diff --git a/unit_tests/operator/Test_MetaOperator.cpp b/unit_tests/operator/Test_MetaOperator.cpp
index 564d4a41692c033540fa74be7ddf8571f039815c..5fc68acf157400728740eb10dc6c1ae962164df7 100644
--- a/unit_tests/operator/Test_MetaOperator.cpp
+++ b/unit_tests/operator/Test_MetaOperator.cpp
@@ -191,9 +191,9 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
std::shared_ptr<Node> myPaddedConv =
PaddedConv(3, 4, {3, 3}, "myPaddedConv", {1, 1}, {1, 1, 1, 1});
}
-
SECTION("LSTM(forward)") {
- auto myLSTM = LSTM(32, 64, 16, true, "ltsm");
+ auto pop = Pop();
+ auto myLSTM = LSTM(32, 64, 0, true, "ltsm");
auto op = std::static_pointer_cast<OperatorTensor>(myLSTM->getOperator());
auto microGraph = std::dynamic_pointer_cast<MetaOperator_Op>(op)->getMicroGraph();
@@ -204,28 +204,55 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
REQUIRE(myLSTM->nbOutputs() == 2);
std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
- Array1D<float, 32>{{1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0}});
+ Array2D<float, 16, 32>{});
std::shared_ptr<Tensor> myInit = std::make_shared<Tensor>(
- Array2D<float, 1, 64>{{0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0}});
+ Array2D<float, 1, 64>{});
+ std::shared_ptr<Tensor> myInitW = std::make_shared<Tensor>(
+ Array2D<float, 64, 32>{});
+ std::shared_ptr<Tensor> myInitR = std::make_shared<Tensor>(
+ Array2D<float, 64, 64>{});
- op->associateInput(0, myInput);
+ pop->addChild(myLSTM, 0, 0);
+ pop->getOperator()->associateInput(0, myInput);
op->associateInput(17, myInit);
op->associateInput(18, myInit);
- op->computeOutputDims();
+ // Weights X
+ myLSTM->input(1).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(2).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(3).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(4).first->getOperator()->setOutput(0, myInitW);
+ // Weights H
+ myLSTM->input(5).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(6).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(7).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(8).first->getOperator()->setOutput(0, myInitR);
+
+ auto g = getConnectedGraphView(myLSTM);
+ g->setDataType(DataType::Float32);
+ g->setBackend("cpu");
+
+ auto scheduler = SequentialScheduler(g);
+ scheduler.forward(true, true);
+
+ g->save("lstm_outside_dims", true, true);
+
microGraph->save("lstm_dims", true, true);
REQUIRE(op->outputDimsForwarded());
- op->setDataType(DataType::Float32);
- op->setBackend("cpu");
-
- op->forward();
auto microGraphScheduler = std::dynamic_pointer_cast<MetaOperator_Op>(op)->getMicroGraphScheduler();
microGraphScheduler->saveSchedulingDiagram("lstm_scheduling");
- }
+ REQUIRE(op->getNbConsumedData(0) == 512);
+ REQUIRE(op->getNbConsumedData(1) == 32768);
+ REQUIRE(op->getNbProducedData(0) == 1088);
+ REQUIRE(op->getNbProducedData(1) == 1088);
+ REQUIRE(microGraphScheduler->getStaticScheduling(0).size() == 26);
+ REQUIRE(microGraphScheduler->getStaticScheduling(1).size() == 24);
+ REQUIRE(microGraphScheduler->getStaticScheduling(15).size() == 24);
+ }
SECTION("LSTM(forward_values)") {
- auto myLSTM = LSTM(2, 3, 1, true, "ltsm");
+ auto myLSTM = LSTM(2, 3, 0, true, "ltsm");
auto op = std::static_pointer_cast<OperatorTensor>(myLSTM->getOperator());
auto microGraph = std::dynamic_pointer_cast<MetaOperator_Op>(op)->getMicroGraph();
@@ -248,30 +275,32 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
op->associateInput(17, myInit);
op->associateInput(18, myInit);
- op->computeOutputDims();
- REQUIRE(op->outputDimsForwarded());
- microGraph->save("lstm_values_dims", false, true);
-
// Weights X
- op->associateInput(1, myInitW);
- op->associateInput(2, myInitW);
- op->associateInput(3, myInitW);
- op->associateInput(4, myInitW);
+ myLSTM->input(1).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(2).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(3).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(4).first->getOperator()->setOutput(0, myInitW);
// Weights H
- op->associateInput(5, myInitR);
- op->associateInput(6, myInitR);
- op->associateInput(7, myInitR);
- op->associateInput(8, myInitR);
+ myLSTM->input(5).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(6).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(7).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(8).first->getOperator()->setOutput(0, myInitR);
- op->setDataType(DataType::Float32);
- op->setBackend("cpu");
+ auto g = getConnectedGraphView(myLSTM);
+ g->setDataType(DataType::Float32);
+ g->setBackend("cpu");
+
+ auto scheduler = SequentialScheduler(g);
+ scheduler.forward();
+
+ microGraph->save("lstm_values_dims", false, true);
std::shared_ptr<Tensor> myHiddenState = std::make_shared<Tensor>(
Array2D<float, 3, 3>{{{0.0952412, 0.0952412, 0.0952412},
{0.25606447, 0.25606447, 0.25606447},
{0.40323776, 0.40323776, 0.40323776}}});
- op->forward();
+
auto microGraphScheduler = std::dynamic_pointer_cast<MetaOperator_Op>(op)->getMicroGraphScheduler();
microGraphScheduler->saveSchedulingDiagram("lstm_values_scheduling");
@@ -280,23 +309,73 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
REQUIRE(approxEq<float>(*(op->getOutput(1)), *myHiddenState));
}
-
SECTION("LSTM(forward_values_seq)") {
+ auto pop = Pop();
+ auto myLSTM = LSTM(2, 3, 2, true, "ltsm");
+ auto myGraph = Sequential({pop, myLSTM});
+ auto op = std::static_pointer_cast<OperatorTensor>(myLSTM->getOperator());
+
+ REQUIRE(myLSTM->nbInputs() == 3 + 8 + 8);
+ REQUIRE(myLSTM->nbData() == 1);
+ REQUIRE(myLSTM->nbOutputs() == 2);
+
+ std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
+ Array3D<float, 2, 3, 2>{{{{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}, {{2.0, 3.0}, {4.0, 5.0}, {6.0, 7.0}}}});
+ std::shared_ptr<Tensor> myInit = std::make_shared<Tensor>(
+ Array2D<float, 3, 3>{{{0.0, 0.0, 0.0}, {0.0, 0.0, 0.0}, {0.0, 0.0, 0.0}}});
+ std::shared_ptr<Tensor> myInitW = std::make_shared<Tensor>(
+ Array2D<float, 3, 2>{{{0.1, 0.1}, {0.1, 0.1}, {0.1, 0.1}}});
+ std::shared_ptr<Tensor> myInitR = std::make_shared<Tensor>(
+ Array2D<float, 3, 3>{{{0.1, 0.1, 0.1}, {0.1, 0.1, 0.1}, {0.1, 0.1, 0.1}}});
+
+ pop->getOperator()->associateInput(0, myInput);
+ op->associateInput(17, myInit);
+ op->associateInput(18, myInit);
+
+ // Weights X
+ myLSTM->input(1).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(2).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(3).first->getOperator()->setOutput(0, myInitW);
+ myLSTM->input(4).first->getOperator()->setOutput(0, myInitW);
+ // Weights H
+ myLSTM->input(5).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(6).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(7).first->getOperator()->setOutput(0, myInitR);
+ myLSTM->input(8).first->getOperator()->setOutput(0, myInitR);
+
+ auto g = getConnectedGraphView(myLSTM);
+ g->setDataType(DataType::Float32);
+ g->setBackend("cpu");
+
+ g->save("lstm_seq", true, true);
+
+ auto scheduler = SequentialScheduler(g);
+ scheduler.forward(true, true);
+ scheduler.saveSchedulingDiagram("lstm_seq_schedule");
+
+ std::shared_ptr<Tensor> myHiddenState = std::make_shared<Tensor>(
+ Array2D<float, 3, 3>{{{0.24439372, 0.24439372, 0.24439372},
+ {0.49801484, 0.49801484, 0.49801484},
+ {0.67162132, 0.67162132, 0.67162132}}});
+
+ myGraph->save("lstm_seq_mygraph", true, true);
+
+ op->getOutput(1)->print();
+ myHiddenState->print();
+
+ REQUIRE(approxEq<float>(*(op->getOutput(1)), *myHiddenState));
+ }
+ SECTION("LSTM(forward_values_seq_flatten)") {
auto pop = Pop();
auto myLSTM = LSTM(2, 3, 2, true, "ltsm");
auto op = std::static_pointer_cast<MetaOperator_Op>(myLSTM->getOperator());
- // NOTE: LSTM really need to be flatten in the graph before execution.
- // Here, we actually don't use the meta-op as a closed black-box,
- // because its scheduling cannot be run independently of the input.
- // Since we use the Pop operator to generate sequential inputs, running
- // the meta-op internal scheduler would not work because it would not
- // update its input!
+ // Here we test LSTM as it is was flatten in the graph.
// We just borrow its micro-graph into our larger myGraph graph.
auto myGraph = std::make_shared<GraphView>();
- myGraph->add(pop);
- myGraph->add(op->getMicroGraph());
pop->addChild(op->getMicroGraph()->getOrderedInputs()[0].first, 0, 0);
+ myGraph->add(op->getMicroGraph());
+ myGraph->add(pop);
REQUIRE(myLSTM->nbInputs() == 3 + 8 + 8);
REQUIRE(myLSTM->nbData() == 1);
@@ -316,18 +395,22 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
op->associateInput(18, myInit);
// Weights X
- op->associateInput(1, myInitW);
- op->associateInput(2, myInitW);
- op->associateInput(3, myInitW);
- op->associateInput(4, myInitW);
+ auto prodX = Producer(myInitW);
+ prodX->addChild(op->getMicroGraph()->getOrderedInputs()[1].first, 0, 1);
+ prodX->addChild(op->getMicroGraph()->getOrderedInputs()[2].first, 0, 1);
+ prodX->addChild(op->getMicroGraph()->getOrderedInputs()[3].first, 0, 1);
+ prodX->addChild(op->getMicroGraph()->getOrderedInputs()[4].first, 0, 1);
// Weights H
- op->associateInput(5, myInitR);
- op->associateInput(6, myInitR);
- op->associateInput(7, myInitR);
- op->associateInput(8, myInitR);
+ auto prodH = Producer(myInitR);
+ prodH->addChild(op->getMicroGraph()->getOrderedInputs()[5].first, 0, 1);
+ prodH->addChild(op->getMicroGraph()->getOrderedInputs()[6].first, 0, 1);
+ prodH->addChild(op->getMicroGraph()->getOrderedInputs()[7].first, 0, 1);
+ prodH->addChild(op->getMicroGraph()->getOrderedInputs()[8].first, 0, 1);
+ myGraph->add({prodX, prodH});
myGraph->setDataType(DataType::Float32);
myGraph->setBackend("cpu");
+ myGraph->save("lstm_seq_flatten", true, true);
std::shared_ptr<Tensor> myHiddenState = std::make_shared<Tensor>(
Array2D<float, 3, 3>{{{0.24439372, 0.24439372, 0.24439372},
@@ -335,10 +418,8 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
{0.67162132, 0.67162132, 0.67162132}}});
auto scheduler = SequentialScheduler(myGraph);
- scheduler.forward();
- scheduler.saveSchedulingDiagram("lstm_seq_schedule");
-
- myGraph->save("lstm_seq", true, true);
+ scheduler.forward(true, true);
+ scheduler.saveSchedulingDiagram("lstm_seq_flatten_schedule");
op->getOutput(1)->print();
myHiddenState->print();
diff --git a/unit_tests/operator/Test_PaddedConv.cpp b/unit_tests/operator/Test_PaddedConv.cpp
index 3baf0a7aa0f366a8f0dd4e3e9df6700a5cdb0cea..03a592e52b7d057065353a7d99c088d9831c67c7 100644
--- a/unit_tests/operator/Test_PaddedConv.cpp
+++ b/unit_tests/operator/Test_PaddedConv.cpp
@@ -150,12 +150,15 @@ TEST_CASE("[cpu/operator] PaddedConv(forward)", "[PaddedConv][CPU]") {
});
myConv->getOperator()->associateInput(0,myInput);
- myConv->getOperator()->associateInput(1,myWeights);
- myConv->getOperator()->associateInput(2,myBias);
- myConv->getOperator()->setDataType(DataType::Int32);
- myConv->getOperator()->setBackend("cpu");
- op->computeOutputDims();
- myConv->forward();
+ myConv->input(1).first->getOperator()->setOutput(0, myWeights);
+ myConv->input(2).first->getOperator()->setOutput(0, myBias);
+
+ auto g = getConnectedGraphView(myConv);
+ g->setDataType(DataType::Int32);
+ g->setBackend("cpu");
+
+ auto scheduler = SequentialScheduler(g);
+ scheduler.forward();
REQUIRE(*(op->getOutput(0)) == *myOutput);
}
@@ -309,12 +312,15 @@ TEST_CASE("[cpu/operator] PaddedConv(forward)", "[PaddedConv][CPU]") {
});
myConv->getOperator()->associateInput(0,myInput);
- myConv->getOperator()->associateInput(1,myWeights);
- myConv->getOperator()->associateInput(2,myBias);
- myConv->getOperator()->setDataType(DataType::Int32);
- myConv->getOperator()->setBackend("cpu");
- op->computeOutputDims();
- myConv->forward();
+ myConv->input(1).first->getOperator()->setOutput(0, myWeights);
+ myConv->input(2).first->getOperator()->setOutput(0, myBias);
+
+ auto g = getConnectedGraphView(myConv);
+ g->setDataType(DataType::Int32);
+ g->setBackend("cpu");
+
+ auto scheduler = SequentialScheduler(g);
+ scheduler.forward();
REQUIRE(*(op->getOutput(0)) == *myOutput);
}
diff --git a/unit_tests/recipies/Test_FuseBatchNorm.cpp b/unit_tests/recipies/Test_FuseBatchNorm.cpp
index c4b3bf18a5f5b68d0e41b9cd40966790a0cf7ff6..5386a23177d589640849bf864985b463645622db 100644
--- a/unit_tests/recipies/Test_FuseBatchNorm.cpp
+++ b/unit_tests/recipies/Test_FuseBatchNorm.cpp
@@ -86,14 +86,11 @@ TEST_CASE("[core/recipies] FuseBatchNorm", "[recipies][FuseBatchNorm]") {
myBNOp -> setInput(4, std::make_shared<Tensor>(Array1D<float,3> {{0.4470, 0.3064, 0.7061}}));
auto g1 = Sequential({
+ myProd,
myConv,
myBN
});
g1 -> setName("fuseBNGraph");
- myProd -> addChild(myConv); // set graph input
-
- myProdOp -> setDataType(DataType::Float32);
- myProdOp -> setBackend("cpu");
g1 -> compile("cpu", DataType::Float32);
auto s = SequentialScheduler(g1);
@@ -107,7 +104,7 @@ TEST_CASE("[core/recipies] FuseBatchNorm", "[recipies][FuseBatchNorm]") {
std::shared_ptr<Tensor> res2 = std::make_shared<Tensor>(*(myConvOp -> getOutput(0)));
REQUIRE(g1 -> outputNodes().size() == 1);
- REQUIRE(g1 -> inputNodes().size() == 1);
+ REQUIRE(g1 -> inputNodes().size() == 0);
bool eq = true;
for (std::size_t i = 0; i < res1->size(); ++i) {
eq &= std::abs(res1->get<float>(i) - res2->get<float>(i)) < 1.0e-06;
diff --git a/unit_tests/scheduler/Test_Scheduler.cpp b/unit_tests/scheduler/Test_Scheduler.cpp
index 2440fa8264aecdbc718f24a44fb6af782cdfa053..8779933fc7fc7c07305f1018f9469895026f05e4 100644
--- a/unit_tests/scheduler/Test_Scheduler.cpp
+++ b/unit_tests/scheduler/Test_Scheduler.cpp
@@ -218,14 +218,15 @@ TEST_CASE("[cpu/scheduler] SequentialScheduler(forward)") {
auto add2 = Add(2, "add2");
auto bias = Producer(biasTensor, "bias");
auto init = Producer(initTensor, "init");
+ auto input = Producer(in, "input");
std::shared_ptr<GraphView> g = Sequential({add1, mem, add2});
init->addChild(mem, 0, 1);
mem->addChild(add1, 1, 1);
bias->addChild(add2, 0, 1);
- add1->getOperator()->setInput(0, in);
+ input->addChild(add1, 0, 0);
// Update GraphView inputs/outputs following previous connections:
- g->add({mem, add1, add2, init, bias});
+ g->add({mem, add1, add2, init, bias, input});
g->setBackend("cpu");
g->setDataType(Aidge::DataType::Int32);