#include <ios>
#include <iomanip>
#include <map>
+#include <ie_algorithm.hpp>
+#include <ie_common.h>
+#include <ie_precision.hpp>
#if defined __INTEL_COMPILER || defined _MSC_VER
#include <malloc.h>
#include "gna_plugin.hpp"
#include "gna_model_serial.hpp"
+#include "serial/headers/latest/gna_model_header.hpp"
+
+using namespace GNAPluginNS;
inline void writeNBytes(const void *ptr, uint32_t size, std::ostream & os) {
os.write(static_cast<const char*>(ptr), size);
const int gna_header_magic = is_little_endian() ? 0x4d414e47 : 0x474e414d;
-ModelHeader GNAModelSerial::ReadHeader(std::istream &is) {
+GNAPluginNS::HeaderLatest::ModelHeader GNAModelSerial::ReadHeader(std::istream &is) {
is.exceptions(std::istream::failbit);
-
- ModelHeader header;
- readBits(header, is);
+ is.seekg(0, is.end);
+ auto stream_len = is.tellg();
+ if (stream_len == -1) {
+ THROW_GNA_EXCEPTION << "Can't open file to import";
+ }
+ is.seekg(0, is.beg);
+
+ HeaderLatest::ModelHeader header;
+ header.version.major = 0u;
+ header.version.minor = 0u;
+ auto size_of_headers_header = sizeof(HeaderLatest::ModelHeader::gnam) + sizeof(HeaderLatest::ModelHeader::headerSize)
+ + sizeof(HeaderLatest::ModelHeader::Version);
+ if (stream_len > size_of_headers_header) {
+ readNBytes(&header, size_of_headers_header, is);
+ } else {
+ readNBytes(&header, stream_len, is);
+ }
if (*reinterpret_cast<int*>(header.gnam) != gna_header_magic) {
THROW_GNA_EXCEPTION << "Imported file unsupported: magic number should be GNAM(0x474e414d), but was 0x"
<< std::setfill('0') <<
std::hex << std::setw(2) << static_cast<short>(header.gnam[2]) <<
std::hex << std::setw(2) << static_cast<short>(header.gnam[3]);
}
- if (header.version.major != HEADER_MAJOR) {
- THROW_GNA_EXCEPTION << "Imported file unsupported: major version should be == " << HEADER_MAJOR;
- }
- if (header.headerSize < sizeof(header)) {
- THROW_GNA_EXCEPTION << "Unsupported header size minimal value is : " << sizeof (header) << ", but read: " << header.headerSize;
+
+ is.seekg(0, is.beg);
+ Header2dot1::ModelHeader tempHeader2dot1;
+ switch (header.version.major) {
+ case 2:
+ switch (header.version.minor) {
+ case 1:
+ readBits(tempHeader2dot1, is);
+ header = Header2dot3::ModelHeader(tempHeader2dot1);
+ break;
+ case 2:
+ case 3:
+ readBits(header, is);
+ break;
+ default:
+ THROW_GNA_EXCEPTION << "Imported file unsupported. minor version should be equal to 1 or 2 and is: " << header.version.minor;
+ }
+ break;
+ default:
+ THROW_GNA_EXCEPTION << "Imported file unsupported. Import for files with major version equal to: " << header.version.major << " is not implemented";
}
+
/*
* extra data need to be added into new header and modify check as appropriate
*/
sizeof(Gna2Shape),
sizeof(Gna2Shape)}},
{Gna2OperationTypeCopy, {sizeof(Gna2Shape)}},
+ {Gna2OperationTypeTransposition, {sizeof(Gna2Shape)}},
};
-void GNAModelSerial::Import(void *basePointer, size_t gnaGraphSize, std::istream & is) {
+void GNAModelSerial::Import(void *basePointer,
+ size_t gnaGraphSize,
+ std::istream & is,
+ std::shared_ptr<GNAPluginNS::InputDesc> inputsDesc,
+ std::vector<GNAPluginNS::OutputDesc> &desc,
+ InferenceEngine::InputsDataMap& inputsDataMap,
+ InferenceEngine::OutputsDataMap& outputsDataMap) {
is.exceptions(std::istream::failbit);
+ if (modelHeader.version.major == 2) {
+ if (modelHeader.version.minor >= 3) {
+ for (auto inputIndex = 0; inputIndex < modelHeader.nInputs; inputIndex++) {
+ uint32_t nameSize = 0;
+ readNBits<32>(nameSize, is);
+ std::string inName(nameSize, '\0');
+ readNBytes(&inName[0], nameSize, is);
+ inputNames.push_back(inName.substr(0, nameSize - 1));
+ }
+ }
+ }
+ ImportInputs(is, basePointer, inputsDesc, inputsDataMap);
+
+ if (modelHeader.version.major == 2) {
+ if (modelHeader.version.minor >= 3) {
+ for (auto inputIndex = 0; inputIndex < modelHeader.nOutputs; inputIndex++) {
+ uint32_t nameSize = 0;
+ readNBits<32>(nameSize, is);
+ std::string outName(nameSize, '\0');
+ readNBytes(&outName[0], nameSize, is);
+ outputNames.push_back(outName.substr(0, nameSize - 1));
+ }
+ }
+ }
+ ImportOutputs(is, basePointer, desc, outputsDataMap);
+
for (auto operation = gna2Model->Operations; operation != gna2Model->Operations + gna2Model->NumberOfOperations; ++operation) {
readNBits<32>(operation->Type, is);
readBits(operation->NumberOfOperands, is);
operation->Operands = static_cast<Gna2Tensor const **>(gnaUserAllocator(sizeof(Gna2Tensor*) * operation->NumberOfOperands));
+ IE_ASSERT(operation->Operands != nullptr);
for (uint32_t i = 0; i < operation->NumberOfOperands; i++) {
Gna2Tensor t{};
readBits(t, is);
case Gna2OperationTypeFullyConnectedAffine:
case Gna2OperationTypeConvolution:
case Gna2OperationTypeCopy:
+ case Gna2OperationTypeTransposition:
break;
case Gna2OperationTypeRecurrent:
THROW_GNA_EXCEPTION << "Importing of recurrent operation not supported";
- case Gna2OperationTypeTransposition:
- THROW_GNA_EXCEPTION << "Importing of transposition operation not supported";
default:
THROW_GNA_EXCEPTION << "Importing of unknown GNA operation type(" << operation->Type << ") not supported";
}
else
operation->Parameters = nullptr;
for (uint32_t i = 0; i < operation->NumberOfParameters; i++) {
- uint32_t paramSize;
+ uint32_t paramSize = 0;
readBits(paramSize, is);
+ IE_ASSERT(operation->Parameters != nullptr);
if (paramSize == 0) {
+ IE_ASSERT(operation->Parameters != nullptr);
operation->Parameters[i] = nullptr;
continue;
}
return out;
};
- auto convert_to_serial = [getOffsetFromBase](const GNAModelSerial::RuntimeEndPoint& ep) {
- ModelHeader::EndPoint out;
+ auto convert_to_serial = [getOffsetFromBase](const HeaderLatest::RuntimeEndPoint& ep) {
+ HeaderLatest::RuntimeEndPoint out;
out.elements_count = ep.elements_count;
- out.element_size = ep.element_size;
out.descriptor_offset = offsetFromBase(ep.descriptor_ptr);
out.scaleFactor = ep.scaleFactor;
+ out.element_size = ep.element_size;
+ out.orientation = ep.orientation;
return out;
};
/**
* writing header
*/
- ModelHeader header;
+ HeaderLatest::ModelHeader header;
header.gnam[0] = 'G';
header.gnam[1] = 'N';
header.gnam[2] = 'A';
header.gnam[3] = 'M';
- header.headerSize = sizeof(ModelHeader);
- header.version.major = HEADER_MAJOR;
- header.version.minor = HEADER_MINOR;
+ header.headerSize = sizeof(HeaderLatest::ModelHeader);
header.gnaMemSize = gnaGraphSize;
header.layersCount = layers.size();
header.nGroup = guessGrouping(*gna2Model);
- header.input = convert_to_serial(input);
- header.output = convert_to_serial(output);
-
+ header.nInputs = inputs.size();
+ header.nOutputs = outputs.size();
header.nRotateRows = nRotateRows;
header.nRotateColumns = nRotateColumns;
+ header.doRotateInput = doRotateInput;
writeBits(header, os);
+ for (auto &name : inputNames) {
+ const auto nameSize = strlen(name.c_str()) + 1;
+ writeBits(static_cast<uint32_t>(nameSize), os);
+ writeNBytes(name.c_str(), nameSize , os);
+ }
+ for (const auto &input : inputs) {
+ writeBits(convert_to_serial(input), os);
+ }
+ for (auto &name : outputNames) {
+ const auto nameSize = strlen(name.c_str()) + 1;
+ writeBits(static_cast<uint32_t>(nameSize), os);
+ writeNBytes(name.c_str(), nameSize, os);
+ }
+ for (const auto &output : outputs) {
+ writeBits(convert_to_serial(output), os);
+ }
+
for (const auto & layer : layers) {
writeBits(static_cast<uint32_t>(layer.Type), os);
writeBits(layer.NumberOfOperands, os);
case Gna2OperationTypeFullyConnectedAffine:
case Gna2OperationTypeConvolution:
case Gna2OperationTypeCopy:
+ case Gna2OperationTypeTransposition:
break;
case Gna2OperationTypeRecurrent:
THROW_GNA_EXCEPTION << "Exporting of recurrent operation not supported";
- case Gna2OperationTypeTransposition:
- THROW_GNA_EXCEPTION << "Exporting of interleave operation not supported";
default:
THROW_GNA_EXCEPTION << "Exporting of unknown GNA operation type(" << layer.Type << ") not supported";
}
}
#else
-void GNAModelSerial::Import(void *basePointer, size_t gnaGraphSize, std::istream & is) {
+void GNAModelSerial::Import(void *basePointer,
+ size_t gnaGraphSize,
+ std::istream & is,
+ std::shared_ptr<GNAPluginNS::InputDesc> inputsDesc,
+ std::vector<GNAPluginNS::OutputDesc> &desc,
+ InferenceEngine::InputsDataMap& inputsDataMap,
+ InferenceEngine::OutputsDataMap& outputsDataMap) {
is.exceptions(std::istream::failbit);
+ ImportInputs(is, basePointer, inputsDesc, inputsDataMap);
+ ImportOutputs(is, basePointer, desc, outputsDataMap);
+
auto readPwl = [&is, basePointer](intel_pwl_func_t & value) {
readBits(value.nSegments, is);
if (value.nSegments != 0) {
}
};
- auto convert_to_serial = [getOffsetFromBase](const GNAModelSerial::RuntimeEndPoint& ep){
- ModelHeader::EndPoint out;
+ auto convert_to_serial = [getOffsetFromBase](const HeaderLatest::RuntimeEndPoint& ep){
+ HeaderLatest::RuntimeEndPoint out;
out.elements_count = ep.elements_count;
out.element_size = ep.element_size;
out.descriptor_offset = offsetFromBase(ep.descriptor_ptr);
out.scaleFactor = ep.scaleFactor;
+ out.orientation = ep.orientation;
return out;
};
/**
* writing header
*/
- ModelHeader header;
+ HeaderLatest::ModelHeader header;
header.gnam[0] = 'G';
header.gnam[1] = 'N';
header.gnam[2] = 'A';
header.gnam[3] = 'M';
- header.version.major = HEADER_MAJOR;
- header.version.minor = HEADER_MINOR;
+ header.version.major = 1u;
+ header.version.minor = 1u;
header.gnaMemSize = gnaGraphSize;
header.layersCount = layers.size();
header.nGroup = ptr_nnet->nGroup;
- header.input = convert_to_serial(input);
- header.output = convert_to_serial(output);
- header.headerSize = sizeof(ModelHeader);
+ header.nInputs = 1;
+ header.nOutputs = 1;
+ header.headerSize = sizeof(HeaderLatest::ModelHeader);
header.nRotateRows = nRotateRows;
header.nRotateColumns = nRotateColumns;
writeBits(header, os);
+ writeBits(convert_to_serial(inputs[0]), os);
+ writeBits(convert_to_serial(outputs[0]), os);
for (auto & layer : layers) {
writeBits(layer.nInputColumns, os);
}
#endif
+
+std::vector<HeaderLatest::RuntimeEndPoint> GNAModelSerial::serializeOutputs(const InferenceEngine::OutputsDataMap& outputsDataMap,
+ const std::vector<GNAPluginNS::OutputDesc>& outputsDesc) {
+ std::vector<HeaderLatest::RuntimeEndPoint> endPoints;
+ std::size_t outputIndex = 0;
+ for (auto const &output : outputsDataMap) {
+ auto outputName = output.first;
+ auto inputDims = output.second->getTensorDesc().getDims();
+ uint32_t elementsCount = static_cast<uint32_t>(InferenceEngine::details::product(inputDims.begin(), inputDims.end()));
+
+ HeaderLatest::RuntimeEndPoint endPoint(outputsDesc[outputIndex].scale_factor,
+ outputsDesc[outputIndex].ptrs[0],
+ outputsDesc[outputIndex].num_bytes_per_element,
+ elementsCount,
+ outputsDesc[outputIndex].orientation);
+ endPoints.push_back(endPoint);
+ outputIndex++;
+ }
+ return endPoints;
+}
+
+std::vector<HeaderLatest::RuntimeEndPoint> GNAModelSerial::serializeInputs(const InferenceEngine::InputsDataMap& inputsDataMap,
+ std::shared_ptr<GNAPluginNS::InputDesc> inputDesc) {
+ std::vector<HeaderLatest::RuntimeEndPoint> endPoints;
+
+ std::size_t inputIndex = 0;
+ for (auto const& input : inputsDataMap) {
+ auto inputName = input.first;
+ auto inputDims = input.second->getTensorDesc().getDims();
+
+ double scaleFactor = inputDesc->getScaleFactor(inputIndex);
+ std::vector<void *> descriptor_ptr = inputDesc->getPtrInputsGlobal(inputName);
+ IE_ASSERT(descriptor_ptr.size() > 0);
+ uint32_t element_size = 2u;
+ uint32_t elementsCount = static_cast<uint32_t>(InferenceEngine::details::product(inputDims.begin(), inputDims.end()));
+ intel_dnn_orientation_t orientation = inputDesc->getOrientation(inputName);
+
+ HeaderLatest::RuntimeEndPoint endPoint(scaleFactor,
+ descriptor_ptr[0],
+ element_size,
+ elementsCount,
+ orientation);
+ endPoints.push_back(endPoint);
+ inputIndex++;
+ }
+ return endPoints;
+}
+
+void GNAModelSerial::ImportInputs(std::istream &is,
+ void* basePtr,
+ std::shared_ptr<GNAPluginNS::InputDesc> inputsDesc,
+ InferenceEngine::InputsDataMap& dataMap) {
+ dataMap.clear();
+
+ for (auto inputIndex = 0; inputIndex < modelHeader.nInputs; inputIndex++) {
+ const std::string& name = (modelHeader.version.major == 2 && modelHeader.version.minor >= 3)
+ ? inputNames.at(inputIndex) : std::string("input" + std::to_string(inputIndex));
+ HeaderLatest::RuntimeEndPoint input;
+ is.read(reinterpret_cast<char *>(&input), sizeof(input));
+ inputsDesc->getPtrInputsGlobal(name).push_back(reinterpret_cast<float*>(reinterpret_cast<uint8_t *> (basePtr) + input.descriptor_offset));
+ inputsDesc->orientation_in[name] = input.orientation;
+ inputsDesc->bytes_allocated_for_input[name] = input.element_size * input.elements_count;
+
+ auto inputDims = InferenceEngine::SizeVector({modelHeader.nGroup, input.elements_count / modelHeader.nGroup});
+
+ dataMap[name] = std::make_shared<InferenceEngine::InputInfo>();
+ dataMap[name]->setInputData(std::make_shared<InferenceEngine::Data>(name,
+ InferenceEngine::TensorDesc(
+ InferenceEngine::Precision::FP32,
+ inputDims,
+ InferenceEngine::Layout::NC)));
+ inputsDesc->inputScaleFactors.push_back(input.scaleFactor);
+ }
+}
+
+void GNAModelSerial::ImportOutputs(std::istream &is,
+ void* basePtr,
+ std::vector<GNAPluginNS::OutputDesc> &desc,
+ InferenceEngine::OutputsDataMap& dataMap) {
+ desc.clear();
+ dataMap.clear();
+ desc.resize(modelHeader.nOutputs);
+
+ for (auto outputIndex = 0; outputIndex < modelHeader.nOutputs; outputIndex++) {
+ const std::string& name = (modelHeader.version.major == 2 && modelHeader.version.minor >= 3)
+ ? outputNames.at(outputIndex) : std::string("input" + std::to_string(outputIndex));
+ HeaderLatest::RuntimeEndPoint output;
+ is.read(reinterpret_cast<char *>(&output), sizeof(output));
+ OutputDesc description;
+ description.ptrs.push_back(reinterpret_cast<float*>(reinterpret_cast<uint8_t *> (basePtr) + output.descriptor_offset));
+ description.orientation = kDnnInterleavedOrientation;
+ description.orientation = output.orientation;
+ description.num_bytes_per_element = output.element_size;
+ description.scale_factor = output.scaleFactor;
+
+ auto outputDims = InferenceEngine::SizeVector({modelHeader.nGroup, output.elements_count / modelHeader.nGroup});
+ dataMap[name] = std::make_shared<InferenceEngine::Data>(name,
+ InferenceEngine::TensorDesc(
+ InferenceEngine::Precision::FP32,
+ outputDims,
+ InferenceEngine::Layout::NC));
+ desc.at(outputIndex) = description;
+ }
+}
+
+void GNAModelSerial::setHeader(HeaderLatest::ModelHeader header) {
+ modelHeader = header;
+}
projects / platform / upstream / dldt.git / blobdiff