0.23727845, -0.89113606,
-1.70909842, 0.26460363, -0.70566808, 2.383518, 1.07024615,
-1.21722605, 0.82919357, 0.55765697, 0.12657686, 0.63432172, 0.75425957, -2.43721014,
-1.24478184, 2.65316853, 1.19509542, -0.95523998, 0.5149006, -0.01151649, 0.68327026,
-0.4589638, -0.46554745, 0.21055324, 0.39266729, 2.05098086, 1.83207919};
-1.21722605, 0.82919357, 0.55765697, 0.12657686, 0.63432172,
0.75425957, -2.43721014, -1.24478184, 2.65316853, 1.19509542,
-0.95523998, 0.5149006, -0.01151649, 0.68327026, -0.4589638,
-0.46554745, 0.21055324, 0.39266729, 2.05098086, 1.83207919};
} // namespace
// clang-format on
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_0) {
auto function = onnx_import::import_onnx_model(file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_axis_0.onnx"));
@ -710,19 +715,24 @@ NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_0) {
auto test_case = test::TestCase(function, s_device);
test_case.add_input<float>(SOFTMAX_INPUT);
// clang-format off
test_case.add_expected_output<float>(
Shape{3, 4, 5},
{0.09683057, 0.00369363, 0.01394559, 0.00329012, 0.00234823, 0.00757665, 0.02449322,
0.02019284, 0.04985249, 0.00592694, 0.00279593, 0.04505148, 0.00641108, 0.00458466,
0.00348007, 0.00172928, 0.00330577, 0.01093237, 0.01554086, 0.10351497,
{0.09683057, 0.00369363, 0.01394559, 0.00329012, 0.00234823,
0.00757665, 0.02449322, 0.02019284, 0.04985249, 0.00592694,
0.00279593, 0.04505148, 0.00641108, 0.00458466, 0.00348007,
0.00172928, 0.00330577, 0.01093237, 0.01554086, 0.10351497,
0.01648154, 0.00583583, 0.00182802, 0.00515374, 0.02491679, 0.00537859, 0.00195794,
0.00802367, 0.00254737, 0.0223216, 0.02893419, 0.0142204, 0.02053893, 0.00778581,
0.00251907, 0.00111174, 0.00800149, 0.0030324, 0.06658917, 0.0179084,
0.01648154, 0.00583583, 0.00182802, 0.00515374, 0.02491679,
0.00537859, 0.00195794, 0.00802367, 0.00254737, 0.0223216,
0.02893419, 0.0142204, 0.02053893, 0.00778581, 0.00251907,
0.00111174, 0.00800149, 0.0030324, 0.06658917, 0.0179084,
0.00181811, 0.01407243, 0.01072611, 0.0069699, 0.01158077, 0.01305647, 0.00053677,
0.0017687, 0.08719896, 0.02028982, 0.00236265, 0.01027717, 0.0060709, 0.01216173,
0.00388087, 0.00385541, 0.00758048, 0.00909469, 0.04775123, 0.03836337});
0.00181811, 0.01407243, 0.01072611, 0.0069699, 0.01158077,
0.01305647, 0.00053677, 0.0017687, 0.08719896, 0.02028982,
0.00236265, 0.01027717, 0.0060709, 0.01216173, 0.00388087,
0.00385541, 0.00758048, 0.00909469, 0.04775123, 0.03836337});
// clang-format on
test_case.run(6);
}
@ -733,35 +743,113 @@ NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_1) {
auto test_case = test::TestCase(function, s_device);
test_case.add_input<float>(SOFTMAX_INPUT);
// clang-format off
test_case.add_expected_output<float>(
Shape{3, 4, 5},
{0.22757064, 0.00868076, 0.03277484, 0.00773243, 0.0055188, 0.0178066, 0.05756383,
0.04745709, 0.11716303, 0.01392945, 0.00657097, 0.10587974, 0.01506727, 0.01077484,
0.00817884, 0.00406413, 0.00776921, 0.0256932, 0.03652405, 0.24328028,
{0.22757064, 0.00868076, 0.03277484, 0.00773243, 0.0055188,
0.0178066, 0.05756383, 0.04745709, 0.11716303, 0.01392945,
0.00657097, 0.10587974, 0.01506727, 0.01077484, 0.00817884,
0.00406413, 0.00776921, 0.0256932, 0.03652405, 0.24328028,
0.06217413, 0.02201481, 0.00689594, 0.01944171, 0.09399488, 0.02028993, 0.00738604,
0.03026811, 0.00960958, 0.08420492, 0.10914991, 0.05364435, 0.07748005, 0.02937079,
0.0095028, 0.00419387, 0.03018442, 0.01143929, 0.2511977, 0.06755678,
0.06217413, 0.02201481, 0.00689594, 0.01944171, 0.09399488,
0.02028993, 0.00738604, 0.03026811, 0.00960958, 0.08420492,
0.10914991, 0.05364435, 0.07748005, 0.02937079, 0.0095028,
0.00419387, 0.03018442, 0.01143929, 0.2511977, 0.06755678,
0.00587593, 0.04548053, 0.0346656, 0.02252594, 0.03742775, 0.04219705, 0.00173478,
0.00571623, 0.2818174, 0.06557446, 0.00763582, 0.03321466, 0.01962049, 0.03930537,
0.01254255, 0.01246025, 0.02449929, 0.02939305, 0.15432668, 0.12398617});
0.00587593, 0.04548053, 0.0346656, 0.02252594, 0.03742775,
0.04219705, 0.00173478, 0.00571623, 0.2818174, 0.06557446,
0.00763582, 0.03321466, 0.01962049, 0.03930537, 0.01254255,
0.01246025, 0.02449929, 0.02939305, 0.15432668, 0.12398617});
// clang-format on
test_case.run(4);
}
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_invalid_axis_1D) {
ASSERT_THROW(
onnx_import::import_onnx_model(file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_invalid_axis_1D.onnx")),
ngraph::ngraph_error)
<< "Softmax model with invalid axis was successfully imported while it should have thrown.";
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_1_opset11) {
auto function =
onnx_import::import_onnx_model(file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_axis_1_opset11.onnx"));
auto test_case = test::TestCase(function, s_device);
test_case.add_input<float>(SOFTMAX_INPUT);
// clang-format off
test_case.add_expected_output<float>(
Shape{3, 4, 5},
{0.88890495, 0.04825497, 0.27088348, 0.04490523, 0.02037154,
0.06955369, 0.31998834, 0.39223197, 0.68041159, 0.05141776,
0.02566661, 0.5885689, 0.12453075, 0.06257374, 0.03019055,
0.01587475, 0.0431878, 0.21235381, 0.21210944, 0.89802015,
0.31752626, 0.19442629, 0.0546935, 0.06279221, 0.36823282,
0.10362164, 0.06523066, 0.24006419, 0.03103672, 0.32987983,
0.55743381, 0.473766, 0.61451431, 0.09486084, 0.03722801,
0.02141829, 0.26657706, 0.090728, 0.81131024, 0.26465935,
0.08619648, 0.43343993, 0.3877785, 0.04523505, 0.15625437,
0.61900597, 0.01653285, 0.06394322, 0.56592636, 0.27376196,
0.11201305, 0.31654337, 0.21947994, 0.07893034, 0.05236297,
0.18278451, 0.23348385, 0.32879834, 0.30990825, 0.5176207});
// clang-format on
test_case.run(4);
}
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_invalid_axis_3D) {
ASSERT_THROW(
onnx_import::import_onnx_model(file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_invalid_axis_3D.onnx")),
ngraph::ngraph_error)
<< "Softmax model with invalid axis was successfully imported while it should have thrown.";
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_negative_1_opset11) {
auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_axis_negative_1_opset11.onnx"));
auto test_case = test::TestCase(function);
test_case.add_input<float>(SOFTMAX_INPUT);
// clang-format off
test_case.add_expected_output<float>(
Shape{3, 4, 5},
{0.88890495, 0.04825497, 0.27088348, 0.04490523, 0.02037154,
0.06955369, 0.31998834, 0.39223197, 0.68041159, 0.05141776,
0.02566661, 0.5885689, 0.12453075, 0.06257374, 0.03019055,
0.01587475, 0.0431878, 0.21235381, 0.21210944, 0.89802015,
0.31752626, 0.19442629, 0.0546935, 0.06279221, 0.36823282,
0.10362164, 0.06523066, 0.24006419, 0.03103672, 0.32987983,
0.55743381, 0.473766, 0.61451431, 0.09486084, 0.03722801,
0.02141829, 0.26657706, 0.090728, 0.81131024, 0.26465935,
0.08619648, 0.43343993, 0.3877785, 0.04523505, 0.15625437,
0.61900597, 0.01653285, 0.06394322, 0.56592636, 0.27376196,
0.11201305, 0.31654337, 0.21947994, 0.07893034, 0.05236297,
0.18278451, 0.23348385, 0.32879834, 0.30990825, 0.5176207});
// clang-format on
test_case.run(6);
}
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_softmax_axis_negative_1_opset13) {
auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/softmax_axis_negative_1_opset13.onnx"));
auto test_case = test::TestCase(function);
test_case.add_input<float>(SOFTMAX_INPUT);
// clang-format off
test_case.add_expected_output<float>(
Shape{3, 4, 5},
{0.88890495, 0.04825497, 0.27088348, 0.04490523, 0.02037154,
0.06955369, 0.31998834, 0.39223197, 0.68041159, 0.05141776,
0.02566661, 0.5885689, 0.12453075, 0.06257374, 0.03019055,
0.01587475, 0.0431878, 0.21235381, 0.21210944, 0.89802015,
0.31752626, 0.19442629, 0.0546935, 0.06279221, 0.36823282,
0.10362164, 0.06523066, 0.24006419, 0.03103672, 0.32987983,
0.55743381, 0.473766, 0.61451431, 0.09486084, 0.03722801,
0.02141829, 0.26657706, 0.090728, 0.81131024, 0.26465935,
0.08619648, 0.43343993, 0.3877785, 0.04523505, 0.15625437,
0.61900597, 0.01653285, 0.06394322, 0.56592636, 0.27376196,
0.11201305, 0.31654337, 0.21947994, 0.07893034, 0.05236297,
0.18278451, 0.23348385, 0.32879834, 0.30990825, 0.5176207});
// clang-format on
test_case.run(6);
}
NGRAPH_TEST(${BACKEND_NAME}, onnx_model_sub) {

16
src/frontends/onnx/frontend/src/core/graph.cpp
View File

@ -199,9 +199,10 @@ void Graph::decode_to_framework_nodes() {
if (node.has_subgraphs()) {
const auto& subgraphs = node.get_subgraphs();
auto inputs = node.get_ng_inputs();
std::vector<std::shared_ptr<Function>> functions;
for (const auto& kv : subgraphs) {
auto& subgraph = kv.second;
subgraph->decode();
functions.push_back(subgraph->decode());
for (const auto& input : subgraph->get_inputs_from_parent()) {
const auto& name = input.get_node()->get_friendly_name();
if (std::find_if(inputs.begin(), inputs.end(), [&name](const Output<ngraph::Node>& n) -> bool {
@ -211,10 +212,9 @@ void Graph::decode_to_framework_nodes() {
}
}
}
framework_node =
std::make_shared<ngraph::frontend::ONNXSubgraphFrameworkNode>(shared_from_this(), node, inputs);
framework_node = std::make_shared<frontend::ONNXSubgraphFrameworkNode>(node, functions, inputs);
} else {
framework_node = std::make_shared<ngraph::frontend::ONNXFrameworkNode>(shared_from_this(), node);
framework_node = std::make_shared<frontend::ONNXFrameworkNode>(node);
}
OutputVector ng_nodes{framework_node->outputs()};
set_friendly_names(node, ng_nodes);
@ -240,7 +240,10 @@ std::shared_ptr<Function> Graph::create_function() {
std::shared_ptr<Function> Graph::decode() {
decode_to_framework_nodes();
return create_function();
auto function = create_function();
auto& rt_info = function->get_rt_info();
rt_info[ONNX_GRAPH_RT_ATTRIBUTE] = shared_from_this();
return function;
}
bool Graph::is_ng_node_in_cache(const std::string& name) const {
@ -399,7 +402,8 @@ void Subgraph::find_inputs_from_parent() {
for (const auto& out_name : node_proto.output()) {
if (m_cache->contains(out_name)) {
auto node_to_replace_input = m_cache->get_node(out_name).get_node();
if (!dynamic_cast<op::util::MultiSubGraphOp*>(node_to_replace_input))
if (!ov::is_type<op::util::MultiSubGraphOp>(node_to_replace_input) &&
!ov::is_type<frontend::ONNXSubgraphFrameworkNode>(node_to_replace_input))
continue;
auto inputs = node_to_replace_input->input_values();
for (size_t i = 0; i < inputs.size(); i++) {

2
src/frontends/onnx/frontend/src/core/graph.hpp
View File

@ -121,6 +121,8 @@ inline std::ostream& operator<<(std::ostream& outs, const Graph& graph) {
return (outs << "<Graph: " << graph.get_name() << ">");
}
static const char* const ONNX_GRAPH_RT_ATTRIBUTE = "onnx_graph";
} // namespace onnx_import
} // namespace ngraph

6
src/frontends/onnx/frontend/src/onnx_framework_node.cpp
View File

@ -21,10 +21,14 @@ namespace frontend {
NGRAPH_RTTI_DEFINITION(ONNXFrameworkNode, "ONNXFrameworkNode", 1);
std::shared_ptr<Node> ONNXFrameworkNode::clone_with_new_inputs(const OutputVector& inputs) const {
return std::make_shared<ONNXFrameworkNode>(m_graph, m_node, inputs);
return std::make_shared<ONNXFrameworkNode>(m_node, inputs);
}
NGRAPH_RTTI_DEFINITION(ONNXSubgraphFrameworkNode, "ONNXSubgraphFrameworkNode", 1);
std::shared_ptr<Node> ONNXSubgraphFrameworkNode::clone_with_new_inputs(const OutputVector& inputs) const {
return std::make_shared<ONNXSubgraphFrameworkNode>(m_node, m_functions, inputs);
}
} // namespace frontend
} // namespace ngraph

42
src/frontends/onnx/frontend/src/onnx_framework_node.hpp
View File

@ -38,20 +38,16 @@ class ONNXFrameworkNode : public ov::op::util::FrameworkNode {
public:
NGRAPH_RTTI_DECLARATION;
ONNXFrameworkNode(std::shared_ptr<onnx_import::Graph> graph, const onnx_import::Node& node)
ONNXFrameworkNode(const onnx_import::Node& node)
: ov::op::util::FrameworkNode(node.get_ng_inputs(), node.get_outputs_size()),
m_node(node),
m_graph(graph) {}
m_node(node) {}
ONNXFrameworkNode(std::shared_ptr<onnx_import::Graph> graph,
const onnx_import::Node& node,
const OutputVector& inputs)
ONNXFrameworkNode(const onnx_import::Node& node, const OutputVector& inputs)
: ov::op::util::FrameworkNode(inputs, node.get_outputs_size()),
m_node(node),
m_graph(graph) {}
m_node(node) {}
OutputVector get_ng_nodes() const {
OutputVector ng_nodes{m_graph->make_ng_nodes(m_node)};
OutputVector get_ng_nodes(const std::shared_ptr<onnx_import::Graph>& graph) const {
OutputVector ng_nodes{graph->make_ng_nodes(m_node)};
if (ng_nodes.size() > get_output_size()) {
ng_nodes.resize(get_output_size());
}
@ -71,35 +67,31 @@ public:
protected:
onnx_import::Node m_node;
private:
std::shared_ptr<onnx_import::Graph> m_graph;
};
class ONNXSubgraphFrameworkNode : public ONNXFrameworkNode {
public:
NGRAPH_RTTI_DECLARATION;
ONNXSubgraphFrameworkNode(std::shared_ptr<onnx_import::Graph> graph,
const onnx_import::Node& node,
ONNXSubgraphFrameworkNode(const onnx_import::Node& node,
const std::vector<std::shared_ptr<Function>>& functions,
const OutputVector& inputs)
: ONNXFrameworkNode(graph, node, inputs) {}
: ONNXFrameworkNode(node, inputs),
m_functions(functions) {}
void infer_inputs_from_parent() {
for (auto& subgraph : m_node.get_subgraphs())
subgraph.second->infer_inputs_from_parent();
}
std::vector<std::shared_ptr<Function>> get_subgraph_functions() const {
std::vector<std::shared_ptr<Function>> ret;
for (const auto& kv : m_node.get_subgraphs()) {
auto& subgraph = kv.second;
ret.push_back(std::make_shared<Function>(subgraph->get_ng_outputs(),
subgraph->get_ng_parameters(),
subgraph->get_name()));
}
return ret;
const std::vector<std::shared_ptr<Function>>& get_subgraph_functions() const {
return m_functions;
}
virtual std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& inputs) const override;
private:
std::vector<std::shared_ptr<Function>> m_functions;
};
} // namespace frontend

25
src/frontends/onnx/frontend/src/op/softmax.cpp
View File

@ -37,17 +37,8 @@ OutputVector softmax(const Node& node) {
result = default_opset::Constant::create(data.get_element_type(), Shape{}, {1});
break;
}
case 1: {
// checks if the axis belongs to the allowed values set (-1 and 0 for 1D)
ngraph::normalize_axis(node.get_description(), axis, data.get_partial_shape().rank());
result = std::make_shared<default_opset::Softmax>(data, 0);
break;
}
default: {
const auto normalized_axis =
ngraph::normalize_axis(node.get_description(), axis, data.get_partial_shape().rank());
result = onnx_softmax(data, normalized_axis);
result = onnx_softmax(data, axis);
break;
}
}
@ -69,17 +60,8 @@ OutputVector softmax(const Node& node) {
result = default_opset::Constant::create(data.get_element_type(), Shape{}, {1});
break;
}
case 1: {
// checks if the axis belongs to the allowed values set (-1 and 0 for 1D)
ngraph::normalize_axis(node.get_description(), axis, data.get_partial_shape().rank());
result = std::make_shared<default_opset::Softmax>(data, 0);
break;
}
default: {
const auto normalized_axis =
ngraph::normalize_axis(node.get_description(), axis, data.get_partial_shape().rank());
result = std::make_shared<default_opset::Softmax>(data, normalized_axis);
result = std::make_shared<ov::op::v8::Softmax>(data, axis);
break;
}
}
@ -92,9 +74,8 @@ OutputVector softmax(const Node& node) {
const auto data = node.get_ng_inputs().at(0);
const auto axis = node.get_attribute_value<int64_t>("axis", -1);
const auto normalized_axis = ngraph::normalize_axis(node.get_description(), axis, data.get_partial_shape().rank());
return {std::make_shared<default_opset::Softmax>(data, normalized_axis)};
return {std::make_shared<ov::op::v8::Softmax>(data, axis)};
}
} // namespace set_13
} // namespace op

9
src/frontends/onnx/frontend/src/utils/onnx_internal.cpp
View File

@ -60,6 +60,12 @@ void apply_transformations(ONNX_NAMESPACE::ModelProto& model_proto, const std::s
} // namespace
void convert_decoded_function(std::shared_ptr<Function> function) {
auto& rt_info = function->get_rt_info();
auto it = rt_info.find(ONNX_GRAPH_RT_ATTRIBUTE);
OPENVINO_ASSERT(it != rt_info.end(),
"Could not find '" + std::string(ONNX_GRAPH_RT_ATTRIBUTE) +
"' attribute in decoded model. Model probably wasn't created by FrontEnd::decode function.");
auto onnx_graph = it->second.as<std::shared_ptr<onnx_import::Graph>>();
for (const auto& node : function->get_ordered_ops()) {
if (auto raw_node = std::dynamic_pointer_cast<frontend::ONNXFrameworkNode>(node)) {
if (auto subgraph_node = std::dynamic_pointer_cast<frontend::ONNXSubgraphFrameworkNode>(node)) {
@ -68,7 +74,7 @@ void convert_decoded_function(std::shared_ptr<Function> function) {
convert_decoded_function(function);
}
}
auto ng_nodes = raw_node->get_ng_nodes();
auto ng_nodes = raw_node->get_ng_nodes(onnx_graph);
replace_node(raw_node, ng_nodes);
} else {
// Have to revalidate node because new intpus can affect shape/type
@ -76,6 +82,7 @@ void convert_decoded_function(std::shared_ptr<Function> function) {
node->revalidate_and_infer_types();
}
}
rt_info.erase(it);
detail::remove_dangling_parameters(function);
detail::remove_dangling_results(function);
}

8
src/inference/include/ie/multi-device/multi_device_config.hpp
View File

@ -34,5 +34,13 @@ namespace MultiDeviceConfigParams {
*/
DECLARE_MULTI_CONFIG_KEY(DEVICE_PRIORITIES);
/**
* @brief network priority config option, the range of value is from 0 to the max integer,
* when there are multi devices, the value is smaller, the priority is higher,
* 0 is the highest priority. Auto plugin dispatch the network to device
* according to priority value. when all devices are free, even if the priority value
* is not 0, the network will be dispatched to the strongest device.
*/
DECLARE_CONFIG_KEY(AUTO_NETWORK_PRIORITY);
} // namespace MultiDeviceConfigParams
} // namespace InferenceEngine

16
src/inference/src/cpp_interfaces/interface/ie_iplugin_internal.cpp
View File

@ -309,16 +309,15 @@ void IInferencePlugin::SetExeNetworkInfo(const std::shared_ptr<IExecutableNetwor
OPENVINO_ASSERT(outputsInfo.size() == function->get_output_size());
for (const auto& param : function->get_parameters()) {
auto new_param = param->copy_with_new_inputs({});
auto new_param = ov::as_type_ptr<ov::op::v0::Parameter>(param->copy_with_new_inputs({}));
new_param->set_friendly_name(param->get_friendly_name());
if (add_operation_names)
new_param->output(0).get_tensor().add_names({new_param->get_friendly_name()});
// WA: use CNNNetwork's precisions since plugins sometimes override their precisions
// after transformation pipeline is run
new_param->set_output_type(
0,
InferenceEngine::details::convertPrecision(inputsInfo.at(new_param->get_friendly_name())->getPrecision()),
new_param->get_output_partial_shape(0));
new_param->set_element_type(
InferenceEngine::details::convertPrecision(inputsInfo.at(new_param->get_friendly_name())->getPrecision()));
new_param->validate_and_infer_types();
const_params.emplace_back(new_param);
}
for (const auto& result : function->get_results()) {
@ -326,10 +325,9 @@ void IInferencePlugin::SetExeNetworkInfo(const std::shared_ptr<IExecutableNetwor
result->get_output_partial_shape(0));
const std::string param_name = ngraph::op::util::create_ie_output_name(result->input_value(0));
fake_param->set_friendly_name(param_name);
fake_param->set_output_type(
0,
InferenceEngine::details::convertPrecision(outputsInfo.at(param_name)->getPrecision()),
fake_param->get_output_partial_shape(0));
fake_param->set_element_type(
InferenceEngine::details::convertPrecision(outputsInfo.at(param_name)->getPrecision()));
fake_param->validate_and_infer_types();
auto new_result = result->copy_with_new_inputs({fake_param});
new_result->set_friendly_name(result->get_friendly_name());
if (add_operation_names) {

72
src/plugins/auto/executable_network.cpp
View File

@ -149,11 +149,13 @@ MultiDeviceExecutableNetwork::MultiDeviceExecutableNetwork(const std::string&
const std::vector<DeviceInformation>& metaDevices,
const std::string& strDevices,
MultiDeviceInferencePlugin* plugin,
const bool needPerfCounters)
const AutoContext& context,
const bool needPerfCounters)
: _devicePriorities{metaDevices}
, _devicePrioritiesInitial{metaDevices}
, _needPerfCounters(needPerfCounters)
, _multiPlugin(plugin)
, _context(context)
, _workModeIsAUTO(true) {
if (_multiPlugin->GetCore() == nullptr) {
IE_THROW() << "Please, work with MULTI device via InferencEngine::Core object";
@ -173,7 +175,8 @@ MultiDeviceExecutableNetwork::MultiDeviceExecutableNetwork(const std::string&
_loadContext[ACTUALDEVICE].isEnabled = true;
_loadContext[ACTUALDEVICE].networkPrecision = GetNetworkPrecision(network);
_loadContext[ACTUALDEVICE].metaDevices = metaDevices;
_loadContext[ACTUALDEVICE].deviceInfo = _multiPlugin->SelectDevice(metaDevices, _loadContext[ACTUALDEVICE].networkPrecision);
_loadContext[ACTUALDEVICE].deviceInfo = _multiPlugin->SelectDevice(metaDevices,
_loadContext[ACTUALDEVICE].networkPrecision, _context.modelPriority);
LOG_INFO("[AUTOPLUGIN]:select device:%s", _loadContext[ACTUALDEVICE].deviceInfo.deviceName.c_str());
bool isActualDevCPU =
_loadContext[ACTUALDEVICE].deviceInfo.deviceName.find("CPU") != std::string::npos;
@ -292,6 +295,13 @@ void MultiDeviceExecutableNetwork::TryToLoadNetWork(AutoLoadContext& context,
return;
}
// need to reload network, unregister it's priority
// there maybe potential issue.
// for example they are dGPU, VPUX, iGPU, customer want to LoadNetwork with
// configure 0 dGPU, 1 VPUX, if dGPU load failed,
// the result will be not sure, maybe two network are loaded into VPUX,
// maybe 0 is loaded to VPUX, 1 is loaded to iGPU
_multiPlugin->UnregisterPriority(_context.modelPriority, context.deviceInfo.uniqueName);
// remove the current device from deviceList
auto eraseDevice = std::find_if(deviceList.begin(), deviceList.end(),
[device](DeviceInformation& d){
@ -305,7 +315,8 @@ void MultiDeviceExecutableNetwork::TryToLoadNetWork(AutoLoadContext& context,
// select next candidate device
try {
context.deviceInfo = _multiPlugin->SelectDevice(deviceList, context.networkPrecision);
context.deviceInfo = _multiPlugin->SelectDevice(deviceList,
context.networkPrecision, _context.modelPriority);
}
catch (const std::exception& e) {
return;
@ -382,7 +393,7 @@ void MultiDeviceExecutableNetwork::WaitActualNetworkReady() const {
// for every MultiDeviceExecutableNetwork instance
std::call_once(_oc, [this] () {
if (_loadContext[ACTUALDEVICE].future.valid()) {
_loadContext[ACTUALDEVICE].future.get();
_loadContext[ACTUALDEVICE].future.wait();
}
// if _loadContext[ACTUALDEVICE] load failed, fall back to _loadContext[CPU]
if (!_loadContext[ACTUALDEVICE].isAlready) {
@ -460,13 +471,17 @@ void MultiDeviceExecutableNetwork::run(Task inferPipelineTask) {
}
MultiDeviceExecutableNetwork::~MultiDeviceExecutableNetwork() {
// this is necessary to guarantee member destroyed after getting future
if (_workModeIsAUTO && _loadContext[CPU].isEnabled) {
_loadContext[CPU].future.get();
WaitActualNetworkReady();
// it's necessary to wait the loading network threads to stop here.
InferenceEngine::ExecutorManager::getInstance()->clear("AutoDeviceAsyncLoad");
_executor.reset();
if (_workModeIsAUTO) {
// this is necessary to guarantee member destroyed after getting future
if (_loadContext[CPU].isEnabled) {
_loadContext[CPU].future.wait();
WaitActualNetworkReady();
// it's necessary to wait the loading network threads to stop here.
InferenceEngine::ExecutorManager::getInstance()->clear("AutoDeviceAsyncLoad");
_executor.reset();
}
_multiPlugin->UnregisterPriority(_context.modelPriority,
_loadContext[ACTUALDEVICE].deviceInfo.uniqueName);
}
{
std::lock_guard<std::mutex> lock(_mutex);
@ -615,33 +630,32 @@ void MultiDeviceExecutableNetwork::SetConfig(const std::map<std::string, Inferen
_devicePriorities = metaDevices;
// update value in config
_confMutex.lock();
std::lock_guard<std::mutex> lockConf(_confMutex);
_config[MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES] = priorities->second;
_confMutex.unlock();
}
}
}
InferenceEngine::Parameter MultiDeviceExecutableNetwork::GetConfig(const std::string &name) const {
_confMutex.lock();
auto it = _config.find(name);
if (it != _config.end()) {
_confMutex.unlock();
return it->second;
} else {
_confMutex.unlock();
// find config key among networks config keys
for (const auto& desc : _networksPerDevice) {
const auto& execNetwork = desc.second;
auto param = execNetwork->GetMetric(METRIC_KEY(SUPPORTED_CONFIG_KEYS));
for (auto &&configKey : param.as<std::vector<std::string>>()) {
if (configKey == name) {
return execNetwork->GetConfig(configKey);
}
{
std::lock_guard<std::mutex> lock(_confMutex);
auto it = _config.find(name);
if (it != _config.end()) {
return it->second;
}
}
// find config key among networks config keys
for (const auto& desc : _networksPerDevice) {
const auto& execNetwork = desc.second;
auto param = execNetwork->GetMetric(METRIC_KEY(SUPPORTED_CONFIG_KEYS));
for (auto &&configKey : param.as<std::vector<std::string>>()) {
if (configKey == name) {
return execNetwork->GetConfig(configKey);
}
}
IE_THROW(NotFound) << name <<" not found in the ExecutableNetwork config";
}
IE_THROW(NotFound) << name <<" not found in the ExecutableNetwork config";
}
InferenceEngine::Parameter MultiDeviceExecutableNetwork::GetMetric(const std::string &name) const {

8
src/plugins/auto/executable_network.hpp
View File

@ -43,6 +43,12 @@ struct DeviceInformation {
std::map<std::string, std::string> config;
int numRequestsPerDevices;
std::string defaultDeviceID;
DeviceName uniqueName;
};
struct AutoContext {
bool needPerfCounters = {false};
unsigned int modelPriority = 0;
};
struct AutoLoadContext {
@ -153,6 +159,7 @@ public:
const std::vector<DeviceInformation>& metaDevices,
const std::string& strDevices,
MultiDeviceInferencePlugin* plugin,
const AutoContext& context,
const bool needPerfCounters = false);
void SetConfig(const std::map<std::string, InferenceEngine::Parameter> &config) override;
@ -202,6 +209,7 @@ private:
std::shared_ptr<InferenceEngine::ICore> _core;
InferenceEngine::IStreamsExecutor::Ptr _executor;
MultiDeviceInferencePlugin* _multiPlugin;
AutoContext _context;
bool _workModeIsAUTO = {false};
mutable std::once_flag _oc;
std::once_flag _firstLoadOC;

238
src/plugins/auto/plugin.cpp
View File

@ -60,10 +60,15 @@ namespace {
res.push_back(CONFIG_KEY_INTERNAL(MULTI_WORK_MODE_AS_AUTO));
res.push_back(PluginConfigParams::KEY_PERF_COUNT);
res.push_back(PluginConfigParams::KEY_EXCLUSIVE_ASYNC_REQUESTS);
res.push_back(MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY);
return res;
}();
} // namespace
std::mutex MultiDeviceInferencePlugin::_mtx;
std::map<unsigned int, std::list<std::string>> MultiDeviceInferencePlugin::_priorityMap;
std::map<std::string, std::string> MultiDeviceInferencePlugin::GetSupportedConfig(
const std::map<std::string, std::string> & config, const std::string & deviceName) const {
std::vector<std::string> supportedConfigKeys = GetCore()->GetMetric(deviceName, METRIC_KEY(SUPPORTED_CONFIG_KEYS));
@ -137,11 +142,31 @@ std::vector<DeviceInformation> MultiDeviceInferencePlugin::ParseMetaDevices(cons
std::string defaultDeviceID = "";
DeviceIDParser parsed{deviceName};
if (parsed.getDeviceID().empty())
std::string deviceid = parsed.getDeviceID();
if (deviceid.empty()) {
defaultDeviceID = getDefaultDeviceID(deviceName);
deviceid = defaultDeviceID;
}
std::string fullDeviceName = "";
std::string uniqueName = "";
if (parsed.getDeviceName() == "GPU") {
std::vector<std::string> supportedMetrics = GetCore()->GetMetric(deviceName, METRIC_KEY(SUPPORTED_METRICS));
if (std::find(supportedMetrics.begin(), supportedMetrics.end(), METRIC_KEY(FULL_DEVICE_NAME)) != supportedMetrics.end()) {
fullDeviceName = GetCore()->GetMetric(deviceName, METRIC_KEY(FULL_DEVICE_NAME)).as<std::string>();
}
}
if (fullDeviceName.empty()) {
uniqueName = parsed.getDeviceName() + "_" + deviceid;
} else {
uniqueName = fullDeviceName + "_" + deviceid;
}
LOG_DEBUG("deviceName:%s, defaultDeviceID:%s, uniqueName:%s",
deviceName.c_str(), defaultDeviceID.c_str(), uniqueName.c_str());
// create meta device
metaDevices.push_back({ deviceName, getDeviceConfig(deviceName), numRequests, defaultDeviceID });
metaDevices.push_back({ deviceName, getDeviceConfig(deviceName), numRequests, defaultDeviceID, uniqueName});
}
return metaDevices;
@ -162,10 +187,9 @@ InferenceEngine::Parameter MultiDeviceInferencePlugin::GetConfig(const std::stri
}
void MultiDeviceInferencePlugin::SetConfig(const std::map<std::string, std::string> & config) {
bool needPerfCounters = false;
AutoContext context;
std::map<std::string, std::string> filterConfig;
CheckConfig(config, needPerfCounters, filterConfig);
CheckConfig(config, context, filterConfig);
for (auto && kvp : config) {
const auto& name = kvp.first;
_config[name] = kvp.second;
@ -237,10 +261,11 @@ IExecutableNetworkInternal::Ptr MultiDeviceInferencePlugin::LoadNetworkImpl(cons
if (workModeAuto) {
// check the configure and check if need to set PerfCounters configure to device
// and set filter configure
OV_ITT_SCOPED_TASK(itt::domains::MULTIPlugin, "MultiDeviceInferencePlugin::LoadNetworkImpl::AutoMode");
bool needPerfCounters = false;
AutoContext context;
std::map<std::string, std::string> filterConfig;
CheckConfig(fullConfig, needPerfCounters, filterConfig);
CheckConfig(fullConfig, context, filterConfig);
// filter the device that supports filter configure
auto strDevices = GetDeviceList(fullConfig);
auto metaDevices = ParseMetaDevices(strDevices, fullConfig);
@ -269,7 +294,7 @@ IExecutableNetworkInternal::Ptr MultiDeviceInferencePlugin::LoadNetworkImpl(cons
strDevices += ((iter + 1) == supportDevices.end()) ? "" : ",";
}
return std::make_shared<MultiDeviceExecutableNetwork>(modelPath, network, supportDevices, strDevices, this, needPerfCounters);
return std::make_shared<MultiDeviceExecutableNetwork>(modelPath, network, supportDevices, strDevices, this, context, context.needPerfCounters);
}
OV_ITT_SCOPED_TASK(itt::domains::MULTIPlugin, "MultiDeviceInferencePlugin::LoadNetworkImpl:MultiMode");
if (priorities == fullConfig.end()) {
@ -377,20 +402,18 @@ QueryNetworkResult MultiDeviceInferencePlugin::QueryNetwork(const CNNNetwork&
return queryResult;
}
DeviceInformation MultiDeviceInferencePlugin::SelectDevice(const std::vector<DeviceInformation>& metaDevices, const std::string& networkPrecision) {
DeviceInformation MultiDeviceInferencePlugin::SelectDevice(const std::vector<DeviceInformation>& metaDevices,
const std::string& networkPrecision, unsigned int priority) {
OV_ITT_SCOPED_TASK(itt::domains::MULTIPlugin, "MultiDeviceInferencePlugin::SelectDevice");
if (metaDevices.empty()) {
IE_THROW(NotFound) << "No available device to select in " << GetName() << " plugin";
}
if (metaDevices.size() == 1) {
return metaDevices.at(0);
}
std::vector<DeviceInformation> CPU;
std::vector<DeviceInformation> dGPU;
std::vector<DeviceInformation> iGPU;
std::vector<DeviceInformation> MYRIAD;
std::vector<DeviceInformation> VPUX;
std::list<DeviceInformation> CPU;
std::list<DeviceInformation> dGPU;
std::list<DeviceInformation> iGPU;
std::list<DeviceInformation> MYRIAD;
std::list<DeviceInformation> VPUX;
for (auto& item : metaDevices) {
if (item.deviceName.find("CPU") == 0) {
@ -406,96 +429,103 @@ DeviceInformation MultiDeviceInferencePlugin::SelectDevice(const std::vector<Dev
continue;
}
if (item.deviceName.find("GPU") == 0) {
auto gpuFullDeviceName = GetCore()->GetMetric(item.deviceName, METRIC_KEY(FULL_DEVICE_NAME)).as<std::string>();
if (gpuFullDeviceName.find("iGPU") != std::string::npos) {
auto& gpuUniqueName = item.uniqueName;
if (gpuUniqueName.find("iGPU") != std::string::npos) {
iGPU.push_back(item);
} else if (gpuFullDeviceName.find("dGPU") != std::string::npos) {
} else if (gpuUniqueName.find("dGPU") != std::string::npos) {
dGPU.push_back(item);
}
continue;
}
}
if (CPU.empty() && dGPU.empty() && iGPU.empty() && MYRIAD.empty() && VPUX.empty()) {
IE_THROW(NotFound) << "No available device found";
}
// Priority of selecting device: dGPU > VPUX > iGPU > MYRIAD > CPU
if (!dGPU.empty()) {
for (auto&& item : dGPU) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), networkPrecision);
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!VPUX.empty()) {
for (auto&& item : VPUX) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), networkPrecision);
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!iGPU.empty()) {
for (auto&& item : iGPU) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), networkPrecision);
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!MYRIAD.empty()) {
for (auto&& item : MYRIAD) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), networkPrecision);
if (supportNetwork != capability.end()) {
return item;
}
}
}
std::list<DeviceInformation> devices;
devices.splice(devices.end(), dGPU);
devices.splice(devices.end(), VPUX);
devices.splice(devices.end(), iGPU);
devices.splice(devices.end(), MYRIAD);
// If network is FP32 but there is no device support FP32, offload FP32 network to device support FP16.
std::list<DeviceInformation> validDevices;
auto selectSupportDev = [this, &devices, &validDevices](const std::string& networkPrecision) {
for (auto iter = devices.begin(); iter != devices.end();) {
std::vector<std::string> capability = GetCore()->GetMetric(iter->deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), (networkPrecision));
if (supportNetwork != capability.end()) {
validDevices.push_back(std::move(*iter));
devices.erase(iter++);
continue;
}
iter++;
}
};
selectSupportDev(networkPrecision);
// If network is FP32, continue to collect the device support FP16 but not support FP32.
if (networkPrecision == "FP32") {
if (!dGPU.empty()) {
for (auto&& item : dGPU) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), "FP16");
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!VPUX.empty()) {
for (auto&& item : VPUX) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), "FP16");
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!iGPU.empty()) {
for (auto&& item : iGPU) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), "FP16");
if (supportNetwork != capability.end()) {
return item;
}
}
} else if (!MYRIAD.empty()) {
for (auto&& item : MYRIAD) {
std::vector<std::string> capability = GetCore()->GetMetric(item.deviceName, METRIC_KEY(OPTIMIZATION_CAPABILITIES));
auto supportNetwork = std::find(capability.begin(), capability.end(), "FP16");
if (supportNetwork != capability.end()) {
return item;
}
const std::string f16 = "FP16";
selectSupportDev(f16);
}
// add cpu devices if exist.
validDevices.splice(validDevices.end(), CPU);
if (validDevices.empty()) {
IE_THROW() << "Cannot select any device";
}
// all available Devices are in validDevices now
// need to remove higher priority devices
// save the last device first
DeviceInformation lastDevice = validDevices.back();
{
// begin to filter devices
std::lock_guard<std::mutex> lck(_mtx);
for (auto && kvp : _priorityMap) {
if (kvp.first >= priority) {
continue;
}
auto& filterDevices = kvp.second;
auto sd = std::remove_if(validDevices.begin(), validDevices.end(), [&filterDevices](DeviceInformation device) {
auto iter = std::find_if(filterDevices.begin(), filterDevices.end(), [&device](std::string uniqueName) {
return (uniqueName == device.uniqueName);
});
return iter != filterDevices.end() ? true : false;
});
validDevices.erase(sd, validDevices.end());
}
}
if (CPU.empty()) {
IE_THROW() << "Cannot select any device";
DeviceInformation* ptrSelectDevice = NULL;
if (validDevices.empty()) {
// after remove higher priority device,but the available devices is null,
// so select the last device of all available Devices.
ptrSelectDevice = &lastDevice;
} else {
// select the first device in the rest of available devices.
ptrSelectDevice = &validDevices.front();
}
return CPU[0];
//recode the device priority
RegisterPriority(priority, ptrSelectDevice->uniqueName);
return *ptrSelectDevice;
}
void MultiDeviceInferencePlugin::UnregisterPriority(const unsigned int& priority,
const std::string& deviceName) {
std::lock_guard<std::mutex> lck(_mtx);
auto& priorityDevices = _priorityMap[priority];
for (auto iter = priorityDevices.begin(); iter != priorityDevices.end();) {
if (*iter == deviceName) {
priorityDevices.erase(iter);
break;
}
iter++;
}
}
void MultiDeviceInferencePlugin::RegisterPriority(const unsigned int& priority,
const std::string& deviceName) {
std::lock_guard<std::mutex> lck(_mtx);
auto& priorityDevices = _priorityMap[priority];
priorityDevices.push_back(deviceName);
}
std::string MultiDeviceInferencePlugin::GetDeviceList(const std::map<std::string, std::string>& config) const {
@ -520,19 +550,17 @@ std::string MultiDeviceInferencePlugin::GetDeviceList(const std::map<std::string
}
void MultiDeviceInferencePlugin::CheckConfig(const std::map<std::string, std::string>& config,
bool& needPerfCounters, std::map<std::string, std::string>& filterConfig) {
AutoContext& context, std::map<std::string, std::string>& filterConfig) {
// TODO need to optimize this code, too much duplicated code
const auto perf_hints_configs = PerfHintsConfig::SupportedKeys();
for (auto&& kvp : config) {
if (kvp.first.find("AUTO_") == 0) {
continue;
} else if (kvp.first == PluginConfigParams::KEY_PERF_COUNT) {
if (kvp.first == PluginConfigParams::KEY_PERF_COUNT) {
if (kvp.second == PluginConfigParams::YES) {
needPerfCounters = true;
context.needPerfCounters = true;
filterConfig.insert({kvp.first, kvp.second});
} else if (kvp.second == PluginConfigParams::NO) {
needPerfCounters = false;
context.needPerfCounters = false;
} else {
IE_THROW() << "Unsupported config value: " << kvp.second
<< " for key: " << kvp.first;
@ -551,10 +579,24 @@ void MultiDeviceInferencePlugin::CheckConfig(const std::map<std::string, std::st
IE_THROW() << "Unsupported config value: " << kvp.second
<< " for key: " << kvp.first;
}
} else if (kvp.first == MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY) {
try {
int priority = std::stoi(kvp.second);
if (priority < 0) {
IE_THROW() << "Unsupported config value: " << kvp.second
<< " for key: " << kvp.first;
}
context.modelPriority = priority;
} catch(...) {
IE_THROW() << "Unsupported config value: " << kvp.second
<< " for key: " << kvp.first;
}
} else if (std::find(perf_hints_configs.begin(), perf_hints_configs.end(), kvp.first) != perf_hints_configs.end()) {
PerfHintsConfig::CheckConfigAndValue(kvp);
} else if (supported_configKeys.end() == std::find(supported_configKeys.begin(), supported_configKeys.end(), kvp.first)) {
IE_THROW() << "Unsupported config key: " << kvp.first;
} else if (kvp.first.find("AUTO_") == 0) {
continue;
}
}
}

10
src/plugins/auto/plugin.hpp
View File

@ -8,6 +8,7 @@
#include <map>
#include <vector>
#include <string>
#include <list>
#include <cpp_interfaces/interface/ie_iplugin_internal.hpp>
#include <cpp_interfaces/interface/ie_internal_plugin_config.hpp>
@ -45,7 +46,10 @@ public:
const std::map<std::string, std::string> & config) const;
std::string GetDeviceList(const std::map<std::string, std::string>& config) const;
MOCKTESTMACRO DeviceInformation SelectDevice(const std::vector<DeviceInformation>& metaDevices, const std::string& networkPrecision = METRIC_VALUE(FP32));
MOCKTESTMACRO DeviceInformation SelectDevice(const std::vector<DeviceInformation>& metaDevices,
const std::string& networkPrecision = METRIC_VALUE(FP32), unsigned int priority = 0);
void UnregisterPriority(const unsigned int& priority, const std::string& deviceName);
void RegisterPriority(const unsigned int& priority, const std::string& deviceName);
protected:
std::map<std::string, std::string> GetSupportedConfig(const std::map<std::string, std::string>& config,
@ -56,10 +60,12 @@ private:
InferenceEngine::CNNNetwork network,
const std::map<std::string, std::string>& config,
const std::string &networkPrecision = METRIC_VALUE(FP32));
static void CheckConfig(const std::map<std::string, std::string>& config, bool& needPerfCounters,
static void CheckConfig(const std::map<std::string, std::string>& config, AutoContext& context,
std::map<std::string, std::string>& filterConfig);
std::vector<DeviceInformation> FilterDevice(const std::vector<DeviceInformation>& metaDevices,
const std::map<std::string, std::string>& config);
static std::mutex _mtx;
static std::map<unsigned int, std::list<std::string>> _priorityMap;
};
} // namespace MultiDevicePlugin

2
src/plugins/intel_gpu/include/intel_gpu/plugin/primitives_list.hpp
View File

@ -216,6 +216,8 @@ REGISTER_FACTORY(v8, GatherND);
REGISTER_FACTORY(v8, DeformableConvolution);
REGISTER_FACTORY(v8, NV12toRGB);
REGISTER_FACTORY(v8, NV12toBGR);
REGISTER_FACTORY(v8, I420toRGB);
REGISTER_FACTORY(v8, I420toBGR);
// --------------------------- Supported internal ops --------------------------- //
REGISTER_FACTORY(internal, NonMaxSuppressionIEInternal);

16
src/plugins/intel_gpu/src/plugin/ops/convert_color.cpp
View File

@ -44,16 +44,28 @@ static void CreateCommonConvertColorOp(Program& p, const std::shared_ptr<ngraph:
static void CreateNV12toRGBOp(Program& p, const std::shared_ptr<ngraph::op::v8::NV12toRGB>& op) {
p.ValidateInputs(op, {1, 2});
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::NV12, cldnn::convert_color::color_format::RGB);
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::NV12, cldnn::convert_color::color_format::RGB);
}
static void CreateNV12toBGROp(Program& p, const std::shared_ptr<ngraph::op::v8::NV12toBGR>& op) {
p.ValidateInputs(op, {1, 2});
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::NV12, cldnn::convert_color::color_format::BGR);
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::NV12, cldnn::convert_color::color_format::BGR);
}
static void CreateI420toRGBOp(Program& p, const std::shared_ptr<ngraph::op::v8::I420toRGB>& op) {
p.ValidateInputs(op, {1, 3});
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::I420, cldnn::convert_color::color_format::RGB);
}
static void CreateI420toBGROp(Program& p, const std::shared_ptr<ngraph::op::v8::I420toBGR>& op) {
p.ValidateInputs(op, {1, 3});
CreateCommonConvertColorOp(p, op, cldnn::convert_color::color_format::I420, cldnn::convert_color::color_format::BGR);
}
REGISTER_FACTORY_IMPL(v8, NV12toRGB);
REGISTER_FACTORY_IMPL(v8, NV12toBGR);
REGISTER_FACTORY_IMPL(v8, I420toRGB);
REGISTER_FACTORY_IMPL(v8, I420toBGR);
} // namespace intel_gpu
} // namespace runtime

10
src/plugins/intel_gpu/src/plugin/ops/convolution.cpp
View File

@ -137,11 +137,12 @@ static void CreateConvolutionBackpropDataOp(Program& p, const std::shared_ptr<ng
auto weightsName = inputs[1];
auto weights_node = op->get_input_node_shared_ptr(1);
// WA: For the cases like Const(weights)->Sub(zp)->Deconv.
bool hasConstantWeights = IsNodeOnConstPath(weights_node);
// WA: For the cases like Const(weights)->Sub(zp)->Deconv. And also for the cases with real runtime weights.
// Dimensions order of weights blob is IOYX, but
// the selected format is OIYX by default. So we need to swap (and transpose) I and O dimensions to match the format
// For Constant node on input transpose is not needed, because the data is transposed on const node creation
if (IsNodeOnConstPath(weights_node) && std::dynamic_pointer_cast<ngraph::op::v0::Constant>(weights_node) == nullptr) {
if ((hasConstantWeights && std::dynamic_pointer_cast<ngraph::op::v0::Constant>(weights_node) == nullptr) || !hasConstantWeights) {
std::string permuteName = layerName + "_cldnn_weights_permute";
auto weights_rank = op->get_input_shape(1).size();
std::vector<uint16_t> permute_order(weights_rank);
@ -195,11 +196,12 @@ static void CreateGroupConvolutionBackpropDataOp(Program& p, const std::shared_p
auto weightsName = inputs[1];
auto weights_node = op->get_input_node_shared_ptr(1);
// WA: For the cases like Const(weights)->Sub(zp)->Deconv.
bool hasConstWeights = IsNodeOnConstPath(weights_node);
// WA: For the cases like Const(weights)->Sub(zp)->Deconv. And also for the cases with real runtime weights.
// Dimensions order of weights blob is IOYX, but
// the selected format is OIYX by default. So we need to swap I and O dimensions to match the format.
// For Constant node on input transpose is not needed, because the data is transposed on const node creation
if (IsNodeOnConstPath(weights_node) && std::dynamic_pointer_cast<ngraph::op::v0::Constant>(weights_node) == nullptr) {
if ((hasConstWeights && std::dynamic_pointer_cast<ngraph::op::v0::Constant>(weights_node) == nullptr) || !hasConstWeights) {
std::string permuteName = layerName + "_cldnn_weights_permute";
auto weights_rank = op->get_input_shape(1).size();
std::vector<uint16_t> permute_order(weights_rank);

4
src/plugins/intel_gpu/src/plugin/ops/parameter.cpp
View File

@ -181,7 +181,9 @@ static void CreateParameterOp(Program& p, const std::shared_ptr<ngraph::op::v0::
bool is_convert_color_input = false;
for (auto& node : op->get_users()) {
is_convert_color_input |= ngraph::is_type<ngraph::op::v8::NV12toRGB>(node) ||
ngraph::is_type<ngraph::op::v8::NV12toBGR>(node);
ngraph::is_type<ngraph::op::v8::NV12toBGR>(node) ||
ngraph::is_type<ngraph::op::v8::I420toRGB>(node) ||
ngraph::is_type<ngraph::op::v8::I420toBGR>(node);
}
if (is_convert_color_input) {

4
src/plugins/intel_gpu/src/plugin/ops/result.cpp
View File

@ -41,7 +41,9 @@ static void CreateResultOp(Program& p, const std::shared_ptr<ngraph::op::v0::Res
auto outputlayout = outputDesc.getLayout();
if (ngraph::is_type<ngraph::op::v8::NV12toRGB>(prev) ||
ngraph::is_type<ngraph::op::v8::NV12toBGR>(prev)) {
ngraph::is_type<ngraph::op::v8::NV12toBGR>(prev) ||
ngraph::is_type<ngraph::op::v8::I420toRGB>(prev) ||
ngraph::is_type<ngraph::op::v8::I420toBGR>(prev)) {
outputlayout = NHWC;
}

2
src/plugins/intel_gpu/src/plugin/program.cpp
View File

@ -346,7 +346,7 @@ bool IsNodeOnConstPath(const std::shared_ptr<ngraph::Node>& node) {
std::function<bool(const std::shared_ptr<ngraph::Node>&)> is_const_node = [&nodes_processed, &is_const_node](const std::shared_ptr<ngraph::Node>& node) {
if (nodes_processed.count(node)) return true;
nodes_processed.insert(node);
// If input is constant, then drop if from the processing list
// If input is constant, then drop it from the processing list
if (std::dynamic_pointer_cast<ngraph::op::v0::Constant>(node) != nullptr)
return true;
// If the node doesn't have any parents and it's not a constant, then we deal with dynamic path

36
src/tests/functional/plugin/cpu/shared_tests_instances/behavior/plugin/configuration_tests.cpp
View File

@ -96,7 +96,13 @@ namespace {
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_LOG_LEVEL, InferenceEngine::PluginConfigParams::LOG_DEBUG}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_LOG_LEVEL, InferenceEngine::PluginConfigParams::LOG_TRACE}}
{InferenceEngine::PluginConfigParams::KEY_LOG_LEVEL, InferenceEngine::PluginConfigParams::LOG_TRACE}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY, "0"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY, "1"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY, "2"}}
};
INSTANTIATE_TEST_SUITE_P(smoke_BehaviorTests, CorrectConfigTests,
@ -149,17 +155,6 @@ namespace {
{InferenceEngine::PluginConfigParams::KEY_DYN_BATCH_LIMIT, "NAN"}}
};
const std::vector<std::map<std::string, std::string>> multiconf = {
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::THROUGHPUT}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::LATENCY}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::LATENCY},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT_NUM_REQUESTS, "1"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU}}
};
const std::vector<std::map<std::string, std::string>> autoinconfigs = {
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, "DOESN'T EXIST"}},
@ -176,7 +171,22 @@ namespace {
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_DYN_BATCH_LIMIT, "NAN"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_LOG_LEVEL, "NAN"}}
{InferenceEngine::MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY, "-1"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::MultiDeviceConfigParams::KEY_AUTO_NETWORK_PRIORITY, "should be int"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_LOG_LEVEL, "NAN"}}
};
const std::vector<std::map<std::string, std::string>> multiconf = {
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::THROUGHPUT}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::LATENCY}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT, InferenceEngine::PluginConfigParams::LATENCY},
{InferenceEngine::PluginConfigParams::KEY_PERFORMANCE_HINT_NUM_REQUESTS, "1"}},
{{InferenceEngine::MultiDeviceConfigParams::KEY_MULTI_DEVICE_PRIORITIES , CommonTestUtils::DEVICE_CPU}}
};
INSTANTIATE_TEST_SUITE_P(smoke_BehaviorTests, IncorrectConfigTests,

11
src/tests/functional/plugin/cpu/shared_tests_instances/blob_tests/set_blob.cpp
View File

@ -8,8 +8,15 @@
using namespace BehaviorTestsDefinitions;
using namespace InferenceEngine;
const std::vector<Precision> precisionSet = {Precision::FP32, Precision::I16, Precision::U8, Precision::I8, Precision::U16, Precision::I32, Precision::BOOL,
Precision::I64, Precision::U64};
const std::vector<Precision> precisionSet = {
Precision::U8, Precision::I8,
Precision::U16, Precision::I16,
Precision::U32, Precision::I32,
Precision::U64, Precision::I64,
Precision::BF16, Precision::FP16,
Precision::FP32, Precision::FP64,
Precision::BOOL
};
const std::vector<setType> typeSet = {setType::INPUT, setType::OUTPUT, setType::BOTH};

28
src/tests/functional/plugin/cpu/shared_tests_instances/single_layer_tests/conversion.cpp
View File

@ -18,24 +18,26 @@ const std::vector<ngraph::helpers::ConversionTypes> conversionOpTypes = {
const std::vector<std::vector<size_t>> inShape = {{1, 2, 3, 4}};
const std::vector<InferenceEngine::Precision> netPrecisions = {
// Ticket: 59594
// InferenceEngine::Precision::I4,
InferenceEngine::Precision::I8,
InferenceEngine::Precision::I16,
InferenceEngine::Precision::I32,
InferenceEngine::Precision::I64,
// Ticket: 59594
// InferenceEngine::Precision::BIN,
// InferenceEngine::Precision::BOOL,
// InferenceEngine::Precision::U4,
InferenceEngine::Precision::U8,
InferenceEngine::Precision::I8,
InferenceEngine::Precision::U16,
// Ticket: 59594
// InferenceEngine::Precision::U32,
InferenceEngine::Precision::I16,
InferenceEngine::Precision::U32,
InferenceEngine::Precision::I32,
InferenceEngine::Precision::U64,
InferenceEngine::Precision::I64,
InferenceEngine::Precision::BF16,
InferenceEngine::Precision::FP16,
InferenceEngine::Precision::FP32};
InferenceEngine::Precision::FP32,
InferenceEngine::Precision::FP64,
InferenceEngine::Precision::BOOL,
InferenceEngine::Precision::MIXED,
InferenceEngine::Precision::Q78,
InferenceEngine::Precision::U4,
InferenceEngine::Precision::I4,
InferenceEngine::Precision::BIN,
InferenceEngine::Precision::CUSTOM,
};
INSTANTIATE_TEST_SUITE_P(smoke_ConversionLayerTest,
ConversionLayerTest,

26
src/tests/functional/plugin/cpu/shared_tests_instances/skip_tests_config.cpp
View File

@ -104,17 +104,6 @@ std::vector<std::string> disabledTestPatterns() {
// CPU plugin does not support some precisions
R"(smoke_CachingSupportCase_CPU/LoadNetworkCacheTestBase.CompareWithRefImpl/ReadConcatSplitAssign_f32_batch1_CPU)",
// CPU plugin does not support some precisions
R"(.*Behavior.*OVExecGraphImportExportTest.*elementType=(i8|u32).*)",
R"(.*Behavior.*OVExecGraphImportExportTest.*elementType=(f16).*)",
R"(.*EltwiseLayerTest.*NetType=f16.*)",
// TODO: CVS-66526 overrides i/o precisions in execution graph
// as WA we used GetInputsInfo() precisions instead of ngraph ones
// R"(.*smoke_BehaviorTests.*OVExecGraphImportExportTest.*importExportedFunction.*type=(i16|u16).*)",
// R"(.*smoke_BehaviorTests.*OVExecGraphImportExportTest.*importExportedFunction.*type=(i64|u64).*)",
// R"(.*smoke_BehaviorTests.*OVExecGraphImportExportTest.*importExportedIENetwork.*type=(i16|u16).*)",
// R"(.*smoke_BehaviorTests.*OVExecGraphImportExportTest.*importExportedIENetwork.*type=(i64|u64).*)",
// CPU does not support dynamic rank
// Issue: CVS-66778
@ -152,6 +141,7 @@ std::vector<std::string> disabledTestPatterns() {
// bad accuracy
R"(.*smoke_FakeQuantizeLayerCPUTest_Decompos.
*IS=_TS=\(\(4\.5\.6\.7\)\)_RS=\(\(1\.1\.6\.1\)\)_\(\(1\.5\.6\.1\)\)_\(\(1\.1\.1\.1\)\)_\(\(1\.1\.6\.1\)\).*)",
// Issue: 71121
R"(.*smoke_Proposal*.*TS=\(2.*)",
// TODO : CVS-69533
@ -165,6 +155,20 @@ std::vector<std::string> disabledTestPatterns() {
// Failure happened on win and macos for current seeds.
R"(.*CTCLossLayerTest.*CMR=1.*)",
R"(.*CTCLossLayerCPUTest.*ctcMergeRepeated=1.*)",
// Issue: 71756
R"(.*Deconv_.*D_(Blocked|DW|1x1)_.*DeconvolutionLayerCPUTest\.CompareWithRefs.*inFmts=(nChw16c|nCdhw16c)_outFmts=(nChw16c|nCdhw16c)_primitive=jit_avx512_.*Fused=Multiply\(PerChannel\)\.Add\(PerChannel\).*)",
R"(.*smoke_GroupDeconv_(2|3)D_Blocked_BF16.*S=(\(2\.2\)|\(2\.2\.2\))_PB=(\(0\.0\)|\(0\.0\.0\))_PE=(\(0\.0\)|\(0\.0\.0\))_D=(\(1\.1\)|\(1\.1\.1\))_.*_O=64_G=4.*)",
// Issue: 72150
R"(.*smoke_SetBlobCPU/SetBlobTest.CompareWithRefs/Type=.*_Device=CPU_PrecisionInNet=BOOL.*)",
// Issue: 59594
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*BOOL.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*MIXED.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*Q78.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*U4.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*I4.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*BIN.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*CUSTOM.*)",
R"(smoke_ConversionLayerTest/ConversionLayerTest.CompareWithRefs.*UNSPECIFIED.*)",
};
#define FIX_62820 0

679
src/tests/functional/plugin/cpu/single_layer_tests/convolution_backprop_data.cpp
View File

@ -5,38 +5,78 @@
#include "test_utils/cpu_test_utils.hpp"
#include "test_utils/convolution_params.hpp"
#include "test_utils/fusing_test_utils.hpp"
#include "shared_test_classes/base/layer_test_utils.hpp"
#include "ngraph_functions/utils/ngraph_helpers.hpp"
#include "shared_test_classes/base/ov_subgraph.hpp"
#include "functional_test_utils/ov_tensor_utils.hpp"
#include "ngraph_functions/builders.hpp"
#include <shared_test_classes/single_layer/convolution_backprop_data.hpp>
#include "openvino/core/preprocess/pre_post_process.hpp"
using namespace InferenceEngine;
using namespace CPUTestUtils;
using namespace ov::test;
namespace CPULayerTestsDefinitions {
using LayerTestsDefinitions::convBackpropDataSpecificParams;
using LayerTestsDefinitions::convBackpropDataLayerTestParamsSet;
typedef std::tuple<
convBackpropDataLayerTestParamsSet,
CPUSpecificParams,
fusingSpecificParams,
std::map<std::string, std::string> > deconvLayerCPUTestParamsSet;
using DeconvSpecParams = LayerTestsDefinitions::convBackpropDataSpecificParams;
class DeconvolutionLayerCPUTest : public testing::WithParamInterface<deconvLayerCPUTestParamsSet>,
virtual public LayerTestsUtils::LayerTestsCommon, public CpuTestWithFusing {
using DeconvInputData = std::tuple<InputShape, // data shape
ngraph::helpers::InputLayerType, // 'output_shape' input type
std::vector<std::vector<int32_t>>>; // values for 'output_shape'
using DeconvLayerCPUTestParamsSet = std::tuple<DeconvSpecParams,
DeconvInputData,
ElementType,
fusingSpecificParams,
CPUSpecificParams,
std::map<std::string, std::string>>;
class DeconvolutionLayerCPUTest : public testing::WithParamInterface<DeconvLayerCPUTestParamsSet>,
virtual public SubgraphBaseTest, public CpuTestWithFusing {
public:
static std::string getTestCaseName(testing::TestParamInfo<deconvLayerCPUTestParamsSet> obj) {
convBackpropDataLayerTestParamsSet basicParamsSet;
CPUSpecificParams cpuParams;
static std::string getTestCaseName(testing::TestParamInfo<DeconvLayerCPUTestParamsSet> obj) {
DeconvSpecParams basicParamsSet;
DeconvInputData inputData;
ElementType prec;
fusingSpecificParams fusingParams;
CPUSpecificParams cpuParams;
std::map<std::string, std::string> additionalConfig;
std::tie(basicParamsSet, cpuParams, fusingParams, additionalConfig) = obj.param;
std::tie(basicParamsSet, inputData, prec, fusingParams, cpuParams, additionalConfig) = obj.param;
ngraph::op::PadType padType;
InferenceEngine::SizeVector kernel, stride, dilation;
std::vector<ptrdiff_t> padBegin, padEnd, outPadding;
size_t convOutChannels;
std::tie(kernel, stride, padBegin, padEnd, dilation, convOutChannels, padType, outPadding) = basicParamsSet;
InputShape inputShape;
ngraph::helpers::InputLayerType outShapeType;
std::vector<std::vector<int32_t>> outShapeData;
std::tie(inputShape, outShapeType, outShapeData) = inputData;
std::ostringstream result;
result << LayerTestsDefinitions::ConvolutionBackpropDataLayerTest::getTestCaseName(testing::TestParamInfo<convBackpropDataLayerTestParamsSet>(
basicParamsSet, 0));
result << "IS=";
result << CommonTestUtils::partialShape2str({inputShape.first}) << "_";
result << "TS=";
for (const auto& shape : inputShape.second) {
result << "(";
result << CommonTestUtils::vec2str(shape);
result << ")_";
}
result << "PRC=" << prec << "_";
result << "K=" << CommonTestUtils::vec2str(kernel) << "_";
result << "S=" << CommonTestUtils::vec2str(stride) << "_";
result << "PB=" << CommonTestUtils::vec2str(padBegin) << "_";
result << "PE=" << CommonTestUtils::vec2str(padEnd) << "_";
result << "D=" << CommonTestUtils::vec2str(dilation) << "_";
result << "OP=" << CommonTestUtils::vec2str(outPadding) << "_";
result << "O=" << convOutChannels << "_";
result << "AP=" << padType << "_";
result << "OUT_SH=" << outShapeType << "_";
result << "OUT_D=";
for (const auto& data : outShapeData) {
result << "(";
result << CommonTestUtils::vec2str(data);
result << ")_";
}
result << CPUTestsBase::getTestCaseName(cpuParams);
result << CpuTestWithFusing::getTestCaseName(fusingParams);
@ -50,53 +90,159 @@ public:
return result.str();
}
void generate_inputs(const std::vector<ngraph::Shape>& targetInputStaticShapes) override {
inputs.clear();
const auto& funcInputs = function->inputs();
for (int i = 0; i < funcInputs.size(); ++i) {
const auto& funcInput = funcInputs[i];
ov::runtime::Tensor tensor;
if (i == 1) {
tensor = ov::runtime::Tensor(funcInput.get_element_type(), targetInputStaticShapes[i], outShapeData[inferRequestNum].data());
} else {
tensor = ov::test::utils::create_and_fill_tensor(funcInput.get_element_type(), targetInputStaticShapes[i], 2560, 0, 256);
}
inputs.insert({funcInput.get_node_shared_ptr(), tensor});
}
inferRequestNum++;
}
void init_ref_function(std::shared_ptr<ov::Model> &funcRef, const std::vector<ov::Shape>& targetInputStaticShapes) override {
if (function->get_parameters().size() == 1) {
ngraph::helpers::resize_function(funcRef, targetInputStaticShapes);
} else {
// WA: output_shape depends on 3rd deconvolution input data
// but the reference implementation doesn't implement shape inference
// so we need to build a new ngraph function and replace the 3rd input parameter with a constant
// to get valid output shapes
funcRef = createGraph({targetInputStaticShapes[0]}, ngraph::helpers::InputLayerType::CONSTANT);
}
}
void validate() override {
if (function->get_parameters().size() == 2) {
auto pos = std::find_if(inputs.begin(), inputs.end(),
[](const std::pair<std::shared_ptr<ov::Node>, ov::runtime::Tensor> &params) {
return params.first->get_friendly_name() == "param_1";
});
IE_ASSERT(pos != inputs.end());
inputs.erase(pos);
}
SubgraphBaseTest::validate();
}
void configure_model() override {
ov::preprocess::PrePostProcessor p(function);
{
auto& params = function->get_parameters();
for (size_t i = 0; i < params.size(); i++) {
if (i > 0) {
continue;
}
if (inType != ov::element::Type_t::undefined) {
p.input(i).tensor().set_element_type(inType);
}
}
}
{
auto results = function->get_results();
for (size_t i = 0; i < results.size(); i++) {
if (outType != ov::element::Type_t::undefined) {
p.output(i).tensor().set_element_type(outType);
}
}
}
function = p.build();
}
std::shared_ptr<ov::Model> createGraph(const std::vector<ov::PartialShape>& inShapes, ngraph::helpers::InputLayerType outShapeType) {
auto params = ngraph::builder::makeDynamicParams(prec, {inShapes.front()});
std::shared_ptr<ov::Node> outShapeNode;
if (!outShapeData.empty()) {
if (outShapeType == ngraph::helpers::InputLayerType::PARAMETER) {
IE_ASSERT(inputDynamicShapes.size() == 2);
auto outShapeParam = std::make_shared<ngraph::opset8::Parameter>(ngraph::element::i32, inputDynamicShapes.back());
params.push_back(outShapeParam);
outShapeNode = outShapeParam;
} else {
outShapeNode = ngraph::opset8::Constant::create(ngraph::element::i32, {outShapeData[inferRequestNum].size()}, outShapeData[inferRequestNum]);
}
}
for (size_t i = 0; i < params.size(); i++) {
params[i]->set_friendly_name(std::string("param_") + std::to_string(i));
}
std::shared_ptr<ov::Node> deconv;
if (!outShapeData.empty()) {
IE_ASSERT(outShapeNode != nullptr);
deconv = ngraph::builder::makeConvolutionBackpropData(params[0], outShapeNode, prec, kernel, stride, padBegin,
padEnd, dilation, padType, convOutChannels);
} else {
deconv = ngraph::builder::makeConvolutionBackpropData(params[0], prec, kernel, stride, padBegin,
padEnd, dilation, padType, convOutChannels, false, outPadding);
}
return makeNgraphFunction(prec, params, deconv, "DeconvCPU");
}
protected:
InferenceEngine::SizeVector kernel, stride;
void SetUp() override {
convBackpropDataLayerTestParamsSet basicParamsSet;
CPUSpecificParams cpuParams;
rel_threshold = 1e-4f;
targetDevice = CommonTestUtils::DEVICE_CPU;
DeconvSpecParams basicParamsSet;
DeconvInputData inputData;
fusingSpecificParams fusingParams;
CPUSpecificParams cpuParams;
std::map<std::string, std::string> additionalConfig;
std::tie(basicParamsSet, cpuParams, fusingParams, additionalConfig) = this->GetParam();
std::tie(basicParamsSet, inputData, prec, fusingParams, cpuParams, additionalConfig) = this->GetParam();
InputShape inputShape;
ngraph::helpers::InputLayerType outShapeType;
std::tie(inputShape, outShapeType, outShapeData) = inputData;
configuration.insert(additionalConfig.begin(), additionalConfig.end());
std::tie(inFmts, outFmts, priority, selectedType) = cpuParams;
std::tie(postOpMgrPtr, fusedOps) = fusingParams;
convBackpropDataSpecificParams convParams;
std::vector<size_t> inputShape;
std::vector<size_t> outputShape;
auto netPrecision = InferenceEngine::Precision::UNSPECIFIED;
std::tie(convParams, netPrecision, inPrc, outPrc, inLayout, outLayout, inputShape, outputShape, targetDevice) = basicParamsSet;
std::tie(kernel, stride, padBegin, padEnd, dilation, convOutChannels, padType, outPadding) = basicParamsSet;
if (inPrc == Precision::UNSPECIFIED) {
selectedType += std::string("_") + Precision(Precision::FP32).name();
if (additionalConfig[InferenceEngine::PluginConfigParams::KEY_ENFORCE_BF16] == InferenceEngine::PluginConfigParams::YES) {
inType = outType = prec = ElementType::bf16;
rel_threshold = 1e-2f;
} else {
selectedType += std::string("_") + inPrc.name();
inType = outType = prec;
}
ngraph::op::PadType padType;
InferenceEngine::SizeVector dilation;
std::vector<ptrdiff_t> padBegin, padEnd, outPadding;
size_t convOutChannels;
std::tie(kernel, stride, padBegin, padEnd, dilation, convOutChannels, padType, outPadding) = convParams;
auto ngPrc = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(netPrecision);
selectedType = makeSelectedTypeStr(selectedType, prec);
auto inputParams = ngraph::builder::makeParams(ngraph::element::f32, { inputShape });
auto paramOuts = ngraph::helpers::convert2OutputVector(ngraph::helpers::castOps2Nodes<ngraph::op::Parameter>(inputParams));
auto deconvolutionNode = ngraph::builder::makeConvolutionBackpropData(paramOuts.front(), ngPrc, kernel, stride, padBegin,
padEnd, dilation, padType, convOutChannels, false, outPadding);
if (!outputShape.empty()) {
auto outShape = ngraph::opset3::Constant::create(ngraph::element::i64, {outputShape.size()}, outputShape);
deconvolutionNode = ngraph::builder::makeConvolutionBackpropData(paramOuts.front(), outShape, ngPrc, kernel, stride, padBegin,
padEnd, dilation, padType, convOutChannels);
std::vector<InputShape> paramsShapes;
paramsShapes.push_back(inputShape);
if (!outShapeData.empty() && outShapeType == ngraph::helpers::InputLayerType::PARAMETER) {
const auto outShapeDims = ov::Shape{outShapeData.front().size()};
paramsShapes.push_back(InputShape{outShapeDims, std::vector<ov::Shape>(inputShape.second.size(), outShapeDims)});
}
function = makeNgraphFunction(ngPrc, inputParams, deconvolutionNode, "convolutionBackpropData");
init_input_shapes(paramsShapes);
function = createGraph(inputDynamicShapes, outShapeType);
}
private:
ElementType prec;
ngraph::op::PadType padType;
InferenceEngine::SizeVector dilation;
std::vector<ptrdiff_t> padBegin, padEnd, outPadding;
size_t convOutChannels;
ngraph::helpers::InputLayerType outShapeType;
std::vector<std::vector<int32_t>> outShapeData;
size_t inferRequestNum = 0;
};
TEST_P(DeconvolutionLayerCPUTest, CompareWithRefs) {
@ -113,7 +259,7 @@ TEST_P(DeconvolutionLayerCPUTest, CompareWithRefs) {
}
}
Run();
run();
CheckPluginRelatedResults(executableNetwork, "Deconvolution");
}
@ -126,29 +272,29 @@ const std::vector<fusingSpecificParams> fusingParamsSet{
};
const std::map<std::string, std::string> cpuEmptyPluginConfig;
const std::map<std::string, std::string> cpuBF16PluginConfig = { { PluginConfigParams::KEY_ENFORCE_BF16, PluginConfigParams::YES } };
const std::vector<SizeVector> emptyOutputShape = { {} };
const std::map<std::string, std::string>cpuBF16PluginConfig = { { InferenceEngine::PluginConfigParams::KEY_ENFORCE_BF16,
InferenceEngine::PluginConfigParams::YES } };
const std::vector<std::vector<ptrdiff_t>> emptyOutputPadding = { {} };
/* ============= Deconvolution params (planar layout) ============= */
const SizeVector numOutChannels_Planar = { 6 };
const InferenceEngine::SizeVector numOutChannels_Planar = { 6 };
/* ============= Deconvolution params (blocked layout) ============= */
const SizeVector numOutChannels_Blocked = { 64 };
const InferenceEngine::SizeVector numOutChannels_Blocked = { 64 };
/* ============= Deconvolution params (2D) ============= */
const std::vector<SizeVector> kernels2d = { {3, 3}, {1, 1} };
const std::vector<SizeVector> strides2d = { {1, 1}, {2, 2} };
const std::vector<InferenceEngine::SizeVector> kernels2d = { {3, 3}, {1, 1} };
const std::vector<InferenceEngine::SizeVector> strides2d = { {1, 1}, {2, 2} };
const std::vector<std::vector<ptrdiff_t>> padBegins2d = { {0, 0} };
const std::vector<std::vector<ptrdiff_t>> padEnds2d = { {0, 0} };
const std::vector<SizeVector> dilations2d = { {1, 1} };
const std::vector<InferenceEngine::SizeVector> dilations2d = { {1, 1} };
/* ============= Deconvolution params (3D) ============= */
const std::vector<SizeVector> kernels3d = { {3, 3, 3}, {1, 1, 1} };
const std::vector<SizeVector> strides3d = { {1, 1, 1}, {2, 2, 2} };
const std::vector<InferenceEngine::SizeVector> kernels3d = { {3, 3, 3}, {1, 1, 1} };
const std::vector<InferenceEngine::SizeVector> strides3d = { {1, 1, 1}, {2, 2, 2} };
const std::vector<std::vector<ptrdiff_t>> padBegins3d = { {0, 0, 0} };
const std::vector<std::vector<ptrdiff_t>> padEnds3d = { {0, 0, 0} };
const std::vector<SizeVector> dilations3d = { {1, 1, 1} };
const std::vector<InferenceEngine::SizeVector> dilations3d = { {1, 1, 1} };
/* ============= */
/* INSTANCES */
@ -164,41 +310,99 @@ const auto convParams_ExplicitPadding_Planar_2D = ::testing::Combine(
::testing::ValuesIn(emptyOutputPadding)
);
const std::vector<DeconvInputData> Planar_2D_inputs_smoke = {
DeconvInputData{
InputShape{{}, {{ 2, 12, 7, 7 }}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 12, -1, -1}, {{ 2, 12, 7, 7}, { 2, 12, 5, 7}, { 1, 12, 9, 4}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{15, 15}, {9, 10}, {9, 9}}
}
};
const std::vector<DeconvInputData> Planar_2D_inputs_nightly = {
DeconvInputData{
InputShape{{-1, 12, -1, -1}, {{ 2, 12, 7, 7}, { 2, 12, 5, 7}, { 1, 12, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 12, -1, -1}, {{ 2, 12, 7, 7}, { 2, 12, 5, 7}, { 1, 12, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{{15, 15}}
}
};
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_Planar_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Planar_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 12, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
convParams_ExplicitPadding_Planar_2D,
::testing::ValuesIn(Planar_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_Planar_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Planar_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::BF16),
::testing::Values(Precision::BF16),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 12, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
convParams_ExplicitPadding_Planar_2D,
::testing::ValuesIn(Planar_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= GroupDeconvolution (Planar 3D) ============= */
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_2D_Planar_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Planar_2D,
::testing::ValuesIn(Planar_2D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_2D_Planar_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Planar_2D,
::testing::ValuesIn(Planar_2D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= Deconvolution (Planar 3D) ============= */
const std::vector<DeconvInputData> Planar_3D_inputs_smoke = {
DeconvInputData{
InputShape{{}, {{ 2, 12, 7, 7, 7 }}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 12, -1, -1, -1}, {{ 2, 12, 7, 7, 7}, { 2, 12, 5, 7, 7}, { 1, 12, 9, 4, 9}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{15, 15, 15}, {9, 10, 10}, {9, 9, 9}}
}
};
const std::vector<DeconvInputData> Planar_3D_inputs_nightly = {
DeconvInputData{
InputShape{{-1, 12, -1, -1, -1}, {{ 2, 12, 7, 7, 7}, { 2, 12, 5, 7, 7}, { 1, 12, 9, 4, 9}}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 12, -1, -1, -1}, {{ 2, 12, 7, 7, 7}, { 2, 12, 5, 7, 7}, { 1, 12, 9, 4, 9}}},
ngraph::helpers::InputLayerType::CONSTANT,
{{15, 15, 15}}
}
};
const auto convParams_ExplicitPadding_Planar_3D = ::testing::Combine(
::testing::ValuesIn(kernels3d),
::testing::ValuesIn(strides3d),
@ -212,39 +416,71 @@ const auto convParams_ExplicitPadding_Planar_3D = ::testing::Combine(
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_3D_Planar_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Planar_3D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 12, 7, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
convParams_ExplicitPadding_Planar_3D,
::testing::ValuesIn(Planar_3D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_3D_Planar_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Planar_3D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::BF16),
::testing::Values(Precision::BF16),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 12, 7, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
convParams_ExplicitPadding_Planar_3D,
::testing::ValuesIn(Planar_3D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= GroupDeconvolution (Blocked 2D) ============= */
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_3D_Planar_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Planar_3D,
::testing::ValuesIn(Planar_3D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_3D_Planar_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Planar_3D,
::testing::ValuesIn(Planar_3D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_3D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= Deconvolution (Blocked 2D) ============= */
const std::vector<DeconvInputData> Blocked_2D_inputs_smoke = {
DeconvInputData{
InputShape{{}, {{ 2, 67, 7, 7 }}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{15, 15}, {9, 10}, {9, 9}}
}
};
const std::vector<DeconvInputData> Blocked_2D_inputs_nightly = {
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{{15, 15}}
}
};
const auto convParams_ExplicitPadding_Blocked_2D = ::testing::Combine(
::testing::ValuesIn(kernels2d),
::testing::ValuesIn(strides2d),
@ -258,92 +494,129 @@ const auto convParams_ExplicitPadding_Blocked_2D = ::testing::Combine(
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_Blocked_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Blocked_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
convParams_ExplicitPadding_Blocked_2D,
::testing::ValuesIn(Blocked_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_Blocked_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Blocked_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::BF16),
::testing::Values(Precision::BF16),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
convParams_ExplicitPadding_Blocked_2D,
::testing::ValuesIn(Blocked_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= GroupDeconvolution (Blocked 3D) ============= */
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_2D_Blocked_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Blocked_2D,
::testing::ValuesIn(Blocked_2D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_2D_Blocked_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Blocked_2D,
::testing::ValuesIn(Blocked_2D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= Deconvolution (Blocked 3D) ============= */
const std::vector<DeconvInputData> Blocked_3D_inputs_smoke = {
DeconvInputData{
InputShape{{}, {{ 2, 35, 7, 7, 7 }}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 35, -1, -1, -1}, {{ 1, 35, 5, 5, 5}, { 2, 35, 5, 7, 5}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{7, 7, 7}, {7, 9, 7}}
}
};
const std::vector<DeconvInputData> Blocked_3D_inputs_nightly = {
DeconvInputData{
InputShape{{-1, 35, -1, -1, -1}, {{ 1, 35, 5, 5, 5}, { 2, 35, 5, 7, 5}}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 35, -1, -1, -1}, {{ 1, 35, 5, 5, 5}, { 2, 35, 5, 7, 5}}},
ngraph::helpers::InputLayerType::CONSTANT,
{{7, 7, 7}}
}
};
const auto convParams_ExplicitPadding_Blocked_3D = ::testing::Combine(
::testing::ValuesIn(kernels3d),
::testing::ValuesIn(strides3d),
::testing::ValuesIn(padBegins3d),
::testing::ValuesIn(padEnds3d),
::testing::ValuesIn(dilations3d),
::testing::ValuesIn(numOutChannels_Blocked),
::testing::Values(32),
::testing::Values(ngraph::op::PadType::EXPLICIT),
::testing::ValuesIn(emptyOutputPadding)
);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_3D_Blocked_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Blocked_3D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
convParams_ExplicitPadding_Blocked_3D,
::testing::ValuesIn(Blocked_3D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_3D_Blocked_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_Blocked_3D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::BF16),
::testing::Values(Precision::BF16),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
convParams_ExplicitPadding_Blocked_3D,
::testing::ValuesIn(Blocked_3D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_3D_Blocked_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Blocked_3D,
::testing::ValuesIn(Blocked_3D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(nightly_Deconv_3D_Blocked_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
convParams_ExplicitPadding_Blocked_3D,
::testing::ValuesIn(Blocked_3D_inputs_nightly),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_3D})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= Kernel_1x1 (2D) ============= */
const auto convParams_ExplicitPadding_1x1_2D = ::testing::Combine(
::testing::Values(SizeVector({1, 1})),
::testing::Values(SizeVector({1, 1})),
::testing::Values(InferenceEngine::SizeVector({1, 1})),
::testing::Values(InferenceEngine::SizeVector({1, 1})),
::testing::Values(std::vector<ptrdiff_t>({0, 0})),
::testing::Values(std::vector<ptrdiff_t>({0, 0})),
::testing::Values(SizeVector({1, 1})),
::testing::Values(InferenceEngine::SizeVector({1, 1})),
::testing::ValuesIn(numOutChannels_Blocked),
::testing::Values(ngraph::op::PadType::EXPLICIT),
::testing::ValuesIn(emptyOutputPadding)
@ -351,39 +624,89 @@ const auto convParams_ExplicitPadding_1x1_2D = ::testing::Combine(
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_1x1_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_1x1_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Precision::UNSPECIFIED),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D_1x1})),
convParams_ExplicitPadding_1x1_2D,
::testing::ValuesIn(Blocked_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D_1x1})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_1x1_BF16, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(
convParams_ExplicitPadding_1x1_2D,
::testing::Values(Precision::FP32),
::testing::Values(Precision::BF16),
::testing::Values(Precision::BF16),
::testing::Values(Layout::ANY),
::testing::Values(Layout::ANY),
::testing::Values(std::vector<size_t >({ 2, 67, 7, 7 })),
::testing::ValuesIn(emptyOutputShape),
::testing::Values(CommonTestUtils::DEVICE_CPU)),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D_1x1})),
convParams_ExplicitPadding_1x1_2D,
::testing::ValuesIn(Blocked_2D_inputs_smoke),
::testing::Values(ElementType::f32),
::testing::ValuesIn(fusingParamsSet),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D_1x1})),
::testing::Values(cpuBF16PluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ========= */
/* ============= Reorder + Deconvolution ============= */
INSTANTIATE_TEST_SUITE_P(smoke_reorder_Deconv_2D, DeconvolutionLayerCPUTest,
::testing::Combine(
::testing::Combine(::testing::ValuesIn(kernels2d),
::testing::Values(InferenceEngine::SizeVector{1, 1}),
::testing::ValuesIn(padBegins2d),
::testing::ValuesIn(padEnds2d),
::testing::ValuesIn(dilations2d),
::testing::ValuesIn(numOutChannels_Blocked),
::testing::Values(ngraph::op::PadType::EXPLICIT),
::testing::ValuesIn(emptyOutputPadding)),
::testing::Values(DeconvInputData{InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{15, 15}, {9, 10}, {9, 9}}}),
::testing::Values(ElementType::f32),
::testing::Values(emptyFusingSpec),
::testing::ValuesIn(filterCPUInfoForDevice({conv_avx512_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
/* ============= Deconvolution auto padding tests ============= */
const std::vector<DeconvInputData> inputs_2D_AutoPadding = {
DeconvInputData{
InputShape{{}, {{ 2, 67, 7, 7 }}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{}
},
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::CONSTANT,
{{15, 15}}
},
DeconvInputData{
InputShape{{-1, 67, -1, -1}, {{ 2, 67, 7, 7}, { 2, 67, 5, 7}, { 1, 67, 9, 4}}},
ngraph::helpers::InputLayerType::PARAMETER,
{{15, 15}, {9, 10}, {9, 9}}
}
};
const auto deconvParams_AutoPadding_2D = ::testing::Combine(
::testing::ValuesIn(kernels2d),
::testing::ValuesIn(strides2d),
::testing::ValuesIn(padBegins2d),
::testing::ValuesIn(padEnds2d),
::testing::ValuesIn(dilations2d),
::testing::ValuesIn(numOutChannels_Blocked),
::testing::Values(ngraph::op::PadType::SAME_UPPER, ngraph::op::PadType::SAME_LOWER),
::testing::ValuesIn(emptyOutputPadding)
);
INSTANTIATE_TEST_SUITE_P(smoke_Deconv_2D_AutoPadding_FP32, DeconvolutionLayerCPUTest,
::testing::Combine(
deconvParams_AutoPadding_2D,
::testing::ValuesIn(inputs_2D_AutoPadding),
::testing::Values(ElementType::f32),
::testing::Values(emptyFusingSpec),
::testing::ValuesIn(filterCPUInfoForDevice({conv_gemm_2D, conv_avx512_2D})),
::testing::Values(cpuEmptyPluginConfig)),
DeconvolutionLayerCPUTest::getTestCaseName);
} // namespace
} // namespace CPULayerTestsDefinitions

Some files were not shown because too many files have changed in this diff Show More