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

[GNA] Add support for future devices with relaxed capabilities (#12000)

Browse Source
This commit is contained in:
Krzysztof Bruniecki 2022-07-13 14:18:34 +02:00 committed by GitHub
parent a185299ba4
commit 9d0e623af6
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
16 changed files with 788 additions and 242 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/plugins/intel_gna/backend/am_intel_dnn.hpp
View File

@ -13,6 +13,7 @@
#include "gna/gna_config.hpp"
#include "gna_plugin_log.hpp"
#include "common/gna_target.hpp"
#include "memory/gna_memory.hpp"
#include <gna2-model-api.h>
@ -296,7 +297,7 @@ public:
void WriteDnnText(const char *filename, intel_dnn_number_type_t logging_precision);
void InitGNAStruct(Gna2Model *gnaModel, const std::string& gnaCompileTarget = InferenceEngine::GNAConfigParams::GNA_TARGET_2_0);
void InitGNAStruct(Gna2Model *gnaModel, const std::string& gnaCompileTarget = common::kGnaTarget2_0);
void DestroyGNAStruct(Gna2Model *gnaModel);

121
src/plugins/intel_gna/backend/gna_limitations.cpp
View File

@ -14,6 +14,7 @@
#include <layers/gna_layer_info.hpp>
#include "gna_graph_tools.hpp"
#include "gna_lib_ver_selector.hpp"
#include "common/gna_target.hpp"
namespace GNAPluginNS {
namespace GNALimitations {
@ -118,55 +119,113 @@ std::string RectLimitByChannelsAndPrecision::GetErrorOrEmpty(const uint32_t h, c
return GetByPrecision(precision).GetErrorOrEmpty(h, w, channels, what);
}
const RangeLimit2D Validator_30::kInputHWLimit{{16, 384, "input height"}, {16, 240, "input width"}};
const RangeMultipleLimit Validator_30::kInputChannelsNumberLimit{{8, 384, "number of input channels"}, 8};
const RangeMultipleLimit Validator_30::kKernelNumberLimit{{8, 1024, "number of kernels"}, 8};
const RectLimitByChannelsAndPrecision Validator_30::kKernelLimit{
{{{96, {7, 7}}, {136, {7, 5}}, {168, {7, 4}}, {240, {7, 3}}, {384, {7, 2}}}},
{{{48, {7, 7}}, {64, {7, 5}}, {80, {7, 4}}, {120, {7, 3}}, {384, {7, 1}}}},
};
const RangeLimit2D Validator_30::kDilationLimit{{convDilationHeight, convDilationHeight, "dilation height"},
{convDilationWidth, convDilationWidth, "dilation width"}};
bool Validator_30::ValidateCnn2D(const std::string &name, const uint32_t inHeight, const uint32_t inWidth,
const uint32_t inChannels, const uint32_t kernelH, const uint32_t kernelW, const uint32_t kernelN,
const uint32_t strideH, const uint32_t strideW, const uint32_t dilationH, const uint32_t dilationW,
OvGnaType inPrecision, bool exception) const {
const OvGnaType inPrecision, const bool throwOnError) const {
const auto& kStrideLimit = kKernelLimit;
auto error = kInputHWLimit.GetErrorOrEmpty(inHeight, inWidth);
const std::string prefix = "Layer Convolution2D: " + name + ":";
auto error = inputHWLimit.GetErrorOrEmpty(inHeight, inWidth);
error += kKernelNumberLimit.GetErrorOrEmpty(kernelN);
error += kInputChannelsNumberLimit.GetErrorOrEmpty(inChannels);
error += kKernelLimit.GetErrorOrEmpty(kernelH, kernelW, inPrecision, inChannels, "kernel");
error += kStrideLimit.GetErrorOrEmpty(strideH, strideW, inPrecision, inChannels, "convolution stride");
error += kernelNumberLimit.GetErrorOrEmpty(kernelN);
error += inputChannelsNumberLimit.GetErrorOrEmpty(inChannels);
error += kernelLimit.GetErrorOrEmpty(kernelH, kernelW, inPrecision, inChannels, "kernel");
error += strideLimit.GetErrorOrEmpty(strideH, strideW, inPrecision, inChannels, "convolution stride");
const RangeLimit kKernelStrideHLimit{1, kernelH, "kernel stride height (must be up to kernel height)"};
const RangeLimit kKernelStrideWLimit{1, kernelW, "kernel stride width (must be up to kernel width)"};
const RangeLimit kernelStrideHLimit{1, kernelH, "kernel stride height (must be up to kernel height)"};
const RangeLimit kernelStrideWLimit{1, kernelW, "kernel stride width (must be up to kernel width)"};
error += kKernelStrideHLimit.GetErrorOrEmpty(strideH);
error += kKernelStrideWLimit.GetErrorOrEmpty(strideW);
error += kernelStrideHLimit.GetErrorOrEmpty(strideH);
error += kernelStrideWLimit.GetErrorOrEmpty(strideW);
error += kDilationLimit.GetErrorOrEmpty(dilationH, dilationW);
error += dilationLimit.GetErrorOrEmpty(dilationH, dilationW);
if (exception)
ThrowIfNotEmpty(prefix, error);
return error.empty() ? true : false;
return ValidationSuccesful(throwOnError, error, name, "Convolution2D");
}
const VectorOrSquareLimit Validator_30::kPoolingWindowLimit{3, 1, 1};
bool Validator_30::ValidatePooling2D(const std::string& name,
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW,
bool exception) const {
const std::string prefix = "Layer Pooling2D: " + name + ":";
auto error = poolingWindowLimit.GetErrorOrEmpty(windowH, windowW, "pooling window");
const bool throwOnError) const {
auto error = kPoolingWindowLimit.GetErrorOrEmpty(windowH, windowW, "pooling window");
const RangeLimit poolingStrideHLimit{ 1, windowH, "pooling stride height (must be up to pooling window height)" };
const RangeLimit poolingStrideWLimit{ 1, windowW, "pooling stride width (must be up to pooling window width)" };
error += poolingStrideHLimit.GetErrorOrEmpty(strideH);
error += poolingStrideWLimit.GetErrorOrEmpty(strideW);
if (exception)
ThrowIfNotEmpty(prefix, error);
return ValidationSuccesful(throwOnError, error, name, "Pooling2D");
}
return error.empty() ? true : false;
bool Validator_30::IsPaddingSupported() const {
return false;
}
const RangeLimit2D Validator_35::kInputHWLimit{{1, 65535, "input height"}, {1, 65535, "input width"}};
const RangeLimit Validator_35::kInputChannelsNumberLimit1B{1, 2048, "number of input channels"};
const RangeLimit Validator_35::kInputChannelsNumberLimit2B{1, 1024, "number of input channels"};
const RangeLimit Validator_35::kKernelNumberLimit{1, 8192, "number of kernels"};
const RangeLimit2D Validator_35::kKerneHWlLimit{{1, 255, "kernel height"}, {1, 256, "kernel width"}};
const RangeLimit2D Validator_35::kStrideHWLimit{{1, 255, "convolution stride height"},
{1, 256, "convolution stride width"}};
const RangeLimit2D Validator_35::kDilationLimit{{convDilationHeight, convDilationHeight, "dilation height"},
{convDilationWidth, convDilationWidth, "dilation width"}};
bool Validator_35::ValidateCnn2D(const std::string& name, const uint32_t inHeight, const uint32_t inWidth,
const uint32_t inChannels, const uint32_t kernelH, const uint32_t kernelW, const uint32_t kernelN,
const uint32_t strideH, const uint32_t strideW, const uint32_t dilationH, const uint32_t dilationW,
const OvGnaType inPrecision, const bool throwOnError) const {
auto error = kInputHWLimit.GetErrorOrEmpty(inHeight, inWidth);
error += kKernelNumberLimit.GetErrorOrEmpty(kernelN);
auto& inputChannelsNumberLimit = inPrecision == OvGnaTypeInt8 ? kInputChannelsNumberLimit1B : kInputChannelsNumberLimit2B;
error += inputChannelsNumberLimit.GetErrorOrEmpty(inChannels);
error += kKerneHWlLimit.GetErrorOrEmpty(kernelH, kernelW);
error += kStrideHWLimit.GetErrorOrEmpty(strideH, strideW);
error += kDilationLimit.GetErrorOrEmpty(dilationH, dilationW);
return ValidationSuccesful(throwOnError, error, name, "Convolution2D");
}
const RangeLimit2D Validator_35::kPoolingWindowHWLimit{{1, 255, "pooling window height"},
{1, 255, "pooling window width"}};
const RangeLimit2D Validator_35::kPoolingStrideHWLimit{{1, 255, "pooling stride height"},
{1, 255, "pooling stride width"}};
bool Validator_35::ValidatePooling2D(const std::string& name,
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW,
const bool throwOnError) const {
auto error = kPoolingWindowHWLimit.GetErrorOrEmpty(windowH, windowW);
error += kPoolingStrideHWLimit.GetErrorOrEmpty(strideH, strideW);
return ValidationSuccesful(throwOnError, error, name, "Pooling2D");
}
bool Validator_35::IsPaddingSupported() const {
return true;
}
std::unique_ptr<AbstractValidator> AbstractValidator::Create(const std::string& target) {
if (target == InferenceEngine::GNAConfigParams::GNA_TARGET_3_0) {
if (target == common::kGnaTarget3_0) {
return tools::make_unique<Validator_30>();
} else if (target == common::kGnaTarget3_5) {
return tools::make_unique<Validator_35>();
}
return nullptr;
}
@ -177,6 +236,18 @@ void AbstractValidator::ThrowIfNotEmpty(const std::string& prefix, const std::st
}
}
bool AbstractValidator::ValidationSuccesful(const bool throwOnError,
const std::string& error,
const std::string& operationName,
const std::string& type) {
if (throwOnError) {
const std::string prefix = "Layer " + type + ": " + operationName + ":";
ThrowIfNotEmpty(prefix, error);
}
return error.empty();
}
} // namespace Cnn2D
IE_SUPPRESS_DEPRECATED_START

65
src/plugins/intel_gna/backend/gna_limitations.hpp
View File

@ -109,6 +109,11 @@ struct RectLimitByChannelsAndPrecision {
class AbstractValidator {
protected:
static void ThrowIfNotEmpty(const std::string& prefix, const std::string& error);
static bool ValidationSuccesful(const bool throwOnError,
const std::string& error,
const std::string& operation,
const std::string& type);
public:
virtual ~AbstractValidator() = default;
virtual bool ValidateCnn2D(const std::string& name, const uint32_t inHeight, const uint32_t inWidth,
@ -120,30 +125,21 @@ public:
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW,
bool exception = true) const = 0;
virtual bool IsPaddingSupported() const = 0;
static std::unique_ptr<AbstractValidator> Create(const std::string&);
};
class Validator_30 : public AbstractValidator {
RangeLimit2D inputHWLimit{ { 16, 384, "input height"} , { 16, 240, "input width"} };
RangeMultipleLimit inputChannelsNumberLimit{ {8, 384, "number of input channels"}, 8 };
static const RangeLimit2D kInputHWLimit;
static const RangeMultipleLimit kInputChannelsNumberLimit;
RangeMultipleLimit kernelNumberLimit{ {8, 1024, "number of kernels"}, 8 };
RectLimitByChannelsAndPrecision kernelLimit {
{ { {96, {7, 7}},
{136, {7, 5}},
{168, {7, 4}},
{240, {7, 3}},
{384, {7, 2}} } },
{ { {48, {7, 7}},
{64, {7, 5}},
{80, {7, 4}},
{120, {7, 3}},
{384, {7, 1}} } },
};
RectLimitByChannelsAndPrecision& strideLimit = kernelLimit;
RangeLimit2D dilationLimit{ {convDilationHeight, convDilationHeight, "dilation height" },
{ convDilationWidth, convDilationWidth, "dilation width" } };
const VectorOrSquareLimit poolingWindowLimit{ 3, 1, 1 };
static const RangeMultipleLimit kKernelNumberLimit;
static const RectLimitByChannelsAndPrecision kKernelLimit;
static const RangeLimit2D kDilationLimit;
static const VectorOrSquareLimit kPoolingWindowLimit;
public:
Validator_30() = default;
@ -157,6 +153,37 @@ public:
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW,
bool exception = true) const override;
bool IsPaddingSupported() const override;
};
class Validator_35 : public AbstractValidator {
static const RangeLimit2D kInputHWLimit;
static const RangeLimit kInputChannelsNumberLimit1B;
static const RangeLimit kInputChannelsNumberLimit2B;
static const RangeLimit kKernelNumberLimit;
static const RangeLimit2D kKerneHWlLimit;
static const RangeLimit2D kStrideHWLimit;
static const RangeLimit2D kDilationLimit;
static const RangeLimit2D kPoolingWindowHWLimit;
static const RangeLimit2D kPoolingStrideHWLimit;
public:
Validator_35() = default;
bool ValidateCnn2D(const std::string& name, const uint32_t inHeight, const uint32_t inWidth,
const uint32_t inChannels, const uint32_t kH, const uint32_t kW, const uint32_t kN,
const uint32_t strideH, const uint32_t strideW, const uint32_t dilationH, const uint32_t dilationW,
OvGnaType inPrecision, bool exception = true) const override;
bool ValidatePooling2D(const std::string& name,
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW,
bool exception = true) const override;
bool IsPaddingSupported() const override;
};
} // namespace Cnn2D

5
src/plugins/intel_gna/common/gna_target.cpp Normal file
View File

@ -0,0 +1,5 @@
// Copyright (C) 2018-2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include "gna_target.hpp"

15
src/plugins/intel_gna/common/gna_target.hpp Normal file
View File

@ -0,0 +1,15 @@
// Copyright (C) 2018-2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#pragma once
namespace GNAPluginNS {
namespace common {
static constexpr const char* kGnaTargetUnspecified = "";
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";
static constexpr const char* kGnaTarget3_5 = "GNA_TARGET_3_5";
} // namespace common
} // namespace GNAPluginNS

157
src/plugins/intel_gna/gna2_model_export_helper.cpp
View File

@ -8,11 +8,14 @@
#include "gna2-model-export-api.h"
#include "gna2-model-suecreek-header.h"
#include "gna2-device-api.h"
#include "gna/gna_config.hpp"
#include "common/gna_target.hpp"
#include "gna2-tlv-writer.h"
#include <cstdint>
#include <fstream>
#include <numeric>
void * ExportSueLegacyUsingGnaApi2(
uint32_t modelId,
@ -55,6 +58,7 @@ void * ExportSueLegacyUsingGnaApi2(
#define Gna2TlvTypeOVInputScaleFactor GNA2_TLV_IMPL_CHAR_TO_TYPE("OVIS")
#define Gna2TlvTypeOVOutputScaleFactor GNA2_TLV_IMPL_CHAR_TO_TYPE("OVOS")
#define Gna2TlvTypeOVString GNA2_TLV_IMPL_CHAR_TO_TYPE("OVSS")
#define Gna2ExportTlv(...) 1
static_assert(std::numeric_limits<float>::is_iec559, "Float is not IEC 559 compatible");
@ -67,17 +71,94 @@ TlvFloatRecord GetFloatInTLV(Gna2TlvType type, float value) {
reinterpret_cast<Gna2TlvRecord*>(r.data())->length = sizeof(float);
*reinterpret_cast<float*>(r.data() + sizeof(Gna2TlvRecord)) = value;
return r;
} // namespace
std::vector<char> GetStringAsTlv(Gna2TlvType type, const std::string& s) {
std::vector<char> record(sizeof(Gna2TlvRecord));
reinterpret_cast<Gna2TlvRecord*>(record.data())->type = type;
std::vector<char> vs(s.begin(), s.end());
vs.resize(vs.size() + (4 - vs.size() % 4) % 4, 0);
reinterpret_cast<Gna2TlvRecord*>(record.data())->length = vs.size();
record.insert(record.end(), vs.begin(), vs.end());
return record;
}
} // namespace
Gna2DeviceVersion getEmbeddedTargetFromCompileTarget(const std::string compileTarget) {
static const std::map<std::string, Gna2DeviceVersion> targetMap = {
{GNAPluginNS::common::kGnaTarget3_1, Gna2DeviceVersionEmbedded3_1},
{GNAPluginNS::common::kGnaTarget3_5, Gna2DeviceVersionEmbedded3_5},
};
auto found = targetMap.find(compileTarget);
if (found == targetMap.end()) {
return Gna2DeviceVersionEmbedded1_0;
}
return found->second;
}
namespace {
Gna2DeviceVersion getTlvTargetFromCompileTarget(const std::string compileTarget) {
const auto target = getEmbeddedTargetFromCompileTarget(compileTarget);
static const std::set<Gna2DeviceVersion> supportedTargets = {
Gna2DeviceVersionEmbedded3_1,
Gna2DeviceVersionEmbedded3_5,
};
const auto found = supportedTargets.count(target) > 0;
if (!found) {
THROW_GNA_EXCEPTION << "Unsupported compile target for TLV export: " << compileTarget << "\n";
}
return target;
}
std::string WriteAllEndpoints(std::ostream& outStream,
const std::vector<GnaEndpoint>& allEndpoints,
const Gna2TlvType sfTlvType,
const GnaAllocation* allocation) {
const std::string endPointType = sfTlvType == Gna2TlvTypeOVInputScaleFactor ? "Input" : "Output";
if (allEndpoints.size() >= 1) {
const auto scaleFactorTlv = GetFloatInTLV(sfTlvType, allEndpoints[0].scaleFactor);
outStream.write(scaleFactorTlv.data(), scaleFactorTlv.size());
}
if (allEndpoints.size() != 1) {
gnawarn() << "Number of endpoints: " << allEndpoints.size() << " for " << endPointType << "\n";
}
std::stringstream stream;
stream << "Endpoints for " << endPointType << ":\n";
for (const auto& endpoint : allEndpoints) {
stream << "name=[" << endpoint.name << "]\n";
stream << "scaleFactor=[" << endpoint.scaleFactor << "]\n";
stream << "byteSize=[" << endpoint.byteSize << "]\n";
stream << "numberOfBytesPerElement=[" << endpoint.numberOfBytesPerElement << "]\n";
if (allocation == nullptr) {
stream << "allocation=[nullptr]\n";
}
if (endpoint.gnaPointer == nullptr) {
stream << "gnaPointer=[nullptr]\n";
}
if (allocation != nullptr && endpoint.gnaPointer != nullptr) {
const auto gnaOffset = allocation->getOffset(endpoint.gnaPointer);
if (!gnaOffset.first) {
stream << "offset=[invalid]\n";
}
stream << "offset=[" << gnaOffset.second << "]\n";
}
}
return stream.str();
}
} // namespace
void ExportTlvModel(uint32_t modelId,
uint32_t deviceIndex,
std::ostream& outStream,
Gna2DeviceVersion deviceVersionToExport,
uint32_t input_size,
uint32_t output_size,
float inputSF,
float outputSF) {
std::string compileTarget,
const std::vector<GnaEndpoint>& allInputs,
const std::vector<GnaEndpoint>& allOutputs,
const GnaAllocations& allAllocations) {
const auto deviceVersionToExport = getTlvTargetFromCompileTarget(compileTarget);
uint32_t exportConfig;
auto status = Gna2ModelExportConfigCreate(gnaUserAllocatorAlignedPage, &exportConfig);
@ -150,6 +231,7 @@ void ExportTlvModel(uint32_t modelId,
const auto gnaLibraryVersion = GNADeviceHelper::GetGnaLibraryVersion();
uint32_t outTlvSize = 0;
auto tlv_status = Gna2ExportTlv(
deviceVersionToExport,
gnaUserAllocator,
@ -162,19 +244,20 @@ void ExportTlvModel(uint32_t modelId,
(const char*)bufferStateRWData,
sizeOfStateRWData,
sizeOfScratchRWData,
input_size,
output_size,
GnaEndpoint::GetTotalByteSize(allInputs),
GnaEndpoint::GetTotalByteSize(allOutputs),
gnaLibraryVersion.c_str(),
nullptr,
0);
if (Gna2TlvStatusSuccess == tlv_status) {
outStream.write(outTlv, outTlvSize);
auto tlvInSF = GetFloatInTLV(Gna2TlvTypeOVInputScaleFactor, inputSF);
auto tlvOutSF = GetFloatInTLV(Gna2TlvTypeOVOutputScaleFactor, outputSF);
outStream.write(tlvInSF.data(), tlvInSF.size());
outStream.write(tlvOutSF.data(), tlvOutSF.size());
auto metadata = WriteAllEndpoints(outStream, allInputs, Gna2TlvTypeOVInputScaleFactor, allAllocations.Get(Gna2MemoryTagInput));
metadata += WriteAllEndpoints(outStream, allOutputs, Gna2TlvTypeOVOutputScaleFactor, allAllocations.Get(Gna2MemoryTagOutput));
auto metadataTlv = GetStringAsTlv(Gna2TlvTypeOVString, metadata);
outStream.write(metadataTlv.data(), metadataTlv.size());
}
gnaUserFree(outTlv);
gnaUserFree(bufferLayerDescriptors);
@ -185,57 +268,9 @@ void ExportTlvModel(uint32_t modelId,
gnaUserFree(bufferInputRWData);
gnaUserFree(bufferOutputRWData);
GNADeviceHelper::checkGna2Status((Gna2Status)status, "ExportTlvModel");
status = Gna2ModelExportConfigRelease(exportConfig);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigRelease");
if (Gna2TlvStatusSuccess != tlv_status) {
THROW_GNA_EXCEPTION << "Not succesfull status returned: " << tlv_status << ", from Gna2ExportTlv() function\n";
}
void ExportLdForDeviceVersion(
uint32_t modelId,
std::ostream & outStream,
const Gna2DeviceVersion deviceVersionToExport) {
uint32_t exportConfig;
auto status = Gna2ModelExportConfigCreate(gnaUserAllocatorAlignedPage, &exportConfig);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigCreate");
status = Gna2ModelExportConfigSetSource(exportConfig, 0, modelId);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigSetSource");
status = Gna2ModelExportConfigSetTarget(exportConfig, deviceVersionToExport);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigSetTarget");
void * ldDump = nullptr;
uint32_t ldDumpSize;
status = Gna2ModelExport(exportConfig,
Gna2ModelExportComponentLayerDescriptors,
&ldDump, &ldDumpSize);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExport(LayerDescriptors)");
outStream.write(static_cast<char*>(ldDump), ldDumpSize);
status = Gna2ModelExportConfigRelease(exportConfig);
GNADeviceHelper::checkGna2Status(status, "Gna2ModelExportConfigRelease");
gnaUserFree(ldDump);
}
void ExportGnaDescriptorPartiallyFilled(uint32_t number_of_layers, std::ostream& outStream) {
const uint32_t scratchPadSize = 0x2000;
const auto constScratchFill = static_cast<char>(-1);
const uint32_t gnaDescSize = 32;
char gd[gnaDescSize] = {};
char gd2[gnaDescSize] = {};
gd[0] = 1;
*reinterpret_cast<uint32_t *>(gd + 4) = number_of_layers;
*reinterpret_cast<uint32_t *>(gd + 8) = 0xffffffff;
*reinterpret_cast<uint32_t *>(gd + 0xC) = 2 * sizeof(gd) + scratchPadSize;
outStream.write(gd, sizeof(gd));
outStream.write(gd2, sizeof(gd2));
// TODO: GNA2: Scratchpad
auto previousCharacter = outStream.fill(constScratchFill);
auto previousWidth = outStream.width(scratchPadSize);
outStream << constScratchFill;
outStream.width(previousWidth);
outStream.fill(previousCharacter);
}

72
src/plugins/intel_gna/gna2_model_export_helper.hpp
View File

@ -4,29 +4,79 @@
#pragma once
#include "gna2-model-export-api.h"
#include "gna2-common-api.h"
#include "gna2-model-suecreek-header.h"
#include "gna_device_allocation.hpp"
#include <algorithm>
#include <cstdint>
#include <list>
#include <map>
#include <numeric>
#include <string>
#include <utility>
#include <vector>
struct GnaEndpoint {
std::string name;
uint32_t byteSize = 0;
uint32_t offset = 0;
uint32_t numberOfBytesPerElement = 0;
float scaleFactor = 0;
void* gnaPointer = nullptr;
template <class T>
static uint32_t GetTotalByteSize(const T& container);
template <class T>
static GnaEndpoint CreateFromDescriptor(const T& descriptor);
template <class T>
static std::vector<GnaEndpoint> CreateFromDescriptorContainer(const T& container);
};
template <class T>
uint32_t GnaEndpoint::GetTotalByteSize(const T& container) {
return std::accumulate(container.begin(), container.end(), 0, [](uint32_t cur, const GnaEndpoint& next) {
return cur + next.byteSize;
});
}
template <class T>
GnaEndpoint GnaEndpoint::CreateFromDescriptor(const T& descriptor) {
GnaEndpoint e;
e.scaleFactor = descriptor.scale_factor;
e.byteSize = descriptor.get_required_size();
e.name = descriptor.name;
e.numberOfBytesPerElement = static_cast<uint32_t>(descriptor.tensor_precision.size());
if (!descriptor.ptrs.empty()) {
e.gnaPointer = descriptor.ptrs.front();
}
return e;
}
template <class T>
std::vector<GnaEndpoint> GnaEndpoint::CreateFromDescriptorContainer(const T& container) {
std::vector<GnaEndpoint> result;
for (const auto& e : container) {
result.push_back(CreateFromDescriptor(e));
}
return result;
}
void * ExportSueLegacyUsingGnaApi2(
uint32_t modelId,
uint32_t deviceIndex,
Gna2ModelSueCreekHeader* modelHeader);
void ExportLdForDeviceVersion(
uint32_t modelId,
std::ostream & outStream,
Gna2DeviceVersion deviceVersionToExport);
Gna2DeviceVersion getEmbeddedTargetFromCompileTarget(const std::string compileTarget);
void ExportTlvModel(uint32_t modelId,
uint32_t deviceIndex,
std::ostream& outStream,
Gna2DeviceVersion deviceVersionToExport,
uint32_t input_size,
uint32_t output_size,
float inputSF,
float outputSF);
void ExportGnaDescriptorPartiallyFilled(uint32_t numberOfLayers, std::ostream & outStream);
std::string compileTarget,
const std::vector<GnaEndpoint>& inputs,
const std::vector<GnaEndpoint>& outputs,
const GnaAllocations& allAllocation);

88
src/plugins/intel_gna/gna_device.cpp
View File

@ -23,6 +23,7 @@
#include "backend/am_intel_dnn.hpp"
#include "backend/gna_limitations.hpp"
#include "common/gna_target.hpp"
#include "gna/gna_config.hpp"
#include "gna_plugin_log.hpp"
#include "layers/gna_convolution_layer.hpp"
@ -31,25 +32,22 @@
std::mutex GNADeviceHelper::acrossPluginsSync{};
GNADeviceHelper::GNADeviceHelper(std::string executionTargetIn,
std::string compileTargetIn,
bool swExactModeIn,
bool isPerformanceMeasuring,
bool deviceEmbedded,
int deviceVersionParsed)
bool deviceEmbedded)
: swExactMode(swExactModeIn),
executionTarget(executionTargetIn),
compileTarget(compileTargetIn),
isPerformanceMeasuring(isPerformanceMeasuring),
nGnaDeviceIndex{selectGnaDevice()},
useDeviceEmbeddedExport(deviceEmbedded),
exportGeneration(static_cast<Gna2DeviceVersion>(deviceVersionParsed)) {
useDeviceEmbeddedExport(deviceEmbedded) {
open();
initGnaPerfCounters();
// check GNA Library version
const auto gnaLibVersion = GetGnaLibraryVersion();
GetGnaLibraryVersion();
maxLayersCount_ = retrieveMaxLayersCount();
}
@ -241,30 +239,16 @@ bool GNADeviceHelper::enforceLegacyCnnNeeded() const {
}
Gna2DeviceVersion GNADeviceHelper::parseTarget(const std::string& target) {
const std::map<std::string, Gna2DeviceVersion> targetMap {
{InferenceEngine::GNAConfigParams::GNA_TARGET_2_0, Gna2DeviceVersion2_0},
{InferenceEngine::GNAConfigParams::GNA_TARGET_3_0, Gna2DeviceVersion3_0},
{"", Gna2DeviceVersionSoftwareEmulation},
static const std::map<std::string, Gna2DeviceVersion> targetMap {
{GNAPluginNS::common::kGnaTarget2_0, Gna2DeviceVersion2_0},
{GNAPluginNS::common::kGnaTarget3_0, Gna2DeviceVersion3_0},
{GNAPluginNS::common::kGnaTargetUnspecified, Gna2DeviceVersionSoftwareEmulation},
};
const auto f = targetMap.find(target);
if (f != targetMap.end()) {
return f->second;
}
THROW_GNA_EXCEPTION << "Unsupported " << "GNAConfigParams::GNA_TARGET = \"" << target << "\"\n";
}
Gna2DeviceVersion GNADeviceHelper::parseDeclaredTarget(std::string target, const bool execTarget) const {
auto parsed = Gna2DeviceVersion2_0;
auto throwUnsupportedGnaTarget = [&](std::string extraSuffix) {
auto key = execTarget ? InferenceEngine::GNAConfigParams::KEY_GNA_EXEC_TARGET : InferenceEngine::GNAConfigParams::KEY_GNA_COMPILE_TARGET;
THROW_GNA_EXCEPTION << "Unsupported " << key << " = \"" << target << "\"" << extraSuffix;
};
if (target == InferenceEngine::GNAConfigParams::GNA_TARGET_3_0) {
parsed = Gna2DeviceVersion3_0;
} else if (target != InferenceEngine::GNAConfigParams::GNA_TARGET_2_0) {
throwUnsupportedGnaTarget("");
}
return parsed;
THROW_GNA_EXCEPTION << "Unsupported " << "GNA target: \"" << target << "\"\n";
}
Gna2DeviceVersion GNADeviceHelper::getDefaultTarget() const {
@ -275,10 +259,10 @@ Gna2DeviceVersion GNADeviceHelper::getDefaultTarget() const {
Gna2DeviceVersion GNADeviceHelper::getTargetDevice(const bool execTarget) const {
const auto declared = execTarget ? executionTarget : compileTarget;
if (declared.empty()) {
if (declared == GNAPluginNS::common::kGnaTargetUnspecified) {
return execTarget ? getDefaultTarget() : getTargetDevice(true);
}
return parseDeclaredTarget(declared, execTarget);
return parseTarget(declared);
}
uint32_t GNADeviceHelper::createRequestConfig(const uint32_t modelID) const {
@ -514,32 +498,18 @@ GNADeviceHelper::DumpResult GNADeviceHelper::dumpXnn(const uint32_t modelId) {
return r;
}
void GNADeviceHelper::dumpXnnForDeviceVersion(
const uint32_t modelId,
void GNADeviceHelper::dumpTLVForDeviceVersion(const uint32_t modelId,
std::ostream& outStream,
const Gna2DeviceVersion targetDeviceVersion) {
Gna2ModelSueCreekHeader sueHeader;
auto ptr = ExportSueLegacyUsingGnaApi2(modelId, nGnaDeviceIndex, &sueHeader);
gnaUserFree(ptr);
ExportGnaDescriptorPartiallyFilled(sueHeader.NumberOfLayers, outStream);
ExportLdForDeviceVersion(modelId, outStream, targetDeviceVersion);
if (dumpXNNROPtr == nullptr) {
THROW_GNA_EXCEPTION << "Bad RO pointer (nullptr)";
}
outStream.write(static_cast<const char*>(dumpXNNROPtr), dumpXNNROSize);
// TODO: GNA2: remove
outStream.write("Gna2ModelSueCreekHeader", 24);
outStream.write(reinterpret_cast<const char*>(&sueHeader), sizeof(sueHeader));
}
void GNADeviceHelper::dumpTLVForDeviceVersion(const uint32_t modelId, std::ostream& outStream,
uint32_t input_size, uint32_t output_size,
float inSF, float outSF) {
ExportTlvModel(modelId, nGnaDeviceIndex, outStream, exportGeneration, input_size, output_size, inSF, outSF);
const std::vector<GnaEndpoint>& inputsContainer,
const std::vector<GnaEndpoint>& outputsContainer) {
const auto compileTarget = GetCompileTarget();
ExportTlvModel(modelId,
nGnaDeviceIndex,
outStream,
compileTarget,
inputsContainer,
outputsContainer,
allAllocations);
}
void GNADeviceHelper::createVirtualDevice(Gna2DeviceVersion devVersion) {
@ -557,6 +527,9 @@ void GNADeviceHelper::open() {
updateGnaDeviceVersion();
const auto gnaExecTarget = parseTarget(executionTarget);
if (useDeviceEmbeddedExport) {
const auto gnaCompileTarget = GetCompileTarget();
const auto exportGeneration = getEmbeddedTargetFromCompileTarget(gnaCompileTarget);
createVirtualDevice(exportGeneration);
updateGnaDeviceVersion();
} else if (!executionTarget.empty() && gnaExecTarget != detectedGnaDevVersion) {
@ -612,17 +585,10 @@ void GNADeviceHelper::getGnaPerfCounters(std::map<std::string, InferenceEngine::
retPerfCounters["1.2 Stall scoring time in HW"] = info;
}
std::string GNADeviceHelper::getEffectiveGnaCompileTarget() const {
if (getTargetDevice(false) == Gna2DeviceVersion3_0) {
return InferenceEngine::GNAConfigParams::GNA_TARGET_3_0;
}
return InferenceEngine::GNAConfigParams::GNA_TARGET_2_0;
}
std::string GNADeviceHelper::GetCompileTarget() const {
static const std::map<Gna2DeviceVersion, std::string> targetMap = {
{Gna2DeviceVersion2_0, InferenceEngine::GNAConfigParams::GNA_TARGET_2_0},
{Gna2DeviceVersion3_0, InferenceEngine::GNAConfigParams::GNA_TARGET_3_0},
{Gna2DeviceVersion2_0, GNAPluginNS::common::kGnaTarget2_0},
{Gna2DeviceVersion3_0, GNAPluginNS::common::kGnaTarget3_0},
};
const auto target = getTargetDevice(false);
auto found = targetMap.find(target);

15
src/plugins/intel_gna/gna_device.hpp
View File

@ -16,6 +16,7 @@
#include <ie_common.h>
#include "gna2_model_export_helper.hpp"
#include "memory/gna_mem_requests.hpp"
#include "gna2-common-api.h"
@ -47,7 +48,6 @@ class GNADeviceHelper : public GNAPluginNS::GNADevice {
std::string executionTarget;
std::string compileTarget;
bool useDeviceEmbeddedExport = false;
Gna2DeviceVersion exportGeneration = Gna2DeviceVersionEmbedded1_0;
uint32_t maxLayersCount_ = 0;
static const uint32_t TotalGna2InstrumentationPoints = 2;
@ -74,8 +74,7 @@ public:
std::string compileTargetIn = "",
bool swExactModeIn = false,
bool isPerformanceMeasuring = false,
bool deviceEmbedded = false,
int deviceVersionParsed = 0);
bool deviceEmbedded = false);
GNADeviceHelper(const GNADeviceHelper&) = delete;
GNADeviceHelper& operator= (const GNADeviceHelper&) = delete;
@ -121,12 +120,10 @@ public:
DumpResult dumpXnn(const uint32_t modelId);
void dumpXnnForDeviceVersion(const uint32_t modelId,
void dumpTLVForDeviceVersion(const uint32_t modelId,
std::ostream& outStream,
Gna2DeviceVersion targetDeviceVersion);
void dumpTLVForDeviceVersion(const uint32_t modelId, std::ostream& outStream,
uint32_t input_size, uint32_t output_size,
float inSF, float outSF);
const std::vector<GnaEndpoint>& inputsContainer,
const std::vector<GnaEndpoint>& outputsContainer);
void free(void * ptr);
@ -134,7 +131,6 @@ public:
void getGnaPerfCounters(std::map<std::string,
InferenceEngine::InferenceEngineProfileInfo>& retPerfCounters);
static std::string GetGnaLibraryVersion();
std::string getEffectiveGnaCompileTarget() const;
std::string GetCompileTarget() const;
const GnaAllocations& getAllAllocations() const {
@ -181,7 +177,6 @@ private:
static void enforceLegacyCnns(Gna2Model& gnaModel);
static void enforceLegacyCnnsWhenNeeded(Gna2Model& gnaModel);
static Gna2DeviceVersion parseTarget(const std::string& target);
Gna2DeviceVersion parseDeclaredTarget(std::string target, const bool execTarget) const;
Gna2DeviceVersion getDefaultTarget() const;
Gna2DeviceVersion getTargetDevice(bool execTarget) const;

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

@ -165,4 +165,13 @@ public:
newAllocation.sizeGranted = sizeGranted;
allocations.push_back(newAllocation);
}
const GnaAllocation* Get(const Gna2MemoryTag tag) const {
for (auto&& a : allocations) {
if (a.isTag(tag)) {
return &a;
}
}
return nullptr;
}
};

49
src/plugins/intel_gna/gna_graph_compiler.cpp
View File

@ -193,24 +193,36 @@ void GNAGraphCompiler::fillSplitConnections(InferenceEngine::CNNLayerPtr layer)
split_connection.emplace(id, layerInfoItem);
}
void GNAPluginNS::GNAGraphCompiler::SetValidatorTarget(std::string target) {
if (InferenceEngine::GNAConfigParams::GNA_TARGET_3_0 == target) {
cnn2dValidator.reset(new GNALimitations::Cnn2D::Validator_30());
}
void GNAPluginNS::GNAGraphCompiler::SetValidatorTarget(const std::string& target) {
auto temp = GNALimitations::Cnn2D::AbstractValidator::Create(target);
cnn2dValidator.reset(temp.release());
}
void GNAPluginNS::GNAGraphCompiler::ValidateCnn2D(std::string name, const uint32_t inHeight, const uint32_t inWidth, const uint32_t inChannels,
const uint32_t kH, const uint32_t kW, const uint32_t kN, const uint32_t strideH, const uint32_t strideW, OvGnaType inPrecision,
const uint32_t dilH, const uint32_t dilW) const {
void GNAPluginNS::GNAGraphCompiler::ValidateCnn2D(const std::string& name,
const uint32_t inHeight,
const uint32_t inWidth,
const uint32_t inChannels,
const uint32_t kH,
const uint32_t kW,
const uint32_t kN,
const uint32_t strideH,
const uint32_t strideW,
const uint32_t dilH,
const uint32_t dilW,
OvGnaType inPrecision) const {
if (cnn2dValidator) {
cnn2dValidator->ValidateCnn2D(name, inHeight, inWidth, inChannels, kH, kW, kN, strideH, strideW, dilH, dilW, inPrecision);
cnn2dValidator
->ValidateCnn2D(name, inHeight, inWidth, inChannels, kH, kW, kN, strideH, strideW, dilH, dilW, inPrecision);
} else {
THROW_GNA_EXCEPTION << "No Cnn2D validator found for layer " << name;
}
}
void GNAPluginNS::GNAGraphCompiler::ValidatePooling2D(std::string name, const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW) const {
void GNAPluginNS::GNAGraphCompiler::ValidatePooling2D(const std::string& name,
const uint32_t windowH,
const uint32_t windowW,
const uint32_t strideH,
const uint32_t strideW) const {
if (cnn2dValidator) {
cnn2dValidator->ValidatePooling2D(name, windowH, windowW, strideH, strideW);
} else {
@ -218,6 +230,14 @@ void GNAPluginNS::GNAGraphCompiler::ValidatePooling2D(std::string name, const ui
}
}
bool GNAPluginNS::GNAGraphCompiler::IsCnn2DInputPaddingSupported(const std::string& name) const {
if (cnn2dValidator) {
return cnn2dValidator->IsPaddingSupported();
} else {
THROW_GNA_EXCEPTION << "No Cnn2D input padding validator found for layer " << name;
}
}
void GNAGraphCompiler::DiagonalPrimitive(InferenceEngine::CNNLayerPtr layer) {
AffinePrimitive(layer, true);
}
@ -587,8 +607,9 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP
auto effectiveInputWidth = in_width;
auto effectiveInputHeight = in_height;
if (convolution._padding_x != 0 || convolution._padding_y != 0 ||
convolution._pads_end.at(X_AXIS) != 0 || convolution._pads_end.at(Y_AXIS) != 0) {
if (!IsCnn2DInputPaddingSupported(convolution.name) &&
(convolution._padding_x != 0 || convolution._padding_y != 0 ||
convolution._pads_end.at(X_AXIS) != 0 || convolution._pads_end.at(Y_AXIS) != 0)) {
THROW_GNA_LAYER_EXCEPTION(layer) << "Convolution's input padding is not supported";
}
@ -598,6 +619,8 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP
if (convolution._padding_y != convolution._pads_end.at(Y_AXIS)) {
THROW_GNA_LAYER_EXCEPTION(layer) << "Convolution's input padding is not symetric along Y axis";
}
convolution._padding_x = convolution._pads_end.at(X_AXIS);
convolution._padding_y = convolution._pads_end.at(Y_AXIS);
if (convolution._kernel_x > effectiveInputWidth ||
convolution._kernel_y > effectiveInputHeight) {
@ -636,7 +659,7 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP
ValidateCnn2D(layer->name,
in_height, in_width, in_channels,
convolution._kernel_y, convolution._kernel_x, filter_n, convolution._stride_y, convolution._stride_x,
inputPrec, convolution._dilation_y, convolution._dilation_x);
convolution._dilation_y, convolution._dilation_x, inputPrec);
float weight_scale_factor = getScaleFactor(layer, QuantizedDataType::weights);
float output_scale_factor = getScaleFactor(layer, QuantizedDataType::output);

14
src/plugins/intel_gna/gna_graph_compiler.hpp
View File

@ -68,17 +68,19 @@ public:
void fillConcatConnections(InferenceEngine::CNNLayerPtr layer);
void fillSplitConnections(InferenceEngine::CNNLayerPtr layer);
void ValidateCnn2D(std::string name, const uint32_t inHeight, const uint32_t inWidth,
void ValidateCnn2D(const std::string& name, const uint32_t inHeight, const uint32_t inWidth,
const uint32_t inChannels, const uint32_t kH, const uint32_t kW, const uint32_t kN,
const uint32_t strideH, const uint32_t strideW, OvGnaType inPrecision,
const uint32_t dilH, const uint32_t dilW) const;
const uint32_t strideH, const uint32_t strideW,
const uint32_t dilH, const uint32_t dilW,
OvGnaType inPrecision) const;
void ValidatePooling2D(std::string name,
void ValidatePooling2D(const std::string& name,
const uint32_t windowH, const uint32_t windowW,
const uint32_t strideH, const uint32_t strideW) const;
void SetValidatorTarget(std::string target);
bool IsCnn2DInputPaddingSupported(const std::string& name) const;
void SetValidatorTarget(const std::string& target);
/**
* Connects either memory output, or generic output to a layer

29
src/plugins/intel_gna/gna_plugin.cpp
View File

@ -24,6 +24,7 @@
#include <gna/gna_config.hpp>
#include "gna_plugin_config.hpp"
#include "gna_plugin.hpp"
#include "common/gna_target.hpp"
#include "optimizer/gna_pass_manager.hpp"
#include "layers/gna_layer_type.hpp"
#include "preprocessing.hpp"
@ -38,6 +39,7 @@
#include "gna_graph_patterns.hpp"
#include "gna_tensor_tools.hpp"
#include "gna_itt.hpp"
#include "gna2_model_export_helper.hpp"
#include "gna2_model_helper.hpp"
#include "request/model_wrapper_factory.hpp"
#include "request/worker_pool_impl.hpp"
@ -339,7 +341,7 @@ std::string GNAPluginNS::GNAPlugin::GetCompileTarget() const {
} else if (!config.gnaCompileTarget.empty()) {
return config.gnaCompileTarget;
}
return InferenceEngine::GNAConfigParams::GNA_TARGET_3_0;
return common::kGnaTarget3_0;
}
GNAPlugin::GNAPlugin(const std::map<std::string, std::string>& configMap) {
@ -371,8 +373,7 @@ void GNAPlugin::InitGNADevice() {
config.gnaCompileTarget,
config.swExactMode,
gnaFlags->performance_counting,
!config.dumpXNNPath.empty(),
GetDeviceVersionFromString(config.gnaCompileTarget));
!config.dumpXNNPath.empty());
size_t page_size_bytes = 4096;
gnamem = std::make_shared<gna_memory_device>(memory::GNAAllocator(gnadevice), page_size_bytes);
if (gnaFlags->log_level == ov::log::Level::DEBUG) {
@ -1206,13 +1207,6 @@ std::shared_ptr<request::ModelWrapper> GNAPluginNS::GNAPlugin::createModelWrappe
return request::ModelWrapperFactory::createWithNumberOfEmptyOperations(numberOfOperations);
}
int GNAPlugin::GetDeviceVersionFromString(const std::string deviceString) {
if (deviceString.empty() || deviceString == InferenceEngine::GNAConfigParams::GNA_TARGET_2_0) {
return static_cast<int>(Gna2DeviceVersionEmbedded1_0);
}
return static_cast<int>(Gna2DeviceVersionEmbedded3_5);
}
void GNAPlugin::DumpXNNToFile() const {
// TODO: output precision as well as pointer might be incorrect, LSTM for sure
// gna looks automatically set layer 0 as output and adjust it's pointer / precision/ size respectively
@ -1236,7 +1230,7 @@ void GNAPlugin::DumpXNNToFile() const {
const auto& inputsDesc = inputs_ptr_->Get();
const auto& outputsDesc = outputs_.Get();
if (InferenceEngine::GNAConfigParams::GNA_TARGET_2_0 == gnadevice->getEffectiveGnaCompileTarget()) {
if (common::kGnaTarget2_0 == gnadevice->GetCompileTarget()) {
auto dump = gnadevice->dumpXnn(modelId);
dump.header.RwRegionSize = gnamem->getRegionBytes(REGION_SCRATCH);
dump.header.InputScalingFactor = inputsDesc.begin()->scale_factor;
@ -1244,16 +1238,9 @@ void GNAPlugin::DumpXNNToFile() const {
dumpStream.write(reinterpret_cast<char*>(&dump.header), sizeof(Gna2ModelSueCreekHeader));
dumpStream.write(reinterpret_cast<char*>(dump.model.get()), dump.header.ModelSize);
} else {
uint32_t input_size = 0;
uint32_t output_size = 0;
for (auto i : inputsDesc)
input_size += i.get_allocated_size();
for (auto o : outputsDesc)
output_size += o.get_required_size();
auto inSF = inputsDesc.begin()->scale_factor;
auto outSF = outputsDesc.front().scale_factor;
gnadevice->dumpTLVForDeviceVersion(modelId, dumpStream,
input_size, output_size, inSF, outSF);
const auto inputsForTlv = GnaEndpoint::CreateFromDescriptorContainer(inputsDesc);
const auto outputsForTlv = GnaEndpoint::CreateFromDescriptorContainer(outputsDesc);
gnadevice->dumpTLVForDeviceVersion(modelId, dumpStream, inputsForTlv, outputsForTlv);
}
gnadevice->releaseModel(modelId);
}

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

@ -52,8 +52,6 @@ protected:
uint32_t dnn_dump_write_index = 0;
intel_dnn_number_type_t output_type = kDnnInt;
static int GetDeviceVersionFromString(const std::string deviceString);
std::shared_ptr<GNADeviceHelper> gnadevice;
std::shared_ptr<request::WorkerPool> requestWorkerPool_;

19
src/plugins/intel_gna/gna_plugin_config.cpp
View File

@ -8,6 +8,7 @@
#include <gna/gna_config.hpp>
#include "gna_plugin.hpp"
#include "gna_plugin_config.hpp"
#include "common/gna_target.hpp"
#include "cpp_interfaces/interface/ie_internal_plugin_config.hpp"
#include "ie_common.h"
#include <caseless.hpp>
@ -39,9 +40,9 @@ static const caseless_unordered_map<std::string, std::pair<Gna2AccelerationMode,
OPENVINO_SUPPRESS_DEPRECATED_END
static const std::set<std::string> supportedTargets = {
GNAConfigParams::GNA_TARGET_2_0,
GNAConfigParams::GNA_TARGET_3_0,
""
common::kGnaTarget2_0,
common::kGnaTarget3_0,
common::kGnaTargetUnspecified
};
void Config::UpdateFromMap(const std::map<std::string, std::string>& config) {
@ -153,9 +154,9 @@ OPENVINO_SUPPRESS_DEPRECATED_END
auto target = ov::util::from_string(value, ov::intel_gna::execution_target);
std::string target_str = "";
if (ov::intel_gna::HWGeneration::GNA_2_0 == target) {
target_str = GNAConfigParams::GNA_TARGET_2_0;
target_str = common::kGnaTarget2_0;
} else if (ov::intel_gna::HWGeneration::GNA_3_0 == target) {
target_str = GNAConfigParams::GNA_TARGET_3_0;
target_str = common::kGnaTarget3_0;
}
set_target(target_str);
} else if (key == GNA_CONFIG_KEY(EXEC_TARGET)) {
@ -373,12 +374,12 @@ Parameter Config::GetParameter(const std::string& name) const {
} else if (name == ov::intel_gna::pwl_design_algorithm) {
return gnaFlags.pwl_design_algorithm;
} else if (name == ov::intel_gna::execution_target) {
return ((gnaExecTarget == GNAConfigParams::GNA_TARGET_2_0) ? ov::intel_gna::HWGeneration::GNA_2_0 :
(gnaExecTarget == GNAConfigParams::GNA_TARGET_3_0) ? ov::intel_gna::HWGeneration::GNA_3_0 :
return ((gnaExecTarget == common::kGnaTarget2_0) ? ov::intel_gna::HWGeneration::GNA_2_0 :
(gnaExecTarget == common::kGnaTarget3_0) ? ov::intel_gna::HWGeneration::GNA_3_0 :
ov::intel_gna::HWGeneration::UNDEFINED);
} else if (name == ov::intel_gna::compile_target) {
return ((gnaCompileTarget == GNAConfigParams::GNA_TARGET_2_0) ? ov::intel_gna::HWGeneration::GNA_2_0 :
(gnaCompileTarget == GNAConfigParams::GNA_TARGET_3_0) ? ov::intel_gna::HWGeneration::GNA_3_0 :
return ((gnaCompileTarget == common::kGnaTarget2_0) ? ov::intel_gna::HWGeneration::GNA_2_0 :
(gnaCompileTarget == common::kGnaTarget3_0) ? ov::intel_gna::HWGeneration::GNA_3_0 :
ov::intel_gna::HWGeneration::UNDEFINED);
} else if (name == ov::hint::performance_mode) {
return performance_mode;

361
src/tests/unit/gna/backend/gna_limitations_test.cpp Normal file
View File

@ -0,0 +1,361 @@
// Copyright (C) 2018-2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include "backend/gna_limitations.hpp"
#include "common/gna_target.hpp"
#include <gtest/gtest.h>
#include <utility>
using namespace GNAPluginNS::GNALimitations;
using GNAPluginNS::common::kGnaTarget3_0;
using GNAPluginNS::common::kGnaTarget3_5;
struct GNACnn2DValidatorTestParam {
std::string target;
std::string whatInvalid;
std::vector<uint32_t> invalid;
};
const std::vector<uint32_t> kInvalidH_30 = {0, 400};
const std::vector<uint32_t> kInvalidH_35 = {0, 65536};
const std::vector<uint32_t> kInvalidW_30 = {0, 400};
const std::vector<uint32_t> kInvalidW_35 = {0, 65536};
const std::vector<uint32_t> kInvalidC_30 = {0, 1, 400};
const std::vector<uint32_t> kInvalidC_35 = {0, 2049};
const std::vector<uint32_t> kInvalidkH_30 = {0, 8, 400};
const std::vector<uint32_t> kInvalidkH_35 = {0, 257, 2049};
const std::vector<uint32_t> kInvalidkW_30 = {0, 8, 400};
const std::vector<uint32_t> kInvalidkW_35 = {0, 257, 2049};
const std::vector<uint32_t> kInvalidkN_30 = {0, 1, 400};
const std::vector<uint32_t> kInvalidkN_35 = {0, 2049};
const std::vector<uint32_t> kInvalidsH_30 = {0, 400};
const std::vector<uint32_t> kInvalidsH_35 = {0, 2049};
const std::vector<uint32_t> kInvalidsW_30 = {0, 400};
const std::vector<uint32_t> kInvalidsW_35 = {0, 2049};
const std::vector<uint32_t> kInvaliddH_30 = {0, 2, 400};
const std::vector<uint32_t> kInvaliddH_35 = {0, 2, 2049};
const std::vector<uint32_t> kInvaliddW_30 = {0, 2, 400};
const std::vector<uint32_t> kInvaliddW_35 = {0, 2, 2049};
const GNACnn2DValidatorTestParam target_30 {
kGnaTarget3_0,
"inH",
kInvalidH_30,
};
const GNACnn2DValidatorTestParam target_35 {
kGnaTarget3_5,
"inH",
kInvalidH_35,
};
const GNACnn2DValidatorTestParam target_30_inW{
kGnaTarget3_0,
"inW",
kInvalidW_30,
};
const GNACnn2DValidatorTestParam target_35_inW{
kGnaTarget3_5,
"inW",
kInvalidW_35,
};
const GNACnn2DValidatorTestParam target_30_inC{
kGnaTarget3_0,
"inC",
kInvalidC_30,
};
const GNACnn2DValidatorTestParam target_35_inC{
kGnaTarget3_5,
"inC",
kInvalidC_35,
};
const GNACnn2DValidatorTestParam target_30_kH{
kGnaTarget3_0,
"kH",
kInvalidkH_30,
};
const GNACnn2DValidatorTestParam target_35_kH{
kGnaTarget3_5,
"kH",
kInvalidkH_35,
};
const GNACnn2DValidatorTestParam target_30_kW{
kGnaTarget3_0,
"kW",
kInvalidkW_30,
};
const GNACnn2DValidatorTestParam target_35_kW{
kGnaTarget3_5,
"kW",
kInvalidkW_35,
};
const GNACnn2DValidatorTestParam target_30_kN{
kGnaTarget3_0,
"inC",
kInvalidkN_30,
};
const GNACnn2DValidatorTestParam target_35_kN{
kGnaTarget3_5,
"inC",
kInvalidkN_35,
};
const GNACnn2DValidatorTestParam target_30_sH{
kGnaTarget3_0,
"sH",
kInvalidsH_30,
};
const GNACnn2DValidatorTestParam target_35_sH{
kGnaTarget3_5,
"sH",
kInvalidsH_35,
};
const GNACnn2DValidatorTestParam target_30_sW{
kGnaTarget3_0,
"sW",
kInvalidsW_30,
};
const GNACnn2DValidatorTestParam target_35_sW{
kGnaTarget3_5,
"sW",
kInvalidsW_35,
};
const GNACnn2DValidatorTestParam target_30_dH{
kGnaTarget3_0,
"dH",
kInvaliddH_30,
};
const GNACnn2DValidatorTestParam target_35_dH{
kGnaTarget3_5,
"dH",
kInvaliddH_35,
};
const GNACnn2DValidatorTestParam target_30_dW{
kGnaTarget3_0,
"dW",
kInvaliddW_30,
};
const GNACnn2DValidatorTestParam target_35_dW{
kGnaTarget3_5,
"dW",
kInvaliddW_35,
};
const std::vector<uint32_t> kInvalidpw_30 = {0, 2, 10};
const GNACnn2DValidatorTestParam target_30_pwH{
kGnaTarget3_0,
"windowH",
kInvalidpw_30,
};
const GNACnn2DValidatorTestParam target_30_pwW{
kGnaTarget3_0,
"windowW",
kInvalidpw_30,
};
const std::vector<uint32_t> kInvalidps_30 = {0, 4, 10};
const GNACnn2DValidatorTestParam target_30_psH{
kGnaTarget3_0,
"strideH",
kInvalidps_30,
};
const GNACnn2DValidatorTestParam target_30_psW{
kGnaTarget3_0,
"strideW",
kInvalidps_30,
};
const std::vector<uint32_t> kInvalidPoolingRange35 = {0, 256};
const GNACnn2DValidatorTestParam target_35_pwH{
kGnaTarget3_5,
"windowH",
kInvalidPoolingRange35,
};
const GNACnn2DValidatorTestParam target_35_pwW{
kGnaTarget3_5,
"windowW",
kInvalidPoolingRange35,
};
const GNACnn2DValidatorTestParam target_35_psH{
kGnaTarget3_5,
"strideH",
kInvalidPoolingRange35,
};
const GNACnn2DValidatorTestParam target_35_psW{
kGnaTarget3_5,
"strideW",
kInvalidPoolingRange35,
};
struct ValidateCnn2DParams {
std::map<std::string, uint32_t> parameters;
OvGnaType precission;
static const bool exceptionMode = false;
static ValidateCnn2DParams GetValid() {
ValidateCnn2DParams v;
v.parameters["inH"] = 16;
v.parameters["inW"] = 16;
v.parameters["inC"] = 16;
v.parameters["kH"] = 2;
v.parameters["kW"] = 2;
v.parameters["kN"] = 8;
v.parameters["sH"] = 1;
v.parameters["sW"] = 1;
v.parameters["dH"] = 1;
v.parameters["dW"] = 1;
v.precission = OvGnaTypeInt16;
return v;
}
static ValidateCnn2DParams GetValidPooling() {
ValidateCnn2DParams v;
v.parameters["windowH"] = 3;
v.parameters["windowW"] = 3;
v.parameters["strideH"] = 3;
v.parameters["strideW"] = 3;
return v;
}
bool ValidateCnn2D(const Cnn2D::AbstractValidator& validator) {
return validator.ValidateCnn2D({},
parameters["inH"],
parameters["inW"],
parameters["inC"],
parameters["kH"],
parameters["kW"],
parameters["kN"],
parameters["sH"],
parameters["sW"],
parameters["dH"],
parameters["dW"],
precission, exceptionMode);
}
bool ValidatePooling2D(const Cnn2D::AbstractValidator& validator) {
return validator.ValidatePooling2D({},
parameters["windowH"],
parameters["windowW"],
parameters["strideH"],
parameters["strideW"],
exceptionMode);
}
void set(const std::string& what, const uint32_t value) {
if (what == "precission") {
precission = static_cast<OvGnaType>(value);
} else {
parameters[what] = value;
}
}
};
class GNACnn2DValidatorTest : public ::testing::TestWithParam<GNACnn2DValidatorTestParam> {
protected:
void SetUp() override {
validator = Cnn2D::AbstractValidator::Create(GetParam().target);
ASSERT_TRUE(validator != nullptr);
}
std::unique_ptr<Cnn2D::AbstractValidator> validator;
};
class GNACnn2DValidatorTestPadding : public GNACnn2DValidatorTest {
protected:
bool isPaddingSupported() {
static const std::set<std::string> supported{kGnaTarget3_5};
return supported.count(GetParam().target);
}
};
class GNACnn2DValidatorTestPooling2D : public GNACnn2DValidatorTest {};
namespace {
TEST_P(GNACnn2DValidatorTestPadding, testPaddingSupported) {
ASSERT_TRUE(validator->IsPaddingSupported() == isPaddingSupported());
}
TEST_P(GNACnn2DValidatorTest, testValidateCnn2DInvalid) {
auto valid = ValidateCnn2DParams::GetValid();
for (const auto invalid : GetParam().invalid) {
valid.set(GetParam().whatInvalid, invalid);
ASSERT_FALSE(valid.ValidateCnn2D(*validator));
}
}
TEST_P(GNACnn2DValidatorTestPooling2D, testValidateCnn2DInvalid) {
auto valid = ValidateCnn2DParams::GetValidPooling();
for (const auto invalid : GetParam().invalid) {
valid.set(GetParam().whatInvalid, invalid);
ASSERT_FALSE(valid.ValidatePooling2D(*validator));
}
}
INSTANTIATE_TEST_SUITE_P(smoke_GNACnn2DValidatorTestPadding,
GNACnn2DValidatorTestPadding,
testing::Values(target_30, target_35));
INSTANTIATE_TEST_SUITE_P(smoke_GNACnn2DValidatorTest,
GNACnn2DValidatorTest,
testing::Values(target_30,
target_35,
target_30_inW,
target_35_inW,
target_30_inC,
target_35_inC,
target_30_kH,
target_35_kH,
target_30_kW,
target_35_kW,
target_30_kN,
target_35_kN,
target_30_sH,
target_30_sW,
target_30_dH,
target_30_dW,
target_35_sH,
target_35_sW,
target_35_dH,
target_35_dW)
);
INSTANTIATE_TEST_SUITE_P(smoke_GNACnn2DValidatorTestPooling2D,
GNACnn2DValidatorTestPooling2D,
testing::Values(target_30_pwH,
target_30_pwW,
target_30_psH,
target_30_psW,
target_35_pwH,
target_35_pwW,
target_35_psH,
target_35_psW));
} // namespace