Scheduler refactor

unit_tests/operator/Test_BatchNormImpl.cpp

@@ -14,6 +14,7 @@
 
 #include "aidge/data/Tensor.hpp"
 #include "aidge/operator/BatchNorm.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
 
 #include "aidge/backend/cpu.hpp"
 

unit_tests/operator/Test_MetaOperator.cpp

@@ -23,6 +23,8 @@
 #include "aidge/operator/MetaOperatorDefs.hpp"
 #include "aidge/operator/Pad.hpp"
 #include "aidge/operator/Pop.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
+#include "aidge/scheduler/ParallelScheduler.hpp"
 
 using namespace Aidge;
 
@@ -206,7 +208,7 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
         std::shared_ptr<Tensor> myInput = std::make_shared<Tensor>(
             Array2D<float, 16, 32>{});
         std::shared_ptr<Tensor> myInit = std::make_shared<Tensor>(
-            Array2D<float, 1, 64>{});
+            Array2D<float, 32, 64>{});
         std::shared_ptr<Tensor> myInitW = std::make_shared<Tensor>(
             Array2D<float, 64, 32>{});
         std::shared_ptr<Tensor> myInitR = std::make_shared<Tensor>(
@@ -233,7 +235,7 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
         g->setBackend("cpu");
 
         auto scheduler = SequentialScheduler(g);
-        scheduler.forward(true, true);
+        scheduler.forward(true);
 
         g->save("lstm_outside_dims", true, true);
 
@@ -245,8 +247,8 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
 
         REQUIRE(op->getNbConsumedData(0) == 512);
         REQUIRE(op->getNbConsumedData(1) == 32768);
-        REQUIRE(op->getNbProducedData(0) == 1088);
-        REQUIRE(op->getNbProducedData(1) == 1088);
+        REQUIRE(op->getNbProducedData(0) == 34816);
+        REQUIRE(op->getNbProducedData(1) == 34816);
         REQUIRE(microGraphScheduler->getStaticScheduling(0).size() == 26);
         REQUIRE(microGraphScheduler->getStaticScheduling(1).size() == 24);
         REQUIRE(microGraphScheduler->getStaticScheduling(15).size() == 24);
@@ -350,7 +352,7 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
         g->save("lstm_seq", true, true);
 
         auto scheduler = SequentialScheduler(g);
-        scheduler.forward(true, true);
+        scheduler.forward(true);
         scheduler.saveSchedulingDiagram("lstm_seq_schedule");
 
         std::shared_ptr<Tensor> myHiddenState = std::make_shared<Tensor>(
@@ -365,7 +367,7 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
 
         REQUIRE(approxEq<float>(*(op->getOutput(0)), *myHiddenState));
     }
-    SECTION("LSTM(forward_values_seq_flatten)") {
+    SECTION("LSTM(forward_values_seq_flatten)(sequential)") {
         auto pop = Pop();
         auto myLSTM = LSTM(2, 3, 2, true, "ltsm");
         auto op = std::static_pointer_cast<MetaOperator_Op>(myLSTM->getOperator());
@@ -418,8 +420,72 @@ TEST_CASE("[cpu/operator] MetaOperator", "[MetaOperator][CPU]") {
                                      {0.67162132, 0.67162132, 0.67162132}}});
 
         auto scheduler = SequentialScheduler(myGraph);
-        scheduler.forward(true, true);
-        scheduler.saveSchedulingDiagram("lstm_seq_flatten_schedule");
+        scheduler.generateScheduling();
+        scheduler.saveStaticSchedulingDiagram("lstm_static_schedule");
+        scheduler.forward(true);
+        scheduler.saveSchedulingDiagram("lstm_seq_flatten_schedule_seq");
+
+        op->getOutput(0)->print();
+        myHiddenState->print();
+
+        REQUIRE(approxEq<float>(*(op->getOutput(0)), *myHiddenState));
+    }
+    SECTION("LSTM(forward_values_seq_flatten)(parallel)") {
+        auto pop = Pop();
+        auto myLSTM = LSTM(2, 3, 2, true, "ltsm");
+        auto op = std::static_pointer_cast<MetaOperator_Op>(myLSTM->getOperator());
+
+        // 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>();
+        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);
+        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
+        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
+        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},
+                                     {0.49801484, 0.49801484, 0.49801484},
+                                     {0.67162132, 0.67162132, 0.67162132}}});
+
+        auto scheduler = ParallelScheduler(myGraph);
+        scheduler.generateScheduling();
+        scheduler.forward(true);
+        scheduler.saveSchedulingDiagram("lstm_seq_flatten_schedule_par");
 
         op->getOutput(0)->print();
         myHiddenState->print();

unit_tests/operator/Test_PaddedConv.cpp

@@ -16,6 +16,7 @@
 #include "aidge/data/Tensor.hpp"
 #include "aidge/operator/MetaOperator.hpp"
 #include "aidge/operator/MetaOperatorDefs.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
 
 #include "aidge/backend/cpu.hpp"
 

unit_tests/recipies/Test_FuseBatchNorm.cpp

@@ -19,7 +19,7 @@
 #include "aidge/operator/BatchNorm.hpp"
 #include "aidge/operator/Producer.hpp"
 #include "aidge/recipes/Recipes.hpp"
-#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
 
 #include "aidge/data/Tensor.hpp"
 

unit_tests/recipies/Test_HorizontalTiling.cpp

@@ -17,7 +17,7 @@
 #include "aidge/operator/Conv.hpp"
 #include "aidge/operator/ReLU.hpp"
 #include "aidge/recipes/Recipes.hpp"
-#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
 #include "aidge/operator/Concat.hpp"
 
 

unit_tests/scheduler/Test_CastMove.cpp

@@ -18,7 +18,7 @@
 #include "aidge/graph/Node.hpp"
 #include "aidge/graph/GraphView.hpp"
 #include "aidge/graph/OpArgs.hpp"
-#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
 #include "aidge/recipes/Recipes.hpp"
 
 #include "aidge/backend/cpu.hpp"

unit_tests/scheduler/Test_Scheduler.cpp

@@ -17,7 +17,8 @@
 #include "aidge/graph/Node.hpp"
 #include "aidge/graph/GraphView.hpp"
 #include "aidge/graph/OpArgs.hpp"
-#include "aidge/scheduler/Scheduler.hpp"
+#include "aidge/scheduler/SequentialScheduler.hpp"
+#include "aidge/scheduler/ParallelScheduler.hpp"
 
 #include "aidge/backend/cpu.hpp"
 
@@ -205,7 +206,7 @@ TEST_CASE("[cpu/scheduler] SequentialScheduler(forward)") {
     SECTION("Test Residual graph") {
     }
 
-    SECTION("Test Recurrent graph") {
+    SECTION("Test Recurrent graph (sequential)") {
         std::shared_ptr<Tensor> in = std::make_shared<Tensor>(
                 Array2D<int, 2, 3>{{{1, 2, 3}, {4, 5, 6}}});
         std::shared_ptr<Tensor> initTensor = std::make_shared<Tensor>(
@@ -232,9 +233,54 @@ TEST_CASE("[cpu/scheduler] SequentialScheduler(forward)") {
         g->setDataType(Aidge::DataType::Int32);
         g->save("graphRecurrent");
         g->forwardDims();
+
         SequentialScheduler scheduler(g);
-        REQUIRE_NOTHROW(scheduler.forward(true, true));
-        scheduler.saveSchedulingDiagram("schedulingRecurrent");
+        REQUIRE_NOTHROW(scheduler.forward(true));
+        scheduler.saveStaticSchedulingDiagram("static_schedule");
+        scheduler.saveSchedulingDiagram("schedulingRecurrent_seq");
+
+        std::shared_ptr<Tensor> expectedOutput = std::make_shared<Tensor>(
+                Array2D<int, 2, 3>{{{5, 6, 9}, {14, 16, 19}}});
+        std::shared_ptr<Tensor> result =
+                std::static_pointer_cast<Tensor>(g->getNode("add2")->getOperator()->getRawOutput(0));
+        result->print();
+        expectedOutput->print();
+        bool equal = (*result == *expectedOutput);
+        REQUIRE(equal);
+    }
+
+
+    SECTION("Test Recurrent graph (parallel)") {
+        std::shared_ptr<Tensor> in = std::make_shared<Tensor>(
+                Array2D<int, 2, 3>{{{1, 2, 3}, {4, 5, 6}}});
+        std::shared_ptr<Tensor> initTensor = std::make_shared<Tensor>(
+                Array2D<int, 2, 3>{{{0, 0, 0}, {1, 1, 1}}});
+        std::shared_ptr<Tensor> biasTensor = std::make_shared<Tensor>(
+                Array2D<int, 2, 3>{{{2, 0, 0}, {1, 0, 0}}});
+
+        auto add1 = Add(2, "add1");
+        auto mem = Memorize(3, "mem1");
+        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);
+        input->addChild(add1, 0, 0);
+        // Update GraphView inputs/outputs following previous connections:
+        g->add({mem, add1, add2, init, bias, input});
+
+        g->setBackend("cpu");
+        g->setDataType(Aidge::DataType::Int32);
+        g->save("graphRecurrent");
+        g->forwardDims();
+
+        ParallelScheduler scheduler(g);
+        REQUIRE_NOTHROW(scheduler.forward(true));
+        scheduler.saveSchedulingDiagram("schedulingRecurrent_par");
 
         std::shared_ptr<Tensor> expectedOutput = std::make_shared<Tensor>(
                 Array2D<int, 2, 3>{{{5, 6, 9}, {14, 16, 19}}});
@@ -299,7 +345,7 @@ TEST_CASE("[cpu/scheduler] SequentialScheduler(forward)") {
         SequentialScheduler scheduler(g);
 
         std::vector<std::shared_ptr<Aidge::Tensor>> dataIn = {inputTensor};
-        REQUIRE_NOTHROW(scheduler.forward(true, false, dataIn));
+        REQUIRE_NOTHROW(scheduler.forward(true, dataIn));
         
         scheduler.saveSchedulingDiagram("schedulingSequential");