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

[CPU] Avoid inserting reorder after RNN in native order case (#10799)

Browse Source
This commit is contained in:
Maxim Andronov authored and GitHub committed 2022-04-01 16:02:50 +03:00
1 parent fca159293d
commit 3d92c8c4c7
4 files changed
+91 -31
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/plugins/intel_cpu/src/graph_optimizer.cpp
+1 -2
View File
@@ -2161,8 +2161,7 @@ void GraphOptimizer::reshapeRnnSeq(Graph &graph) {
if (node->type != Type::RNNSeq)
return false;
auto rnnNode = std::dynamic_pointer_cast<RNN>(node);
return rnnNode && (!rnnNode->hasNativeOrder() || node->isDynamicNode()) && node->outputShapes[0].getRank() == 4 &&
node->outputShapes[0].getDims()[1] == 1;
return rnnNode && !rnnNode->hasNativeOrder() && node->outputShapes[0].getRank() == 4 && node->outputShapes[0].getDims()[1] == 1;
};
for (size_t i = 0; i < graphNodes.size(); i++) {
src/plugins/intel_cpu/src/nodes/reorder.cpp
+1 -1
View File
@@ -105,7 +105,7 @@ void Reorder::initSupportedPrimitiveDescriptors() {
isDynamic = !(config.inConfs[0].getMemDesc()->isDefined() && config.outConfs[0].getMemDesc()->isDefined());
if (isDynamic && (config.inConfs[0].getMemDesc()->getShape().getRank() != config.outConfs[0].getMemDesc()->getShape().getRank()))
IE_THROW() << "Reorder node doesn't support case when input and output shapes have different rank and dynamic";
IE_THROW() << "Reorder node with name: " << getName() << " doesn't support case when input and output shapes have different rank and dynamic";
if (!isOptimized) {
const auto &inShape = getInputShapeAtPort(0);
if (one_of(inShape.getRank(), 4, 5) &&
src/plugins/intel_cpu/src/nodes/rnn.cpp
+2 -3
View File
@@ -481,8 +481,7 @@ void RNN::fillSequenceDesc() {
outCandidate.reserve(3);
if (nativeOrder) {
outCandidate.emplace_back(std::make_shared<DnnlBlockedMemoryDesc>(Shape{{T.minVal, N.minVal, SC}, {T.maxVal, N.maxVal, SC}},
dataType, memory::format_tag::tnc));
outCandidate.emplace_back(std::make_shared<DnnlBlockedMemoryDesc>(getOutputShapeAtPort(0), dataType, memory::format_tag::abcd));
} else if (N.isStatic() && N.maxVal == 1) {
// WA to avoid reorder after sequence for some models
outCandidate.emplace_back(std::make_shared<DnnlBlockedMemoryDesc>(shapeNTSC, dataType, memory::format_tag::tnc));
@@ -926,7 +925,7 @@ std::vector<VectorDims> RNN::shapeInfer() const {
auto originOutputShapes = Node::shapeInfer();
// Graph optimizer makes the same optimization. So this is required to make shapes compatible.
if (getType() == Type::RNNSeq && originOutputShapes[0].size() == 4lu && originOutputShapes[0][1] == 1lu) {
if (getType() == Type::RNNSeq && !hasNativeOrder() && originOutputShapes[0].size() == 4lu && originOutputShapes[0][1] == 1lu) {
originOutputShapes[0].erase(originOutputShapes[0].begin() + 1);
}
return originOutputShapes;
src/tests/functional/plugin/cpu/subgraph_tests/src/seq_native_order.cpp
+87 -25
View File
@@ -33,7 +33,8 @@ using SeqParams = std::tuple<SEQ_TYPE, // node type
float, // Clip
bool, // Linear_before_reset
ov::op::RecurrentSequenceDirection, // Direction
ElementType>; // Network precision
ElementType, // Network precision
ngraph::helpers::InputLayerType>; // 'sequence_lengths' input type
class SequenceCPUTest : public testing::WithParamInterface<SeqParams>, virtual public ov::test::SubgraphBaseTest, public CPUTestsBase {
public:
@@ -46,8 +47,9 @@ public:
bool linearBeforeReset;
ov::op::RecurrentSequenceDirection direction;
ElementType netPrecision;
ngraph::helpers::InputLayerType seqInType;
std::tie(seqType, hidden_size, input_size, inShapeParams, activations, clip, linearBeforeReset, direction, netPrecision) = obj.param;
std::tie(seqType, hidden_size, input_size, inShapeParams, activations, clip, linearBeforeReset, direction, netPrecision, seqInType) = obj.param;
std::vector<ov::Dimension> bounds;
std::vector<TargetShapeParams> targetShapes;
@@ -76,7 +78,8 @@ public:
result << "clip=" << clip << "_";
result << "linear=" << linearBeforeReset << "_";
result << "direction=" << direction << "_";
result << "netPrec=" << netPrecision;
result << "netPrec=" << netPrecision << "_";
result << "seqInType=" << seqInType << "_";
return result.str();
}
@@ -95,7 +98,7 @@ protected:
ov::op::RecurrentSequenceDirection direction;
ElementType netPrecision;
std::tie(seqType, hidden_size, input_size, inShapeParams, activations, clip, linearBeforeReset, direction, netPrecision) = this->GetParam();
std::tie(seqType, hidden_size, input_size, inShapeParams, activations, clip, linearBeforeReset, direction, netPrecision, seqInType) = this->GetParam();
std::vector<ov::Dimension> bounds;
std::vector<TargetShapeParams> targetShapes;
@@ -116,8 +119,6 @@ protected:
if (seqType == SEQ_TYPE::LSTM) {
inputDynamicShapes.push_back(second_in_shape);
}
ov::PartialShape seq_len_shape(std::vector<ov::Dimension>{bounds[batch_size_pos]});
inputDynamicShapes.push_back(seq_len_shape);
auto hidden_size_weight = hidden_size;
if (seqType == SEQ_TYPE::GRU) {
@@ -139,6 +140,14 @@ protected:
}
weightShape.push_back(B_shape);
ov::PartialShape seq_len_shape(std::vector<ov::Dimension>{bounds[batch_size_pos]});
if (seqInType == ngraph::helpers::InputLayerType::PARAMETER) {
inputDynamicShapes.push_back(seq_len_shape);
} else {
IE_ASSERT(seq_len_shape.is_static());
weightShape.push_back(seq_len_shape.to_shape());
}
// target shape
for (const auto &ts : targetShapes) {
std::vector<ov::Shape> currTS;
@@ -151,13 +160,17 @@ protected:
if (seqType == SEQ_TYPE::LSTM) {
currTS.emplace_back(std::vector<size_t>{bs, numDirections, hidden_size});
}
currTS.emplace_back(std::vector<size_t>{bs});
if (seqInType == ngraph::helpers::InputLayerType::PARAMETER) {
currTS.emplace_back(std::vector<size_t>{bs});
}
targetStaticShapes.push_back(currTS);
}
// funciton creation
std::vector<ov::element::Type> types(inputDynamicShapes.size(), netPrecision);
types.back() = ElementType::i64;
if (seqInType == ngraph::helpers::InputLayerType::PARAMETER) {
types.back() = ElementType::i64;
}
auto params = ngraph::builder::makeDynamicParams(types, inputDynamicShapes);
std::vector<int64_t> order_ref_before = {1, 0, 2};
@@ -184,7 +197,9 @@ protected:
linearBeforeReset,
true,
direction,
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM);
(seqInType == ngraph::helpers::InputLayerType::CONSTANT ?
ngraph::helpers::SequenceTestsMode::CONVERT_TO_TI_MAX_SEQ_LEN_CONST :
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM));
} else if (seqType == SEQ_TYPE::LSTM) {
seq_node = ngraph::builder::makeLSTM(inputs,
weightShape,
@@ -195,7 +210,9 @@ protected:
clip,
true,
direction,
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM);
(seqInType == ngraph::helpers::InputLayerType::CONSTANT ?
ngraph::helpers::SequenceTestsMode::CONVERT_TO_TI_MAX_SEQ_LEN_CONST :
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM));
} else if (seqType == SEQ_TYPE::RNN) {
seq_node = ngraph::builder::makeRNN(inputs,
weightShape,
@@ -206,7 +223,9 @@ protected:
clip,
true,
direction,
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM);
(seqInType == ngraph::helpers::InputLayerType::CONSTANT ?
ngraph::helpers::SequenceTestsMode::CONVERT_TO_TI_MAX_SEQ_LEN_CONST :
ngraph::helpers::SequenceTestsMode::PURE_SEQ_RAND_SEQ_LEN_PARAM));
} else {
IE_THROW() << "Unsupported seq type";
}
@@ -232,16 +251,19 @@ protected:
const size_t batchSize = targetInputStaticShapes[0][1];
const int64_t maxSeqLen = targetInputStaticShapes[0][0];
const auto& funcInputs = function->inputs();
const auto& seqLenInput = inputs.find(funcInputs[seqLengthInIdx].get_node_shared_ptr());
if (seqLenInput == inputs.end())
throw std::runtime_error("Could not find Sequence length input.");
if (seqInType == ngraph::helpers::InputLayerType::PARAMETER) {
const auto& funcInputs = function->inputs();
const auto& seqLenInput = inputs.find(funcInputs[seqLengthInIdx].get_node_shared_ptr());
if (seqLenInput == inputs.end())
throw std::runtime_error("Could not find Sequence length input.");
auto lenData = seqLenInput->second.data<ov::element_type_traits<ElementType::i64>::value_type>();
std::fill(lenData, lenData + batchSize, maxSeqLen);
auto lenData = seqLenInput->second.data<ov::element_type_traits<ElementType::i64>::value_type>();
std::fill(lenData, lenData + batchSize, maxSeqLen);
}
}
private:
ngraph::helpers::InputLayerType seqInType;
size_t seqLengthInIdx = 2;
};
@@ -265,14 +287,27 @@ const std::vector<size_t> inputSizes = {
1, 10
};
const std::vector<InputShapeParams> inShapeParams = {
const std::vector<InputShapeParams> inShapeParams_dynamic = {
InputShapeParams{std::vector<ov::Dimension>{-1, -1}, std::vector<TargetShapeParams>{TargetShapeParams{3, 8},
TargetShapeParams{10, 2}}},
InputShapeParams{std::vector<ov::Dimension>{{1, 15}, {1, 15}}, std::vector<TargetShapeParams>{TargetShapeParams{3, 8},
TargetShapeParams{10, 2}}}
InputShapeParams{std::vector<ov::Dimension>{{1, 15}, {1, 5}}, std::vector<TargetShapeParams>{TargetShapeParams{7, 5},
TargetShapeParams{10, 2}}},
InputShapeParams{std::vector<ov::Dimension>{{1, 8}, 9}, std::vector<TargetShapeParams>{TargetShapeParams{7, 9},
TargetShapeParams{8, 9}}},
InputShapeParams{std::vector<ov::Dimension>{6, {1, 5}}, std::vector<TargetShapeParams>{TargetShapeParams{6, 5},
TargetShapeParams{6, 2}}},
};
std::vector<std::vector<std::string>> activations = {
const std::vector<InputShapeParams> inShapeParams_static = {
InputShapeParams{std::vector<ov::Dimension>{10, 2}, std::vector<TargetShapeParams>{TargetShapeParams{10, 2},
TargetShapeParams{10, 2}}}
};
std::vector<std::vector<std::string>> activations_gru_support = {
{"sigmoid", "tanh"}
};
std::vector<std::vector<std::string>> activations_lstm_support = {
{"sigmoid", "tanh", "tanh"}
};
@@ -284,16 +319,43 @@ std::vector<bool> linearBeforeReset = {true, false};
std::vector<ElementType> netPrecisions = { ElementType::f32 };
INSTANTIATE_TEST_SUITE_P(smoke_SequenceCPUTest, SequenceCPUTest,
INSTANTIATE_TEST_SUITE_P(smoke_SequenceCPUTest_dynamic, SequenceCPUTest,
::testing::Combine(::testing::ValuesIn(nodeType),
::testing::ValuesIn(hiddenSizes),
::testing::ValuesIn(inputSizes),
::testing::ValuesIn(inShapeParams),
::testing::ValuesIn(activations),
::testing::ValuesIn(inShapeParams_dynamic),
::testing::ValuesIn(activations_lstm_support),
::testing::ValuesIn(clip),
::testing::ValuesIn(linearBeforeReset),
::testing::ValuesIn(direction),
::testing::ValuesIn(netPrecisions)),
::testing::ValuesIn(netPrecisions),
::testing::Values(ngraph::helpers::InputLayerType::PARAMETER)),
SequenceCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_SequenceCPUTest_static_gru, SequenceCPUTest,
::testing::Combine(::testing::Values(SEQ_TYPE::GRU),
::testing::ValuesIn(hiddenSizes),
::testing::ValuesIn(inputSizes),
::testing::ValuesIn(inShapeParams_static),
::testing::ValuesIn(activations_gru_support),
::testing::ValuesIn(clip),
::testing::ValuesIn(linearBeforeReset),
::testing::ValuesIn(direction),
::testing::ValuesIn(netPrecisions),
::testing::Values(ngraph::helpers::InputLayerType::CONSTANT)),
SequenceCPUTest::getTestCaseName);
INSTANTIATE_TEST_SUITE_P(smoke_SequenceCPUTest_static_rnn_lstm, SequenceCPUTest,
::testing::Combine(::testing::Values(SEQ_TYPE::LSTM, SEQ_TYPE::RNN),
::testing::ValuesIn(hiddenSizes),
::testing::ValuesIn(inputSizes),
::testing::ValuesIn(inShapeParams_static),
::testing::ValuesIn(activations_lstm_support),
::testing::ValuesIn(clip),
::testing::ValuesIn(linearBeforeReset),
::testing::ValuesIn(direction),
::testing::ValuesIn(netPrecisions),
::testing::Values(ngraph::helpers::InputLayerType::CONSTANT)),
SequenceCPUTest::getTestCaseName);
} // namespace SubgraphTestsDefinitions