[CPU] SoftMax cache (#9480)
* [CPUCache]SoftMax cache * [CpuCache]fix bf16 tests * [CPUCache]apply review comments * [CPUCache]fix compilation
This commit is contained in:
Zhang Yi
GitHub
parent
af105b86f8
commit
c1206ef447
@ -10,11 +10,45 @@
|
||||
#include <memory_desc/cpu_memory_desc_utils.h>
|
||||
#include <ngraph/opsets/opset1.hpp>
|
||||
#include "memory_desc/dnnl_blocked_memory_desc.h"
|
||||
#include <common/primitive_hashing_utils.hpp>
|
||||
|
||||
using namespace mkldnn;
|
||||
using namespace MKLDNNPlugin;
|
||||
using namespace InferenceEngine;
|
||||
|
||||
namespace {
|
||||
struct SoftmaxKey {
|
||||
DnnlMemoryDescCPtr inp0;
|
||||
impl_desc_type implType;
|
||||
size_t axis;
|
||||
|
||||
size_t hash() const;
|
||||
bool operator==(const SoftmaxKey& rhs) const;
|
||||
};
|
||||
|
||||
size_t SoftmaxKey::hash() const {
|
||||
using namespace dnnl::impl;
|
||||
using namespace dnnl::impl::primitive_hashing;
|
||||
|
||||
size_t seed = 0;
|
||||
|
||||
seed = hash_combine(seed, get_md_hash(inp0->getDnnlDesc().data));
|
||||
seed = hash_combine(seed, implType);
|
||||
seed = hash_combine(seed, axis);
|
||||
return seed;
|
||||
}
|
||||
|
||||
bool SoftmaxKey::operator==(const SoftmaxKey& rhs) const {
|
||||
bool retVal = true;
|
||||
if (inp0 != rhs.inp0) {
|
||||
retVal = retVal && inp0 && rhs.inp0 && inp0->getDnnlDesc() == rhs.inp0->getDnnlDesc();
|
||||
}
|
||||
|
||||
retVal = retVal && implType == rhs.implType && axis == rhs.axis;
|
||||
return retVal;
|
||||
}
|
||||
} // namespace
|
||||
|
||||
bool MKLDNNSoftMaxNode::isSupportedOperation(const std::shared_ptr<const ngraph::Node>& op, std::string& errorMessage) noexcept {
|
||||
try {
|
||||
if (!std::dynamic_pointer_cast<const ngraph::opset1::Softmax>(op)) {
|
||||
@ -108,32 +142,44 @@ void MKLDNNSoftMaxNode::createDescriptor(const std::vector<MemoryDescPtr> &input
|
||||
|
||||
void MKLDNNSoftMaxNode::prepareParams() {
|
||||
auto inpDesc = getParentEdgeAt(0)->getMemory().GetDescWithType<DnnlMemoryDesc>();
|
||||
const auto& in_candidate = inpDesc->getDnnlDesc();
|
||||
MKLDNNDescriptor desc(std::shared_ptr<softmax_forward::desc>(
|
||||
new softmax_forward::desc(prop_kind::forward_scoring, in_candidate, axis)));
|
||||
|
||||
const NodeDesc* selected_pd = getSelectedPrimitiveDescriptor();
|
||||
|
||||
if (selected_pd == nullptr)
|
||||
IE_THROW() << "Preferable primitive descriptor is not set for node " << getName() << ".";
|
||||
|
||||
SoftmaxKey key = {inpDesc, selected_pd->getImplementationType(), axis};
|
||||
auto engine = getEngine();
|
||||
auto builder = [&engine](const SoftmaxKey& key) -> std::shared_ptr<mkldnn::primitive> {
|
||||
softmax_forward::primitive_desc prim_desc;
|
||||
primitive_desc_iterator itpd = desc.createPrimitiveDescriptorIterator(getEngine());
|
||||
MKLDNNDescriptor desc(std::shared_ptr<softmax_forward::desc>(
|
||||
new softmax_forward::desc(prop_kind::forward_scoring, key.inp0->getDnnlDesc(), key.axis)));
|
||||
primitive_desc_iterator itpd = desc.createPrimitiveDescriptorIterator(engine);
|
||||
|
||||
while (itpd) {
|
||||
impl_desc_type impl_type = parse_impl_name(itpd.impl_info_str());
|
||||
if (impl_type == selected_pd->getImplementationType() ||
|
||||
// At least for oneDNN v2.4 the softmax primitive is optimized for the cases where the dimension of the softmax axis is physically dense.
|
||||
// There could be situations where it is not possible to detect the optimized case in advance in case of dynamic shapes, but
|
||||
// in runtime the shape could be suitable for the optimized implementation, so we have to select the optimized one.
|
||||
(ref_any == selected_pd->getImplementationType() && (impl_type & jit))) {
|
||||
if (impl_type == key.implType ||
|
||||
// At least for oneDNN v2.4 the softmax primitive is optimized for the cases where the dimension of the
|
||||
// softmax axis is physically dense. There could be situations where it is not possible to detect the
|
||||
// optimized case in advance in case of dynamic shapes, but in runtime the shape could be suitable for
|
||||
// the optimized implementation, so we have to select the optimized one.
|
||||
(ref_any == key.implType && (impl_type & jit))) {
|
||||
prim_desc = itpd.get();
|
||||
break;
|
||||
}
|
||||
if (!itpd.next_impl())
|
||||
return nullptr;
|
||||
}
|
||||
return std::make_shared<softmax_forward>(prim_desc);
|
||||
};
|
||||
|
||||
auto cache = getRuntimeCache();
|
||||
auto result = cache->getOrCreate(key, builder);
|
||||
|
||||
if (!result.first) {
|
||||
IE_THROW() << "Primitive descriptor was not found for node " << getName() << ".";
|
||||
}
|
||||
|
||||
prim.reset(new softmax_forward(prim_desc));
|
||||
prim = result.first;
|
||||
|
||||
auto src = getParentEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
|
||||
auto dst = getChildEdgesAtPort(0)[0]->getMemoryPtr()->GetPrimitive();
|
||||
|
||||
@ -1,222 +1,213 @@
|
||||
//// Copyright (C) 2018-2021 Intel Corporation
|
||||
//// SPDX-License-Identifier: Apache-2.0
|
||||
////
|
||||
// Copyright (C) 2018-2021 Intel Corporation
|
||||
// SPDX-License-Identifier: Apache-2.0
|
||||
//
|
||||
//#include <ngraph_functions/builders.hpp>
|
||||
//#include "test_utils/cpu_test_utils.hpp"
|
||||
//
|
||||
//using namespace InferenceEngine;
|
||||
//using namespace CPUTestUtils;
|
||||
//
|
||||
//namespace CPULayerTestsDefinitions {
|
||||
//using ShapesDefenition = std::pair<std::vector<ngraph::PartialShape>, std::vector<std::vector<ngraph::Shape>>>;
|
||||
//
|
||||
//struct SoftMaxConfig {
|
||||
// ShapesDefenition inputShapes;
|
||||
// size_t axis;
|
||||
//};
|
||||
//
|
||||
//typedef std::tuple<
|
||||
// InferenceEngine::Precision, // netPrecision
|
||||
// SoftMaxConfig, // softmaxTestConfig
|
||||
// std::string, // targetDevice
|
||||
// CPUSpecificParams
|
||||
//> softmaxCPUTestParams;
|
||||
//
|
||||
//class SoftMaxLayerCPUTest : public testing::WithParamInterface<softmaxCPUTestParams>,
|
||||
// virtual public LayerTestsUtils::LayerTestsCommon, public CPUTestsBase {
|
||||
//public:
|
||||
// static std::string getTestCaseName(const testing::TestParamInfo<softmaxCPUTestParams>& obj) {
|
||||
// CPUSpecificParams cpuParams;
|
||||
// InferenceEngine::Precision netPrecision;
|
||||
// SoftMaxConfig config;
|
||||
// std::string targetDevice;
|
||||
// std::tie(netPrecision, config, targetDevice, cpuParams) = obj.param;
|
||||
//
|
||||
// std::ostringstream result;
|
||||
// result << "netPRC=" << netPrecision.name() << "_";
|
||||
// if (!config.inputShapes.first.empty()) {
|
||||
// result << "IS=" << CommonTestUtils::partialShape2str(config.inputShapes.first) << "_";
|
||||
// }
|
||||
// result << "TS=";
|
||||
// for (const auto& shape : config.inputShapes.second) {
|
||||
// result << "(";
|
||||
// if (!shape.empty()) {
|
||||
// auto itr = shape.begin();
|
||||
// do {
|
||||
// result << CommonTestUtils::vec2str(*itr);
|
||||
// } while (++itr != shape.end() && result << "_");
|
||||
// }
|
||||
// result << ")_";
|
||||
// }
|
||||
// result << "axis=" << config.axis << "_";
|
||||
// result << "trgDev=" << targetDevice;
|
||||
// result << CPUTestsBase::getTestCaseName(cpuParams);
|
||||
//
|
||||
// return result.str();
|
||||
// }
|
||||
//
|
||||
//protected:
|
||||
// void SetUp() override {
|
||||
// InferenceEngine::Precision netPrecision;
|
||||
// SoftMaxConfig config;
|
||||
// CPUSpecificParams cpuParams;
|
||||
// std::tie(netPrecision, config, targetDevice, cpuParams) = this->GetParam();
|
||||
//
|
||||
// inPrc = outPrc = netPrecision;
|
||||
//
|
||||
// std::tie(inFmts, outFmts, priority, selectedType) = cpuParams;
|
||||
// if (selectedType.empty()) {
|
||||
// selectedType = getPrimitiveType();
|
||||
// }
|
||||
// selectedType.push_back('_');
|
||||
// selectedType += inPrc.name();
|
||||
//
|
||||
// const auto ngPrc = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(netPrecision);
|
||||
//
|
||||
// targetStaticShapes = config.inputShapes.second;
|
||||
// inputDynamicShapes = config.inputShapes.first;
|
||||
//
|
||||
// auto inputShape = targetStaticShapes.front().front();
|
||||
//
|
||||
// auto params = ngraph::builder::makeParams(ngPrc, {inputShape});
|
||||
//
|
||||
// const auto paramOuts =
|
||||
// ngraph::helpers::convert2OutputVector(ngraph::helpers::castOps2Nodes<ngraph::op::Parameter>(params));
|
||||
//
|
||||
// const auto softMax = std::make_shared<ngraph::opset1::Softmax>(paramOuts.at(0), config.axis);
|
||||
//
|
||||
// function = makeNgraphFunction(ngPrc, params, softMax, "SoftMax");
|
||||
// }
|
||||
//};
|
||||
//
|
||||
//TEST_P(SoftMaxLayerCPUTest, CompareWithRefs) {
|
||||
// SKIP_IF_CURRENT_TEST_IS_DISABLED()
|
||||
//
|
||||
// Run();
|
||||
// CheckPluginRelatedResults(executableNetwork, "Softmax");
|
||||
//}
|
||||
//
|
||||
//namespace {
|
||||
////not optimized cpu spec
|
||||
//const auto notOptimizedCPUSpec = CPUSpecificParams{{}, {}, {}, "ref_any"};
|
||||
//
|
||||
//const std::vector<SoftMaxConfig> optimizedConfigsFP32 = {
|
||||
// //Static shapes
|
||||
// {ShapesDefenition{{}, {{{1, 100}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{10, 10}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{100, 1}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{100, 1}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 1}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 1, 1}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 1}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5}}}}, 3},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 1, 1, 1}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 1, 1}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 1, 1}}}}, 3},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5}}}}, 3},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 1}}}}, 4},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5}}}}, 4},
|
||||
// //Dynamic shapes
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {-1, -1}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{10, 10}},
|
||||
// {{15, 15}},
|
||||
// {{10, 5}}
|
||||
// }}, 1},
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {{1, 100}, {1, 100}}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{10, 10}},
|
||||
// {{15, 15}},
|
||||
// {{10, 5}}
|
||||
// }}, 1},
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {-1, -1, 1, 1, 1}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{5, 5, 1, 1, 1}},
|
||||
// {{10, 7, 1, 1, 1}}
|
||||
// }}, 1},
|
||||
//};
|
||||
//
|
||||
//const std::vector<SoftMaxConfig> notOptimizedConfigsFP32 {
|
||||
// //Static shapes
|
||||
// {ShapesDefenition{{}, {{{1, 100}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{10, 10}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{10, 10, 10}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{10, 10, 10}}}}, 1},
|
||||
// //Dynamic shapes
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {-1, -1}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{10, 1}}, {{15, 15}}, {{10, 5}}
|
||||
// }}, 0},
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {{1, 100}, {1, 100}, -1}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{10, 10, 10}}, {{10, 10, 1}}, {{10, 5, 10}}
|
||||
// }}, 1},
|
||||
//};
|
||||
//
|
||||
//const std::vector<SoftMaxConfig> unsupportedConfigsFP32 {
|
||||
// //Static shapes
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 0},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 1},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 2},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 3},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 4},
|
||||
// {ShapesDefenition{{}, {{{5, 5, 5, 5, 5, 5}}}}, 5},
|
||||
// //Dynamic shapes
|
||||
// {ShapesDefenition{
|
||||
// { //dynamic shape
|
||||
// {-1, -1, -1, -1, -1, -1}
|
||||
// },
|
||||
// { //target static shapes
|
||||
// {{5, 5, 5, 5, 5, 5}}, {{7, 7, 7, 7, 7, 7}}
|
||||
// }}, 4},
|
||||
//};
|
||||
//
|
||||
//const auto OptimizedParams = testing::Combine(
|
||||
// testing::Values(Precision::FP32, Precision::BF16),
|
||||
// testing::ValuesIn(optimizedConfigsFP32),
|
||||
// testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
// testing::Values(emptyCPUSpec));
|
||||
//
|
||||
//INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_Optimized_CPU, SoftMaxLayerCPUTest, OptimizedParams, SoftMaxLayerCPUTest::getTestCaseName);
|
||||
//
|
||||
//const auto NotOptimizedParams = testing::Combine(
|
||||
// testing::Values(Precision::FP32, Precision::BF16),
|
||||
// testing::ValuesIn(notOptimizedConfigsFP32),
|
||||
// testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
// testing::Values(notOptimizedCPUSpec));
|
||||
//
|
||||
//INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_CPU, SoftMaxLayerCPUTest, NotOptimizedParams, SoftMaxLayerCPUTest::getTestCaseName);
|
||||
//
|
||||
//const auto UnsupportedParams = testing::Combine(
|
||||
// testing::Values(Precision::FP32, Precision::BF16),
|
||||
// testing::ValuesIn(unsupportedConfigsFP32),
|
||||
// testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
// testing::Values(notOptimizedCPUSpec));
|
||||
//
|
||||
//INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_Unsupported_CPU, SoftMaxLayerCPUTest, UnsupportedParams, SoftMaxLayerCPUTest::getTestCaseName);
|
||||
//
|
||||
//} // namespace
|
||||
//} // namespace CPULayerTestsDefinitions
|
||||
|
||||
#include <ngraph_functions/builders.hpp>
|
||||
|
||||
#include "shared_test_classes/base/ov_subgraph.hpp"
|
||||
#include "test_utils/cpu_test_utils.hpp"
|
||||
|
||||
using namespace InferenceEngine;
|
||||
using namespace CPUTestUtils;
|
||||
using namespace ov::test;
|
||||
|
||||
namespace CPULayerTestsDefinitions {
|
||||
|
||||
struct SoftMaxConfig {
|
||||
ov::test::InputShape inputShape;
|
||||
size_t axis;
|
||||
};
|
||||
|
||||
typedef std::tuple<ElementType, // netPrecision
|
||||
SoftMaxConfig, // softmaxTestConfig
|
||||
std::string, // targetDevice
|
||||
CPUSpecificParams>
|
||||
softmaxCPUTestParams;
|
||||
|
||||
class SoftMaxLayerCPUTest : public testing::WithParamInterface<softmaxCPUTestParams>,
|
||||
virtual public SubgraphBaseTest,
|
||||
public CPUTestsBase {
|
||||
public:
|
||||
static std::string getTestCaseName(const testing::TestParamInfo<softmaxCPUTestParams>& obj) {
|
||||
CPUSpecificParams cpuParams;
|
||||
ElementType inType;
|
||||
SoftMaxConfig config;
|
||||
std::string targetDevice;
|
||||
std::tie(inType, config, targetDevice, cpuParams) = obj.param;
|
||||
|
||||
std::ostringstream result;
|
||||
result << "netPRC=" << inType << "_";
|
||||
result << "IS=" << CommonTestUtils::partialShape2str({config.inputShape.first}) << "_";
|
||||
result << "TS=";
|
||||
for (const auto& shape : config.inputShape.second) {
|
||||
result << "(";
|
||||
result << CommonTestUtils::vec2str(shape);
|
||||
result << ")_";
|
||||
}
|
||||
result << "axis=" << config.axis << "_";
|
||||
result << "trgDev=" << targetDevice;
|
||||
result << CPUTestsBase::getTestCaseName(cpuParams);
|
||||
|
||||
return result.str();
|
||||
}
|
||||
|
||||
protected:
|
||||
void SetUp() override {
|
||||
ElementType inType;
|
||||
SoftMaxConfig config;
|
||||
CPUSpecificParams cpuParams;
|
||||
std::tie(inType, config, targetDevice, cpuParams) = this->GetParam();
|
||||
|
||||
std::tie(inFmts, outFmts, priority, selectedType) = cpuParams;
|
||||
if (selectedType.empty()) {
|
||||
selectedType = getPrimitiveType();
|
||||
}
|
||||
|
||||
if (inType == ElementType::bf16) {
|
||||
rel_threshold = 1e-2f;
|
||||
}
|
||||
selectedType = makeSelectedTypeStr(selectedType, inType);
|
||||
init_input_shapes({config.inputShape});
|
||||
auto params = ngraph::builder::makeDynamicParams(inType, inputDynamicShapes);
|
||||
|
||||
const auto paramOuts =
|
||||
ngraph::helpers::convert2OutputVector(ngraph::helpers::castOps2Nodes<ngraph::op::Parameter>(params));
|
||||
|
||||
const auto softMax = std::make_shared<ngraph::opset1::Softmax>(paramOuts.at(0), config.axis);
|
||||
|
||||
function = makeNgraphFunction(inType, params, softMax, "SoftMax");
|
||||
}
|
||||
};
|
||||
|
||||
TEST_P(SoftMaxLayerCPUTest, CompareWithRefs) {
|
||||
SKIP_IF_CURRENT_TEST_IS_DISABLED()
|
||||
run();
|
||||
CheckPluginRelatedResults(executableNetwork, "Softmax");
|
||||
}
|
||||
|
||||
namespace {
|
||||
// not optimized cpu spec
|
||||
const auto notOptimizedCPUSpec = CPUSpecificParams{{}, {}, {"ref_any"}, "ref_any"};
|
||||
|
||||
const std::vector<SoftMaxConfig> optimizedConfigsFP32 = {
|
||||
// Static shapes
|
||||
{ov::test::InputShape{ov::PartialShape{1, 100}, {ov::Shape{1, 100}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{10, 10}, {ov::Shape{10, 10}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{100, 1}, {ov::Shape{100, 1}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{100, 1}, {ov::Shape{100, 1}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 1}, {ov::Shape{5, 5, 1}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5}, {ov::Shape{5, 5, 5}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 1, 1}, {ov::Shape{5, 5, 1, 1}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 1}, {ov::Shape{5, 5, 5, 1}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5}}}, 3},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 1, 1, 1}, {ov::Shape{5, 5, 1, 1, 1}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 1, 1}, {ov::Shape{5, 5, 5, 1, 1}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 1, 1}, {ov::Shape{5, 5, 5, 1, 1}}}, 3},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5}}}, 3},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 1}, {ov::Shape{5, 5, 5, 5, 1}}}, 4},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5}}}, 4},
|
||||
// Dynamic shapes
|
||||
{ov::test::InputShape{// dynamic shape
|
||||
ov::PartialShape{-1, -1},
|
||||
{// target static shapes
|
||||
ov::Shape{10, 10},
|
||||
ov::Shape{15, 15},
|
||||
ov::Shape{10, 10},
|
||||
ov::Shape{10, 5}}},
|
||||
1},
|
||||
{ov::test::InputShape{// dynamic shape
|
||||
ov::PartialShape{-1, -1, 1, 1, 1},
|
||||
{// target static shapes
|
||||
ov::Shape{5, 5, 1, 1, 1},
|
||||
ov::Shape{10, 7, 1, 1, 1},
|
||||
ov::Shape{5, 5, 1, 1, 1}}},
|
||||
1},
|
||||
};
|
||||
|
||||
const std::vector<SoftMaxConfig> notOptimizedConfigsFP32{
|
||||
// Static shapes
|
||||
{ov::test::InputShape{ov::PartialShape{1, 100}, {ov::Shape{1, 100}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{10, 10}, {ov::Shape{10, 10}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{10, 10, 10}, {ov::Shape{10, 10, 10}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{10, 10, 10}, {ov::Shape{10, 10, 10}}}, 1},
|
||||
// Dynamic shapes
|
||||
{ov::test::InputShape{// dynamic shape
|
||||
ov::PartialShape{-1, -1},
|
||||
{// target static shapes
|
||||
ov::Shape{10, 1},
|
||||
ov::Shape{15, 15},
|
||||
ov::Shape{10, 5},
|
||||
ov::Shape{15, 15}}},
|
||||
0},
|
||||
{ov::test::InputShape{// dynamic shape
|
||||
ov::PartialShape{ov::Dimension{1, 100}, ov::Dimension{1, 100}, -1},
|
||||
{// target static shapes
|
||||
ov::Shape{10, 10, 10},
|
||||
ov::Shape{10, 10, 1},
|
||||
ov::Shape{10, 5, 10},
|
||||
ov::Shape{10, 10, 1}}},
|
||||
1},
|
||||
};
|
||||
|
||||
const std::vector<SoftMaxConfig> unsupportedConfigsFP32{
|
||||
// Static shapes
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 0},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 1},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 2},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 3},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 4},
|
||||
{ov::test::InputShape{ov::PartialShape{5, 5, 5, 5, 5, 5}, {ov::Shape{5, 5, 5, 5, 5, 5}}}, 5},
|
||||
// Dynamic shapes
|
||||
{ov::test::InputShape{// dynamic shape
|
||||
ov::PartialShape{-1, -1, -1, -1, -1, -1},
|
||||
{// target static shapes
|
||||
ov::Shape{5, 5, 5, 5, 5, 5},
|
||||
ov::Shape{7, 7, 7, 7, 7, 7},
|
||||
ov::Shape{5, 5, 5, 5, 5, 5}}},
|
||||
4},
|
||||
};
|
||||
|
||||
const auto avx512 = CPUSpecificParams{{}, {}, {"jit_avx512"}, "jit_avx512"};
|
||||
const auto avx2 = CPUSpecificParams{{}, {}, {"jit_avx2"}, "jit_avx2"};
|
||||
const auto sse42 = CPUSpecificParams{{}, {}, {"jit_sse42"}, "jit_sse42"};
|
||||
const auto ref = CPUSpecificParams{{}, {}, {"ref_any"}, "ref_any"};
|
||||
|
||||
const std::vector<CPUSpecificParams> vecCpuConfigs = {ref, sse42, avx2, avx512};
|
||||
const auto OptimizedParams = testing::Combine(testing::Values(ElementType::f32, ElementType::bf16),
|
||||
testing::ValuesIn(optimizedConfigsFP32),
|
||||
testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
testing::ValuesIn(filterCPUInfoForDevice(vecCpuConfigs)));
|
||||
|
||||
INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_Optimized_CPU,
|
||||
SoftMaxLayerCPUTest,
|
||||
OptimizedParams,
|
||||
SoftMaxLayerCPUTest::getTestCaseName);
|
||||
|
||||
const auto NotOptimizedParams = testing::Combine(testing::Values(ElementType::f32, ElementType::bf16),
|
||||
testing::ValuesIn(notOptimizedConfigsFP32),
|
||||
testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
testing::Values(notOptimizedCPUSpec));
|
||||
|
||||
INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_CPU,
|
||||
SoftMaxLayerCPUTest,
|
||||
NotOptimizedParams,
|
||||
SoftMaxLayerCPUTest::getTestCaseName);
|
||||
|
||||
const auto UnsupportedParams = testing::Combine(testing::Values(ElementType::f32, ElementType::bf16),
|
||||
testing::ValuesIn(unsupportedConfigsFP32),
|
||||
testing::Values(CommonTestUtils::DEVICE_CPU),
|
||||
testing::Values(notOptimizedCPUSpec));
|
||||
|
||||
INSTANTIATE_TEST_SUITE_P(smoke_SoftMax_Unsupported_CPU,
|
||||
SoftMaxLayerCPUTest,
|
||||
UnsupportedParams,
|
||||
SoftMaxLayerCPUTest::getTestCaseName);
|
||||
|
||||
} // namespace
|
||||
} // namespace CPULayerTestsDefinitions
|
||||
|
||||
Loading…
Reference in New Issue
Block a user