From f1afe39803d20c1eaaf49529a453d9a058d5d15b Mon Sep 17 00:00:00 2001
From: Olivier BICHLER <olivier.bichler@cea.fr>
Date: Mon, 5 Feb 2024 14:34:29 +0100
Subject: [PATCH] Working concept
---
include/aidge/operator/Memorize.hpp | 101 ++++++++++++++++++++++++++
src/graph/GraphView.cpp | 10 ++-
src/operator/Memorize.cpp | 52 ++++++++++++++
src/scheduler/Scheduler.cpp | 105 +++++++++++++++++++++++++---
4 files changed, 256 insertions(+), 12 deletions(-)
create mode 100644 include/aidge/operator/Memorize.hpp
create mode 100644 src/operator/Memorize.cpp
diff --git a/include/aidge/operator/Memorize.hpp b/include/aidge/operator/Memorize.hpp
new file mode 100644
index 000000000..53340f866
--- /dev/null
+++ b/include/aidge/operator/Memorize.hpp
@@ -0,0 +1,101 @@
+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#ifndef AIDGE_CORE_OPERATOR_MEMORIZE_H_
+#define AIDGE_CORE_OPERATOR_MEMORIZE_H_
+
+#include <cassert>
+#include <memory>
+#include <vector>
+
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/operator/OperatorTensor.hpp"
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/data/Tensor.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/utils/StaticAttributes.hpp"
+
+namespace Aidge {
+enum class MemorizeAttr { ScheduleStep, ForwardStep, EndStep };
+
+class Memorize_Op : public OperatorTensor,
+ public Registrable<Memorize_Op, std::string, std::unique_ptr<OperatorImpl>(const Memorize_Op&)>,
+ public StaticAttributes<MemorizeAttr, unsigned int, unsigned int, unsigned int> {
+public:
+ static const std::string Type;
+
+ using Attributes_ = StaticAttributes<MemorizeAttr, unsigned int, unsigned int, unsigned int>;
+ template <MemorizeAttr e>
+ using attr = typename Attributes_::template attr<e>;
+
+ Memorize_Op(const unsigned int endStep)
+ : OperatorTensor(Type, 2, 0, 2),
+ Attributes_(attr<MemorizeAttr::ScheduleStep>(0),
+ attr<MemorizeAttr::ForwardStep>(0),
+ attr<MemorizeAttr::EndStep>(endStep))
+ {
+ mOutputs[1] = mOutputs[0];
+ }
+
+ /**
+ * @brief Copy-constructor. Copy the operator attributes and its output tensor(s), but not its input tensors (the new operator has no input associated).
+ * @param op Operator to copy.
+ */
+ Memorize_Op(const Memorize_Op& op)
+ : OperatorTensor(op),
+ Attributes_(op)
+ {
+ mImpl = op.mImpl ? Registrar<Memorize_Op>::create(op.mOutputs[0]->getImpl()->backend())(*this) : nullptr;
+ mOutputs[1] = mOutputs[0];
+ }
+
+ /**
+ * @brief Clone the operator using its copy-constructor.
+ * @see Operator::Memorize_Op
+ */
+ std::shared_ptr<Operator> clone() const override {
+ return std::make_shared<Memorize_Op>(*this);
+ }
+
+ void setBackend(const std::string& name, DeviceIdx_t device = 0) override {
+ mImpl = Registrar<Memorize_Op>::create({name})(*this);
+ mOutputs[0]->setBackend(name, device);
+ }
+
+ void computeOutputDims() override;
+ bool outputDimsForwarded() const override;
+ void updateConsummerProducer() override;
+ void forward() override;
+
+ static const std::vector<std::string> getInputsName(){
+ return {"data_input", "data_input_init"};
+ }
+ static const std::vector<std::string> getOutputsName(){
+ return {"data_output", "data_output_rec"};
+ }
+};
+
+inline std::shared_ptr<Node> Memorize(const unsigned int endStep, const std::string& name = "") {
+ return std::make_shared<Node>(std::make_shared<Memorize_Op>(endStep), name);
+}
+} // namespace Aidge
+
+namespace {
+template <>
+const char *const EnumStrings<Aidge::MemorizeAttr>::data[] = {
+ "ScheduleStep",
+ "ForwardStep",
+ "EndStep"
+};
+}
+
+#endif /* AIDGE_CORE_OPERATOR_MEMORIZE_H_ */
\ No newline at end of file
diff --git a/src/graph/GraphView.cpp b/src/graph/GraphView.cpp
index 968e98e75..5c7a524af 100644
--- a/src/graph/GraphView.cpp
+++ b/src/graph/GraphView.cpp
@@ -306,7 +306,12 @@ void Aidge::GraphView::_forwardDims(std::set<std::shared_ptr<Node>> listNodes) {
nextList.insert(nodePtr);
} else { // compute output dimensions of children
std::set<std::shared_ptr<Node>> children = nodePtr->getChildren();
- nextList.insert(children.begin(), children.end());
+ for (auto child : children) {
+ const auto childOp = std::static_pointer_cast<OperatorTensor>(child->getOperator());
+ if (!childOp->outputDimsForwarded()) {
+ nextList.insert(child);
+ }
+ }
}
}
}
@@ -319,6 +324,9 @@ void Aidge::GraphView::_forwardDims(std::set<std::shared_ptr<Node>> listNodes) {
}
}
}
+
+ AIDGE_INTERNAL_ASSERT(nextList != listNodes);
+
if (!nextList.empty()) {
_forwardDims(nextList);
}
diff --git a/src/operator/Memorize.cpp b/src/operator/Memorize.cpp
new file mode 100644
index 000000000..714bf97fc
--- /dev/null
+++ b/src/operator/Memorize.cpp
@@ -0,0 +1,52 @@
+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/Memorize.hpp"
+
+const std::string Aidge::Memorize_Op::Type = "Memorize";
+
+void Aidge::Memorize_Op::computeOutputDims() {
+ // Only require input #1 dims (initialization of the Memorize operator)
+ // Otherwise, forwardDims() won't converge!
+ bool associated = (nbInputs() > 0); // do not compute anything if no input
+ if (!getInput(1)) {
+ AIDGE_THROW_OR_ABORT(std::runtime_error, "Input #1 should be associated with a Tensor");
+ }
+ associated &= !(getInput(1)->empty());
+
+ if (associated) {
+ const auto expectedDims = getInput(1)->dims();
+ mOutputs[0]->resize(expectedDims);
+ }
+}
+
+bool Aidge::Memorize_Op::outputDimsForwarded() const {
+ // Only check the output dims
+ bool forwarded = true;
+ // check outputs have been filled
+ for (IOIndex_t i = 0; i < nbOutputs(); ++i) {
+ forwarded &= !(getOutput(i)->empty());
+ }
+ return forwarded;
+}
+
+void Aidge::Memorize_Op::updateConsummerProducer() {
+ Operator::updateConsummerProducer();
+ ++this->template getAttr<MemorizeAttr::ScheduleStep>();
+ this->template getAttr<MemorizeAttr::ForwardStep>() = 0;
+}
+
+void Aidge::Memorize_Op::forward() {
+ Operator::forward();
+ ++this->template getAttr<MemorizeAttr::ForwardStep>();
+ this->template getAttr<MemorizeAttr::ScheduleStep>() = 0;
+}
diff --git a/src/scheduler/Scheduler.cpp b/src/scheduler/Scheduler.cpp
index 3afbcd044..953804f62 100644
--- a/src/scheduler/Scheduler.cpp
+++ b/src/scheduler/Scheduler.cpp
@@ -20,6 +20,7 @@
#include "aidge/graph/Node.hpp"
#include "aidge/utils/Types.h"
#include "aidge/operator/OperatorTensor.hpp"
+#include "aidge/operator/Producer.hpp"
void drawProgressBar(double progress, int barWidth, const std::string& additionalInfo = "") {
putchar('[');
@@ -43,7 +44,7 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
// setup initial producers list
std::set<std::shared_ptr<Node>> producers;
for (const std::shared_ptr<Node>& nodePtr : mGraphView->getNodes()) {
- if (nodePtr->type() == "Producer") {
+ if (nodePtr->type() == Producer_Op::Type) {
producers.insert(nodePtr);
}
}
@@ -64,7 +65,62 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
/* It may not be necessary to initialize producer */
std::set<std::shared_ptr<Node>> consumers = mGraphView->inputNodes();
+ std::set<std::shared_ptr<Node>> frozenConsumers;
do {
+ // Check required producers
+ std::set<std::shared_ptr<Node>> requiredProducers;
+ if (verbose) printf("Required producers:\n");
+
+ for (const auto& consumer : consumers) {
+ if (verbose) {
+ printf("\t- consumer: "
+ "\x1b[1;37m"
+ "%s"
+ "\x1b[0m"
+ "\n",
+ (consumer->type() + "_" + std::to_string(reinterpret_cast<uintptr_t>(consumer.get()))).c_str());
+ }
+
+ std::set<std::shared_ptr<Node>> consumerRequiredProducers;
+ bool requiredProducerOnly = true;
+ IOIndex_t inputIdx = 0;
+ for (const auto& consumerParent : consumer->inputs()) {
+ if (verbose) printf("\t\t#%u: ", inputIdx);
+
+ if (consumerParent.first &&
+ (consumer->getOperator()->getNbConsumedData(inputIdx) + consumer->getOperator()->getNbRequiredData(inputIdx)) >
+ consumerParent.first->getOperator()->getNbProducedData(consumerParent.second)) {
+ if (verbose) printf("required data from %s: C%zu + R%zu > P%zu\n",
+ consumerParent.first->type().c_str(),
+ consumer->getOperator()->getNbConsumedData(inputIdx),
+ consumer->getOperator()->getNbRequiredData(inputIdx),
+ consumerParent.first->getOperator()->getNbProducedData(consumerParent.second));
+
+ if (consumerParent.first->type() == Producer_Op::Type) {
+ consumerRequiredProducers.insert(consumerParent.first);
+ }
+ else {
+ requiredProducerOnly = false;
+ break;
+ }
+ }
+ else {
+ if (verbose) printf("no data required\n");
+ }
+ ++inputIdx;
+ }
+
+ if (requiredProducerOnly) {
+ requiredProducers.insert(consumerRequiredProducers.begin(), consumerRequiredProducers.end());
+ }
+ }
+
+ // Make producers generate the required data
+ for (const auto& requiredProducer : requiredProducers) {
+ requiredProducer->getOperator()->updateConsummerProducer();
+ mStaticSchedule.push_back(requiredProducer);
+ }
+
// find runnable consumers
std::set<std::shared_ptr<Node>> runnableConsumers;
if (verbose) printf("List of layers receiving data:\n");
@@ -74,7 +130,7 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
"\x1b[1;37m"
"%s"
"\x1b[0m"
- "\n\t\tR/C:\t",
+ "\n\t\tC/R:\t",
(consumer->type() + "_" + std::to_string(reinterpret_cast<uintptr_t>(consumer.get()))).c_str());
for (IOIndex_t inId = 0; inId < consumer->nbInputs() - 1; ++inId) {
printf("%zu/%zu\n\t\t\t", consumer->getOperator()->getNbConsumedData(inId),
@@ -89,18 +145,25 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
printf("%zu", consumer->getOperator()->getNbProducedData(static_cast<IOIndex_t>(consumer->nbOutputs()) - 1));
printf("\n");
}
+
bool isRunnable = true;
- IOIndex_t parentID = 0; // FIXME: handle this correctly
+ IOIndex_t inputIdx = 0; // FIXME: handle this correctly
// Check every input has got enought data to run
- for (const auto& consumerParent : consumer->dataInputs()) {
+ for (const auto& consumerParent : consumer->inputs()) {
if (consumerParent.first &&
- consumer->getOperator()->getNbRequiredData(parentID++) >
+ (consumer->getOperator()->getNbConsumedData(inputIdx) + consumer->getOperator()->getNbRequiredData(inputIdx)) >
consumerParent.first->getOperator()->getNbProducedData(consumerParent.second)) {
+ if (verbose) printf(" not runnable: C%zu + R%zu > P%zu\n",
+ consumer->getOperator()->getNbConsumedData(inputIdx),
+ consumer->getOperator()->getNbRequiredData(inputIdx),
+ consumerParent.first->getOperator()->getNbProducedData(consumerParent.second));
+
// not enough data to run
isRunnable = false;
break;
}
+ ++inputIdx;
}
if (isRunnable) {
@@ -115,13 +178,22 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
mStaticSchedule.push_back(runnable);
}
+ if (runnableConsumers.empty()) {
+ if (frozenConsumers.empty()) {
+ frozenConsumers = consumers;
+ }
+ }
+ else {
+ frozenConsumers.clear();
+ }
+
// update producers and consumers list
if (verbose) printf("Updating producer and consumer lists...\n");
const auto oldConsumers = consumers;
for (const auto& consumer : oldConsumers) {
if (verbose) {
- printf("\t- consumer: %s\n\t\tR/C:\t",
+ printf("\t- consumer: %s\n\t\tC/R:\t",
(consumer->type() + "_" + std::to_string(reinterpret_cast<uintptr_t>(consumer.get()))).c_str());
for (IOIndex_t inId = 0; inId < consumer->nbInputs() - 1; ++inId) {
printf("%ld/%ld\n\t\t\t", consumer->getOperator()->getNbConsumedData(inId),
@@ -138,16 +210,21 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
}
bool isStillConsumer = false;
- IOIndex_t parentID = 0; // FIXME: handle this correctly
+ IOIndex_t inputIdx = 0; // FIXME: handle this correctly
// should we check input or dataInput ?
for (const auto& consumerParent : consumer->inputs()) {
if (consumerParent.first &&
- consumer->getOperator()->getNbConsumedData(parentID++) <
+ consumer->getOperator()->getNbConsumedData(inputIdx) <
consumerParent.first->getOperator()->getNbProducedData(consumerParent.second)) {
+ if (verbose) printf(" still consumer: C%zu < P%zu\n",
+ consumer->getOperator()->getNbConsumedData(inputIdx),
+ consumerParent.first->getOperator()->getNbProducedData(consumerParent.second));
+
// there is still data to consume
isStillConsumer = true;
break;
}
+ ++inputIdx;
}
for (IOIndex_t outId = 0; outId < consumer->nbOutputs(); ++outId) {
@@ -169,9 +246,15 @@ void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
}
}
- if (verbose) printf("*************\n");
- } while (!consumers.empty());
+ if (verbose) printf("********************\n");
+ } while (!consumers.empty() && consumers != frozenConsumers);
+ if (verbose) {
+ if (!consumers.empty()) {
+ printf("*** Frozen state ***\n");
+ printf("********************\n");
+ }
+ }
}
// TODO: handle multiple inputs/outputs
@@ -183,7 +266,7 @@ void Aidge::SequentialScheduler::forward(bool forwardDims, bool verbose) {
// If scheduling was already generated (in one or several steps, i.e. one or
// several successive call to generateScheduling()), do not generate it twice
if (mStaticSchedule.empty()) {
- this->generateScheduling();
+ this->generateScheduling(verbose);
}
// Clear previous scheduling results
--
GitLab