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

[GNA] Fix for GeminiLake detection (#18653) 

* [GNA] Fix for GeminiLake detection

* Added HWGeneration::GNA_1_0_E enumerator
Added DeviceVersion::GNAEmbedded1_0 enumerator, changed the meaning of DeviceVersion::GNA1_0.
Updated ConvLowPrecision test with all supported targets

* [GNA] Extended a few tests with GNA1.0
This commit is contained in:
Marcin Kacprzak 2023-07-28 08:18:22 +02:00 committed by GitHub
parent c4f8afb741
commit e1ec17013d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
15 changed files with 68 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
samples/cpp/speech_sample/main.cpp
View File

@ -223,6 +223,7 @@ int main(int argc, char* argv[]) {
gnaPluginConfig[ov::hint::inference_precision.name()] = (FLAGS_qb == 8) ? ov::element::i8 : ov::element::i16;
const std::unordered_map<std::string, ov::intel_gna::HWGeneration> StringHWGenerationMap{
{"GNA_TARGET_1_0", ov::intel_gna::HWGeneration::GNA_1_0},
{"GNA_TARGET_1_0_E", ov::intel_gna::HWGeneration::GNA_1_0_E},
{"GNA_TARGET_2_0", ov::intel_gna::HWGeneration::GNA_2_0},
{"GNA_TARGET_3_0", ov::intel_gna::HWGeneration::GNA_3_0},
{"GNA_TARGET_3_1", ov::intel_gna::HWGeneration::GNA_3_1},

5
src/inference/include/openvino/runtime/intel_gna/properties.hpp
View File

@ -118,6 +118,7 @@ inline std::istream& operator>>(std::istream& is, ExecutionMode& execution_mode)
enum class HWGeneration {
UNDEFINED = 0, //!< GNA HW generation is undefined
GNA_1_0 = 4, //!< GNA HW generation 1.0
GNA_1_0_E = 9, //!< GNA HW generation 1.0 embedded
GNA_2_0 = 1, //!< GNA HW generation 2.0
GNA_3_0 = 2, //!< GNA HW generation 3.0
GNA_3_1 = 5, //!< GNA HW generation 3.1
@ -134,6 +135,8 @@ inline std::ostream& operator<<(std::ostream& os, const HWGeneration& hw_generat
return os << "UNDEFINED";
case HWGeneration::GNA_1_0:
return os << "GNA_1_0";
case HWGeneration::GNA_1_0_E:
return os << "GNA_1_0_E";
case HWGeneration::GNA_2_0:
return os << "GNA_2_0";
case HWGeneration::GNA_3_0:
@ -160,6 +163,8 @@ inline std::istream& operator>>(std::istream& is, HWGeneration& hw_generation) {
hw_generation = HWGeneration::UNDEFINED;
} else if (str == "GNA_1_0") {
hw_generation = HWGeneration::GNA_1_0;
} else if (str == "GNA_1_0_E") {
hw_generation = HWGeneration::GNA_1_0_E;
} else if (str == "GNA_2_0") {
hw_generation = HWGeneration::GNA_2_0;
} else if (str == "GNA_3_0") {

6
src/plugins/intel_gna/src/backend/gna_limitations.cpp
View File

@ -680,8 +680,9 @@ constexpr uint32_t Limitations::kMemoryPageSize;
thread_local std::shared_ptr<Limitations> Limitations::k_instance{nullptr};
Limitations::Limitations(const DeviceVersion& target) {
m_use_only_16bit_conv_weights = (target == DeviceVersion::GNA1_0 || target == DeviceVersion::GNA2_0 ||
target == DeviceVersion::GNA3_0 || target == DeviceVersion::GNA3_1);
m_use_only_16bit_conv_weights =
(target == DeviceVersion::GNA1_0 || target == DeviceVersion::GNAEmbedded1_0 ||
target == DeviceVersion::GNA2_0 || target == DeviceVersion::GNA3_0 || target == DeviceVersion::GNA3_1);
m_mem_alignment = get_memory_alignment_bytes(target);
m_cnn_validator = cnn2d::AbstractValidator::Create(target);
@ -698,6 +699,7 @@ size_t Limitations::get_min_batch_to_fit_in_buffer(InferenceEngine::DataPtr inpu
size_t Limitations::get_memory_alignment_bytes(const DeviceVersion& target) const {
static const std::unordered_map<DeviceVersion, size_t> mem_alignment_map{{DeviceVersion::GNA1_0, 64},
{DeviceVersion::GNAEmbedded1_0, 64},
{DeviceVersion::GNA2_0, 64},
{DeviceVersion::GNA3_0, 64},
{DeviceVersion::GNA3_1, 64},

8
src/plugins/intel_gna/src/common/gna_target.cpp
View File

@ -18,6 +18,7 @@ namespace target {
static constexpr const char* kGnaTargetUnspecified = "";
static constexpr const char* kGnaTargetSoftwareEmulation = "GNA_SW_EMULATION";
static constexpr const char* kGnaTarget1_0 = "GNA_TARGET_1_0";
static constexpr const char* kGnaTarget1_0_e = "GNA_TARGET_1_0_E";
static constexpr const char* kGnaTarget2_0 = "GNA_TARGET_2_0";
static constexpr const char* kGnaTarget3_0 = "GNA_TARGET_3_0";
static constexpr const char* kGnaTarget3_1 = "GNA_TARGET_3_1";
@ -28,6 +29,7 @@ static constexpr const char* kGnaTarget4_0 = "GNA_TARGET_4_0";
static const std::unordered_map<HWGeneration, DeviceVersion> HWGenerationDeviceMap{
{HWGeneration::GNA_1_0, DeviceVersion::GNA1_0},
{HWGeneration::GNA_1_0_E, DeviceVersion::GNAEmbedded1_0},
{HWGeneration::GNA_2_0, DeviceVersion::GNA2_0},
{HWGeneration::GNA_3_0, DeviceVersion::GNA3_0},
{HWGeneration::GNA_3_1, DeviceVersion::GNA3_1},
@ -38,7 +40,8 @@ static const std::unordered_map<HWGeneration, DeviceVersion> HWGenerationDeviceM
{HWGeneration::UNDEFINED, DeviceVersion::NotSet}};
static const std::unordered_map<Gna2DeviceVersion, DeviceVersion> GnaDeviceMap{
{Gna2DeviceVersionEmbedded1_0, DeviceVersion::GNA1_0},
{Gna2DeviceVersion1_0, DeviceVersion::GNA1_0},
{Gna2DeviceVersionEmbedded1_0, DeviceVersion::GNAEmbedded1_0},
{Gna2DeviceVersion2_0, DeviceVersion::GNA2_0},
{Gna2DeviceVersion3_0, DeviceVersion::GNA3_0},
{Gna2DeviceVersionEmbedded3_1, DeviceVersion::GNA3_1},
@ -50,6 +53,7 @@ static const std::unordered_map<Gna2DeviceVersion, DeviceVersion> GnaDeviceMap{
static const std::unordered_map<std::string, DeviceVersion> StringDeviceMap{
{kGnaTarget1_0, DeviceVersion::GNA1_0},
{kGnaTarget1_0_e, DeviceVersion::GNAEmbedded1_0},
{kGnaTarget2_0, DeviceVersion::GNA2_0},
{kGnaTarget3_0, DeviceVersion::GNA3_0},
{kGnaTarget3_1, DeviceVersion::GNA3_1},
@ -60,7 +64,7 @@ static const std::unordered_map<std::string, DeviceVersion> StringDeviceMap{
{kGnaTargetSoftwareEmulation, DeviceVersion::SoftwareEmulation},
{kGnaTargetUnspecified, DeviceVersion::NotSet}};
static const std::vector<DeviceVersion> EmbeddedDevices{DeviceVersion::GNA1_0,
static const std::vector<DeviceVersion> EmbeddedDevices{DeviceVersion::GNAEmbedded1_0,
DeviceVersion::GNA3_1,
DeviceVersion::GNAEmbedded3_5,
DeviceVersion::GNA3_6,

3
src/plugins/intel_gna/src/common/gna_target.hpp
View File

@ -16,7 +16,8 @@ namespace target {
enum class DeviceVersion {
NotSet = -1,
SoftwareEmulation = 0,
GNA1_0 = 0x10e,
GNA1_0 = 0x10,
GNAEmbedded1_0 = 0x10e,
GNA2_0 = 0x20,
GNA3_0 = 0x30,
GNA3_1 = 0x31e,

4
src/plugins/intel_gna/src/gna2_model_export_helper.cpp
View File

@ -105,7 +105,7 @@ void ExportTlvModel(uint32_t modelId,
const std::vector<GnaEndpoint>& allInputs,
const std::vector<GnaEndpoint>& allOutputs,
const GnaAllocations& allAllocations) {
if (compile_target == target::DeviceVersion::GNA1_0) {
if (compile_target == target::DeviceVersion::GNAEmbedded1_0) {
THROW_GNA_EXCEPTION << "Unsupported compile target for TLV export: GNA Embedded 1.0" << std::endl;
}
@ -229,7 +229,7 @@ void* ExportSueLegacyUsingGnaApi2(uint32_t modelId, uint32_t deviceIndex, Gna2Mo
status = Gna2ModelExportConfigSetSource(exportConfig, deviceIndex, modelId);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigSetSource");
status = Gna2ModelExportConfigSetTarget(exportConfig, DeviceToGna(target::DeviceVersion::GNA1_0));
status = Gna2ModelExportConfigSetTarget(exportConfig, DeviceToGna(target::DeviceVersion::GNAEmbedded1_0));
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigSetTarget");
void* bufferSueCreekHeader = nullptr;

1
src/plugins/intel_gna/src/gna_device.cpp
View File

@ -571,6 +571,7 @@ uint32_t GNADeviceHelper::retrieveMaxLayersCount() {
switch (target->get_effective_execution_target()) {
case DeviceVersion::GNA1_0:
case DeviceVersion::GNAEmbedded1_0:
case DeviceVersion::GNA2_0:
return Limitations::kMaxLayersCountGNA2_0;
case DeviceVersion::GNA3_0:

9
src/plugins/intel_gna/src/gna_device.hpp
View File

@ -172,7 +172,14 @@ private:
static void enforceLegacyCnns(Gna2Model& gnaModel);
static void enforceLegacyCnnsWhenNeeded(Gna2Model& gnaModel);
static bool is_up_to_20_hw(const target::DeviceVersion device_version) {
return device_version <= target::DeviceVersion::GNA2_0 && is_hw_target(device_version);
switch (device_version) {
case target::DeviceVersion::GNA1_0:
case target::DeviceVersion::GNAEmbedded1_0:
case target::DeviceVersion::GNA2_0:
return true;
default:
return false;
}
}
void createVirtualDevice(const target::DeviceVersion& devVersion);
void updateGnaDeviceVersion();

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

@ -780,7 +780,7 @@ void GNAPlugin::DumpXNNToFile() const {
const auto& inputsDesc = inputs_ptr_->Get();
const auto& outputsDesc = outputs_.Get();
if (config.target->get_effective_compile_target() == target::DeviceVersion::GNA1_0) {
if (config.target->get_effective_compile_target() == target::DeviceVersion::GNAEmbedded1_0) {
auto dump = gnadevice->dumpXnn(modelId);
dump.header.RwRegionSize = static_cast<uint32_t>(gnamem->getRegionBytes(REGION_SCRATCH));
dump.header.InputScalingFactor = inputsDesc.begin()->scale_factor;

2
src/plugins/intel_gna/src/request/subrequest_impl.cpp
View File

@ -29,7 +29,7 @@ RequestStatus SubrequestImpl::wait(int64_t timeoutMilliseconds) {
try {
status_ = waitHandler_(requestID_, timeoutMilliseconds);
} catch (const std::exception& e) {
ov::intel_gna::log::error() << "Exception when executiong wait: " << e.what() << std::endl;
ov::intel_gna::log::error() << "Exception when execution wait: " << e.what() << std::endl;
status_ = RequestStatus::kCompletedWithError;
}

1
src/plugins/intel_gna/tests/functional/pass_tests/concat_transpose.cpp
View File

@ -120,6 +120,7 @@ TEST_P(TransposesConcatTest, CompareWithRefs) {
std::vector<std::map<std::string, std::string>> configs = {{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}}};
std::vector<std::map<std::string, std::string>> target_configs = {{{"GNA_DEVICE_MODE", "GNA_SW_FP32"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_1_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}}};

1
src/plugins/intel_gna/tests/functional/pass_tests/convert_padded_to_valid_conv.cpp
View File

@ -225,6 +225,7 @@ const std::vector<InferenceEngine::Precision> netPrecisions = {InferenceEngine::
InferenceEngine::Precision::FP16};
const std::vector<std::map<std::string, std::string>> configs1D = {
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_SCALE_FACTOR_0", "1"}, {"GNA_EXEC_TARGET", "GNA_TARGET_1_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_SCALE_FACTOR_0", "1"}, {"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_SCALE_FACTOR_0", "1"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_SCALE_FACTOR_0", "1"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}}};

1
src/plugins/intel_gna/tests/functional/preprocess_tests/gather_transpose_cpu.cpp
View File

@ -199,6 +199,7 @@ TEST_P(PrePostProcessBaseTest, CompareWithRefs) {
std::vector<std::map<std::string, std::string>> configs = {{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}}};
std::vector<std::map<std::string, std::string>> target_configs = {{{"GNA_DEVICE_MODE", "GNA_SW_FP32"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_1_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}}};

45
src/plugins/intel_gna/tests/functional/shared_tests_instances/single_layer_tests/conv_low_precision.cpp
View File

@ -17,15 +17,16 @@
#include "ngraph_functions/builders.hpp"
#include "ngraph_functions/pass/convert_prc.hpp"
#include "ngraph_functions/utils/ngraph_helpers.hpp"
#include "openvino/opsets/opset12.hpp"
#include "shared_test_classes/base/layer_test_utils.hpp"
namespace ConvLowPrecicionTestNs {
using namespace ngraph;
using namespace ov;
using namespace ngraph::builder;
using namespace ngraph::element;
using namespace ngraph::op;
using namespace ngraph::opset1;
using namespace ov::element;
using namespace ov::op;
using namespace ov::opset12;
using namespace std;
using ConvLowPrecisionTestParams = tuple<InferenceEngine::Precision, // Network Precision
@ -128,15 +129,21 @@ TEST_P(ConvLowPrecisionTest, CompareWithRefs) {
const vector<InferenceEngine::Precision> netPrecisions = {InferenceEngine::Precision::FP16,
InferenceEngine::Precision::FP32};
const vector<map<string, string>> configs_3_X = {
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}}};
const vector<map<string, string>> configs_1 = {
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_1_0"}},
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_1_0"}},
};
const vector<map<string, string>> configs_2_0 = {
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
const vector<map<string, string>> configs_2 = {
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
};
const vector<map<string, string>> configs_3 = {
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I8"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}},
{{"GNA_DEVICE_MODE", "GNA_AUTO"}, {"GNA_PRECISION", "I16"}, {"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}},
};
const Shape conv1D = {1, 8, 1, 16};
@ -148,11 +155,21 @@ const vector<pair<float, float>> fqMinMax = {{-1.0f, 1.0f}};
const vector<std::size_t> levels = {numeric_limits<uint8_t>::max(), numeric_limits<uint16_t>::max()};
INSTANTIATE_TEST_SUITE_P(smoke_LowPrecision10,
ConvLowPrecisionTest,
::testing::Combine(::testing::ValuesIn(netPrecisions),
::testing::Values(ov::test::utils::DEVICE_GNA),
::testing::ValuesIn(configs_1),
::testing::Values(conv1D),
::testing::ValuesIn(fqMinMax),
::testing::ValuesIn(levels)),
ConvLowPrecisionTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_LowPrecision20,
ConvLowPrecisionTest,
::testing::Combine(::testing::ValuesIn(netPrecisions),
::testing::Values(ov::test::utils::DEVICE_GNA),
::testing::ValuesIn(configs_2_0),
::testing::ValuesIn(configs_2),
::testing::Values(conv1D),
::testing::ValuesIn(fqMinMax),
::testing::ValuesIn(levels)),
@ -162,7 +179,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_LowPrecision3X,
ConvLowPrecisionTest,
::testing::Combine(::testing::ValuesIn(netPrecisions),
::testing::Values(ov::test::utils::DEVICE_GNA),
::testing::ValuesIn(configs_3_X),
::testing::ValuesIn(configs_3),
::testing::ValuesIn(inputShapes),
::testing::ValuesIn(fqMinMax),
::testing::ValuesIn(levels)),

3
src/plugins/intel_gna/tests/unit/gna_plugin_config_test.cpp
View File

@ -184,6 +184,9 @@ IE_SUPPRESS_DEPRECATED_END
TEST_F(GNAPluginConfigTest, GnaConfigGnaExecTargetTest) {
SetAndCompare(GNA_CONFIG_KEY(EXEC_TARGET), "GNA_TARGET_1_0");
EXPECT_EQ(config.target->get_user_set_execution_target(), DeviceVersion::GNA1_0);
SetAndCompare(GNA_CONFIG_KEY(EXEC_TARGET), "GNA_TARGET_1_0_E");
EXPECT_EQ(config.target->get_user_set_execution_target(), DeviceVersion::GNAEmbedded1_0);
SetAndCompare(GNA_CONFIG_KEY(EXEC_TARGET), "GNA_TARGET_2_0");
EXPECT_EQ(config.target->get_user_set_execution_target(), DeviceVersion::GNA2_0);
SetAndCompare(GNA_CONFIG_KEY(EXEC_TARGET), "GNA_TARGET_3_0");