Skip to content
This repository was archived by the owner on Jul 18, 2024. It is now read-only.

refactoring work in nn-hal #117

Open
jxzgithub wants to merge 6 commits into
base: chromeos-gna
Choose a base branch
from
Open

refactoring work in nn-hal #117

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
e263d71
Remove duplicated allocation in blob objects
jxzgithub Jul 26, 2021
483a2e9
Output a log line when deallocation returns false
jxzgithub Jul 26, 2021
bdbf2be
refactoring: remove reset buffer ptr
jxzgithub Jul 26, 2021
8361f7c
remove mOpIndex2BlobMap and the closure using it
jxzgithub Jul 28, 2021
afe114a
remove unused helper function addToBlobLayerMap()
jxzgithub Jul 28, 2021
29e9215
amend logic in GnaPreparedModel::operationTANH()
jxzgithub Jul 30, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
162 changes: 49 additions & 113 deletions intel_nn_hal/GnaPreparedModel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,7 +197,8 @@ bool GnaPreparedModel::constructGNAGraph(std::pair<int, int> indices) {
gnaPluginPtr->loadNetwork(passed_network, isDecoderNw);
#endif
for (auto item:mModelIRBlobs) {
item->deallocate();
if (item->deallocate() == false)
ALOGI("%s deallocating IR an IR blob returns false", __func__);
}
gnaPluginPtr->queryState();
gnaPluginPtr->reset();
Expand Down Expand Up @@ -1593,23 +1594,11 @@ bool GnaPreparedModel::operationAdd(const Operation& operation) {
return (int)op.lifetime;
};

auto getIRBlobFromOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];

auto blob = GetConstOperandAsTensor(idx, offset);
if (op.lifetime == V1_3_OperandLifeTime::SUBGRAPH_INPUT)
{
mOpIndex2BlobMap[idx] = blob;
}

return blob;
};

IRBuilder::BuilderADDLayer::AddParams params;
params.input1.data = getIRBlobFromOperand(operation.inputs[0], 0);
params.input1.data = GetConstOperandAsTensor(operation.inputs[0], 0);
params.input1.lifeTime = getV1_3_OperandLifeTime(operation.inputs[0]);

params.input2.data = getIRBlobFromOperand(operation.inputs[1], 1);
params.input2.data = GetConstOperandAsTensor(operation.inputs[1], 1);
params.input2.lifeTime = getV1_3_OperandLifeTime(operation.inputs[1]);

auto input1Dims = params.input1.data->getTensorDesc().getDims();
Expand Down Expand Up @@ -1691,20 +1680,8 @@ bool GnaPreparedModel::operationTANH(const Operation& operation) {
return (int)op.lifetime;
};

auto getIRBlobFromOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];

auto blob = GetConstOperandAsTensor(idx, offset);
if (op.lifetime == V1_3_OperandLifeTime::SUBGRAPH_INPUT)
{
mOpIndex2BlobMap[idx] = blob;
}

return blob;
};

IRBuilder::BuilderTANHLayer::TanhParams params;
params.input.data = getIRBlobFromOperand(operation.inputs[0], 0);
params.input.data = GetConstOperandAsTensor(operation.inputs[0], 0);
params.input.lifeTime = getV1_3_OperandLifeTime(operation.inputs[0]);

auto inputDims = params.input.data->getTensorDesc().getDims();
Expand All @@ -1717,7 +1694,7 @@ bool GnaPreparedModel::operationTANH(const Operation& operation) {

if (mBuilderModel == nullptr) {
VLOG(L1, "mBuilder = nullptr !!!");
// ASSERT
return false;
}

std::vector<std::string> inLayers;
Expand Down Expand Up @@ -1783,17 +1760,6 @@ bool GnaPreparedModel::operationFullyConnected(const Operation& operation) {
return (int)op.lifetime;
};

auto getIRBlobFromOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];
auto blob = GetConstOperandAsTensor(idx, offset);
if (op.lifetime == V1_3_OperandLifeTime::SUBGRAPH_INPUT)
{
mOpIndex2BlobMap[idx] = blob;
}

return blob;
};

auto validateOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];
auto len_out = op.location.length;
Expand All @@ -1805,12 +1771,12 @@ bool GnaPreparedModel::operationFullyConnected(const Operation& operation) {

IRBuilder::BuilderFCLayer::FCParams params;
params.input.lifeTime = getV1_3_OperandLifeTime(operation.inputs[0]);
params.input.data = getIRBlobFromOperand(operation.inputs[0], 0);
params.input.data = GetConstOperandAsTensor(operation.inputs[0], 0);

params.weights.data = getIRBlobFromOperand(operation.inputs[1], 1);
params.weights.data = GetConstOperandAsTensor(operation.inputs[1], 1);
params.weights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[1]);

params.bias.data = getIRBlobFromOperand(operation.inputs[2], 2);
params.bias.data = GetConstOperandAsTensor(operation.inputs[2], 2);
params.bias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[2]);

uint32_t len;
Expand Down Expand Up @@ -1974,19 +1940,6 @@ bool GnaPreparedModel::operationLSTM(const Operation& operation)
return (int)op.lifetime;
};

auto getIRBlobFromOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];

auto blob = GetConstOperandAsTensor(idx, offset);
if (op.lifetime == V1_3_OperandLifeTime::SUBGRAPH_INPUT)
{
mOpIndex2BlobMap[idx] = blob;
VLOG(L1, "blob idx=%d (model_input) ptr=%p", idx, blob.get());
}

return blob;
};

IRBuilder::LstmLayer::LstmCellDescription lstmDesc;
lstmDesc.clippingThresholdCellState = 0;
lstmDesc.clippingThresholdProjState = 0;
Expand Down Expand Up @@ -2043,65 +1996,65 @@ bool GnaPreparedModel::operationLSTM(const Operation& operation)

VLOG(L1, "Lstm cell description %s", lstmDescription.c_str());

params.input.data = getIRBlobFromOperand(operation.inputs[0], 0);
params.input.data = GetConstOperandAsTensor(operation.inputs[0], 0);
params.input.lifeTime = getV1_3_OperandLifeTime(operation.inputs[0]);

params.outputState.data = getIRBlobFromOperand(operation.inputs[18], 18);
params.outputState.data = GetConstOperandAsTensor(operation.inputs[18], 18);
params.outputState.lifeTime = getV1_3_OperandLifeTime(operation.inputs[18]);

params.cellState.data = getIRBlobFromOperand(operation.inputs[19], 19);
params.cellState.data = GetConstOperandAsTensor(operation.inputs[19], 19);
params.cellState.lifeTime = getV1_3_OperandLifeTime(operation.inputs[19]);

params.input2inputWeights.data = getIRBlobFromOperand(operation.inputs[1], 1);
params.input2inputWeights.data = GetConstOperandAsTensor(operation.inputs[1], 1);
params.input2inputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[1]);

params.input2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[2], 2);
params.input2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[2], 2);
params.input2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[2]);

params.input2CellWeights.data = getIRBlobFromOperand(operation.inputs[3], 3);
params.input2CellWeights.data = GetConstOperandAsTensor(operation.inputs[3], 3);
params.input2CellWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[3]);

params.input2OutputWeights.data = getIRBlobFromOperand(operation.inputs[4], 4);
params.input2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[4], 4);
params.input2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[4]);

params.recurrant2inputWeights.data = getIRBlobFromOperand(operation.inputs[5], 5);
params.recurrant2inputWeights.data = GetConstOperandAsTensor(operation.inputs[5], 5);
params.recurrant2inputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[5]);

params.recurrant2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[6], 6);
params.recurrant2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[6], 6);
params.recurrant2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[6]);

params.recurrant2CellWeights.data = getIRBlobFromOperand(operation.inputs[7], 7);
params.recurrant2CellWeights.data = GetConstOperandAsTensor(operation.inputs[7], 7);
params.recurrant2CellWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[7]);

params.recurrant2OutputWeights.data = getIRBlobFromOperand(operation.inputs[8], 8);
params.recurrant2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[8], 8);
params.recurrant2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[8]);

params.cell2InputWeights.data = getIRBlobFromOperand(operation.inputs[9], 9);
params.cell2InputWeights.data = GetConstOperandAsTensor(operation.inputs[9], 9);
params.cell2InputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[9]);

params.cell2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[10], 10);
params.cell2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[10], 10);
params.cell2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[10]);

params.cell2OutputWeights.data = getIRBlobFromOperand(operation.inputs[11], 11);
params.cell2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[11], 11);
params.cell2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[11]);

params.inputGateBias.data = getIRBlobFromOperand(operation.inputs[12], 12);
params.inputGateBias.data = GetConstOperandAsTensor(operation.inputs[12], 12);
params.inputGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[12]);

params.forgetGateBias.data = getIRBlobFromOperand(operation.inputs[13], 13);
params.forgetGateBias.data = GetConstOperandAsTensor(operation.inputs[13], 13);
params.forgetGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[13]);

params.cellBias.data = getIRBlobFromOperand(operation.inputs[14], 14);
params.cellBias.data = GetConstOperandAsTensor(operation.inputs[14], 14);
params.cellBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[14]);

params.outputGateBias.data = getIRBlobFromOperand(operation.inputs[15], 15);
params.outputGateBias.data = GetConstOperandAsTensor(operation.inputs[15], 15);
params.outputGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[15]);

if (lstmDesc.projectionLayerEnabled) {
params.projectionWeights.data = getIRBlobFromOperand(operation.inputs[16], 16);
params.projectionWeights.data = GetConstOperandAsTensor(operation.inputs[16], 16);
params.projectionWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[16]);

params.projectionBias.data = getIRBlobFromOperand(operation.inputs[17], 17);
params.projectionBias.data = GetConstOperandAsTensor(operation.inputs[17], 17);
params.projectionBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[17]);
}

Expand Down Expand Up @@ -2201,18 +2154,6 @@ bool GnaPreparedModel::operationQuantizedLSTM(const Operation& operation)
return (int)op.lifetime;
};

auto getIRBlobFromOperand = [&](uint32_t idx, uint32_t offset) {
const auto op = mModel.main.operands[idx];

auto blob = GetConstOperandAsTensor(idx, offset);
if (op.lifetime == V1_3_OperandLifeTime::SUBGRAPH_INPUT)
{
mOpIndex2BlobMap[idx] = blob;
}

return blob;
};

IRBuilder::LstmLayer::LstmCellDescription lstmDesc;
lstmDesc.clippingThresholdCellState = 0;
lstmDesc.clippingThresholdProjState = 0;
Expand Down Expand Up @@ -2254,67 +2195,67 @@ bool GnaPreparedModel::operationQuantizedLSTM(const Operation& operation)
params.useLayerNorm = true;
params.useBatchedLayerNorm = true;

params.input.data = getIRBlobFromOperand(operation.inputs[0], 0);
params.input.data = GetConstOperandAsTensor(operation.inputs[0], 0);
params.input.lifeTime = getV1_3_OperandLifeTime(operation.inputs[0]);

params.outputState.data = getIRBlobFromOperand(operation.inputs[18], 18);
params.outputState.data = GetConstOperandAsTensor(operation.inputs[18], 18);
params.outputState.lifeTime = getV1_3_OperandLifeTime(operation.inputs[18]);

params.cellState.data = getIRBlobFromOperand(operation.inputs[19], 19);
params.cellState.data = GetConstOperandAsTensor(operation.inputs[19], 19);
params.cellState.lifeTime = getV1_3_OperandLifeTime(operation.inputs[19]);

if (!lstmDesc.cifgEnabled) {
params.input2inputWeights.data = getIRBlobFromOperand(operation.inputs[1], 1);
params.input2inputWeights.data = GetConstOperandAsTensor(operation.inputs[1], 1);
params.input2inputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[1]);

params.recurrant2inputWeights.data = getIRBlobFromOperand(operation.inputs[5], 5);
params.recurrant2inputWeights.data = GetConstOperandAsTensor(operation.inputs[5], 5);
params.recurrant2inputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[5]);

params.inputGateBias.data = getIRBlobFromOperand(operation.inputs[12], 12);
params.inputGateBias.data = GetConstOperandAsTensor(operation.inputs[12], 12);
params.inputGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[12]);
}

params.input2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[2], 2);
params.input2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[2], 2);
params.input2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[2]);

params.input2CellWeights.data = getIRBlobFromOperand(operation.inputs[3], 3);
params.input2CellWeights.data = GetConstOperandAsTensor(operation.inputs[3], 3);
params.input2CellWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[3]);

params.input2OutputWeights.data = getIRBlobFromOperand(operation.inputs[4], 4);
params.input2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[4], 4);
params.input2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[4]);

params.recurrant2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[6], 6);
params.recurrant2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[6], 6);
params.recurrant2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[6]);

params.recurrant2CellWeights.data = getIRBlobFromOperand(operation.inputs[7], 7);
params.recurrant2CellWeights.data = GetConstOperandAsTensor(operation.inputs[7], 7);
params.recurrant2CellWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[7]);

params.recurrant2OutputWeights.data = getIRBlobFromOperand(operation.inputs[8], 8);
params.recurrant2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[8], 8);
params.recurrant2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[8]);

params.cell2InputWeights.data = getIRBlobFromOperand(operation.inputs[9], 9);
params.cell2InputWeights.data = GetConstOperandAsTensor(operation.inputs[9], 9);
params.cell2InputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[9]);

params.cell2ForgetWeights.data = getIRBlobFromOperand(operation.inputs[10], 10);
params.cell2ForgetWeights.data = GetConstOperandAsTensor(operation.inputs[10], 10);
params.cell2ForgetWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[10]);

params.cell2OutputWeights.data = getIRBlobFromOperand(operation.inputs[11], 11);
params.cell2OutputWeights.data = GetConstOperandAsTensor(operation.inputs[11], 11);
params.cell2OutputWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[11]);

params.forgetGateBias.data = getIRBlobFromOperand(operation.inputs[13], 13);
params.forgetGateBias.data = GetConstOperandAsTensor(operation.inputs[13], 13);
params.forgetGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[13]);

params.cellBias.data = getIRBlobFromOperand(operation.inputs[14], 14);
params.cellBias.data = GetConstOperandAsTensor(operation.inputs[14], 14);
params.cellBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[14]);

params.outputGateBias.data = getIRBlobFromOperand(operation.inputs[15], 15);
params.outputGateBias.data = GetConstOperandAsTensor(operation.inputs[15], 15);
params.outputGateBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[15]);

if (lstmDesc.projectionLayerEnabled) {
params.projectionWeights.data = getIRBlobFromOperand(operation.inputs[16], 16);
params.projectionWeights.data = GetConstOperandAsTensor(operation.inputs[16], 16);
params.projectionWeights.lifeTime = getV1_3_OperandLifeTime(operation.inputs[16]);

params.projectionBias.data = getIRBlobFromOperand(operation.inputs[17], 17);
params.projectionBias.data = GetConstOperandAsTensor(operation.inputs[17], 17);
params.projectionBias.lifeTime = getV1_3_OperandLifeTime(operation.inputs[17]);
}

Expand Down Expand Up @@ -2756,7 +2697,6 @@ IRBlob::Ptr GnaPreparedModel::GetConstWeightsOperandAsTensor(uint32_t index)
else {
InferenceEngine::TBlob<float>::Ptr blob =
std::make_shared<InferenceEngine::TBlob<float>>(td, (float *)buf, len);
blob->allocate();
}
return blob;
} else {
Expand Down Expand Up @@ -2883,11 +2823,9 @@ IRBlob::Ptr GnaPreparedModel::GetConstOperandAsTensor(int operand_index, int ope
#endif
} else {
blob = std::make_shared<InferenceEngine::TBlob<float>>(td, (float *)buf, len);
blob->allocate();
}
if (op.lifetime == V1_3_OperandLifeTime::CONSTANT_COPY || op.lifetime == V1_3_OperandLifeTime::CONSTANT_REFERENCE) {
mModelIRBlobs.push_back(blob);
buf = nullptr;
}
return blob;
} else {
Expand Down Expand Up @@ -2918,7 +2856,6 @@ IRBlob::Ptr GnaPreparedModel::GetConstOperandAsTensor(int operand_index, int ope
}
if (op.lifetime == V1_3_OperandLifeTime::CONSTANT_COPY || op.lifetime == V1_3_OperandLifeTime::CONSTANT_REFERENCE) {
mModelIRBlobs.push_back(blob);
buf = nullptr;
}
return blob;
}
Expand Down Expand Up @@ -2987,7 +2924,6 @@ Blob::Ptr GnaPreparedModel::GetInOutOperandAsBlob(RunTimeOperandInfo& op, const
#endif
} else {
blob = std::make_shared<InferenceEngine::TBlob<float>>(td, (float *)buf, len);
blob->allocate();
}
return blob;
} else {
Expand Down
1 change: 0 additions & 1 deletion intel_nn_hal/PreparedModel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -259,7 +259,6 @@ class PreparedModel : public V1_3::IPreparedModel {
uint32_t mPadreq;

InferenceEngine::ICNNNetwork *mCnnNetbuilder;
std::map<int, IRBlob::Ptr> mOpIndex2BlobMap;
};

class VpuPreparedModel : public PreparedModel {
Expand Down
4 changes: 0 additions & 4 deletions intel_nn_hal/graphAPI/IRBuilder.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -166,10 +166,6 @@ class ModelBuilder
std::vector<std::string>& inputLayerNames);
std::shared_ptr<InferenceEngine::ICNNNetwork> convertBuilder();

void addToBlobLayerMap(IRBlob::Ptr blob, int index)
{
mBlob2LayerIdxMap[blob] = index;
}
int layer_name_count = 0;

std::map<IRBlob::Ptr, int> mBlob2LayerIdxMap;
Expand Down